Add Pyrefly config and fix type errors

ci: remove frontend validation PR comment after successful check (#6360 )
# Description of Changes - Added a GitHub Actions step to remove the frontend validation check comment when the frontend check succeeds on pull requests. - The step searches for an existing PR comment containing the `` marker and deletes it if found. - The change was made to keep pull request discussions clean by removing stale frontend check comments once the validation passes. --- ## Checklist ### General - [ ] I have read the [Contribution Guidelines](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/CONTRIBUTING.md) - [ ] I have read the [Stirling-PDF Developer Guide](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/DeveloperGuide.md) (if applicable) - [ ] I have read the [How to add new languages to Stirling-PDF](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/devGuide/HowToAddNewLanguage.md) (if applicable) - [ ] I have performed a self-review of my own code - [ ] My changes generate no new warnings ### Documentation - [ ] I have updated relevant docs on [Stirling-PDF's doc repo](https://github.com/Stirling-Tools/Stirling-Tools.github.io/blob/main/docs/) (if functionality has heavily changed) - [ ] I have read the section [Add New Translation Tags](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/devGuide/HowToAddNewLanguage.md#add-new-translation-tags) (for new translation tags only) ### Translations (if applicable) - [ ] I ran [`scripts/counter_translation.py`](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/docs/counter_translation.md) ### UI Changes (if applicable) - [ ] Screenshots or videos demonstrating the UI changes are attached (e.g., as comments or direct attachments in the PR) ### Testing (if applicable) - [ ] I have run `task check` to verify linters, typechecks, and tests pass - [ ] I have tested my changes locally. Refer to the [Testing Guide](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/DeveloperGuide.md#7-testing) for more details.
2026-05-14 15:09:04 +01:00 · 2026-05-14 13:19:55 +00:00 · 2026-05-14 13:19:38 +00:00 · 2026-05-14 12:31:45 +00:00 · 2026-05-13 16:30:58 +01:00 · 2026-05-12 13:22:15 +01:00
154 changed files with 6511 additions and 1773 deletions
@@ -22,7 +22,7 @@ runs:
  steps:
    - name: Generate a GitHub App Token
      id: generate-token
-      uses: actions/create-github-app-token@df432ceedc7162793a195dd1713ff69aefc7379e # v2.0.6
+      uses: actions/create-github-app-token@bcd2ba49218906704ab6c1aa796996da409d3eb1 # v3.2.0
      with:
        app-id: ${{ inputs.app-id }}
        private-key: ${{ inputs.private-key }}
@@ -1,6 +1,6 @@
 # Maintainer: Stirling PDF Inc <contact@stirlingpdf.com>
 pkgname=stirling-pdf-desktop
-pkgver=2.10.1
+pkgver=2.11.0
 pkgrel=1
 pkgdesc="Locally hosted, web-based PDF manipulation tool (Tauri desktop app, official Stirling PDF Inc build)"
 arch=('x86_64')
@@ -1,6 +1,6 @@
 # Maintainer: Stirling PDF Inc <contact@stirlingpdf.com>
 pkgname=stirling-pdf-server-bin
-pkgver=2.10.1
+pkgver=2.11.0
 pkgrel=1
 pkgdesc="Locally hosted, web-based PDF manipulation tool (server JAR, prebuilt)"
 arch=('any')
@@ -35,7 +35,7 @@ jobs:
      pr_ref: ${{ steps.resolve.outputs.ref }}
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -111,7 +111,7 @@ jobs:

    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -359,7 +359,7 @@ jobs:

    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -70,7 +70,7 @@ jobs:
      enable_prototypes: ${{ steps.check-prototypes-flag.outputs.enable_prototypes }}
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -183,7 +183,7 @@ jobs:

    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -474,7 +474,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -21,7 +21,7 @@ jobs:

    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -19,7 +19,7 @@ jobs:
      pull-requests: write
    steps:
      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -19,7 +19,7 @@ jobs:

    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -26,7 +26,7 @@ jobs:
      jar_sha256: ${{ steps.hashes.outputs.jar_sha256 }}
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -70,7 +70,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -106,7 +106,7 @@ jobs:

      - name: Publish stirling-pdf-desktop to AUR
        if: ${{ github.event_name == 'release' || inputs.dry_run == false }}
-        uses: KSXGitHub/github-actions-deploy-aur@v4.1.2
+        uses: KSXGitHub/github-actions-deploy-aur@v4.1.3
        with:
          pkgname: stirling-pdf-desktop
          pkgbuild: .github/aur/stirling-pdf-desktop/PKGBUILD
@@ -16,7 +16,7 @@ jobs:
      pull-requests: write
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -22,7 +22,7 @@ jobs:
        spring-security: [true, false]
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit
      - name: Checkout repository
@@ -35,7 +35,7 @@ jobs:
          distribution: "temurin"

      - name: Cache Gradle dependency artifacts
-        uses: actions/cache@668228422ae6a00e4ad889ee87cd7109ec5666a7 # v5.0.4
+        uses: actions/cache@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5
        with:
          path: |
            ~/.gradle/wrapper
@@ -44,7 +44,7 @@ jobs:
      SYSTEM_ENABLEANALYTICS: "false"
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit
      - name: Checkout repository
@@ -63,7 +63,15 @@ jobs:
      - name: Install Task
        uses: go-task/setup-task@3be4020d41929789a01026e0e427a4321ce0ad44 # v2.0.0
      - name: Install Playwright (chromium only)
-        run: task frontend:test:e2e:install -- chromium
+        run: task e2e:install -- chromium
+      - name: Build frontend (needed for playwright's vite preview webServer)
+        # Enterprise tests target :8080 (Spring Boot's bundled frontend), but
+        # playwright's webServer config still launches `vite preview --port 5173`
+        # before any test run, which needs dist/ to exist. VITE_BUILD_FOR_PREVIEW
+        # forces absolute asset paths so vite preview can serve deep SPA routes.
+        env:
+          VITE_BUILD_FOR_PREVIEW: "1"
+        run: task frontend:build

      - name: Resolve kubernetes.docker.internal to localhost
        # The compose stacks set KC_HOSTNAME=kubernetes.docker.internal so
@@ -140,7 +148,7 @@ jobs:
          wait_for_backend
      - name: Run enterprise OAuth Playwright tests
        id: oauth-tests
-        run: task frontend:test:e2e -- --project=enterprise --grep "OAuth"
+        run: task e2e:enterprise -- --grep "OAuth"
      - name: Stop backend + tear down OAuth Keycloak
        if: always()
        run: |
@@ -213,7 +221,7 @@ jobs:
          wait_for_backend
      - name: Run enterprise SAML Playwright tests
        id: saml-tests
-        run: task frontend:test:e2e -- --project=enterprise --grep "SAML"
+        run: task e2e:enterprise -- --grep "SAML"
      - name: Stop backend + tear down SAML Keycloak
        if: always()
        run: |
@@ -243,7 +251,7 @@ jobs:
          wait_for_backend
      - name: Run enterprise feature Playwright tests
        id: feature-tests
-        run: task frontend:test:e2e -- --project=enterprise --grep "Enterprise license"
+        run: task e2e:enterprise -- --grep "Enterprise license"
      - name: Print backend log on failure
        if: failure()
        run: |
@@ -46,7 +46,7 @@ jobs:
      proprietary: ${{ steps.changes.outputs.proprietary }}
    steps:
      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit
      - name: Checkout repository
@@ -13,7 +13,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -27,7 +27,7 @@ jobs:
          distribution: "temurin"

      - name: Cache Gradle dependency artifacts
-        uses: actions/cache@668228422ae6a00e4ad889ee87cd7109ec5666a7 # v5.0.4
+        uses: actions/cache@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5
        with:
          path: |
            ~/.gradle/wrapper
@@ -14,7 +14,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -28,7 +28,7 @@ jobs:
          distribution: "temurin"

      - name: Cache Gradle dependency artifacts
-        uses: actions/cache@668228422ae6a00e4ad889ee87cd7109ec5666a7 # v5.0.4
+        uses: actions/cache@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5
        with:
          path: |
            ~/.gradle/wrapper
@@ -27,7 +27,7 @@ jobs:
      pull-requests: write # Allow writing to pull requests
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -14,7 +14,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -18,7 +18,7 @@ jobs:

    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -25,7 +25,7 @@ jobs:

    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -39,7 +39,7 @@ jobs:
          distribution: "temurin"

      - name: Cache Gradle dependency artifacts
-        uses: actions/cache@668228422ae6a00e4ad889ee87cd7109ec5666a7 # v5.0.4
+        uses: actions/cache@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5
        with:
          path: |
            ~/.gradle/wrapper
@@ -15,7 +15,7 @@ jobs:
    timeout-minutes: 30
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit
      - name: Checkout repository
@@ -34,58 +34,26 @@ jobs:
      - name: Install Task
        uses: go-task/setup-task@3be4020d41929789a01026e0e427a4321ce0ad44 # v2.0.0
      - name: Install Playwright (chromium only)
-        run: task frontend:test:e2e:install -- chromium
-      - name: Start Spring Boot backend (background)
+        run: task e2e:install -- chromium
+      - name: Build frontend (production bundle for vite preview)
        env:
-          # Suppress the analytics opt-in modal that fires on first admin login when
-          # enableAnalytics is null (see Onboarding.tsx). The modal renders a Mantine
-          # overlay that intercepts pointer events on every tool page until dismissed,
-          # which causes every "click run button" assertion in the live suite to fail.
-          SYSTEM_ENABLEANALYTICS: "false"
-        # NOTE: SECURITY_INITIALLOGIN_USERNAME/PASSWORD are intentionally NOT set.
-        # The live-setup project's bootstrap spec performs the real first-login
-        # flow against the backend's default admin/stirling user, exercising the
-        # forced-password-change UI and leaving the DB at admin/adminadmin for
-        # the rest of the live suite. This is both real coverage of the first-
-        # login flow and a stronger seed than env-var-driven user creation.
-        run: |
-          nohup ./gradlew :stirling-pdf:bootRun > /tmp/backend.log 2>&1 &
-          echo $! > /tmp/backend.pid
-      - name: Wait for backend to become ready
-        run: |
-          start=$SECONDS
-          # 300 iterations × 2s = 10 minute ceiling
-          for i in $(seq 1 300); do
-            if curl -fsS http://localhost:8080/api/v1/info/status >/dev/null 2>&1; then
-              echo "Backend up after $((SECONDS - start))s"
-              exit 0
-            fi
-            sleep 2
-          done
-          echo "Backend did not become ready in $((SECONDS - start))s"
-          tail -200 /tmp/backend.log || true
-          exit 1
+          VITE_BUILD_FOR_PREVIEW: "1"
+        run: task frontend:build
      - name: Run live E2E tests (chromium)
        id: live-tests
-        run: task frontend:test:e2e -- --project=live
+        run: task e2e:live
      - name: Print backend log on failure
        if: failure() && steps.live-tests.conclusion == 'failure'
        run: |
          echo "::group::Spring Boot backend log (last 500 lines)"
-          tail -500 /tmp/backend.log || echo "no backend log found"
+          tail -500 .test-state/playwright/backend.log || echo "no backend log found"
          echo "::endgroup::"
-      - name: Stop backend
-        if: always()
-        run: |
-          if [ -f /tmp/backend.pid ]; then
-            kill "$(cat /tmp/backend.pid)" 2>/dev/null || true
-          fi
      - name: Upload backend log
        if: always()
        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
        with:
          name: backend-log-live-${{ github.run_id }}
-          path: /tmp/backend.log
+          path: .test-state/playwright/backend.log
          retention-days: 7
      - name: Upload Playwright report
        if: always()
@@ -14,7 +14,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit
      - name: Checkout repository
@@ -28,9 +28,13 @@ jobs:
      - name: Install Task
        uses: go-task/setup-task@3be4020d41929789a01026e0e427a4321ce0ad44 # v2.0.0
      - name: Install Playwright (chromium only)
-        run: task frontend:test:e2e:install -- chromium
+        run: task e2e:install -- chromium
+      - name: Build frontend (production bundle for vite preview)
+        env:
+          VITE_BUILD_FOR_PREVIEW: "1"
+        run: task frontend:build
      - name: Run stubbed E2E tests (chromium)
-        run: task frontend:test:e2e -- --project=stubbed --workers=3
+        run: task e2e:stubbed -- --workers=3
      - name: Upload Playwright report
        if: always()
        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
@@ -28,7 +28,7 @@ jobs:
      licenses-backend: ${{ steps.changes.outputs.licenses-backend }}
    steps:
      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -52,7 +52,7 @@ jobs:
      repository-projects: write # Required for enabling automerge
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -316,7 +316,7 @@ jobs:
      repository-projects: write # Required for enabling automerge
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -15,7 +15,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit
      - name: Checkout repository
@@ -85,6 +85,26 @@ jobs:
          echo "to see what still needs fixing manually."
          echo "============================================"
          exit 1
+      - name: Remove frontend check comment on success
+        if: steps.frontend-check.outcome == 'success' && github.event_name == 'pull_request'
+        continue-on-error: true
+        uses: actions/github-script@3a2844b7e9c422d3c10d287c895573f7108da1b3 # v9.0.0
+        with:
+          script: |
+            const marker = '<!-- frontend-check -->';
+            const { data: comments } = await github.rest.issues.listComments({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              issue_number: context.issue.number,
+            });
+            const existing = comments.find(c => c.body.includes(marker));
+            if (existing) {
+              await github.rest.issues.deleteComment({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                comment_id: existing.id,
+              });
+            }
      - name: Upload frontend build artifacts
        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
        with:
@@ -15,7 +15,7 @@ jobs:
      issues: write
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -44,7 +44,7 @@ jobs:
      version: ${{ steps.versionNumber.outputs.versionNumber }}
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -57,7 +57,7 @@ jobs:
          distribution: "temurin"

      - name: Cache Gradle dependencies
-        uses: actions/cache@668228422ae6a00e4ad889ee87cd7109ec5666a7 # v5.0.4
+        uses: actions/cache@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5
        with:
          path: |
            ~/.gradle/caches
@@ -127,7 +127,7 @@ jobs:
            file_suffix: "-server"
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -191,7 +191,7 @@ jobs:
      RELEASE_GPG_PRIVATE_KEY: ${{ secrets.RELEASE_GPG_PRIVATE_KEY }}
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit
          allowed-endpoints: >
@@ -606,7 +606,7 @@ jobs:
      contents: write
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -18,7 +18,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -35,10 +35,10 @@ jobs:
      - name: Install Task
        uses: go-task/setup-task@3be4020d41929789a01026e0e427a4321ce0ad44 # v2.0.0
      - name: Install all Playwright browsers
-        run: task frontend:test:e2e:install
+        run: task e2e:install

      - name: Run E2E tests (all browsers)
-        run: task frontend:test:e2e
+        run: task e2e:cross-browser

      - name: Upload Playwright report
        if: always()
@@ -28,7 +28,7 @@ jobs:
      jar_sha256: ${{ steps.hashes.outputs.jar_sha256 }}
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -79,7 +79,7 @@ jobs:
      contents: write
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -19,7 +19,7 @@ jobs:
      PIP_DISABLE_PIP_VERSION_CHECK: "1"
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -43,7 +43,7 @@ jobs:
          echo "version=${VERSION}" >> $GITHUB_OUTPUT

      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -2,6 +2,22 @@ name: Push Docker Image with VersionNumber

 on:
  workflow_dispatch:
+    inputs:
+      build_main_app:
+        description: "Build & push the main Stirling-PDF image (latest, fat, ultra-lite)."
+        required: false
+        type: boolean
+        default: true
+      build_unoserver:
+        description: "Build & push the standalone stirling-unoserver image."
+        required: false
+        type: boolean
+        default: true
+      force_unoserver_rebuild:
+        description: "Rebuild stirling-unoserver even if its source hash is unchanged."
+        required: false
+        type: boolean
+        default: false
  push:
    branches:
      - master
@@ -31,9 +47,13 @@ jobs:
    permissions:
      packages: write
      id-token: write
+    # On push events these stay 'true'; on workflow_dispatch they follow the inputs.
+    env:
+      RUN_MAIN_APP: ${{ github.event_name != 'workflow_dispatch' || inputs.build_main_app }}
+      RUN_UNOSERVER: ${{ github.event_name != 'workflow_dispatch' || inputs.build_unoserver }}
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -46,7 +66,7 @@ jobs:
          distribution: "temurin"

      - name: Cache Gradle dependencies
-        uses: actions/cache@668228422ae6a00e4ad889ee87cd7109ec5666a7 # v5.0.4
+        uses: actions/cache@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5
        with:
          path: |
            ~/.gradle/caches
@@ -108,6 +128,7 @@ jobs:

      - name: Generate tags for latest
        id: meta
+        if: env.RUN_MAIN_APP == 'true'
        uses: docker/metadata-action@030e881283bb7a6894de51c315a6bfe6a94e05cf # v6.0.0
        with:
          images: |
@@ -122,6 +143,8 @@ jobs:

      - name: Build and push Unified Dockerfile (latest variant)
        id: build-push-latest
+        # Empty-tag guard: build-push-action errors when asked to push with no tags.
+        if: env.RUN_MAIN_APP == 'true' && steps.meta.outputs.tags != ''
        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
        with:
          builder: ${{ steps.buildx.outputs.name }}
@@ -140,7 +163,7 @@ jobs:
          sbom: true

      - name: Sign regular images
-        if: github.ref == 'refs/heads/master' || github.ref == 'refs/heads/V2-master'
+        if: env.RUN_MAIN_APP == 'true' && (github.ref == 'refs/heads/master' || github.ref == 'refs/heads/V2-master') && steps.build-push-latest.outputs.digest != ''
        env:
          DIGEST: ${{ steps.build-push-latest.outputs.digest }}
          TAGS: ${{ steps.meta.outputs.tags }}
@@ -156,7 +179,7 @@ jobs:
      - name: Generate tags for latest-fat
        id: meta-fat
        uses: docker/metadata-action@030e881283bb7a6894de51c315a6bfe6a94e05cf # v6.0.0
-        if: github.ref != 'refs/heads/main' && github.ref != 'refs/heads/testMain'
+        if: env.RUN_MAIN_APP == 'true' && github.ref != 'refs/heads/main' && github.ref != 'refs/heads/testMain'
        with:
          images: |
            ${{ secrets.DOCKER_HUB_USERNAME }}/s-pdf
@@ -170,7 +193,7 @@ jobs:
      - name: Build and push Unified Dockerfile (fat variant)
        id: build-push-fat
        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
-        if: github.ref != 'refs/heads/main' && github.ref != 'refs/heads/testMain'
+        if: env.RUN_MAIN_APP == 'true' && github.ref != 'refs/heads/main' && github.ref != 'refs/heads/testMain' && steps.meta-fat.outputs.tags != ''
        with:
          builder: ${{ steps.buildx.outputs.name }}
          context: .
@@ -186,7 +209,7 @@ jobs:
          sbom: true

      - name: Sign fat images
-        if: github.ref == 'refs/heads/master' || github.ref == 'refs/heads/V2-master'
+        if: env.RUN_MAIN_APP == 'true' && (github.ref == 'refs/heads/master' || github.ref == 'refs/heads/V2-master') && steps.build-push-fat.outputs.digest != ''
        env:
          DIGEST: ${{ steps.build-push-fat.outputs.digest }}
          TAGS: ${{ steps.meta-fat.outputs.tags }}
@@ -200,7 +223,7 @@ jobs:
      - name: Generate tags for ultra-lite
        id: meta-lite
        uses: docker/metadata-action@030e881283bb7a6894de51c315a6bfe6a94e05cf # v6.0.0
-        if: github.ref != 'refs/heads/main' && github.ref != 'refs/heads/testMain'
+        if: env.RUN_MAIN_APP == 'true' && github.ref != 'refs/heads/main' && github.ref != 'refs/heads/testMain'
        with:
          images: |
            ${{ secrets.DOCKER_HUB_USERNAME }}/s-pdf
@@ -214,7 +237,7 @@ jobs:
      - name: Build and push Unified Dockerfile (ultra-lite variant)
        id: build-push-lite
        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
-        if: github.ref != 'refs/heads/main' && github.ref != 'refs/heads/testMain'
+        if: env.RUN_MAIN_APP == 'true' && github.ref != 'refs/heads/main' && github.ref != 'refs/heads/testMain' && steps.meta-lite.outputs.tags != ''
        with:
          builder: ${{ steps.buildx.outputs.name }}
          context: .
@@ -230,7 +253,7 @@ jobs:
          sbom: true

      - name: Sign ultra-lite images
-        if: github.ref == 'refs/heads/master' || github.ref == 'refs/heads/V2-master'
+        if: env.RUN_MAIN_APP == 'true' && (github.ref == 'refs/heads/master' || github.ref == 'refs/heads/V2-master') && steps.build-push-lite.outputs.digest != ''
        env:
          DIGEST: ${{ steps.build-push-lite.outputs.digest }}
          TAGS: ${{ steps.meta-lite.outputs.tags }}
@@ -240,3 +263,136 @@ jobs:
          echo "$TAGS" | tr ',' '\n' | while read -r tag; do
            cosign sign --key env://COSIGN_PRIVATE_KEY --yes "${tag}@${DIGEST}"
          done
+
+      # Standalone unoserver image — versioned independently via
+      # docker/unoserver/VERSION. master/V2-master: publish <version>+latest
+      # only when the version is new. main/testMain: republish :alpha only
+      # when the source hash differs from the published image's annotation.
+      - name: Read unoserver image version
+        id: unoserverVersion
+        if: env.RUN_UNOSERVER == 'true'
+        run: |
+          version=$(tr -d '[:space:]' < docker/unoserver/VERSION)
+          if [ -z "$version" ]; then
+            echo "docker/unoserver/VERSION is empty"; exit 1
+          fi
+          echo "version=${version}" >> "$GITHUB_OUTPUT"
+          echo "Unoserver image version (from file): ${version}"
+
+      - name: Compute unoserver image source hash
+        id: unoserverHash
+        if: env.RUN_UNOSERVER == 'true'
+        run: |
+          set -eu
+          hash=$(cat \
+            docker/unoserver/Dockerfile \
+            docker/unoserver/entrypoint.sh \
+            docker/unoserver/healthcheck.sh \
+            docker/unoserver/VERSION \
+            | sha256sum | cut -d' ' -f1)
+          echo "hash=${hash}" >> "$GITHUB_OUTPUT"
+          echo "Unoserver source hash: ${hash}"
+
+      - name: Decide whether to publish unoserver image
+        id: unoserverDecision
+        if: env.RUN_UNOSERVER == 'true'
+        env:
+          UNOSERVER_VERSION: ${{ steps.unoserverVersion.outputs.version }}
+          UNOSERVER_HASH: ${{ steps.unoserverHash.outputs.hash }}
+          UNOSERVER_IMAGE: ghcr.io/${{ steps.repoowner.outputs.lowercase }}/stirling-unoserver
+          UNOSERVER_HASH_ANNOTATION: org.stirlingpdf.unoserver-source-hash
+          FORCE_REBUILD: ${{ inputs.force_unoserver_rebuild }}
+          GH_REF: ${{ github.ref }}
+          EVENT_NAME: ${{ github.event_name }}
+        run: |
+          set -eu
+          mode="skip"
+          tags=""
+
+          read_published_hash() {
+            local ref="$1"
+            docker buildx imagetools inspect "$ref" --raw 2>/dev/null \
+              | jq -r --arg key "$UNOSERVER_HASH_ANNOTATION" \
+                  '.annotations[$key] // empty' \
+              2>/dev/null || true
+          }
+
+          # Manual dispatch from any branch routes to the :alpha publish path.
+          EFFECTIVE_REF="$GH_REF"
+          if [ "$EVENT_NAME" = "workflow_dispatch" ]; then
+            EFFECTIVE_REF="refs/heads/testMain"
+          fi
+
+          case "$EFFECTIVE_REF" in
+            refs/heads/master|refs/heads/V2-master)
+              if [ "${FORCE_REBUILD}" = "true" ]; then
+                echo "force_unoserver_rebuild=true — building stable regardless"
+                mode="stable"
+                tags="${UNOSERVER_IMAGE}:${UNOSERVER_VERSION},${UNOSERVER_IMAGE}:latest"
+              elif docker manifest inspect "${UNOSERVER_IMAGE}:${UNOSERVER_VERSION}" >/dev/null 2>&1; then
+                echo "stirling-unoserver:${UNOSERVER_VERSION} already on GHCR — skipping"
+              else
+                echo "stirling-unoserver:${UNOSERVER_VERSION} is new — will publish"
+                mode="stable"
+                tags="${UNOSERVER_IMAGE}:${UNOSERVER_VERSION},${UNOSERVER_IMAGE}:latest"
+              fi
+              ;;
+            refs/heads/main|refs/heads/testMain)
+              published_hash=$(read_published_hash "${UNOSERVER_IMAGE}:alpha")
+              if [ "${FORCE_REBUILD}" = "true" ]; then
+                echo "force_unoserver_rebuild=true — rebuilding :alpha regardless"
+                mode="alpha"
+                tags="${UNOSERVER_IMAGE}:alpha"
+              elif [ -n "$published_hash" ] && [ "$published_hash" = "$UNOSERVER_HASH" ]; then
+                echo "Published :alpha source hash matches (${published_hash}) — skipping"
+              else
+                if [ -z "$published_hash" ]; then
+                  echo ":alpha has no source-hash annotation (first publish or pre-tracking image) — will publish"
+                else
+                  echo "Source hash changed (was ${published_hash}, now ${UNOSERVER_HASH}) — will publish"
+                fi
+                mode="alpha"
+                tags="${UNOSERVER_IMAGE}:alpha"
+              fi
+              ;;
+            *)
+              echo "Branch ${GH_REF} does not publish unoserver image"
+              ;;
+          esac
+          echo "mode=${mode}" >> "$GITHUB_OUTPUT"
+          echo "tags=${tags}" >> "$GITHUB_OUTPUT"
+
+      - name: Build and push unoserver image
+        id: build-push-unoserver
+        if: env.RUN_UNOSERVER == 'true' && steps.unoserverDecision.outputs.mode != 'skip'
+        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
+        with:
+          builder: ${{ steps.buildx.outputs.name }}
+          context: .
+          file: ./docker/unoserver/Dockerfile
+          push: true
+          cache-from: type=gha,scope=stirling-unoserver
+          cache-to: type=gha,mode=max,scope=stirling-unoserver
+          tags: ${{ steps.unoserverDecision.outputs.tags }}
+          # Manifest annotation read by the decision step above to detect drift.
+          annotations: |
+            index:org.stirlingpdf.unoserver-source-hash=${{ steps.unoserverHash.outputs.hash }}
+          platforms: linux/amd64,linux/arm64/v8
+          provenance: true
+          sbom: true
+
+      - name: Sign unoserver image
+        if: env.RUN_UNOSERVER == 'true' && steps.unoserverDecision.outputs.mode == 'stable'
+        env:
+          DIGEST: ${{ steps.build-push-unoserver.outputs.digest }}
+          TAGS: ${{ steps.unoserverDecision.outputs.tags }}
+          COSIGN_PRIVATE_KEY: ${{ secrets.COSIGN_PRIVATE_KEY }}
+          COSIGN_PASSWORD: ${{ secrets.COSIGN_PASSWORD }}
+        run: |
+          if [ -n "$COSIGN_PRIVATE_KEY" ]; then
+            echo "$TAGS" | tr ',' '\n' | while read -r tag; do
+              cosign sign --key env://COSIGN_PRIVATE_KEY --yes "${tag}@${DIGEST}"
+            done
+          else
+            echo "Warning: COSIGN_PRIVATE_KEY not set, skipping unoserver image signing"
+          fi
@@ -18,7 +18,7 @@ jobs:
      packages: write
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -35,7 +35,7 @@ jobs:

    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -75,6 +75,6 @@ jobs:

      # Upload the results to GitHub's code scanning dashboard.
      - name: "Upload to code-scanning"
-        uses: github/codeql-action/upload-sarif@95e58e9a2cdfd71adc6e0353d5c52f41a045d225 # v3.29.5
+        uses: github/codeql-action/upload-sarif@e46ed2cbd01164d986452f91f178727624ae40d7 # v3.29.5
        with:
          sarif_file: results.sarif
@@ -17,7 +17,7 @@ jobs:
      pull-requests: write
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -27,7 +27,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -35,7 +35,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -37,7 +37,7 @@ jobs:
      matrix: ${{ steps.set-matrix.outputs.matrix }}
    steps:
      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -86,7 +86,7 @@ jobs:
      RELEASE_GPG_PRIVATE_KEY: ${{ secrets.RELEASE_GPG_PRIVATE_KEY }}
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -557,7 +557,7 @@ jobs:
      pull-requests: write
    steps:
      - name: Harden the runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -648,7 +648,7 @@ jobs:
    if: always()
    steps:
      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -48,7 +48,7 @@ jobs:
            cache-scope: stirling-pdf-fat
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -80,7 +80,7 @@ jobs:
          distribution: "temurin"

      - name: Cache Gradle dependency artifacts
-        uses: actions/cache@668228422ae6a00e4ad889ee87cd7109ec5666a7 # v5.0.4
+        uses: actions/cache@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5
        with:
          path: |
            ~/.gradle/wrapper
@@ -160,3 +160,58 @@ jobs:
            build/reports/problems/
          retention-days: 3
          if-no-files-found: warn
+
+  test-build-unoserver-image:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Harden Runner
+        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        with:
+          egress-policy: audit
+
+      - name: Checkout Repository
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@ce360397dd3f832beb865e1373c09c0e9f86d70a # v4.0.0
+
+      - name: Set up Docker Buildx
+        id: buildx
+        uses: docker/setup-buildx-action@4d04d5d9486b7bd6fa91e7baf45bbb4f8b9deedd # v4.0.0
+
+      - name: Build docker/unoserver/Dockerfile
+        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
+        with:
+          builder: ${{ steps.buildx.outputs.name }}
+          context: .
+          file: ./docker/unoserver/Dockerfile
+          push: false
+          load: true
+          cache-from: type=gha,scope=stirling-unoserver
+          cache-to: type=gha,mode=max,scope=stirling-unoserver
+          platforms: linux/amd64
+          tags: stirling-unoserver:pr-test
+          provenance: false
+          sbom: false
+
+      - name: Smoke test the built image
+        run: |
+          set -eu
+          docker run -d --name unoserver-smoke \
+            -e UNOSERVER_RECYCLE_INTERVAL_SECONDS=0 \
+            stirling-unoserver:pr-test
+          deadline=$((SECONDS + 60))
+          while [ $SECONDS -lt $deadline ]; do
+            status=$(docker inspect -f '{{.State.Health.Status}}' unoserver-smoke 2>/dev/null || echo "starting")
+            if [ "$status" = "healthy" ]; then
+              echo "unoserver became healthy"
+              docker logs unoserver-smoke | tail -30
+              docker rm -f unoserver-smoke
+              exit 0
+            fi
+            sleep 3
+          done
+          echo "unoserver did not become healthy in time"
+          docker logs unoserver-smoke || true
+          docker rm -f unoserver-smoke || true
+          exit 1
@@ -25,7 +25,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -131,7 +131,7 @@ jobs:
      frontend: ${{ steps.changes.outputs.frontend }}
    steps:
      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -150,7 +150,7 @@ jobs:

    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -186,7 +186,7 @@ jobs:

    steps:
      - name: Harden Runner
-        uses: step-security/harden-runner@8d3c67de8e2fe68ef647c8db1e6a09f647780f40 # v2.19.0
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
        with:
          egress-policy: audit

@@ -32,6 +32,7 @@ exampleYmlFiles/stirling/
 /testing/cucumber/junit/
 /testing/cucumber/report.html
 /testing/.failed_tests
+/.test-state/
 SwaggerDoc.json

 # Runtime storage for uploaded files and user data (not Java source code)
@@ -1,5 +1,8 @@
 version: '3'

+# All gradle invocations go through `bash gradlew` (the sh wrapper) so they work
+# uniformly on Linux/macOS (native bash) and Windows-with-Git-Bash.
+
 tasks:
  dev:
    desc: "Start backend dev server"
@@ -10,58 +13,43 @@ tasks:
    env:
      SERVER_PORT: '{{.PORT}}'
    cmds:
-      - cmd: '{{if .AIENGINE_URL}}set "AIENGINE_URL={{.AIENGINE_URL}}" && set "AIENGINE_ENABLED=true" && {{end}}cmd /c gradlew.bat :stirling-pdf:bootRun'
-        platforms: [windows]
-      - cmd: '{{if .AIENGINE_URL}}AIENGINE_URL={{.AIENGINE_URL}} AIENGINE_ENABLED=true {{end}}./gradlew :stirling-pdf:bootRun'
-        platforms: [linux, darwin]
+      - '{{if .AIENGINE_URL}}AIENGINE_URL={{.AIENGINE_URL}} AIENGINE_ENABLED=true {{end}}bash gradlew :stirling-pdf:bootRun'
+
+  dev:bundled:
+    desc: "Clean + bootRun with frontend bundled into the backend (single :8080 server)"
+    ignore_error: true
+    cmds:
+      - bash gradlew clean bootRun -PbuildWithFrontend=true

  build:
    desc: "Full backend build"
    cmds:
-      - cmd: cmd /c gradlew.bat clean build
-        platforms: [windows]
-      - cmd: ./gradlew clean build
-        platforms: [linux, darwin]
+      - bash gradlew clean build

  build:fast:
    desc: "Build without tests"
    cmds:
-      - cmd: cmd /c gradlew.bat clean build -x test
-        platforms: [windows]
-      - cmd: ./gradlew clean build -x test
-        platforms: [linux, darwin]
+      - bash gradlew clean build -x test

  build:ci:
    desc: "Build for CI (formatting checked separately)"
    cmds:
-      - cmd: cmd /c gradlew.bat build -PnoSpotless
-        platforms: [windows]
-      - cmd: ./gradlew build -PnoSpotless
-        platforms: [linux, darwin]
+      - bash gradlew build -PnoSpotless

  test:
    desc: "Run backend tests"
    cmds:
-      - cmd: cmd /c gradlew.bat test
-        platforms: [windows]
-      - cmd: ./gradlew test
-        platforms: [linux, darwin]
+      - bash gradlew test

  format:
    desc: "Auto-fix code formatting"
    cmds:
-      - cmd: cmd /c gradlew.bat spotlessApply
-        platforms: [windows]
-      - cmd: ./gradlew spotlessApply
-        platforms: [linux, darwin]
+      - bash gradlew spotlessApply

  format:check:
    desc: "Check code formatting"
    cmds:
-      - cmd: cmd /c gradlew.bat spotlessCheck
-        platforms: [windows]
-      - cmd: ./gradlew spotlessCheck
-        platforms: [linux, darwin]
+      - bash gradlew spotlessCheck

  fix:
    desc: "Auto-fix backend"
@@ -71,10 +59,7 @@ tasks:
  swagger:
    desc: "Generate OpenAPI docs"
    cmds:
-      - cmd: cmd /c gradlew.bat :stirling-pdf:copySwaggerDoc
-        platforms: [windows]
-      - cmd: ./gradlew :stirling-pdf:copySwaggerDoc
-        platforms: [linux, darwin]
+      - bash gradlew :stirling-pdf:copySwaggerDoc
    sources:
      - app/core/src/main/java/**/*.java
      - app/proprietary/src/main/java/**/*.java
@@ -92,31 +77,19 @@ tasks:
    desc: "Print project version"
    silent: true
    cmds:
-      - cmd: cmd /c gradlew.bat printVersion --quiet | tail -1
-        platforms: [windows]
-      - cmd: ./gradlew printVersion --quiet | tail -1
-        platforms: [linux, darwin]
+      - bash gradlew printVersion --quiet | tail -1

  licenses:check:
    desc: "Check dependency licenses"
    cmds:
-      - cmd: cmd /c gradlew.bat checkLicense --no-parallel
-        platforms: [windows]
-      - cmd: ./gradlew checkLicense --no-parallel
-        platforms: [linux, darwin]
+      - bash gradlew checkLicense --no-parallel

  licenses:generate:
    desc: "Check and generate dependency license report"
    cmds:
-      - cmd: cmd /c gradlew.bat checkLicense generateLicenseReport --no-parallel
-        platforms: [windows]
-      - cmd: ./gradlew checkLicense generateLicenseReport --no-parallel
-        platforms: [linux, darwin]
+      - bash gradlew checkLicense generateLicenseReport --no-parallel

  clean:
    desc: "Clean build artifacts"
    cmds:
-      - cmd: cmd /c gradlew.bat clean
-        platforms: [windows]
-      - cmd: ./gradlew clean
-        platforms: [linux, darwin]
+      - bash gradlew clean
@@ -1,12 +1,7 @@
 version: '3'

 vars:
-  # Single source of truth for the jlink module list (also read by
-  # frontend/scripts/build-universal-mac-jre.sh during CI). The desktop
-  # taskfile runs with dir: frontend (see Taskfile.yml), so this path
-  # is relative to that directory.
-  JLINK_MODULES:
-    sh: tr -d '\n' < src-tauri/jlink-modules.txt
+  JLINK_MODULES: "java.base,java.compiler,java.desktop,java.instrument,java.logging,java.management,java.naming,java.net.http,java.prefs,java.rmi,java.scripting,java.security.jgss,java.security.sasl,java.sql,java.transaction.xa,java.xml,java.xml.crypto,jdk.crypto.ec,jdk.crypto.cryptoki,jdk.unsupported"

 tasks:
  prepare:
@@ -55,3 +55,9 @@ tasks:
    desc: "Tail docker compose logs"
    cmds:
      - docker compose -f {{.COMPOSE_DIR}}/docker-compose.yml logs -f
+
+  test:
+    desc: "Run full Docker integration test suite (builds all variants and tests them)"
+    ignore_error: true
+    cmds:
+      - bash testing/test.sh {{.CLI_ARGS}}
@@ -0,0 +1,204 @@
+version: '3'
+
+tasks:
+  install:
+    desc: "Install Playwright browsers"
+    dir: frontend
+    deps: [ ':frontend:install' ]
+    cmds:
+      - npx playwright install {{.CLI_ARGS}} --with-deps
+
+  stubbed:
+    desc: "Run stubbed E2E tests"
+    dir: frontend
+    deps: [ ':frontend:prepare' ]
+    cmds:
+      - npx playwright test --project=stubbed {{.CLI_ARGS}}
+
+  live:
+    desc: "Run live E2E tests"
+    summary: |
+      Auto-spawns a Spring Boot backend with isolated state under
+      .test-state/playwright/ (purged on every run) so the suite doesn't
+      touch your local DB, settings.yml, backups or customFiles. The
+      backend is stopped automatically when the runner exits.
+
+      Pass extra Playwright flags via -- :
+        task e2e:live -- --headed --grep merge
+    deps:
+      - live:backend
+      - live:runner
+
+  live:backend:
+    internal: true
+    ignore_error: true
+    vars:
+      BASE_DIR: '{{.ROOT_DIR}}/.test-state/playwright'
+    env:
+      STIRLING_BASE_PATH: '{{.BASE_DIR}}'
+      # Suppress the analytics opt-in modal that fires on first admin login.
+      # It renders a Mantine overlay that intercepts pointer events and blocks
+      # FirstLoginSlide from rendering, so the bootstrap spec times out waiting
+      # for the password-change prompt.
+      SYSTEM_ENABLEANALYTICS: "false"
+      # NOTE: SECURITY_INITIALLOGIN_USERNAME/PASSWORD are intentionally NOT set.
+      # The live-setup project's bootstrap spec performs the real first-login
+      # flow against the backend's default admin/stirling user, exercising the
+      # forced-password-change UI and leaving the DB at admin/adminadmin for
+      # the rest of the live suite. This is both real coverage of the first-
+      # login flow and a stronger seed than env-var-driven user creation.
+    cmds:
+      # Wrapped in `bash -c` because Task's embedded shell (mvdan/sh) doesn't
+      # support `&` + `$!` + `wait` reliably (gradle was never actually
+      # backgrounded, so `$!` recorded a stale PID and the runner saw an
+      # immediate "backend died"). Real bash backgrounds gradle properly and
+      # gives us a stable PID to clean up.
+      - |
+        bash -c '
+          set -e
+          rm -rf "{{.BASE_DIR}}"
+          mkdir -p "{{.BASE_DIR}}"
+          # Background gradle and record its PID so the runner can clean up
+          # the exact process tree (wrapper + forked Spring Boot JVM) without
+          # resorting to fuzzy `pkill -f` patterns. `wait` keeps this script
+          # alive for the lifetime of gradle so Task'"'"'s parallel deps stay
+          # synchronised.
+          bash gradlew :stirling-pdf:bootRun > "{{.BASE_DIR}}/backend.log" 2>&1 &
+          GRADLE_PID=$!
+          echo $GRADLE_PID > "{{.BASE_DIR}}/backend.pid"
+          wait $GRADLE_PID
+        '
+
+  live:runner:
+    internal: true
+    deps: [ ':frontend:prepare' ]
+    dir: frontend
+    vars:
+      BASE_DIR: '{{.ROOT_DIR}}/.test-state/playwright'
+    cmds:
+      # Wrapped in `bash -c` because Task's embedded shell (mvdan/sh) only
+      # supports `EXIT`/`ERR` in `trap`, not `INT`/`TERM`. CLI_ARGS are
+      # passed positionally so quoted args like --grep "foo bar" survive.
+      - |
+        bash -c '
+          set +e
+          # bootRun forks a separate Spring Boot JVM as a child of the gradle
+          # wrapper; killing only the wrapper leaves that JVM orphaned holding
+          # :8080. Walk the recorded PID'"'"'s process tree via pgrep so we kill
+          # every descendant, then lsof as a last-resort backstop.
+          kill_tree() {
+            local pid=$1
+            for child in $(pgrep -P "$pid" 2>/dev/null); do
+              kill_tree "$child"
+            done
+            kill "$pid" 2>/dev/null || true
+          }
+          cleanup() {
+            if [ -f "{{.BASE_DIR}}/backend.pid" ]; then
+              echo "Stopping backend..."
+              kill_tree "$(cat "{{.BASE_DIR}}/backend.pid")"
+              sleep 1
+              lsof -ti:8080 2>/dev/null | xargs kill 2>/dev/null || true
+              echo "Backend stopped"
+            fi
+          }
+          # Trap so cleanup runs on every exit path: tests pass, tests fail,
+          # backend never starts, user Ctrl-Cs. Without this the gradle JVM
+          # is left orphaned holding :8080 whenever the wait loop times out.
+          trap cleanup EXIT INT TERM
+
+          echo "Waiting for backend on :8080 (up to 10min; first compile can be slow)..."
+          WAITED=0
+          while [ $WAITED -lt 600 ]; do
+            if curl -sf http://localhost:8080/api/v1/info/status 2>/dev/null | grep -q UP; then
+              echo "Backend ready, starting Playwright"
+              break
+            fi
+            # Bail early if the backend already died: no point waiting 10min.
+            if [ -f "{{.BASE_DIR}}/backend.pid" ]; then
+              BACKEND_PID=$(cat "{{.BASE_DIR}}/backend.pid")
+              if ! kill -0 "$BACKEND_PID" 2>/dev/null; then
+                echo "Backend process exited before becoming ready; last 100 log lines:"
+                tail -100 "{{.BASE_DIR}}/backend.log" 2>/dev/null || true
+                exit 1
+              fi
+            fi
+            sleep 2
+            WAITED=$((WAITED + 2))
+          done
+          if [ $WAITED -ge 600 ]; then
+            echo "Backend did not become ready within 10min, aborting"
+            exit 1
+          fi
+
+          npx playwright test --project=live "$@"
+        ' bash {{.CLI_ARGS}}
+
+  enterprise:
+    desc: "Run enterprise E2E tests"
+    summary: |
+      Requires an already-running keycloak compose stack on :8080. Bring one
+      up first with:
+        task e2e:oauth:up
+        task e2e:saml:up
+    dir: frontend
+    deps: [ ':frontend:prepare' ]
+    cmds:
+      - npx playwright test --project=enterprise {{.CLI_ARGS}}
+
+  cross-browser:
+    desc: "Run stubbed E2E tests on Chromium, Firefox, and WebKit"
+    dir: frontend
+    deps: [ ':frontend:prepare' ]
+    cmds:
+      - npx playwright test --project=stubbed --project=stubbed-firefox --project=stubbed-webkit {{.CLI_ARGS}}
+
+  check:
+    desc: "Run all E2E tests not requiring Docker"
+    cmds:
+      - task: stubbed
+      - task: live
+
+  check:all:
+    desc: "Run all E2E tests"
+    summary: |
+      Includes the enterprise project, which requires a keycloak compose
+      stack on :8080. Bring one up first with:
+        task e2e:oauth:up
+        task e2e:saml:up
+    cmds:
+      - task: stubbed
+      - task: live
+      - task: enterprise
+
+  oauth:up:
+    desc: "Start the OAuth keycloak test environment"
+    summary: |
+      Set LICENSE_KEY=<KEY> to skip the interactive license prompt:
+        task e2e:oauth:up LICENSE_KEY=abc123
+      Pass extra flags via -- :
+        task e2e:oauth:up -- --auto --nobuild
+    ignore_error: true
+    cmds:
+      - bash testing/compose/start-oauth-test.sh {{if .LICENSE_KEY}}--license-key "{{.LICENSE_KEY}}"{{end}} {{.CLI_ARGS}}
+
+  oauth:down:
+    desc: "Stop the OAuth keycloak test environment"
+    cmds:
+      - docker compose -f testing/compose/docker-compose-keycloak-oauth.yml down -v
+
+  saml:up:
+    desc: "Start the SAML keycloak test environment"
+    summary: |
+      Set LICENSE_KEY=<KEY> to skip the interactive license prompt:
+        task e2e:saml:up LICENSE_KEY=abc123
+      Pass extra flags via -- :
+        task e2e:saml:up -- --auto --with-storage --nobuild --language sv-SE
+    ignore_error: true
+    cmds:
+      - bash testing/compose/start-saml-test.sh {{if .LICENSE_KEY}}--license-key "{{.LICENSE_KEY}}"{{end}} {{.CLI_ARGS}}
+
+  saml:down:
+    desc: "Stop the SAML keycloak test environment"
+    cmds:
+      - docker compose -f testing/compose/docker-compose-keycloak-saml.yml down -v
@@ -75,7 +75,7 @@ tasks:
    desc: "Run type checking"
    deps: [install]
    cmds:
-      - uv run pyright . --warnings
+      - uv run pyrefly check

  test:
    desc: "Run tests"
@@ -284,18 +284,6 @@ tasks:
    cmds:
      - npx vitest --coverage

-  test:e2e:
-    desc: "Run E2E tests"
-    deps: [prepare]
-    cmds:
-      - npx playwright test {{.CLI_ARGS}}
-
-  test:e2e:install:
-    desc: "Install E2E test browsers"
-    deps: [install]
-    cmds:
-      - npx playwright install {{.CLI_ARGS}} --with-deps
-
  # ============================================================
  # Code Generation
  # ============================================================
@@ -22,6 +22,9 @@ includes:
  desktop:
    taskfile: .taskfiles/desktop.yml
    dir: frontend
+  e2e:
+    taskfile: .taskfiles/e2e.yml
+    dir: .

 tasks:
  # ============================================================
@@ -59,6 +59,18 @@ public class InstallationPathConfig {
    }

    private static String initializeBasePath() {
+        // Allow tests / harnesses to redirect the entire state tree (configs,
+        // backups, customFiles, pipeline, logs) to an isolated location via
+        // -Dstirling.base-path=... or STIRLING_BASE_PATH=... so a Playwright
+        // run never touches a developer's working state.
+        String override = System.getProperty("stirling.base-path");
+        if (override == null || override.isBlank()) {
+            override = System.getenv("STIRLING_BASE_PATH");
+        }
+        if (override != null && !override.isBlank()) {
+            boolean hasTrailingSeparator = override.endsWith("/") || override.endsWith("\\");
+            return hasTrailingSeparator ? override : override + File.separator;
+        }
        return "." + File.separator;
    }

@@ -76,6 +76,7 @@ public class ApplicationProperties {
    private ProcessExecutor processExecutor = new ProcessExecutor();
    private PdfEditor pdfEditor = new PdfEditor();
    private AiEngine aiEngine = new AiEngine();
+    private InternalApi internalApi = new InternalApi();

    @Bean
    public PropertySource<?> dynamicYamlPropertySource(ConfigurableEnvironment environment)
@@ -246,6 +247,19 @@ public class ApplicationProperties {
        private int longRunningTimeoutSeconds = 600;
    }

+    /**
+     * HTTP timeouts for loopback calls to internal Stirling API endpoints, used by the AI workflow
+     * executor and the pipeline processor. A bounded read timeout prevents a hung tool (e.g. an
+     * infinite loop in a PDF processing service) from stalling the entire chat workflow forever.
+     * Tools that legitimately need longer than the read timeout should be invoked through the async
+     * job executor instead of synchronously.
+     */
+    @Data
+    public static class InternalApi {
+        private int connectTimeoutSeconds = 10;
+        private int readTimeoutSeconds = 300;
+    }
+
    @Data
    public static class Legal {
        private String termsAndConditions;
@@ -0,0 +1,19 @@
+package stirling.software.common.model.api.general;
+
+import io.swagger.v3.oas.annotations.media.Schema;
+
+import lombok.Data;
+import lombok.EqualsAndHashCode;
+
+@Data
+@EqualsAndHashCode
+public class EditTextOperation {
+
+    @Schema(description = "The literal text to find.", requiredMode = Schema.RequiredMode.REQUIRED)
+    private String find;
+
+    @Schema(
+            description = "The replacement text. May be empty to delete the matched text.",
+            requiredMode = Schema.RequiredMode.REQUIRED)
+    private String replace;
+}
@@ -5,6 +5,7 @@ import java.io.UncheckedIOException;
 import java.net.URLDecoder;
 import java.nio.charset.StandardCharsets;
 import java.nio.file.Files;
+import java.time.Duration;
 import java.util.regex.Pattern;

 import org.springframework.beans.factory.annotation.Autowired;
@@ -12,15 +13,18 @@ import org.springframework.core.env.Environment;
 import org.springframework.core.io.FileSystemResource;
 import org.springframework.core.io.Resource;
 import org.springframework.http.*;
+import org.springframework.http.client.SimpleClientHttpRequestFactory;
 import org.springframework.stereotype.Service;
 import org.springframework.util.MultiValueMap;
 import org.springframework.web.client.RequestCallback;
+import org.springframework.web.client.ResourceAccessException;
 import org.springframework.web.client.RestTemplate;

 import jakarta.servlet.ServletContext;

 import lombok.extern.slf4j.Slf4j;

+import stirling.software.common.model.ApplicationProperties;
 import stirling.software.common.model.enumeration.Role;
 import stirling.software.common.util.TempFile;
 import stirling.software.common.util.TempFileManager;
@@ -50,16 +54,29 @@ public class InternalApiClient {
    private final UserServiceInterface userService;
    private final TempFileManager tempFileManager;
    private final Environment environment;
+    private final Duration readTimeout;
+    private final RestTemplate restTemplate;

    public InternalApiClient(
            ServletContext servletContext,
            @Autowired(required = false) UserServiceInterface userService,
            TempFileManager tempFileManager,
-            Environment environment) {
+            Environment environment,
+            ApplicationProperties applicationProperties) {
        this.servletContext = servletContext;
        this.userService = userService;
        this.tempFileManager = tempFileManager;
        this.environment = environment;
+        ApplicationProperties.InternalApi internalApi = applicationProperties.getInternalApi();
+        // A bounded read timeout is what protects the workflow when an internal tool hangs
+        // (e.g. an infinite loop in a PDF processing service). The connect timeout is short
+        // because this is a loopback call; if connecting takes longer than a few seconds the
+        // local server is itself unhealthy.
+        this.readTimeout = Duration.ofSeconds(internalApi.getReadTimeoutSeconds());
+        SimpleClientHttpRequestFactory factory = new SimpleClientHttpRequestFactory();
+        factory.setConnectTimeout(Duration.ofSeconds(internalApi.getConnectTimeoutSeconds()));
+        factory.setReadTimeout(readTimeout);
+        this.restTemplate = new RestTemplate(factory);
    }

    /**
@@ -74,7 +91,6 @@ public class InternalApiClient {
        validateUrl(endpointPath);
        String url = getBaseUrl() + endpointPath;

-        RestTemplate restTemplate = new RestTemplate();
        HttpHeaders headers = new HttpHeaders();
        String apiKey = getApiKeyForUser();
        if (apiKey != null && !apiKey.isEmpty()) {
@@ -84,26 +100,38 @@ public class InternalApiClient {
        HttpEntity<MultiValueMap<String, Object>> entity = new HttpEntity<>(body, headers);
        RequestCallback requestCallback = restTemplate.httpEntityCallback(entity, Resource.class);

-        return restTemplate.execute(
-                url,
-                HttpMethod.POST,
-                requestCallback,
-                response -> {
-                    try {
-                        TempFile tempFile = tempFileManager.createManagedTempFile("internal-api");
-                        Files.copy(
-                                response.getBody(),
-                                tempFile.getPath(),
-                                java.nio.file.StandardCopyOption.REPLACE_EXISTING);
-                        String filename = extractFilename(response.getHeaders());
-                        TempFileResource resource = new TempFileResource(tempFile, filename);
-                        return ResponseEntity.status(response.getStatusCode())
-                                .headers(response.getHeaders())
-                                .body(resource);
-                    } catch (IOException e) {
-                        throw new UncheckedIOException(e);
-                    }
-                });
+        try {
+            return restTemplate.execute(
+                    url,
+                    HttpMethod.POST,
+                    requestCallback,
+                    response -> {
+                        try {
+                            TempFile tempFile =
+                                    tempFileManager.createManagedTempFile("internal-api");
+                            Files.copy(
+                                    response.getBody(),
+                                    tempFile.getPath(),
+                                    java.nio.file.StandardCopyOption.REPLACE_EXISTING);
+                            String filename = extractFilename(response.getHeaders());
+                            TempFileResource resource = new TempFileResource(tempFile, filename);
+                            return ResponseEntity.status(response.getStatusCode())
+                                    .headers(response.getHeaders())
+                                    .body(resource);
+                        } catch (IOException e) {
+                            throw new UncheckedIOException(e);
+                        }
+                    });
+        } catch (ResourceAccessException e) {
+            // RestTemplate wraps low-level I/O failures in ResourceAccessException. Only the
+            // SocketTimeoutException-rooted case is a real timeout; other I/O failures (connection
+            // refused, DNS, etc.) propagate as-is so the upstream generic handler can describe
+            // them accurately.
+            if (e.getCause() instanceof java.net.SocketTimeoutException) {
+                throw new InternalApiTimeoutException(endpointPath, readTimeout, e);
+            }
+            throw e;
+        }
    }

    /**
@@ -0,0 +1,36 @@
+package stirling.software.common.service;
+
+import java.time.Duration;
+
+/**
+ * Thrown when an internal Stirling tool invocation exceeds its configured read timeout or otherwise
+ * fails at the transport layer. Distinguishes a hung/timed-out tool from a tool that returned a
+ * non-2xx HTTP response (which is reported via the response status itself), so callers can present
+ * a clear "the tool didn't respond in time" message to the user instead of a generic stack trace.
+ */
+public class InternalApiTimeoutException extends RuntimeException {
+
+    private final String endpointPath;
+    private final Duration readTimeout;
+
+    public InternalApiTimeoutException(String endpointPath, Duration readTimeout, Throwable cause) {
+        super(buildMessage(endpointPath, readTimeout, cause), cause);
+        this.endpointPath = endpointPath;
+        this.readTimeout = readTimeout;
+    }
+
+    public String getEndpointPath() {
+        return endpointPath;
+    }
+
+    public Duration getReadTimeout() {
+        return readTimeout;
+    }
+
+    private static String buildMessage(String endpointPath, Duration readTimeout, Throwable cause) {
+        String reason = cause != null && cause.getMessage() != null ? cause.getMessage() : "";
+        return String.format(
+                "Internal tool %s did not respond within %ds (%s)",
+                endpointPath, readTimeout.toSeconds(), reason);
+    }
+}
@@ -15,6 +15,7 @@ import java.util.Set;
 import java.util.concurrent.ConcurrentHashMap;
 import java.util.concurrent.Semaphore;
 import java.util.concurrent.TimeUnit;
+import java.util.concurrent.TimeoutException;

 import io.github.pixee.security.BoundedLineReader;

@@ -202,7 +203,15 @@ public class ProcessExecutor {
        boolean useSemaphore = true;
        List<String> commandToRun = command;
        if (shouldUseUnoServerPool(command)) {
-            unoLease = unoServerPool.acquireEndpoint();
+            try {
+                unoLease = unoServerPool.acquireEndpoint(timeoutDuration, TimeUnit.MINUTES);
+            } catch (TimeoutException e) {
+                throw new IOException(
+                        "All unoserver endpoints busy; request timed out after "
+                                + timeoutDuration
+                                + " minutes",
+                        e);
+            }
            commandToRun = applyUnoServerEndpoint(command, unoLease.getEndpoint());
            useSemaphore = false;
        }
@@ -3,8 +3,11 @@ package stirling.software.common.util;
 import java.util.ArrayList;
 import java.util.Collections;
 import java.util.List;
+import java.util.Locale;
 import java.util.concurrent.BlockingQueue;
 import java.util.concurrent.LinkedBlockingQueue;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.TimeoutException;
 import java.util.concurrent.atomic.AtomicBoolean;

 import stirling.software.common.model.ApplicationProperties;
@@ -42,6 +45,27 @@ public class UnoServerPool {
        return new UnoServerLease(endpoints.get(index), index, this);
    }

+    /** Fail-fast variant; non-positive timeout falls back to unbounded acquire. */
+    public UnoServerLease acquireEndpoint(long timeout, TimeUnit unit)
+            throws InterruptedException, TimeoutException {
+        if (endpoints.isEmpty()) {
+            return new UnoServerLease(defaultEndpoint(), null, this);
+        }
+        if (timeout <= 0) {
+            return acquireEndpoint();
+        }
+
+        Integer index = availableIndices.poll(timeout, unit);
+        if (index == null) {
+            throw new TimeoutException(
+                    "Timed out waiting for a free unoserver endpoint after "
+                            + timeout
+                            + " "
+                            + unit.name().toLowerCase(Locale.ROOT));
+        }
+        return new UnoServerLease(endpoints.get(index), index, this);
+    }
+
    private void releaseEndpoint(Integer index) {
        if (index != null) {
            availableIndices.offer(index);
@@ -6,15 +6,18 @@ import java.util.List;

 import lombok.extern.slf4j.Slf4j;

-import stirling.software.common.model.api.security.RedactionArea;
-
-import tools.jackson.core.type.TypeReference;
 import tools.jackson.databind.DeserializationFeature;
+import tools.jackson.databind.JavaType;
 import tools.jackson.databind.ObjectMapper;
 import tools.jackson.databind.json.JsonMapper;

+/**
+ * Spring property editor that decodes a JSON string into a typed {@link ArrayList}. Used to bind
+ * complex list parameters (e.g. {@code List<RedactionArea>}, {@code List<EditTextOperation>}) from
+ * multipart form fields, where Spring's default binding cannot deserialize a JSON array.
+ */
@Slf4j
-public class StringToArrayListPropertyEditor extends PropertyEditorSupport {
+public class StringToArrayListPropertyEditor<T> extends PropertyEditorSupport {

    private final ObjectMapper objectMapper =
            JsonMapper.builder()
@@ -22,6 +25,12 @@ public class StringToArrayListPropertyEditor extends PropertyEditorSupport {
                    .disable(DeserializationFeature.FAIL_ON_UNKNOWN_PROPERTIES)
                    .build();

+    private final Class<T> elementType;
+
+    public StringToArrayListPropertyEditor(Class<T> elementType) {
+        this.elementType = elementType;
+    }
+
    @Override
    public void setAsText(String text) throws IllegalArgumentException {
        if (text == null || text.trim().isEmpty()) {
@@ -29,8 +38,11 @@ public class StringToArrayListPropertyEditor extends PropertyEditorSupport {
            return;
        }
        try {
-            TypeReference<ArrayList<RedactionArea>> typeRef = new TypeReference<>() {};
-            List<RedactionArea> list = objectMapper.readValue(text, typeRef);
+            JavaType listType =
+                    objectMapper
+                            .getTypeFactory()
+                            .constructCollectionType(ArrayList.class, elementType);
+            List<T> list = objectMapper.readValue(text, listType);
            setValue(list);
        } catch (Exception e) {
            log.error("Exception while converting {}", e);
@@ -22,11 +22,13 @@ import org.springframework.mock.env.MockEnvironment;
 import org.springframework.util.LinkedMultiValueMap;
 import org.springframework.util.MultiValueMap;
 import org.springframework.web.client.RequestCallback;
+import org.springframework.web.client.ResourceAccessException;
 import org.springframework.web.client.ResponseExtractor;
 import org.springframework.web.client.RestTemplate;

 import jakarta.servlet.ServletContext;

+import stirling.software.common.model.ApplicationProperties;
 import stirling.software.common.util.TempFile;
 import stirling.software.common.util.TempFileManager;

@@ -42,8 +44,19 @@ class InternalApiClientTest {
    @BeforeEach
    void setUp() {
        lenient().when(servletContext.getContextPath()).thenReturn("");
+        client = newClient();
+    }
+
+    /**
+     * Build a fresh client. Tests that use {@link org.mockito.Mockito#mockConstruction} to
+     * intercept {@link RestTemplate} must call this from inside their {@code mockConstruction}
+     * block, since the client now caches one RestTemplate per instance at construction time.
+     */
+    private InternalApiClient newClient() {
        MockEnvironment environment = new MockEnvironment().withProperty("server.port", "8080");
-        client = new InternalApiClient(servletContext, userService, tempFileManager, environment);
+        ApplicationProperties applicationProperties = new ApplicationProperties();
+        return new InternalApiClient(
+                servletContext, userService, tempFileManager, environment, applicationProperties);
    }

    @Test
@@ -75,7 +88,10 @@ class InternalApiClientTest {
                                    .thenAnswer(inv -> fakeOkResponse(inv.getArgument(3)));
                        })) {

-            ResponseEntity<Resource> response = client.post("/api/v1/general/merge-pdfs", body);
+            // Reconstruct the client so its cached RestTemplate is the mocked one.
+            InternalApiClient mockedClient = newClient();
+            ResponseEntity<Resource> response =
+                    mockedClient.post("/api/v1/general/merge-pdfs", body);

            assertNotNull(response);
            assertEquals(HttpStatus.OK, response.getStatusCode());
@@ -86,6 +102,72 @@ class InternalApiClientTest {
        }
    }

+    @Test
+    void postWrapsSocketTimeoutAsInternalApiTimeoutException() {
+        // Simulates the read-timeout case: RestTemplate wraps SocketTimeoutException in
+        // ResourceAccessException when the underlying socket times out. The client must repackage
+        // that into a typed timeout exception that carries the failing endpoint and the configured
+        // read timeout, so the workflow layer can surface a clean "tool didn't respond" message
+        // to the user.
+        MultiValueMap<String, Object> body = new LinkedMultiValueMap<>();
+        body.add("fileInput", namedResource("input.pdf", "data"));
+
+        try (var ignored =
+                mockConstruction(
+                        RestTemplate.class,
+                        (rt, ctx) -> {
+                            when(rt.httpEntityCallback(any(), eq(Resource.class)))
+                                    .thenAnswer(inv -> (RequestCallback) req -> {});
+                            when(rt.execute(anyString(), eq(HttpMethod.POST), any(), any()))
+                                    .thenThrow(
+                                            new ResourceAccessException(
+                                                    "I/O error on POST request: Read timed out",
+                                                    new java.net.SocketTimeoutException(
+                                                            "Read timed out")));
+                        })) {
+
+            InternalApiClient mockedClient = newClient();
+            InternalApiTimeoutException thrown =
+                    assertThrows(
+                            InternalApiTimeoutException.class,
+                            () -> mockedClient.post("/api/v1/general/merge-pdfs", body));
+
+            assertEquals("/api/v1/general/merge-pdfs", thrown.getEndpointPath());
+            assertNotNull(thrown.getReadTimeout());
+            assertTrue(thrown.getMessage().contains("/api/v1/general/merge-pdfs"));
+        }
+    }
+
+    @Test
+    void postRethrowsNonTimeoutResourceAccessExceptionAsIs() {
+        // ResourceAccessException covers more than just timeouts (e.g. connection refused, DNS
+        // failure). Only SocketTimeoutException-rooted failures are timeouts; everything else
+        // must propagate so the upstream generic handler can label it correctly instead of lying
+        // about a "tool didn't respond" timeout.
+        MultiValueMap<String, Object> body = new LinkedMultiValueMap<>();
+        body.add("fileInput", namedResource("input.pdf", "data"));
+
+        try (var ignored =
+                mockConstruction(
+                        RestTemplate.class,
+                        (rt, ctx) -> {
+                            when(rt.httpEntityCallback(any(), eq(Resource.class)))
+                                    .thenAnswer(inv -> (RequestCallback) req -> {});
+                            when(rt.execute(anyString(), eq(HttpMethod.POST), any(), any()))
+                                    .thenThrow(
+                                            new ResourceAccessException(
+                                                    "I/O error on POST request: Connection refused",
+                                                    new java.net.ConnectException(
+                                                            "Connection refused")));
+                        })) {
+
+            InternalApiClient mockedClient = newClient();
+            assertThrows(
+                    ResourceAccessException.class,
+                    () -> mockedClient.post("/api/v1/general/merge-pdfs", body));
+        }
+    }
+
    @Test
    void postRejectsDisallowedPath() {
        MultiValueMap<String, Object> body = new LinkedMultiValueMap<>();
@@ -125,8 +207,10 @@ class InternalApiClientTest {
                                    .thenAnswer(inv -> fakeOkResponse(inv.getArgument(3)));
                        })) {

+            // Reconstruct the client so its cached RestTemplate is the mocked one.
+            InternalApiClient mockedClient = newClient();
            ResponseEntity<Resource> response =
-                    client.post("/api/v1/ai/tools/pdf-comment-agent", body);
+                    mockedClient.post("/api/v1/ai/tools/pdf-comment-agent", body);

            assertNotNull(response);
            assertEquals(HttpStatus.OK, response.getStatusCode());
@@ -9,6 +9,7 @@ import java.util.concurrent.CountDownLatch;
 import java.util.concurrent.ExecutorService;
 import java.util.concurrent.Executors;
 import java.util.concurrent.TimeUnit;
+import java.util.concurrent.TimeoutException;
 import java.util.concurrent.atomic.AtomicInteger;

 import org.junit.jupiter.api.Test;
@@ -173,6 +174,70 @@ public class UnoServerPoolTest {
        assertEquals(port1, port2, "Should reuse the same endpoint after release");
    }

+    @Test
+    void testAcquireWithTimeoutFailsFast() throws InterruptedException {
+        List<ApplicationProperties.ProcessExecutor.UnoServerEndpoint> endpoints =
+                createEndpoints(1);
+        UnoServerPool pool = new UnoServerPool(endpoints);
+
+        UnoServerPool.UnoServerLease held = pool.acquireEndpoint();
+
+        long start = System.nanoTime();
+        assertThrows(
+                TimeoutException.class,
+                () -> pool.acquireEndpoint(100, TimeUnit.MILLISECONDS),
+                "Should time out when no endpoint available");
+        long elapsedMs = (System.nanoTime() - start) / 1_000_000;
+
+        assertTrue(
+                elapsedMs >= 100 && elapsedMs < 1000,
+                "Should fail-fast after roughly the timeout, got " + elapsedMs + "ms");
+
+        held.close();
+    }
+
+    @Test
+    void testAcquireWithTimeoutSucceedsWhenAvailable()
+            throws InterruptedException, TimeoutException {
+        List<ApplicationProperties.ProcessExecutor.UnoServerEndpoint> endpoints =
+                createEndpoints(1);
+        UnoServerPool pool = new UnoServerPool(endpoints);
+
+        try (UnoServerPool.UnoServerLease lease = pool.acquireEndpoint(1, TimeUnit.SECONDS)) {
+            assertNotNull(lease);
+            assertEquals(2003, lease.getEndpoint().getPort());
+        }
+    }
+
+    @Test
+    void testAcquireWithZeroTimeoutBlocksUnbounded() throws InterruptedException {
+        List<ApplicationProperties.ProcessExecutor.UnoServerEndpoint> endpoints =
+                createEndpoints(1);
+        UnoServerPool pool = new UnoServerPool(endpoints);
+        UnoServerPool.UnoServerLease held = pool.acquireEndpoint();
+
+        AtomicInteger acquired = new AtomicInteger(0);
+        Thread t =
+                Thread.ofVirtual()
+                        .start(
+                                () -> {
+                                    try {
+                                        UnoServerPool.UnoServerLease lease =
+                                                pool.acquireEndpoint(0, TimeUnit.MILLISECONDS);
+                                        acquired.incrementAndGet();
+                                        lease.close();
+                                    } catch (Exception e) {
+                                        fail("unexpected: " + e);
+                                    }
+                                });
+
+        Thread.sleep(150);
+        assertEquals(0, acquired.get(), "Should still be blocked");
+        held.close();
+        t.join(1000);
+        assertEquals(1, acquired.get(), "Should acquire after release");
+    }
+
    @Test
    void testHostLocationAndProtocol() throws InterruptedException {
        List<ApplicationProperties.ProcessExecutor.UnoServerEndpoint> endpoints = new ArrayList<>();
@@ -11,11 +11,11 @@ import stirling.software.common.model.api.security.RedactionArea;

 class StringToArrayListPropertyEditorTest {

-    private StringToArrayListPropertyEditor editor;
+    private StringToArrayListPropertyEditor<RedactionArea> editor;

    @BeforeEach
    void setUp() {
-        editor = new StringToArrayListPropertyEditor();
+        editor = new StringToArrayListPropertyEditor<>(RedactionArea.class);
    }

    @Test
@@ -65,7 +65,7 @@ dependencies {
        exclude group: 'com.fasterxml.jackson.jaxrs'
        exclude group: 'com.fasterxml.jackson.module', module: 'jackson-module-jaxb-annotations'
    }
-    implementation 'commons-io:commons-io:2.21.0'
+    implementation 'commons-io:commons-io:2.22.0'
    implementation "org.bouncycastle:bcprov-jdk18on:$bouncycastleVersion"
    implementation "org.bouncycastle:bcpkix-jdk18on:$bouncycastleVersion"
    implementation 'io.micrometer:micrometer-core'
@@ -0,0 +1,333 @@
+package stirling.software.SPDF.controller.api;
+
+import java.io.OutputStream;
+import java.nio.file.Files;
+import java.util.ArrayList;
+import java.util.Collections;
+import java.util.HashSet;
+import java.util.List;
+import java.util.Objects;
+import java.util.Set;
+import java.util.regex.Matcher;
+import java.util.regex.Pattern;
+
+import org.springframework.core.io.Resource;
+import org.springframework.http.ResponseEntity;
+import org.springframework.web.bind.WebDataBinder;
+import org.springframework.web.bind.annotation.InitBinder;
+import org.springframework.web.bind.annotation.ModelAttribute;
+import org.springframework.web.multipart.MultipartFile;
+
+import io.github.pixee.security.Filenames;
+import io.swagger.v3.oas.annotations.Operation;
+
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+
+import stirling.software.SPDF.config.swagger.StandardPdfResponse;
+import stirling.software.SPDF.model.api.general.EditTextRequest;
+import stirling.software.SPDF.model.json.PdfJsonDocument;
+import stirling.software.SPDF.model.json.PdfJsonPage;
+import stirling.software.SPDF.model.json.PdfJsonTextElement;
+import stirling.software.SPDF.service.PdfJsonConversionService;
+import stirling.software.common.annotations.AutoJobPostMapping;
+import stirling.software.common.annotations.api.GeneralApi;
+import stirling.software.common.model.api.general.EditTextOperation;
+import stirling.software.common.util.ExceptionUtils;
+import stirling.software.common.util.GeneralUtils;
+import stirling.software.common.util.TempFile;
+import stirling.software.common.util.TempFileManager;
+import stirling.software.common.util.WebResponseUtils;
+import stirling.software.common.util.propertyeditor.StringToArrayListPropertyEditor;
+
+/**
+ * Find/replace text editing for PDFs. Round-trips through {@link PdfJsonConversionService}: the
+ * input PDF is parsed into the editable JSON model, find/replace operations are applied to the
+ * {@code text} field of each text element, and the mutated model is rebuilt into a PDF.
+ *
+ * <p>Matching joins all text elements on a page into a single string before searching, so find
+ * strings can span multiple visual runs (titles split per word, kerning-broken phrases, etc.).
+ *
+ * <p>For cross-element matches the entire replacement is written into the first matched element,
+ * any intermediate elements are emptied, and the suffix of the last matched element is preserved.
+ * This anchors the new text at the leftmost match position and lets the font lay it out as one run,
+ * which is more visually reliable than redistributing words to the original per-element X positions
+ * (those positions are calibrated to the original word widths and rarely match the replacement
+ * widths). Centered or tracked layouts will become left-aligned at the original first-word position
+ * when their text is replaced - the trade-off is accepted in favour of avoiding glyph overlaps and
+ * gaps from variable-width word distribution.
+ */
+@Slf4j
+@GeneralApi
+@RequiredArgsConstructor
+public class EditTextController {
+
+    private static final Pattern FILE_EXTENSION_PATTERN = Pattern.compile("[.][^.]+$");
+
+    private final PdfJsonConversionService pdfJsonConversionService;
+    private final TempFileManager tempFileManager;
+
+    @InitBinder
+    public void initBinder(WebDataBinder binder) {
+        binder.registerCustomEditor(
+                List.class,
+                "edits",
+                new StringToArrayListPropertyEditor<>(EditTextOperation.class));
+    }
+
+    @AutoJobPostMapping(consumes = "multipart/form-data", value = "/edit-text")
+    @StandardPdfResponse
+    @Operation(
+            summary = "Edit text in a PDF via find and replace",
+            description =
+                    "Applies an ordered list of find/replace operations to the text in a PDF and"
+                            + " returns the edited PDF. Useful for find-and-replace, bulk renames"
+                            + " (e.g. updating a company name throughout a document), and copy"
+                            + " editing where the AI agent has identified specific replacements."
+                            + " Matching is performed against the joined text of each page, so"
+                            + " find strings can span multiple visual runs (titles split per word,"
+                            + " kerning-broken phrases). Cross-element matches are written as a"
+                            + " single replacement run anchored at the leftmost matched position;"
+                            + " centered or tracked text may shift left when its content changes."
+                            + " Input:PDF Output:PDF Type:SISO")
+    public ResponseEntity<Resource> editText(@ModelAttribute EditTextRequest request)
+            throws Exception {
+        MultipartFile inputFile = request.getFileInput();
+        if (inputFile == null) {
+            throw ExceptionUtils.createFileNullOrEmptyException();
+        }
+        List<EditTextOperation> edits = request.getEdits();
+        if (edits == null || edits.isEmpty()) {
+            throw ExceptionUtils.createIllegalArgumentException(
+                    "error.editText.no.edits",
+                    "No find/replace operations provided for text editing");
+        }
+        for (EditTextOperation edit : edits) {
+            if (edit == null || edit.getFind() == null || edit.getFind().isEmpty()) {
+                throw ExceptionUtils.createIllegalArgumentException(
+                        "error.editText.empty.find", "Each edit must have a non-empty find string");
+            }
+        }
+
+        boolean wholeWordSearch = Boolean.TRUE.equals(request.getWholeWordSearch());
+        List<CompiledEdit> compiledEdits = compileEdits(edits, wholeWordSearch);
+
+        PdfJsonDocument document = pdfJsonConversionService.convertPdfToJsonDocument(inputFile);
+        Set<Integer> pageFilter = resolvePageFilter(request, document);
+        int modifiedSpans = applyEdits(document, compiledEdits, pageFilter);
+        log.info(
+                "edit-text: modified {} text span(s) using {} edit(s) on {} page(s)",
+                modifiedSpans,
+                compiledEdits.size(),
+                pageFilter == null
+                        ? document.getPages() == null ? 0 : document.getPages().size()
+                        : pageFilter.size());
+
+        String docName = buildOutputFilename(inputFile);
+        TempFile tempOut = tempFileManager.createManagedTempFile(".pdf");
+        try (OutputStream os = Files.newOutputStream(tempOut.getPath())) {
+            pdfJsonConversionService.convertJsonToPdf(document, os);
+        } catch (Exception e) {
+            tempOut.close();
+            throw e;
+        }
+        return WebResponseUtils.pdfFileToWebResponse(tempOut, docName);
+    }
+
+    private List<CompiledEdit> compileEdits(
+            List<EditTextOperation> edits, boolean wholeWordSearch) {
+        return edits.stream().map(edit -> compileEdit(edit, wholeWordSearch)).toList();
+    }
+
+    private CompiledEdit compileEdit(EditTextOperation edit, boolean wholeWordSearch) {
+        // Always treat the user-supplied find string as a literal: Pattern.quote escapes any
+        // regex metacharacters, so the constructed pattern can only ever do a literal match
+        // (optionally bounded by our own word-boundary anchors). This rules out catastrophic
+        // backtracking from a malicious find string.
+        String findRaw = edit.getFind();
+        String replacement = Objects.toString(edit.getReplace(), "");
+        String regex = Pattern.quote(findRaw);
+        if (wholeWordSearch) {
+            // Use lookarounds rather than \b so the bound works even when the find string starts
+            // or ends with a non-word character (e.g. "-foo" or "$id"). \b only fires at a
+            // word/non-word transition, which would never trigger for a find that starts with a
+            // non-word char and was preceded by another non-word char in the source text.
+            regex = "(?<!\\w)(?:" + regex + ")(?!\\w)";
+        }
+        Pattern pattern = Pattern.compile(regex);
+        String safeReplacement = Matcher.quoteReplacement(replacement);
+        return new CompiledEdit(pattern, safeReplacement);
+    }
+
+    private Set<Integer> resolvePageFilter(EditTextRequest request, PdfJsonDocument document) {
+        String pageNumbers = request.getPageNumbers();
+        int totalPages = document.getPages() == null ? 0 : document.getPages().size();
+        if (totalPages == 0) {
+            return Collections.emptySet();
+        }
+        if (pageNumbers == null || pageNumbers.isBlank() || "all".equalsIgnoreCase(pageNumbers)) {
+            return null;
+        }
+        List<Integer> pages = GeneralUtils.parsePageList(pageNumbers, totalPages, true);
+        return new HashSet<>(pages);
+    }
+
+    private int applyEdits(
+            PdfJsonDocument document, List<CompiledEdit> edits, Set<Integer> pageFilter) {
+        List<PdfJsonPage> pages = document.getPages();
+        if (pages == null) {
+            return 0;
+        }
+        int modifiedSpans = 0;
+        // The page filter is built from parsePageList against pages.size(), so it returns 1-based
+        // positional page numbers. Match against the same positional numbering here rather than
+        // the per-page pageNumber field, which can become non-sequential after split/merge round
+        // trips and would misalign the filter.
+        for (int i = 0; i < pages.size(); i++) {
+            int positionalPageNumber = i + 1;
+            if (pageFilter != null && !pageFilter.contains(positionalPageNumber)) {
+                continue;
+            }
+            modifiedSpans += applyEditsToPage(pages.get(i), edits);
+        }
+        return modifiedSpans;
+    }
+
+    private int applyEditsToPage(PdfJsonPage page, List<CompiledEdit> edits) {
+        List<PdfJsonTextElement> elements = page.getTextElements();
+        if (elements == null || elements.isEmpty()) {
+            return 0;
+        }
+        Set<Integer> modifiedIndices = new HashSet<>();
+        for (CompiledEdit edit : edits) {
+            applyEditToPage(elements, edit, modifiedIndices);
+        }
+        for (Integer index : modifiedIndices) {
+            // Char codes were captured for the original glyph sequence; clear them so the rebuild
+            // re-encodes from the new text via the font.
+            elements.get(index).setCharCodes(null);
+        }
+        return modifiedIndices.size();
+    }
+
+    /**
+     * Apply a single edit across the page by matching against the concatenation of all element
+     * texts, then writing the replacement back into the originating element(s).
+     */
+    private void applyEditToPage(
+            List<PdfJsonTextElement> elements, CompiledEdit edit, Set<Integer> modifiedIndices) {
+        StringBuilder joined = new StringBuilder();
+        int[] starts = new int[elements.size()];
+        int[] ends = new int[elements.size()];
+        for (int i = 0; i < elements.size(); i++) {
+            starts[i] = joined.length();
+            String text = elements.get(i).getText();
+            if (text != null) {
+                joined.append(text);
+            }
+            ends[i] = joined.length();
+        }
+
+        Matcher matcher = edit.pattern().matcher(joined);
+        List<MatchSpan> spans = new ArrayList<>();
+        StringBuffer interpolation = new StringBuffer();
+        int previousAppendPosition = 0;
+        while (matcher.find()) {
+            if (matcher.start() == matcher.end()) {
+                // Skip zero-length matches (e.g. /a*/ on empty input) — they cannot be applied.
+                continue;
+            }
+            int sizeBefore = interpolation.length();
+            matcher.appendReplacement(interpolation, edit.replacement());
+            int prefixLength = matcher.start() - previousAppendPosition;
+            String actualReplacement =
+                    interpolation.substring(sizeBefore + prefixLength, interpolation.length());
+            spans.add(new MatchSpan(matcher.start(), matcher.end(), actualReplacement));
+            previousAppendPosition = matcher.end();
+        }
+
+        // Apply right-to-left so earlier match positions stay valid as we mutate elements.
+        for (int i = spans.size() - 1; i >= 0; i--) {
+            MatchSpan span = spans.get(i);
+            int firstElement = findElementForCharIndex(starts, ends, span.start());
+            int lastElement = findElementForCharIndex(starts, ends, span.end() - 1);
+            if (firstElement < 0 || lastElement < 0) {
+                continue;
+            }
+            applyMatchToElements(
+                    elements, starts, span, firstElement, lastElement, modifiedIndices);
+        }
+    }
+
+    /**
+     * Find the element whose text covers the character at {@code charIndex} in the joined string.
+     * Returns -1 if no element covers that index (which should not happen for valid match spans).
+     */
+    private static int findElementForCharIndex(int[] starts, int[] ends, int charIndex) {
+        for (int i = 0; i < starts.length; i++) {
+            if (starts[i] <= charIndex && charIndex < ends[i]) {
+                return i;
+            }
+        }
+        return -1;
+    }
+
+    private static void applyMatchToElements(
+            List<PdfJsonTextElement> elements,
+            int[] starts,
+            MatchSpan span,
+            int firstElement,
+            int lastElement,
+            Set<Integer> modifiedIndices) {
+        if (firstElement == lastElement) {
+            PdfJsonTextElement element = elements.get(firstElement);
+            String text = nullToEmpty(element.getText());
+            int matchStartInElement = span.start() - starts[firstElement];
+            int matchEndInElement = span.end() - starts[firstElement];
+            element.setText(
+                    text.substring(0, matchStartInElement)
+                            + span.replacement()
+                            + text.substring(matchEndInElement));
+            modifiedIndices.add(firstElement);
+            return;
+        }
+
+        // Cross-element match: write the whole replacement into the first matched element, empty
+        // any intermediate elements, and keep only the suffix of the last matched element. The
+        // JSON->PDF rebuild concatenates per-token text, so the font lays out the replacement as
+        // one continuous run anchored at the first element's X position.
+        String firstText = nullToEmpty(elements.get(firstElement).getText());
+        int firstSplit = span.start() - starts[firstElement];
+        elements.get(firstElement).setText(firstText.substring(0, firstSplit) + span.replacement());
+        modifiedIndices.add(firstElement);
+
+        for (int mid = firstElement + 1; mid < lastElement; mid++) {
+            elements.get(mid).setText("");
+            modifiedIndices.add(mid);
+        }
+
+        String lastText = nullToEmpty(elements.get(lastElement).getText());
+        int lastSplit = span.end() - starts[lastElement];
+        elements.get(lastElement).setText(lastText.substring(lastSplit));
+        modifiedIndices.add(lastElement);
+    }
+
+    private static String nullToEmpty(String value) {
+        return value != null ? value : "";
+    }
+
+    private String buildOutputFilename(MultipartFile inputFile) {
+        String originalName = inputFile.getOriginalFilename();
+        String baseName =
+                (originalName != null && !originalName.isBlank())
+                        ? FILE_EXTENSION_PATTERN
+                                .matcher(Filenames.toSimpleFileName(originalName))
+                                .replaceFirst("")
+                        : "document";
+        return baseName + "_edited.pdf";
+    }
+
+    private record CompiledEdit(Pattern pattern, String replacement) {}
+
+    private record MatchSpan(int start, int end, String replacement) {}
+}
@@ -96,7 +96,9 @@ public class RedactController {
    @InitBinder
    public void initBinder(WebDataBinder binder) {
        binder.registerCustomEditor(
-                List.class, "redactions", new StringToArrayListPropertyEditor());
+                List.class,
+                "redactions",
+                new StringToArrayListPropertyEditor<>(RedactionArea.class));
    }

    @AutoJobPostMapping(value = "/redact", consumes = MediaType.MULTIPART_FORM_DATA_VALUE)
@@ -682,7 +682,7 @@ public class GlobalExceptionHandler {
     *
     * @param ex the MaxUploadSizeExceededException
     * @param request the HTTP servlet request
-     * @return ProblemDetail with HTTP 413 PAYLOAD_TOO_LARGE
+     * @return ProblemDetail with HTTP 413 CONTENT_TOO_LARGE
     */
    @ExceptionHandler(MaxUploadSizeExceededException.class)
    public ResponseEntity<ProblemDetail> handleMaxUploadSize(
@@ -706,7 +706,7 @@ public class GlobalExceptionHandler {
                getLocalizedMessage("error.fileTooLarge.title", ErrorTitles.FILE_TOO_LARGE_DEFAULT);

        ProblemDetail problemDetail =
-                createBaseProblemDetail(HttpStatus.PAYLOAD_TOO_LARGE, message, request);
+                createBaseProblemDetail(HttpStatus.CONTENT_TOO_LARGE, message, request);
        problemDetail.setType(URI.create(ErrorTypes.FILE_TOO_LARGE));
        problemDetail.setTitle(title);
        problemDetail.setProperty("title", title); // Ensure serialization
@@ -724,7 +724,7 @@ public class GlobalExceptionHandler {
        problemDetail.setProperty(
                "actionRequired", "Reduce the file size to be within the upload limit.");

-        return ResponseEntity.status(HttpStatus.PAYLOAD_TOO_LARGE)
+        return ResponseEntity.status(HttpStatus.CONTENT_TOO_LARGE)
                .contentType(PROBLEM_JSON)
                .body(problemDetail);
    }
@@ -0,0 +1,32 @@
+package stirling.software.SPDF.model.api.general;
+
+import java.util.List;
+
+import io.swagger.v3.oas.annotations.media.Schema;
+
+import lombok.Data;
+import lombok.EqualsAndHashCode;
+
+import stirling.software.SPDF.model.api.PDFWithPageNums;
+import stirling.software.common.model.api.general.EditTextOperation;
+
+@Data
+@EqualsAndHashCode(callSuper = true)
+public class EditTextRequest extends PDFWithPageNums {
+
+    @Schema(
+            description =
+                    "Ordered list of find/replace operations. Each replaces every occurrence on"
+                            + " the selected pages, in order; later operations see the result of"
+                            + " earlier ones (so 'foo'->'foos' then 'foos'->'bars' turns 'foo'"
+                            + " into 'bars').",
+            requiredMode = Schema.RequiredMode.REQUIRED)
+    private List<EditTextOperation> edits;
+
+    @Schema(
+            description =
+                    "Whether matches must be whole words (boundaries determined by non-word"
+                            + " characters)",
+            defaultValue = "false")
+    private Boolean wholeWordSearch;
+}
@@ -252,6 +252,17 @@ public class PdfJsonConversionService {
        convertPdfToJson(file, null, false, out);
    }

+    /**
+     * Converts a PDF to the editable {@link PdfJsonDocument} model. Convenience wrapper around
+     * {@link #convertPdfToJson(MultipartFile, OutputStream)} for callers that need to inspect or
+     * mutate the document in memory before round-tripping it back to PDF.
+     */
+    public PdfJsonDocument convertPdfToJsonDocument(MultipartFile file) throws IOException {
+        ByteArrayOutputStream buffer = new ByteArrayOutputStream();
+        convertPdfToJson(file, null, false, buffer);
+        return objectMapper.readValue(buffer.toByteArray(), PdfJsonDocument.class);
+    }
+
    public void convertPdfToJson(MultipartFile file, boolean lightweight, OutputStream out)
            throws IOException {
        convertPdfToJson(file, null, lightweight, out);
@@ -620,6 +631,13 @@ public class PdfJsonConversionService {
        }
        byte[] jsonBytes = file.getBytes();
        PdfJsonDocument pdfJson = objectMapper.readValue(jsonBytes, PdfJsonDocument.class);
+        convertJsonToPdf(pdfJson, out);
+    }
+
+    public void convertJsonToPdf(PdfJsonDocument pdfJson, OutputStream out) throws IOException {
+        if (pdfJson == null) {
+            throw ExceptionUtils.createNullArgumentException("document");
+        }

        List<PdfJsonFont> fontModels = pdfJson.getFonts();
        if (fontModels == null) {
@@ -2076,14 +2094,29 @@ public class PdfJsonConversionService {
            byte[] toUnicodeBytes = Base64.getDecoder().decode(toUnicodeBase64);
            String toUnicodeStr = new String(toUnicodeBytes, StandardCharsets.UTF_8);

-            // Parse ToUnicode CMap for bfchar and bfrange
+            // Parse ToUnicode CMap for bfchar and bfrange. Both sides go through
+            // parseToUnicodeCodepoint so an 8-hex-char surrogate-pair value on either side
+            // (e.g. "D837DF0E") is decoded as a single supplementary codepoint instead of
+            // overflowing Integer.parseInt and aborting the whole mapping build.
            java.util.regex.Pattern bfcharPattern =
                    java.util.regex.Pattern.compile("<([0-9A-Fa-f]+)>\\s*<([0-9A-Fa-f]+)>");
            java.util.regex.Matcher matcher = bfcharPattern.matcher(toUnicodeStr);
            while (matcher.find()) {
-                int charCode = Integer.parseInt(matcher.group(1), 16);
-                int unicode = Integer.parseInt(matcher.group(2), 16);
-                charCodeToUnicode.put(charCode, unicode);
+                try {
+                    int charCode = parseToUnicodeCodepoint(matcher.group(1));
+                    int unicode = parseToUnicodeCodepoint(matcher.group(2));
+                    charCodeToUnicode.put(charCode, unicode);
+                } catch (NumberFormatException entryEx) {
+                    // Tolerate a single malformed entry: log and skip rather than aborting the
+                    // entire ToUnicode CMap (which would force the whole font onto the slow
+                    // raw-bytes fallback path).
+                    log.debug(
+                            "Skipping malformed ToUnicode entry <{}> <{}> in font {}: {}",
+                            matcher.group(1),
+                            matcher.group(2),
+                            font.getName(),
+                            entryEx.getMessage());
+                }
            }

            // Build JSON mapping: CharCode → CID → GID → Unicode
@@ -2136,6 +2169,42 @@ public class PdfJsonConversionService {
        }
    }

+    /**
+     * Parse a hex string from a PDF ToUnicode CMap into a single Unicode codepoint. Handles three
+     * cases: a single BMP code unit (4 hex chars), a UTF-16 surrogate pair encoding a supplementary
+     * codepoint above U+FFFF (8 hex chars, e.g. {@code D837DF0E} for U+1F40E), and multi-codepoint
+     * mappings (longer; returns the first codepoint as a best-effort representative).
+     *
+     * <p>Without this, {@code Integer.parseInt("D837DF0E", 16)} overflows because the value is ~3.6
+     * billion, throwing {@link NumberFormatException} and forcing the conversion to fall back to a
+     * raw ToUnicode payload that the JSON&rarr;PDF rebuild then fails to use efficiently.
+     */
+    static int parseToUnicodeCodepoint(String hex) {
+        if (hex == null || hex.isEmpty()) {
+            throw new NumberFormatException("Empty ToUnicode hex value");
+        }
+        if (hex.length() <= 4) {
+            return Integer.parseInt(hex, 16);
+        }
+        // Treat the hex string as UTF-16BE: pairs of hex digits form bytes, four hex digits form
+        // one UTF-16 code unit. The PDF ToUnicode CMap convention requires an even number of bytes
+        // (i.e. a multiple of four hex characters) for multi-unit values.
+        if (hex.length() % 4 != 0) {
+            throw new NumberFormatException(
+                    "ToUnicode hex value not a multiple of 4 chars: " + hex);
+        }
+        int unitCount = hex.length() / 4;
+        char[] units = new char[unitCount];
+        for (int i = 0; i < unitCount; i++) {
+            units[i] = (char) Integer.parseInt(hex.substring(i * 4, i * 4 + 4), 16);
+        }
+        // codePointAt assembles a surrogate pair into a supplementary codepoint when the
+        // high/low surrogates appear in sequence; for any other multi-unit sequence it returns
+        // the first BMP codepoint, which is the right best-effort fallback for ligature
+        // decompositions (one charCode -> several Unicode chars).
+        return new String(units).codePointAt(0);
+    }
+
    private PdfJsonFontCidSystemInfo extractCidSystemInfo(COSDictionary fontDictionary) {
        if (fontDictionary == null) {
            return null;
@@ -4256,12 +4325,8 @@ public class PdfJsonConversionService {
            return 0;
        }
        if (font != null) {
-            try (InputStream inputStream = new ByteArrayInputStream(value.getBytes())) {
-                int count = 0;
-                int code;
-                while ((code = font.readCode(inputStream)) != -1) {
-                    count++;
-                }
+            try (ByteArrayInputStream inputStream = new ByteArrayInputStream(value.getBytes())) {
+                int count = countCodesProtected(inputStream, font::readCode);
                if (count > 0) {
                    return count;
                }
@@ -4273,6 +4338,49 @@ public class PdfJsonConversionService {
        return Math.max(1, bytes.length);
    }

+    /**
+     * Functional accessor for {@link PDFont#readCode(InputStream)} so the bounded counting loop can
+     * be exercised in isolation without instantiating a {@link PDFont}.
+     */
+    @FunctionalInterface
+    interface CodeReader {
+        int readCode(InputStream stream) throws IOException;
+    }
+
+    /**
+     * Count how many codes the supplied {@code reader} can extract from {@code inputStream}, with
+     * two safety nets that PDFBox's raw {@link PDFont#readCode(InputStream)} loop lacks:
+     *
+     * <ol>
+     *   <li>Stop when the stream is empty (a corrupt CMap can otherwise loop forever returning
+     *       successfully-matched zero-bytes from an exhausted {@link ByteArrayInputStream}).
+     *   <li>Stop when a {@code readCode} call did not consume any bytes, even if it returned a
+     *       non-{@code -1} value.
+     * </ol>
+     *
+     * <p>Both conditions were observed in the wild on round-tripped fallback fonts where the
+     * embedded ToUnicode CMap matched 0x00 sequences, hanging the JSON&rarr;PDF rebuild.
+     */
+    static int countCodesProtected(ByteArrayInputStream inputStream, CodeReader reader)
+            throws IOException {
+        int count = 0;
+        int previousAvailable = inputStream.available();
+        while (previousAvailable > 0) {
+            int code = reader.readCode(inputStream);
+            if (code == -1) {
+                break;
+            }
+            int currentAvailable = inputStream.available();
+            if (currentAvailable >= previousAvailable) {
+                // No progress made; break to avoid infinite loop on corrupt CMaps.
+                break;
+            }
+            count++;
+            previousAvailable = currentAvailable;
+        }
+        return count;
+    }
+
    private MergedText mergeText(List<PdfJsonTextElement> elements) {
        StringBuilder builder = new StringBuilder();
        List<Integer> combinedCodes = new ArrayList<>();
@@ -0,0 +1,664 @@
+package stirling.software.SPDF.controller.api;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertNotNull;
+import static org.junit.jupiter.api.Assertions.assertNull;
+import static org.junit.jupiter.api.Assertions.assertThrows;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+import static org.mockito.ArgumentMatchers.any;
+import static org.mockito.ArgumentMatchers.anyString;
+import static org.mockito.Mockito.lenient;
+import static org.mockito.Mockito.mock;
+import static org.mockito.Mockito.when;
+
+import java.io.ByteArrayOutputStream;
+import java.io.File;
+import java.io.OutputStream;
+import java.nio.file.Files;
+import java.util.ArrayList;
+import java.util.List;
+
+import org.apache.pdfbox.pdmodel.PDDocument;
+import org.apache.pdfbox.pdmodel.PDPage;
+import org.junit.jupiter.api.BeforeEach;
+import org.junit.jupiter.api.Test;
+import org.junit.jupiter.api.extension.ExtendWith;
+import org.mockito.ArgumentCaptor;
+import org.mockito.InjectMocks;
+import org.mockito.Mock;
+import org.mockito.junit.jupiter.MockitoExtension;
+import org.springframework.core.io.Resource;
+import org.springframework.http.HttpStatus;
+import org.springframework.http.ResponseEntity;
+import org.springframework.mock.web.MockMultipartFile;
+import org.springframework.web.multipart.MultipartFile;
+
+import stirling.software.SPDF.model.api.general.EditTextRequest;
+import stirling.software.SPDF.model.json.PdfJsonDocument;
+import stirling.software.SPDF.model.json.PdfJsonPage;
+import stirling.software.SPDF.model.json.PdfJsonTextElement;
+import stirling.software.SPDF.service.PdfJsonConversionService;
+import stirling.software.common.model.api.general.EditTextOperation;
+import stirling.software.common.util.TempFile;
+import stirling.software.common.util.TempFileManager;
+
+@ExtendWith(MockitoExtension.class)
+class EditTextControllerTest {
+
+    @Mock private PdfJsonConversionService pdfJsonConversionService;
+    @Mock private TempFileManager tempFileManager;
+
+    @InjectMocks private EditTextController controller;
+
+    @BeforeEach
+    void setUp() throws Exception {
+        lenient()
+                .when(tempFileManager.createManagedTempFile(anyString()))
+                .thenAnswer(
+                        inv -> {
+                            File f =
+                                    Files.createTempFile("test", inv.<String>getArgument(0))
+                                            .toFile();
+                            TempFile tf = mock(TempFile.class);
+                            lenient().when(tf.getFile()).thenReturn(f);
+                            lenient().when(tf.getPath()).thenReturn(f.toPath());
+                            return tf;
+                        });
+    }
+
+    private static MultipartFile pdfFile() {
+        return new MockMultipartFile(
+                "fileInput", "doc.pdf", "application/pdf", "stub-pdf-bytes".getBytes());
+    }
+
+    private static EditTextOperation edit(String find, String replace) {
+        EditTextOperation op = new EditTextOperation();
+        op.setFind(find);
+        op.setReplace(replace);
+        return op;
+    }
+
+    private static PdfJsonTextElement textElement(String text) {
+        PdfJsonTextElement el = new PdfJsonTextElement();
+        el.setText(text);
+        el.setCharCodes(new int[] {1, 2, 3});
+        return el;
+    }
+
+    private static PdfJsonDocument documentWithElements(List<PdfJsonTextElement> elements) {
+        PdfJsonDocument doc = new PdfJsonDocument();
+        PdfJsonPage page = new PdfJsonPage();
+        page.setPageNumber(1);
+        page.setTextElements(new ArrayList<>(elements));
+        doc.setPages(new ArrayList<>(List.of(page)));
+        return doc;
+    }
+
+    private static PdfJsonDocument documentWithText(String... textsByPage) {
+        PdfJsonDocument doc = new PdfJsonDocument();
+        List<PdfJsonPage> pages = new ArrayList<>();
+        for (int i = 0; i < textsByPage.length; i++) {
+            PdfJsonPage page = new PdfJsonPage();
+            page.setPageNumber(i + 1);
+            PdfJsonTextElement el = new PdfJsonTextElement();
+            el.setText(textsByPage[i]);
+            el.setCharCodes(new int[] {1, 2, 3});
+            page.setTextElements(new ArrayList<>(List.of(el)));
+            pages.add(page);
+        }
+        doc.setPages(pages);
+        return doc;
+    }
+
+    /** Build a minimal real PDF so tests can run end-to-end without mocking the conversion. */
+    private static byte[] buildEmptyPdf() throws Exception {
+        try (PDDocument doc = new PDDocument();
+                ByteArrayOutputStream baos = new ByteArrayOutputStream()) {
+            doc.addPage(new PDPage());
+            doc.save(baos);
+            return baos.toByteArray();
+        }
+    }
+
+    @Test
+    void editText_nullFileInputThrows() {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(null);
+        request.setEdits(List.of(edit("foo", "bar")));
+
+        assertThrows(IllegalArgumentException.class, () -> controller.editText(request));
+    }
+
+    @Test
+    void editText_emptyEditsThrows() {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of());
+
+        assertThrows(IllegalArgumentException.class, () -> controller.editText(request));
+    }
+
+    @Test
+    void editText_nullEditsThrows() {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(null);
+
+        assertThrows(IllegalArgumentException.class, () -> controller.editText(request));
+    }
+
+    @Test
+    void editText_emptyFindStringThrows() {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("", "replacement")));
+
+        assertThrows(IllegalArgumentException.class, () -> controller.editText(request));
+    }
+
+    @Test
+    void editText_findStringWithRegexMetacharsIsTreatedLiterally() throws Exception {
+        // Find strings are always treated as literals (Pattern.quote'd internally) — no ReDoS
+        // exposure from regex metacharacters supplied by the caller.
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("(unclosed", "fixed")));
+
+        PdfJsonDocument input = documentWithText("text with (unclosed paren");
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        assertEquals(
+                "text with fixed paren",
+                captor.getValue().getPages().get(0).getTextElements().get(0).getText());
+    }
+
+    @Test
+    void editText_literalFindReplace_mutatesMatchingSpansAndClearsCharCodes() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("foo", "bar")));
+
+        PdfJsonDocument input = documentWithText("foo and foo", "no match here");
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        ResponseEntity<Resource> response = controller.editText(request);
+
+        assertEquals(HttpStatus.OK, response.getStatusCode());
+        assertNotNull(response.getBody());
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        PdfJsonDocument mutated = captor.getValue();
+
+        assertEquals("bar and bar", mutated.getPages().get(0).getTextElements().get(0).getText());
+        assertNull(mutated.getPages().get(0).getTextElements().get(0).getCharCodes());
+
+        assertEquals("no match here", mutated.getPages().get(1).getTextElements().get(0).getText());
+        // Char codes preserved on unmodified spans.
+        assertNotNull(mutated.getPages().get(1).getTextElements().get(0).getCharCodes());
+    }
+
+    @Test
+    void editText_wholeWordSearch_doesNotMatchInsideWords() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("cat", "dog")));
+        request.setWholeWordSearch(true);
+
+        PdfJsonDocument input = documentWithText("the cat is in the catalogue");
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        assertEquals(
+                "the dog is in the catalogue",
+                captor.getValue().getPages().get(0).getTextElements().get(0).getText());
+    }
+
+    @Test
+    void editText_wholeWordSearch_matchesFindStartingWithNonWordChar() throws Exception {
+        // Regression: \b only fires on a word/non-word transition. A find that starts with a
+        // non-word char (e.g. "-foo") preceded by another non-word char in the source (a space)
+        // would never match under \b. The lookaround-based bound handles this correctly.
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("-foo", "-bar")));
+        request.setWholeWordSearch(true);
+
+        PdfJsonDocument input = documentWithText("space then -foo here");
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        assertEquals(
+                "space then -bar here",
+                captor.getValue().getPages().get(0).getTextElements().get(0).getText());
+    }
+
+    @Test
+    void editText_wholeWordSearch_doesNotMatchWhenAdjacentToWordChar() throws Exception {
+        // The lookaround bound must still reject matches that are part of a larger word.
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("-foo", "-bar")));
+        request.setWholeWordSearch(true);
+
+        PdfJsonDocument input = documentWithText("inline-foo should not match");
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        // "inline-foo" has 'e' (word) before '-foo' so the lookbehind blocks the match. The
+        // trailing 'o' is followed by a space (non-word) so the trailing lookahead would pass on
+        // its own; the leading lookbehind is what rejects it.
+        assertEquals(
+                "inline-foo should not match",
+                captor.getValue().getPages().get(0).getTextElements().get(0).getText());
+    }
+
+    @Test
+    void editText_pageFilter_onlyAffectsListedPages() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("foo", "bar")));
+        request.setPageNumbers("2");
+
+        PdfJsonDocument input = documentWithText("foo on page 1", "foo on page 2", "foo on page 3");
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        PdfJsonDocument mutated = captor.getValue();
+
+        assertEquals("foo on page 1", mutated.getPages().get(0).getTextElements().get(0).getText());
+        assertEquals("bar on page 2", mutated.getPages().get(1).getTextElements().get(0).getText());
+        assertEquals("foo on page 3", mutated.getPages().get(2).getTextElements().get(0).getText());
+    }
+
+    @Test
+    void editText_pageRange_appliesToAllPagesInRange() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("foo", "bar")));
+        request.setPageNumbers("2-3");
+
+        PdfJsonDocument input =
+                documentWithText("foo page 1", "foo page 2", "foo page 3", "foo page 4");
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        PdfJsonDocument mutated = captor.getValue();
+
+        assertTrue(mutated.getPages().get(0).getTextElements().get(0).getText().startsWith("foo"));
+        assertTrue(mutated.getPages().get(1).getTextElements().get(0).getText().startsWith("bar"));
+        assertTrue(mutated.getPages().get(2).getTextElements().get(0).getText().startsWith("bar"));
+        assertTrue(mutated.getPages().get(3).getTextElements().get(0).getText().startsWith("foo"));
+    }
+
+    @Test
+    void editText_orderedEdits_applyInSequence() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("foo", "bar"), edit("bar", "baz")));
+
+        PdfJsonDocument input = documentWithText("foo");
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        assertEquals("baz", captor.getValue().getPages().get(0).getTextElements().get(0).getText());
+    }
+
+    @Test
+    void editText_replaceWithEmptyString_deletesMatch() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("DRAFT ", "")));
+
+        PdfJsonDocument input = documentWithText("DRAFT Confidential Memo");
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        assertEquals(
+                "Confidential Memo",
+                captor.getValue().getPages().get(0).getTextElements().get(0).getText());
+    }
+
+    @Test
+    void editText_noMatches_returnsPdfWithoutMutation() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("notfound", "replacement")));
+
+        PdfJsonDocument input = documentWithText("nothing to match");
+        int[] originalCodes = input.getPages().get(0).getTextElements().get(0).getCharCodes();
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        ResponseEntity<Resource> response = controller.editText(request);
+
+        assertEquals(HttpStatus.OK, response.getStatusCode());
+        // Char codes left intact when nothing was replaced.
+        assertEquals(
+                originalCodes, input.getPages().get(0).getTextElements().get(0).getCharCodes());
+        assertEquals(
+                "nothing to match", input.getPages().get(0).getTextElements().get(0).getText());
+    }
+
+    @Test
+    void editText_dollarInLiteralReplacement_isQuoted() throws Exception {
+        // Without quoting, '$1' would be interpreted as a backreference and crash.
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("price", "$100")));
+
+        PdfJsonDocument input = documentWithText("the price is set");
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        assertEquals(
+                "the $100 is set",
+                captor.getValue().getPages().get(0).getTextElements().get(0).getText());
+    }
+
+    @Test
+    void editText_emptyDocument_returnsResponseWithoutErrors() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("foo", "bar")));
+
+        PdfJsonDocument input = new PdfJsonDocument();
+        input.setPages(new ArrayList<>());
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        ResponseEntity<Resource> response = controller.editText(request);
+
+        assertEquals(HttpStatus.OK, response.getStatusCode());
+        assertNotNull(response.getBody());
+    }
+
+    @Test
+    void editText_textElementWithNullText_isSkipped() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("foo", "bar")));
+
+        PdfJsonDocument input = new PdfJsonDocument();
+        PdfJsonPage page = new PdfJsonPage();
+        page.setPageNumber(1);
+        PdfJsonTextElement nullText = new PdfJsonTextElement();
+        nullText.setText(null);
+        PdfJsonTextElement realText = new PdfJsonTextElement();
+        realText.setText("foo here");
+        page.setTextElements(new ArrayList<>(List.of(nullText, realText)));
+        input.setPages(new ArrayList<>(List.of(page)));
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        PdfJsonDocument mutated = captor.getValue();
+        assertNull(mutated.getPages().get(0).getTextElements().get(0).getText());
+        assertEquals("bar here", mutated.getPages().get(0).getTextElements().get(1).getText());
+    }
+
+    @Test
+    void editText_crossElement_matchSpansTwoElements() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("Hello World", "Goodbye Earth")));
+
+        PdfJsonDocument input =
+                documentWithElements(List.of(textElement("Hello "), textElement("World")));
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        List<PdfJsonTextElement> elements = captor.getValue().getPages().get(0).getTextElements();
+
+        // Whole replacement lands in the first matched element; the second is emptied. The
+        // JSON->PDF rebuild concatenates per-token text so the font lays out the replacement as
+        // one run anchored at the first element's X position.
+        assertEquals("Goodbye Earth", elements.get(0).getText());
+        assertEquals("", elements.get(1).getText());
+        assertNull(elements.get(0).getCharCodes());
+        assertNull(elements.get(1).getCharCodes());
+    }
+
+    @Test
+    void editText_crossElement_matchSpansFiveElementsLikeFragmentedTitle() throws Exception {
+        // Reproduces the real-world case: a multi-word title fragmented one word per text span.
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(
+                List.of(edit("The Free Adobe Acrobat Alternative", "The PDF automation pipeline")));
+
+        PdfJsonDocument input =
+                documentWithElements(
+                        List.of(
+                                textElement("The "),
+                                textElement("Free "),
+                                textElement("Adobe "),
+                                textElement("Acrobat "),
+                                textElement("Alternative")));
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        List<PdfJsonTextElement> elements = captor.getValue().getPages().get(0).getTextElements();
+
+        // Whole replacement is written into the first matched element; the remaining four are
+        // emptied. Centered titles will become left-aligned at the original first-word X position.
+        assertEquals("The PDF automation pipeline", elements.get(0).getText());
+        for (int i = 1; i < elements.size(); i++) {
+            assertEquals("", elements.get(i).getText(), "element " + i + " should be empty");
+        }
+    }
+
+    @Test
+    void editText_crossElement_preservesPrefixAndSuffix() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("Hello World", "Goodbye Earth")));
+
+        PdfJsonDocument input =
+                documentWithElements(
+                        List.of(textElement("Greeting: Hello "), textElement("World! And more")));
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        List<PdfJsonTextElement> elements = captor.getValue().getPages().get(0).getTextElements();
+
+        // First element keeps its prefix and gets the entire replacement; last element keeps its
+        // suffix only.
+        assertEquals("Greeting: Goodbye Earth", elements.get(0).getText());
+        assertEquals("! And more", elements.get(1).getText());
+    }
+
+    @Test
+    void editText_matchInOneElementOfMany_onlyTouchesThatElement() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("World", "Earth")));
+
+        PdfJsonDocument input =
+                documentWithElements(
+                        List.of(
+                                textElement("Hello "),
+                                textElement("World!"),
+                                textElement(" Goodbye")));
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        List<PdfJsonTextElement> elements = captor.getValue().getPages().get(0).getTextElements();
+
+        assertEquals("Hello ", elements.get(0).getText());
+        assertEquals("Earth!", elements.get(1).getText());
+        assertEquals(" Goodbye", elements.get(2).getText());
+        // Only the modified element's char codes get cleared.
+        assertNotNull(elements.get(0).getCharCodes());
+        assertNull(elements.get(1).getCharCodes());
+        assertNotNull(elements.get(2).getCharCodes());
+    }
+
+    @Test
+    void editText_crossElement_multipleMatchesAppliedRightToLeft() throws Exception {
+        // Two matches in the same page text. Right-to-left application keeps earlier indices
+        // valid as later matches are written.
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("foo bar", "X")));
+
+        PdfJsonDocument input =
+                documentWithElements(
+                        List.of(
+                                textElement("foo "),
+                                textElement("bar baz "),
+                                textElement("foo "),
+                                textElement("bar")));
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        List<PdfJsonTextElement> elements = captor.getValue().getPages().get(0).getTextElements();
+
+        // Joined was "foo bar baz foo bar"; both "foo bar" runs replaced with "X".
+        StringBuilder joined = new StringBuilder();
+        for (PdfJsonTextElement el : elements) {
+            joined.append(el.getText() == null ? "" : el.getText());
+        }
+        assertEquals("X baz X", joined.toString());
+    }
+
+    @Test
+    void editText_subWordFragmentation_writesIntoFirstElement() throws Exception {
+        // When the matched text is split into many character-level spans (typical of Type3 glyph
+        // runs), the whole replacement is dumped into the first matched element and the rest are
+        // emptied. The font lays out the replacement as one run at the first glyph's X position.
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(pdfFile());
+        request.setEdits(List.of(edit("Hello World", "Goodbye Earth")));
+
+        // 11 sub-word elements covering "Hello World".
+        PdfJsonDocument input =
+                documentWithElements(
+                        List.of(
+                                textElement("H"),
+                                textElement("e"),
+                                textElement("l"),
+                                textElement("l"),
+                                textElement("o"),
+                                textElement(" "),
+                                textElement("W"),
+                                textElement("o"),
+                                textElement("r"),
+                                textElement("l"),
+                                textElement("d")));
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        controller.editText(request);
+
+        ArgumentCaptor<PdfJsonDocument> captor = ArgumentCaptor.forClass(PdfJsonDocument.class);
+        org.mockito.Mockito.verify(pdfJsonConversionService)
+                .convertJsonToPdf(captor.capture(), any(OutputStream.class));
+        List<PdfJsonTextElement> elements = captor.getValue().getPages().get(0).getTextElements();
+
+        // Entire replacement goes into the first matched element; the other 10 are emptied.
+        assertEquals("Goodbye Earth", elements.get(0).getText());
+        for (int i = 1; i < elements.size(); i++) {
+            assertEquals("", elements.get(i).getText(), "element " + i + " should be empty");
+        }
+    }
+
+    @Test
+    void editText_outputFilenameDerivedFromInput() throws Exception {
+        EditTextRequest request = new EditTextRequest();
+        request.setFileInput(
+                new MockMultipartFile(
+                        "fileInput", "report.pdf", "application/pdf", buildEmptyPdf()));
+        request.setEdits(List.of(edit("anything", "x")));
+
+        PdfJsonDocument input = documentWithText("nothing matches");
+        when(pdfJsonConversionService.convertPdfToJsonDocument(any(MultipartFile.class)))
+                .thenReturn(input);
+
+        ResponseEntity<Resource> response = controller.editText(request);
+        String contentDisposition =
+                response.getHeaders()
+                        .getFirst(org.springframework.http.HttpHeaders.CONTENT_DISPOSITION);
+        assertNotNull(contentDisposition);
+        assertTrue(contentDisposition.contains("report_edited.pdf"));
+        assertFalse(contentDisposition.contains(".pdf_edited.pdf"));
+    }
+}
@@ -177,14 +177,14 @@ class GlobalExceptionHandlerTest {
    void handleMaxUploadSize_returns_413() {
        MaxUploadSizeExceededException ex = new MaxUploadSizeExceededException(10485760);
        ResponseEntity<ProblemDetail> resp = handler.handleMaxUploadSize(ex, request);
-        assertEquals(HttpStatus.PAYLOAD_TOO_LARGE, resp.getStatusCode());
+        assertEquals(HttpStatus.CONTENT_TOO_LARGE, resp.getStatusCode());
    }

    @Test
    void handleMaxUploadSize_unknown_limit() {
        MaxUploadSizeExceededException ex = new MaxUploadSizeExceededException(-1);
        ResponseEntity<ProblemDetail> resp = handler.handleMaxUploadSize(ex, request);
-        assertEquals(HttpStatus.PAYLOAD_TOO_LARGE, resp.getStatusCode());
+        assertEquals(HttpStatus.CONTENT_TOO_LARGE, resp.getStatusCode());
        assertNull(resp.getBody().getProperties().get("maxSizeBytes"));
    }

@@ -0,0 +1,161 @@
+package stirling.software.SPDF.service;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertThrows;
+import static org.junit.jupiter.api.Assertions.assertTimeoutPreemptively;
+
+import java.io.ByteArrayInputStream;
+import java.io.IOException;
+import java.io.InputStream;
+import java.time.Duration;
+
+import org.junit.jupiter.api.Test;
+
+/**
+ * Unit tests for {@link PdfJsonConversionService#parseToUnicodeCodepoint(String)}.
+ *
+ * <p>The function exists because PDF ToUnicode CMap entries can encode supplementary-plane
+ * codepoints (above U+FFFF) as UTF-16 surrogate pairs, e.g. {@code <D837DF0E>} for U+1F40E. The
+ * naive implementation that called {@link Integer#parseInt(String, int)} on the whole hex string
+ * threw {@link NumberFormatException} for any 8-char value because they overflow {@code int}. That
+ * triggered a fallback that left fonts with no proper CID mapping and made downstream JSON&rarr;PDF
+ * rebuilds slow/hung when rendering text with those fonts.
+ */
+class PdfJsonConversionServiceUnicodeParsingTest {
+
+    @Test
+    void parsesSingleBmpCodeUnit() {
+        // Latin capital A.
+        assertEquals(0x41, PdfJsonConversionService.parseToUnicodeCodepoint("0041"));
+    }
+
+    @Test
+    void parsesShortBmpHexValue() {
+        // Some ToUnicode entries omit leading zeros for low codepoints.
+        assertEquals(0x41, PdfJsonConversionService.parseToUnicodeCodepoint("41"));
+    }
+
+    @Test
+    void parsesSupplementaryCodepointFromSurrogatePair() {
+        // U+1F40E HORSE encoded as UTF-16 surrogate pair D83D DC0E. The bug we are fixing was
+        // that Integer.parseInt("D83DDC0E", 16) overflows because D83DDC0E > Integer.MAX_VALUE.
+        assertEquals(0x1F40E, PdfJsonConversionService.parseToUnicodeCodepoint("D83DDC0E"));
+    }
+
+    @Test
+    void parsesSurrogatePairFromUserReportedHang() {
+        // The exact hex value from the user's hang reproducer log:
+        // "Failed to build Unicode mapping for font NotoSans-Regular: For input string:
+        //  \"D837DF0E\" under radix 16"
+        // D837 DF0E decodes to U+1DF0E (CJK supplementary). Important assertion: it returns a
+        // valid codepoint instead of overflowing Integer.parseInt and throwing.
+        int expected = new String(new char[] {(char) 0xD837, (char) 0xDF0E}).codePointAt(0);
+        assertEquals(0x1DF0E, expected); // sanity check on the test setup itself
+        assertEquals(expected, PdfJsonConversionService.parseToUnicodeCodepoint("D837DF0E"));
+    }
+
+    @Test
+    void parsesLigatureDecompositionAsFirstCodepoint() {
+        // A ToUnicode entry can map one charCode to multiple Unicode chars (a ligature). PDF
+        // spec allows e.g. <0041 0042> for "AB". Our best-effort behavior is to return the
+        // first codepoint so the mapping is at least roughly meaningful for search/copy.
+        assertEquals(0x41, PdfJsonConversionService.parseToUnicodeCodepoint("00410042"));
+    }
+
+    @Test
+    void rejectsEmptyHex() {
+        assertThrows(
+                NumberFormatException.class,
+                () -> PdfJsonConversionService.parseToUnicodeCodepoint(""));
+    }
+
+    @Test
+    void rejectsNullHex() {
+        assertThrows(
+                NumberFormatException.class,
+                () -> PdfJsonConversionService.parseToUnicodeCodepoint(null));
+    }
+
+    @Test
+    void rejectsOddByteCountAboveBmp() {
+        // 6 hex chars is 3 bytes — not a valid UTF-16BE sequence.
+        assertThrows(
+                NumberFormatException.class,
+                () -> PdfJsonConversionService.parseToUnicodeCodepoint("D83DDC"));
+    }
+
+    @Test
+    void countCodesProtectedTerminatesWhenReaderMakesNoProgress() {
+        // Reproduces the user's hang: PDFBox's CMap.readCode can return a successful code (0)
+        // from a stream where no bytes were consumed (corrupt codespace matching 0x00 bytes from
+        // the buffer's uninitialized region after EOF). Without the no-progress guard, the
+        // counting loop in countGlyphs ran forever.
+        ByteArrayInputStream stream = new ByteArrayInputStream(new byte[] {1, 2, 3, 4});
+        PdfJsonConversionService.CodeReader reader = in -> 0; // never reads, always "succeeds"
+
+        int count =
+                assertTimeoutPreemptively(
+                        Duration.ofSeconds(2),
+                        () -> PdfJsonConversionService.countCodesProtected(stream, reader));
+
+        // First iteration sees no progress and breaks immediately.
+        assertEquals(0, count);
+    }
+
+    @Test
+    void countCodesProtectedTerminatesOnEmptyStream() {
+        ByteArrayInputStream stream = new ByteArrayInputStream(new byte[0]);
+        PdfJsonConversionService.CodeReader reader =
+                in -> {
+                    throw new AssertionError("reader must not be called when stream is empty");
+                };
+
+        int count =
+                assertTimeoutPreemptively(
+                        Duration.ofSeconds(2),
+                        () -> PdfJsonConversionService.countCodesProtected(stream, reader));
+
+        assertEquals(0, count);
+    }
+
+    @Test
+    void countCodesProtectedHonorsExplicitMinusOneReturn() throws IOException {
+        ByteArrayInputStream stream = new ByteArrayInputStream(new byte[] {1, 2, 3});
+        PdfJsonConversionService.CodeReader reader =
+                in -> {
+                    int b = in.read();
+                    return b == -1 ? -1 : b;
+                };
+
+        int count = PdfJsonConversionService.countCodesProtected(stream, reader);
+
+        assertEquals(3, count);
+    }
+
+    @Test
+    void countCodesProtectedTerminatesIfReaderReadsThenStops() throws IOException {
+        // A reader that consumes one byte then hits a corrupt-CMap pattern returning 0 without
+        // consuming further must still terminate after counting the consumed bytes.
+        ByteArrayInputStream stream = new ByteArrayInputStream(new byte[] {1, 2, 3, 4});
+        PdfJsonConversionService.CodeReader reader =
+                new PdfJsonConversionService.CodeReader() {
+                    boolean firstCall = true;
+
+                    @Override
+                    public int readCode(InputStream in) throws IOException {
+                        if (firstCall) {
+                            firstCall = false;
+                            return in.read();
+                        }
+                        return 0; // simulates corrupt CMap thereafter
+                    }
+                };
+
+        int count =
+                assertTimeoutPreemptively(
+                        Duration.ofSeconds(2),
+                        () -> PdfJsonConversionService.countCodesProtected(stream, reader));
+
+        assertEquals(1, count);
+    }
+}
@@ -30,6 +30,7 @@ import lombok.extern.slf4j.Slf4j;
 import stirling.software.common.model.job.ResultFile;
 import stirling.software.common.service.JobOwnershipService;
 import stirling.software.common.service.TaskManager;
+import stirling.software.proprietary.model.api.ai.AiWorkflowProgressEvent;
 import stirling.software.proprietary.model.api.ai.AiWorkflowRequest;
 import stirling.software.proprietary.model.api.ai.AiWorkflowResponse;
 import stirling.software.proprietary.model.api.ai.AiWorkflowResultFile;
@@ -140,13 +141,32 @@ public class AiEngineController {
    }

    private void runOrchestrationStream(AiWorkflowRequest request, SseEmitter emitter) {
+        AiWorkflowService.ProgressListener listener =
+                new AiWorkflowService.ProgressListener() {
+                    @Override
+                    public void onProgress(AiWorkflowProgressEvent event) {
+                        sendEvent(emitter, "progress", event);
+                    }
+
+                    @Override
+                    public void onHeartbeat() {
+                        // Forward upstream heartbeats so the SSE pipe stays visibly alive between
+                        // real progress events; if the frontend has gone away, sendEvent throws,
+                        // which propagates up through the stream consumer and closes our upstream
+                        // engine connection so the engine can cancel its in-flight workflow.
+                        sendEvent(emitter, "heartbeat", Map.of());
+                    }
+                };
        try {
-            AiWorkflowResponse result =
-                    aiWorkflowService.orchestrate(
-                            request, progress -> sendEvent(emitter, "progress", progress));
+            AiWorkflowResponse result = aiWorkflowService.orchestrate(request, listener);
            registerFileResultAsJob(result);
            sendEvent(emitter, "result", result);
            emitter.complete();
+        } catch (ClientDisconnectedException e) {
+            // The frontend gave up mid-stream. The exception unwinding through orchestrate()
+            // already closed the upstream engine connection (engine sees disconnect and cancels).
+            // The emitter is already toast; nothing useful left to send.
+            log.debug("Client disconnected mid-stream; aborting workflow", e);
        } catch (Exception e) {
            log.error("AI orchestration stream failed", e);
            // Emit an error frame for the frontend and then complete normally. Using
@@ -192,7 +212,21 @@ public class AiEngineController {
        try {
            emitter.send(SseEmitter.event().name(name).data(data, MediaType.APPLICATION_JSON));
        } catch (IOException e) {
-            log.debug("Failed to send SSE event (client may have disconnected)", e);
+            // Surface the disconnect so the streaming pipeline unwinds: callers higher up close
+            // the upstream engine connection, which lets the engine cancel its in-flight workflow.
+            // Without this, the engine would keep producing (and billing for) tokens whose results
+            // nobody is reading.
+            throw new ClientDisconnectedException("Client disconnected from SSE stream", e);
+        }
+    }
+
+    /**
+     * Thrown by {@link #sendEvent} when the SSE emitter's underlying connection is gone. Treated as
+     * a signal to abort the workflow, not as an error to report.
+     */
+    private static final class ClientDisconnectedException extends RuntimeException {
+        ClientDisconnectedException(String message, Throwable cause) {
+            super(message, cause);
        }
    }

@@ -0,0 +1,24 @@
+package stirling.software.proprietary.model.api.ai;
+
+import java.util.List;
+
+import lombok.AllArgsConstructor;
+import lombok.Data;
+import lombok.NoArgsConstructor;
+
+/**
+ * Body for {@code POST /api/v1/documents} on the AI engine. Sent by Java when the engine reports
+ * {@code need_ingest} and the requested document's extracted content must be stored before the
+ * workflow can continue.
+ */
+@Data
+@NoArgsConstructor
+@AllArgsConstructor
+public class AiDocumentIngestRequest {
+
+    private String documentId;
+
+    private String source;
+
+    private List<AiPageText> pageText;
+}
@@ -0,0 +1,59 @@
+package stirling.software.proprietary.model.api.ai;
+
+import com.fasterxml.jackson.annotation.JsonIgnoreProperties;
+import com.fasterxml.jackson.annotation.JsonSubTypes;
+import com.fasterxml.jackson.annotation.JsonTypeInfo;
+
+/**
+ * Typed engine-emitted progress detail, mirroring the Python {@code ProgressEvent} discriminated
+ * union (see {@code engine/src/stirling/contracts/progress.py}). Carried inside {@link
+ * AiWorkflowProgressEvent#getEngineDetail()} for {@link AiWorkflowPhase#ENGINE_PROGRESS} events.
+ *
+ * <p>Sealed so adding a new engine-side phase forces a matching subtype on the Java side instead of
+ * silently passing through as an opaque map. The {@code phase} string is the discriminator and
+ * stays on the wire so the frontend (which doesn't know about Java's class hierarchy) can switch on
+ * it.
+ */
+@JsonTypeInfo(
+        use = JsonTypeInfo.Id.NAME,
+        include = JsonTypeInfo.As.EXISTING_PROPERTY,
+        property = "phase",
+        visible = true)
+@JsonSubTypes({
+    @JsonSubTypes.Type(
+            value = AiEngineProgressDetail.WholeDocReadStarted.class,
+            name = "whole_doc_read_started"),
+    @JsonSubTypes.Type(
+            value = AiEngineProgressDetail.WholeDocSliceDone.class,
+            name = "whole_doc_slice_done"),
+    @JsonSubTypes.Type(
+            value = AiEngineProgressDetail.WholeDocCompressionRound.class,
+            name = "whole_doc_compression_round"),
+    @JsonSubTypes.Type(
+            value = AiEngineProgressDetail.WholeDocReadDone.class,
+            name = "whole_doc_read_done"),
+})
+@JsonIgnoreProperties(ignoreUnknown = true)
+public sealed interface AiEngineProgressDetail {
+
+    String phase();
+
+    record WholeDocReadStarted(String phase, String question, int pages, int slices)
+            implements AiEngineProgressDetail {}
+
+    record WholeDocSliceDone(
+            String phase,
+            int completed,
+            int total,
+            String pages,
+            int durationMs,
+            int excerpts,
+            int facts)
+            implements AiEngineProgressDetail {}
+
+    record WholeDocCompressionRound(String phase, int roundNumber, int notesIn, int groups)
+            implements AiEngineProgressDetail {}
+
+    record WholeDocReadDone(String phase, int completed, int slices, double durationSeconds)
+            implements AiEngineProgressDetail {}
+}
@@ -4,11 +4,11 @@ import lombok.AllArgsConstructor;
 import lombok.Data;
 import lombok.NoArgsConstructor;

-/** A single page of extracted text for RAG ingest requests. */
+/** A single page of extracted text for document ingest requests. */
@Data
@NoArgsConstructor
@AllArgsConstructor
-public class AiRagPageText {
+public class AiPageText {

    private int pageNumber;

@@ -1,24 +0,0 @@
-package stirling.software.proprietary.model.api.ai;
-
-import java.util.List;
-
-import lombok.AllArgsConstructor;
-import lombok.Data;
-import lombok.NoArgsConstructor;
-
-/**
- * Body for {@code POST /api/v1/rag/documents} on the AI engine. Sent by Java when the engine
- * reports {@code need_ingest} and the requested document's extracted content must be stored before
- * the workflow can continue.
- */
-@Data
-@NoArgsConstructor
-@AllArgsConstructor
-public class AiRagIngestRequest {
-
-    private String documentId;
-
-    private String source;
-
-    private List<AiRagPageText> pageText;
-}
@@ -9,7 +9,13 @@ public enum AiWorkflowPhase {
    CALLING_ENGINE("calling_engine"),
    EXTRACTING_CONTENT("extracting_content"),
    EXECUTING_TOOL("executing_tool"),
-    PROCESSING("processing");
+    PROCESSING("processing"),
+    /**
+     * Generic engine-emitted progress event (e.g. chunked-reasoner slice progress). The original
+     * engine event JSON is carried in {@link AiWorkflowProgressEvent#getEngineDetail()}, including
+     * a specific {@code phase} string the frontend can switch on.
+     */
+    ENGINE_PROGRESS("engine_progress");

    private final String value;

@@ -23,8 +23,17 @@ public class AiWorkflowProgressEvent {
    /** Total number of plan steps, for {@link AiWorkflowPhase#EXECUTING_TOOL} events. */
    private Integer stepCount;

+    /**
+     * Engine-emitted event payload, for {@link AiWorkflowPhase#ENGINE_PROGRESS} events. The payload
+     * is a typed subtype keyed on its {@code phase} string (e.g. {@code "whole_doc_slice_done"})
+     * carrying phase-specific fields (slice index, page range, durations, etc.) that the frontend
+     * can render as detailed progress.
+     */
+    private AiEngineProgressDetail engineDetail;
+
    public static AiWorkflowProgressEvent of(AiWorkflowPhase phase) {
-        return new AiWorkflowProgressEvent(phase, System.currentTimeMillis(), null, null, null);
+        return new AiWorkflowProgressEvent(
+                phase, System.currentTimeMillis(), null, null, null, null);
    }

    public static AiWorkflowProgressEvent executingTool(String tool, int stepIndex, int stepCount) {
@@ -33,6 +42,17 @@ public class AiWorkflowProgressEvent {
                System.currentTimeMillis(),
                tool,
                stepIndex,
-                stepCount);
+                stepCount,
+                null);
+    }
+
+    public static AiWorkflowProgressEvent engineProgress(AiEngineProgressDetail detail) {
+        return new AiWorkflowProgressEvent(
+                AiWorkflowPhase.ENGINE_PROGRESS,
+                System.currentTimeMillis(),
+                null,
+                null,
+                null,
+                detail);
    }
 }
@@ -7,6 +7,8 @@ import java.net.http.HttpRequest;
 import java.net.http.HttpResponse;
 import java.net.http.HttpTimeoutException;
 import java.time.Duration;
+import java.util.function.Consumer;
+import java.util.stream.Stream;

 import org.springframework.beans.factory.annotation.Autowired;
 import org.springframework.http.HttpStatus;
@@ -83,6 +85,70 @@ public class AiEngineClient {
        return response.body();
    }

+    /**
+     * POST a JSON body and consume the response as a stream of NDJSON lines. Each line is passed to
+     * {@code lineConsumer} in arrival order; the call returns when the engine closes the stream.
+     *
+     * <p>This is the right shape for long-running orchestrator calls that emit incremental
+     * progress. The total HTTP timeout is the long-running timeout (typically 600s+), but in
+     * practice line arrival keeps the connection logically alive: as long as the engine emits
+     * events, the work is progressing. Genuine engine hangs still hit the total timeout.
+     */
+    public void streamPost(String path, String jsonBody, Consumer<String> lineConsumer)
+            throws IOException {
+        ApplicationProperties.AiEngine config = applicationProperties.getAiEngine();
+        if (!config.isEnabled()) {
+            throw new ResponseStatusException(
+                    HttpStatus.SERVICE_UNAVAILABLE, "AI engine is not enabled");
+        }
+
+        String url = config.getUrl().stripTrailing() + path;
+        Duration timeout = Duration.ofSeconds(config.getLongRunningTimeoutSeconds());
+        log.debug(
+                "Proxying AI engine streaming request to {} (timeout {}s)",
+                url,
+                timeout.toSeconds());
+
+        HttpRequest request =
+                HttpRequest.newBuilder()
+                        .uri(URI.create(url))
+                        .header("Content-Type", "application/json")
+                        .header("Accept", "application/x-ndjson")
+                        .timeout(timeout)
+                        .POST(HttpRequest.BodyPublishers.ofString(jsonBody))
+                        .build();
+
+        HttpResponse<Stream<String>> response;
+        try {
+            response = httpClient.send(request, HttpResponse.BodyHandlers.ofLines());
+        } catch (HttpTimeoutException e) {
+            throw new ResponseStatusException(HttpStatus.GATEWAY_TIMEOUT, "AI engine timed out", e);
+        } catch (IOException e) {
+            throw new ResponseStatusException(
+                    HttpStatus.SERVICE_UNAVAILABLE, "AI engine unreachable: " + e.getMessage(), e);
+        } catch (InterruptedException e) {
+            Thread.currentThread().interrupt();
+            throw new ResponseStatusException(
+                    HttpStatus.SERVICE_UNAVAILABLE, "AI engine request was interrupted");
+        }
+
+        int status = response.statusCode();
+        if (status >= 400) {
+            throw new ResponseStatusException(
+                    HttpStatus.valueOf(status >= 500 ? 502 : status),
+                    "AI engine returned error: " + status);
+        }
+
+        try (Stream<String> lines = response.body()) {
+            lines.forEach(
+                    line -> {
+                        if (!line.isEmpty()) {
+                            lineConsumer.accept(line);
+                        }
+                    });
+        }
+    }
+
    public String get(String path) throws IOException {
        ApplicationProperties.AiEngine config = applicationProperties.getAiEngine();
        if (!config.isEnabled()) {
@@ -30,15 +30,17 @@ import lombok.extern.slf4j.Slf4j;
 import stirling.software.common.service.CustomPDFDocumentFactory;
 import stirling.software.common.service.FileStorage;
 import stirling.software.common.service.InternalApiClient;
+import stirling.software.common.service.InternalApiTimeoutException;
 import stirling.software.common.service.ToolMetadataService;
 import stirling.software.common.util.ExceptionUtils;
 import stirling.software.common.util.TempFile;
 import stirling.software.common.util.TempFileManager;
 import stirling.software.common.util.ZipExtractionUtils;
 import stirling.software.proprietary.model.api.ai.AiConversationMessage;
+import stirling.software.proprietary.model.api.ai.AiDocumentIngestRequest;
+import stirling.software.proprietary.model.api.ai.AiEngineProgressDetail;
 import stirling.software.proprietary.model.api.ai.AiFile;
-import stirling.software.proprietary.model.api.ai.AiRagIngestRequest;
-import stirling.software.proprietary.model.api.ai.AiRagPageText;
+import stirling.software.proprietary.model.api.ai.AiPageText;
 import stirling.software.proprietary.model.api.ai.AiWorkflowFileInput;
 import stirling.software.proprietary.model.api.ai.AiWorkflowFileRequest;
 import stirling.software.proprietary.model.api.ai.AiWorkflowOutcome;
@@ -60,7 +62,7 @@ import tools.jackson.databind.ObjectMapper;
@RequiredArgsConstructor
 public class AiWorkflowService {

-    private static final String RAG_DOCUMENTS_ENDPOINT = "/api/v1/rag/documents";
+    private static final String DOCUMENTS_ENDPOINT = "/api/v1/documents";

    private final CustomPDFDocumentFactory pdfDocumentFactory;
    private final AiEngineClient aiEngineClient;
@@ -76,6 +78,14 @@ public class AiWorkflowService {
    @FunctionalInterface
    public interface ProgressListener {
        void onProgress(AiWorkflowProgressEvent event);
+
+        /**
+         * Called when the engine emits a keep-alive heartbeat. Default is a no-op; consumers that
+         * forward to a downstream connection (e.g. an SSE emitter) override this to push a
+         * heartbeat through, so the next downstream-disconnect surfaces immediately rather than
+         * waiting for the next real progress event.
+         */
+        default void onHeartbeat() {}
    }

    private static final ProgressListener NOOP_LISTENER = event -> {};
@@ -143,7 +153,7 @@ public class AiWorkflowService {
            ProgressListener listener)
            throws IOException {
        listener.onProgress(AiWorkflowProgressEvent.of(AiWorkflowPhase.CALLING_ENGINE));
-        AiWorkflowResponse response = invokeOrchestrator(request);
+        AiWorkflowResponse response = invokeOrchestrator(request, listener);
        return switch (response.getOutcome()) {
            case NEED_CONTENT -> onNeedContent(response, filesById, request, listener);
            case NEED_INGEST -> onNeedIngest(response, filesById, request, listener);
@@ -277,22 +287,22 @@ public class AiWorkflowService {
    }

    private void ingestFile(AiFile file, MultipartFile multipartFile) throws IOException {
-        List<AiRagPageText> pages = new ArrayList<>();
+        List<AiPageText> pages = new ArrayList<>();
        try (PDDocument document = pdfDocumentFactory.load(multipartFile, true)) {
            int pageCount = document.getNumberOfPages();
            for (int pageNumber = 1; pageNumber <= pageCount; pageNumber++) {
                String pageText = pdfContentExtractor.extractPageTextRaw(document, pageNumber);
                if (pageText != null && !pageText.isBlank()) {
-                    pages.add(new AiRagPageText(pageNumber, pageText));
+                    pages.add(new AiPageText(pageNumber, pageText));
                }
            }
        }
-        AiRagIngestRequest ingestRequest =
-                new AiRagIngestRequest(file.getId(), file.getName(), pages);
+        AiDocumentIngestRequest ingestRequest =
+                new AiDocumentIngestRequest(file.getId(), file.getName(), pages);
        String body = objectMapper.writeValueAsString(ingestRequest);
-        aiEngineClient.postLongRunning(RAG_DOCUMENTS_ENDPOINT, body);
+        aiEngineClient.postLongRunning(DOCUMENTS_ENDPOINT, body);
        log.debug(
-                "Ingested file into RAG: id={}, name={}, pages={}",
+                "Ingested document: id={}, name={}, pages={}",
                file.getId(),
                file.getName(),
                pages.size());
@@ -323,10 +333,12 @@ public class AiWorkflowService {
                            result.files(),
                            inputFileNames(filesById),
                            result.report()));
+        } catch (InternalApiTimeoutException e) {
+            log.error("Tool {} timed out: {}", endpointPath, e.getMessage());
+            return new WorkflowState.Terminal(cannotContinue(toolTimeoutMessage(endpointPath, e)));
        } catch (Exception e) {
            log.error("Failed to execute tool {}: {}", endpointPath, e.getMessage(), e);
-            return new WorkflowState.Terminal(
-                    cannotContinue("Tool execution failed: " + e.getMessage()));
+            return new WorkflowState.Terminal(cannotContinue(toolFailureMessage(endpointPath, e)));
        }
    }

@@ -414,6 +426,10 @@ public class AiWorkflowService {
            return new WorkflowState.Terminal(
                    buildCompletedResponse(
                            summary, currentFiles, inputFileNames(filesById), lastReport));
+        } catch (InternalApiTimeoutException e) {
+            log.error("Plan step on tool {} timed out: {}", e.getEndpointPath(), e.getMessage());
+            return new WorkflowState.Terminal(
+                    cannotContinue(toolTimeoutMessage(e.getEndpointPath(), e)));
        } catch (Exception e) {
            log.error("Failed to execute plan: {}", e.getMessage(), e);
            return new WorkflowState.Terminal(
@@ -425,6 +441,20 @@ public class AiWorkflowService {
        return filesById.values().stream().map(MultipartFile::getOriginalFilename).toList();
    }

+    private static String toolTimeoutMessage(String endpointPath, InternalApiTimeoutException e) {
+        return String.format(
+                "The %s tool did not respond within %d seconds and was aborted. The underlying"
+                        + " operation may be hung; try again, run on a smaller file, or use a"
+                        + " different approach.",
+                endpointPath, e.getReadTimeout().toSeconds());
+    }
+
+    private static String toolFailureMessage(String endpointPath, Throwable cause) {
+        String reason =
+                cause.getMessage() != null ? cause.getMessage() : cause.getClass().getSimpleName();
+        return String.format("The %s tool failed: %s", endpointPath, reason);
+    }
+
    /**
     * Execute a single tool step. If the endpoint accepts multiple files, all files are sent in one
     * call. Otherwise, the endpoint is called once per file. ZIP responses are unpacked so each
@@ -477,8 +507,15 @@ public class AiWorkflowService {
        }
        for (Map.Entry<String, Object> entry : parameters.entrySet()) {
            if (entry.getValue() instanceof List<?> list) {
-                for (Object item : list) {
-                    body.add(entry.getKey(), item);
+                if (containsStructuredElements(list)) {
+                    // Endpoints binding lists of structured objects (e.g. /security/redact's
+                    // redactions, /general/edit-text's edits) parse a single JSON string field via
+                    // a property editor. Pre-serialize the whole list so binding succeeds.
+                    body.add(entry.getKey(), objectMapper.writeValueAsString(list));
+                } else {
+                    for (Object item : list) {
+                        body.add(entry.getKey(), item);
+                    }
                }
            } else {
                body.add(entry.getKey(), entry.getValue());
@@ -529,6 +566,15 @@ public class AiWorkflowService {
        }
    }

+    private static boolean containsStructuredElements(List<?> list) {
+        for (Object item : list) {
+            if (item instanceof Map<?, ?> || item instanceof List<?>) {
+                return true;
+            }
+        }
+        return false;
+    }
+
    private List<Resource> toResources(Map<String, MultipartFile> filesById) throws IOException {
        List<Resource> resources = new ArrayList<>();
        for (MultipartFile file : filesById.values()) {
@@ -614,10 +660,57 @@ public class AiWorkflowService {
        return response;
    }

-    private AiWorkflowResponse invokeOrchestrator(WorkflowTurnRequest request) throws IOException {
+    /**
+     * Drive the engine's streaming orchestrator endpoint. Progress events are forwarded to {@code
+     * listener} as they arrive (each one keeps the SSE connection to the frontend alive too). The
+     * final {@code result} event carries the full {@link AiWorkflowResponse}; an {@code error}
+     * event surfaces engine-side failures.
+     */
+    private AiWorkflowResponse invokeOrchestrator(
+            WorkflowTurnRequest request, ProgressListener listener) throws IOException {
        String requestBody = objectMapper.writeValueAsString(request);
-        String responseBody = aiEngineClient.post("/api/v1/orchestrator", requestBody);
-        return objectMapper.readValue(responseBody, AiWorkflowResponse.class);
+        AiWorkflowResponse[] resultHolder = new AiWorkflowResponse[1];
+        String[] errorHolder = new String[1];
+
+        aiEngineClient.streamPost(
+                "/api/v1/orchestrator",
+                requestBody,
+                line -> handleStreamLine(line, listener, resultHolder, errorHolder));
+
+        if (errorHolder[0] != null) {
+            throw new IOException("AI engine returned error: " + errorHolder[0]);
+        }
+        if (resultHolder[0] == null) {
+            throw new IOException("AI engine stream ended without a result");
+        }
+        return resultHolder[0];
+    }
+
+    private void handleStreamLine(
+            String line,
+            ProgressListener listener,
+            AiWorkflowResponse[] resultHolder,
+            String[] errorHolder) {
+        try {
+            JsonNode node = objectMapper.readTree(line);
+            String event = node.path("event").asText();
+            switch (event) {
+                case "progress" -> {
+                    AiEngineProgressDetail detail =
+                            objectMapper.treeToValue(node, AiEngineProgressDetail.class);
+                    listener.onProgress(AiWorkflowProgressEvent.engineProgress(detail));
+                }
+                case "result" -> {
+                    JsonNode response = node.path("response");
+                    resultHolder[0] = objectMapper.treeToValue(response, AiWorkflowResponse.class);
+                }
+                case "error" -> errorHolder[0] = node.path("message").asText("unknown error");
+                case "heartbeat" -> listener.onHeartbeat();
+                default -> log.warn("Ignoring unknown engine stream event: {}", event);
+            }
+        } catch (JacksonException e) {
+            log.warn("Failed to parse engine stream line: {}", line, e);
+        }
    }

    @Data
@@ -2,11 +2,13 @@ package stirling.software.proprietary.service;

 import static org.junit.jupiter.api.Assertions.assertEquals;
 import static org.junit.jupiter.api.Assertions.assertNotNull;
+import static org.junit.jupiter.api.Assertions.assertTrue;
 import static org.mockito.ArgumentMatchers.any;
 import static org.mockito.ArgumentMatchers.anyBoolean;
 import static org.mockito.ArgumentMatchers.anyInt;
 import static org.mockito.ArgumentMatchers.anyString;
 import static org.mockito.ArgumentMatchers.eq;
+import static org.mockito.Mockito.doAnswer;
 import static org.mockito.Mockito.lenient;
 import static org.mockito.Mockito.never;
 import static org.mockito.Mockito.times;
@@ -21,6 +23,7 @@ import java.nio.file.Path;
 import java.util.ArrayList;
 import java.util.List;
 import java.util.concurrent.atomic.AtomicInteger;
+import java.util.function.Consumer;
 import java.util.zip.ZipEntry;
 import java.util.zip.ZipOutputStream;

@@ -30,6 +33,7 @@ import org.junit.jupiter.api.BeforeEach;
 import org.junit.jupiter.api.Test;
 import org.junit.jupiter.api.extension.ExtendWith;
 import org.junit.jupiter.api.io.TempDir;
+import org.mockito.ArgumentCaptor;
 import org.mockito.Mock;
 import org.mockito.junit.jupiter.MockitoExtension;
 import org.springframework.core.io.ByteArrayResource;
@@ -44,6 +48,7 @@ import stirling.software.common.service.CustomPDFDocumentFactory;
 import stirling.software.common.service.FileStorage;
 import stirling.software.common.service.FileStorage.StoredFile;
 import stirling.software.common.service.InternalApiClient;
+import stirling.software.common.service.InternalApiTimeoutException;
 import stirling.software.common.service.ToolMetadataService;
 import stirling.software.common.util.TempFileManager;
 import stirling.software.common.util.TempFileRegistry;
@@ -251,6 +256,158 @@ class AiWorkflowServiceTest {
        verify(internalApiClient, times(1)).post(eq(COMPRESS_ENDPOINT), any());
    }

+    @Test
+    void parameterListOfStructuredObjectsIsJsonEncodedAsSingleField() throws IOException {
+        // Endpoints like /general/edit-text and /security/redact bind a List<StructuredObject>
+        // from a single JSON form field via a property editor. The plan executor must
+        // pre-serialize such lists rather than splitting them into repeated form fields.
+        String editTextEndpoint = "/api/v1/general/edit-text";
+        MockMultipartFile input = pdf("input.pdf", "bytes");
+        stubOrchestrator(
+                """
+                {
+                  "outcome":"plan",
+                  "summary":"Find and replace",
+                  "steps":[
+                    {"tool":"%s","parameters":{
+                      "edits":[{"find":"foo","replace":"bar"},{"find":"baz","replace":"qux"}],
+                      "useRegex":false
+                    }}
+                  ]
+                }
+                """
+                        .formatted(editTextEndpoint));
+        when(toolMetadataService.isMultiInput(editTextEndpoint)).thenReturn(false);
+        when(toolMetadataService.shouldUnpackZipResponse(editTextEndpoint)).thenReturn(false);
+        stubEndpoint(editTextEndpoint, pdfResource("edited", "edited.pdf"));
+        stubFileStorage();
+
+        service.orchestrate(requestFor(input, "find and replace"));
+
+        @SuppressWarnings("unchecked")
+        ArgumentCaptor<MultiValueMap<String, Object>> bodyCaptor =
+                ArgumentCaptor.forClass(MultiValueMap.class);
+        verify(internalApiClient).post(eq(editTextEndpoint), bodyCaptor.capture());
+        MultiValueMap<String, Object> body = bodyCaptor.getValue();
+
+        // The structured-list field must be serialized as ONE JSON-string entry, not multiple
+        // entries.
+        List<Object> editsValues = body.get("edits");
+        assertNotNull(editsValues);
+        assertEquals(1, editsValues.size());
+        assertEquals(
+                "[{\"find\":\"foo\",\"replace\":\"bar\"},{\"find\":\"baz\",\"replace\":\"qux\"}]",
+                editsValues.get(0));
+
+        // Primitive fields keep the original behavior (single value, not JSON-wrapped).
+        List<Object> useRegex = body.get("useRegex");
+        assertNotNull(useRegex);
+        assertEquals(1, useRegex.size());
+        assertEquals(false, useRegex.get(0));
+    }
+
+    @Test
+    void parameterListOfPrimitivesIsSentAsRepeatedFormFields() throws IOException {
+        // Endpoints like /misc/ocr-pdf bind List<String> via Spring's repeated-form-field
+        // convention. The executor must preserve that behavior for primitive lists.
+        String ocrEndpoint = "/api/v1/misc/ocr-pdf";
+        MockMultipartFile input = pdf("input.pdf", "bytes");
+        stubOrchestrator(
+                """
+                {
+                  "outcome":"plan",
+                  "summary":"OCR the document",
+                  "steps":[
+                    {"tool":"%s","parameters":{"languages":["eng","fra","deu"]}}
+                  ]
+                }
+                """
+                        .formatted(ocrEndpoint));
+        when(toolMetadataService.isMultiInput(ocrEndpoint)).thenReturn(false);
+        when(toolMetadataService.shouldUnpackZipResponse(ocrEndpoint)).thenReturn(false);
+        stubEndpoint(ocrEndpoint, pdfResource("ocred", "ocred.pdf"));
+        stubFileStorage();
+
+        service.orchestrate(requestFor(input, "ocr"));
+
+        @SuppressWarnings("unchecked")
+        ArgumentCaptor<MultiValueMap<String, Object>> bodyCaptor =
+                ArgumentCaptor.forClass(MultiValueMap.class);
+        verify(internalApiClient).post(eq(ocrEndpoint), bodyCaptor.capture());
+        List<Object> languages = bodyCaptor.getValue().get("languages");
+        assertNotNull(languages);
+        assertEquals(List.of("eng", "fra", "deu"), languages);
+    }
+
+    @Test
+    void toolTimeoutSurfacesAsCannotContinueWithFriendlyMessage() throws IOException {
+        // When an internal tool hangs, InternalApiClient throws InternalApiTimeoutException after
+        // its read timeout fires. The workflow must convert that into a clean CANNOT_CONTINUE
+        // outcome so the user sees an actionable message rather than the request hanging forever.
+        MockMultipartFile input = pdf("input.pdf", "bytes");
+        stubOrchestrator(
+                """
+                {"outcome":"tool_call","tool":"%s","parameters":{"angle":90},"rationale":"Rotating"}
+                """
+                        .formatted(ROTATE_ENDPOINT));
+        when(toolMetadataService.isMultiInput(ROTATE_ENDPOINT)).thenReturn(false);
+        when(internalApiClient.post(eq(ROTATE_ENDPOINT), any()))
+                .thenThrow(
+                        new InternalApiTimeoutException(
+                                ROTATE_ENDPOINT,
+                                java.time.Duration.ofSeconds(300),
+                                new java.io.IOException("Read timed out")));
+
+        AiWorkflowResponse result = service.orchestrate(requestFor(input, "rotate"));
+
+        assertEquals(AiWorkflowOutcome.CANNOT_CONTINUE, result.getOutcome());
+        assertNotNull(result.getReason());
+        assertTrue(
+                result.getReason().contains(ROTATE_ENDPOINT),
+                "reason should mention the failing tool path: " + result.getReason());
+        assertTrue(
+                result.getReason().contains("300"),
+                "reason should mention the timeout duration: " + result.getReason());
+        verify(internalApiClient, times(1)).post(eq(ROTATE_ENDPOINT), any());
+    }
+
+    @Test
+    void planTimeoutOnLaterStepStillSurfacesCleanly() throws IOException {
+        // Multi-step plans must also handle a hung tool gracefully (no partial leaks) and the
+        // failure message must identify which step failed so the user can iterate.
+        MockMultipartFile input = pdf("input.pdf", "bytes");
+        stubOrchestrator(
+                """
+                {
+                  "outcome":"plan",
+                  "summary":"Rotate then compress",
+                  "steps":[
+                    {"tool":"%s","parameters":{"angle":90}},
+                    {"tool":"%s","parameters":{}}
+                  ]
+                }
+                """
+                        .formatted(ROTATE_ENDPOINT, COMPRESS_ENDPOINT));
+        when(toolMetadataService.isMultiInput(anyString())).thenReturn(false);
+        when(toolMetadataService.shouldUnpackZipResponse(anyString())).thenReturn(false);
+        stubEndpoint(ROTATE_ENDPOINT, pdfResource("rotated", "rotated.pdf"));
+        // No fileStorage stub here: the second step times out before any output reaches storage.
+        when(internalApiClient.post(eq(COMPRESS_ENDPOINT), any()))
+                .thenThrow(
+                        new InternalApiTimeoutException(
+                                COMPRESS_ENDPOINT,
+                                java.time.Duration.ofSeconds(300),
+                                new java.io.IOException("Read timed out")));
+
+        AiWorkflowResponse result = service.orchestrate(requestFor(input, "rotate then compress"));
+
+        assertEquals(AiWorkflowOutcome.CANNOT_CONTINUE, result.getOutcome());
+        assertTrue(result.getReason().contains(COMPRESS_ENDPOINT));
+        // The first step actually executed; the failure happened on the second.
+        verify(internalApiClient, times(1)).post(eq(ROTATE_ENDPOINT), any());
+        verify(internalApiClient, times(1)).post(eq(COMPRESS_ENDPOINT), any());
+    }
+
    @Test
    void toolCallWithoutEndpointFallsBackToCannotContinue() throws IOException {
        MockMultipartFile input = pdf("input.pdf", "bytes");
@@ -264,7 +421,7 @@ class AiWorkflowServiceTest {
    }

    @Test
-    void needIngestExtractsPageTextAndPostsToRagThenRetries() throws IOException {
+    void needIngestExtractsPageTextAndPostsThenRetries() throws IOException {
        MockMultipartFile input = pdf("report.pdf", "bytes");
        when(fileIdStrategy.idFor(any())).thenReturn("report-id");

@@ -276,37 +433,64 @@ class AiWorkflowServiceTest {
                .thenReturn("page content");

        int[] orchestratorCalls = {0};
-        when(aiEngineClient.post(eq("/api/v1/orchestrator"), anyString()))
-                .thenAnswer(
+        doAnswer(
                        inv -> {
                            orchestratorCalls[0]++;
+                            String responseJson;
                            if (orchestratorCalls[0] == 1) {
-                                return """
-                                       {
-                                         "outcome":"need_ingest",
-                                         "resumeWith":"pdf_question",
-                                         "reason":"ingest first",
-                                         "filesToIngest":[{"id":"report-id","name":"report.pdf"}],
-                                         "contentTypes":["page_text"]
-                                       }
-                                       """;
+                                responseJson =
+                                        """
+                                        {
+                                          "outcome":"need_ingest",
+                                          "resumeWith":"pdf_question",
+                                          "reason":"ingest first",
+                                          "filesToIngest":[{"id":"report-id","name":"report.pdf"}],
+                                          "contentTypes":["page_text"]
+                                        }
+                                        """;
+                            } else {
+                                responseJson =
+                                        """
+                                        {"outcome":"answer","answer":"done","evidence":[]}
+                                        """;
                            }
-                            return """
-                                   {"outcome":"answer","answer":"done","evidence":[]}
-                                   """;
-                        });
+                            Consumer<String> consumer = inv.getArgument(2);
+                            consumer.accept(wrapAsResultEvent(responseJson));
+                            return null;
+                        })
+                .when(aiEngineClient)
+                .streamPost(eq("/api/v1/orchestrator"), anyString(), any());

        AiWorkflowResponse result = service.orchestrate(requestFor(input, "summarise this"));

        assertEquals(AiWorkflowOutcome.ANSWER, result.getOutcome());
-        verify(aiEngineClient, times(1)).postLongRunning(eq("/api/v1/rag/documents"), anyString());
-        verify(aiEngineClient, times(2)).post(eq("/api/v1/orchestrator"), anyString());
+        verify(aiEngineClient, times(1)).postLongRunning(eq("/api/v1/documents"), anyString());
+        verify(aiEngineClient, times(2)).streamPost(eq("/api/v1/orchestrator"), anyString(), any());
    }

    // --- helpers ---

    private void stubOrchestrator(String responseJson) throws IOException {
-        when(aiEngineClient.post(eq("/api/v1/orchestrator"), anyString())).thenReturn(responseJson);
+        doAnswer(
+                        inv -> {
+                            Consumer<String> consumer = inv.getArgument(2);
+                            consumer.accept(wrapAsResultEvent(responseJson));
+                            return null;
+                        })
+                .when(aiEngineClient)
+                .streamPost(eq("/api/v1/orchestrator"), anyString(), any());
+    }
+
+    /**
+     * Wrap a unary orchestrator response JSON as the NDJSON "result" event the streaming endpoint
+     * emits, so existing tests can keep their per-outcome JSON literals.
+     */
+    private String wrapAsResultEvent(String responseJson) throws IOException {
+        return objectMapper
+                .createObjectNode()
+                .put("event", "result")
+                .set("response", objectMapper.readTree(responseJson))
+                .toString();
    }

    private void stubEndpoint(String endpoint, Resource body) {
@@ -20,10 +20,10 @@ import org.gradle.api.tasks.testing.Test
 import org.gradle.jvm.toolchain.JavaLanguageVersion

 ext {
-    springBootVersion = "4.0.5"
+    springBootVersion = "4.0.6"
    pdfboxVersion = "3.0.7"
    imageioVersion = "3.13.1"
-    lombokVersion = "1.18.44"
+    lombokVersion = "1.18.46"
    bouncycastleVersion = "1.83"
    springSecuritySamlVersion = "7.0.5"
    openSamlVersion = "5.2.1"
@@ -78,7 +78,7 @@ springBoot {

 allprojects {
    group = 'stirling.software'
-    version = '2.10.1'
+    version = '2.11.0'

    configurations.configureEach {
        exclude group: "org.springframework.boot", module: "spring-boot-starter-tomcat"
@@ -101,11 +101,14 @@ tasks.register('syncAppVersion') {

        def tauriConfigFile = new File(tauriConfigPath)
        if (tauriConfigFile.exists()) {
-            def parsed = new groovy.json.JsonSlurper().parse(tauriConfigFile)
-            parsed.version = appVersionStr
-            def formatted = groovy.json.JsonOutput.prettyPrint(groovy.json.JsonOutput.toJson(parsed)) + System.lineSeparator()
-            if (tauriConfigFile.getText('UTF-8') != formatted) {
-                tauriConfigFile.write(formatted, 'UTF-8')
+            def content = tauriConfigFile.getText('UTF-8')
+            def matcher = (content =~ /(?m)^(\s*"version":\s*")([^"]*)(")/)
+            if (!matcher.find()) {
+                throw new GradleException("Could not locate version in ${tauriConfigFile} for synchronization")
+            }
+            def updatedContent = matcher.replaceFirst("${matcher.group(1)}${appVersionStr}${matcher.group(3)}")
+            if (content != updatedContent) {
+                tauriConfigFile.write(updatedContent, 'UTF-8')
            }
        }

@@ -97,7 +97,7 @@ To add a 3rd UNO server:
 ```yaml
  unoserver3:
    container_name: UNO-Server-3
-    image: ghcr.io/unoconv/unoserver-docker:0.4.4
+    image: ghcr.io/stirling-tools/stirling-unoserver:latest
    # ... same config as unoserver1/2
 ```

@@ -57,38 +57,34 @@ services:

  unoserver1:
    container_name: UNO-Server-1
-    # NOTE: This image needs to be updated to unoserver 3.6 to match Stirling-PDF's client version
-    # Current :latest uses 3.4 which causes API mismatch errors
-    image: ghcr.io/unoconv/unoserver-docker:latest
+    # See docker/unoserver/VERSION for available tags.
+    image: ghcr.io/stirling-tools/stirling-unoserver:latest
+    environment:
+      UNOSERVER_PORT: "2003"
+      UNOSERVER_UNO_PORT: "2002"
+      UNOSERVER_CONVERSION_TIMEOUT: "1800"
+      # Recycle hourly to bound LibreOffice memory; 0 disables, floor 60s.
+      UNOSERVER_RECYCLE_INTERVAL_SECONDS: "3600"
    volumes:
      - stirling-tmp:/tmp/stirling-pdf:rw
    expose:
      - "2003"
-    healthcheck:
-      test: ["CMD-SHELL", "timeout 2 bash -c 'cat < /dev/null > /dev/tcp/localhost/2003' || exit 1"]
-      interval: 10s
-      timeout: 5s
-      retries: 5
-      start_period: 30s
    networks:
      - stirling-network
    restart: on-failure:5

  unoserver2:
    container_name: UNO-Server-2
-    # NOTE: This image needs to be updated to unoserver 3.6 to match Stirling-PDF's client version
-    # Current :latest uses 3.4 which causes API mismatch errors
-    image: ghcr.io/unoconv/unoserver-docker:latest
+    image: ghcr.io/stirling-tools/stirling-unoserver:latest
+    environment:
+      UNOSERVER_PORT: "2003"
+      UNOSERVER_UNO_PORT: "2002"
+      UNOSERVER_CONVERSION_TIMEOUT: "1800"
+      UNOSERVER_RECYCLE_INTERVAL_SECONDS: "3600"
    volumes:
      - stirling-tmp:/tmp/stirling-pdf:rw
    expose:
      - "2003"
-    healthcheck:
-      test: ["CMD-SHELL", "timeout 2 bash -c 'cat < /dev/null > /dev/tcp/localhost/2003' || exit 1"]
-      interval: 10s
-      timeout: 5s
-      retries: 5
-      start_period: 30s
    networks:
      - stirling-network
    restart: on-failure:5
@@ -0,0 +1,106 @@
+# Standalone unoserver image for Stirling-PDF remote UNO mode.
+# Pinned to unoserver 3.6 to match Stirling-PDF's client (avoids 3.4/3.6 wire mismatch).
+
+FROM ubuntu:noble@sha256:c4a8d5503dfb2a3eb8ab5f807da5bc69a85730fb49b5cfca2330194ebcc41c7b
+
+ARG UNOSERVER_VERSION=3.6
+# ~120 MB of CJK fonts — opt-in.
+ARG INSTALL_CJK_FONTS=false
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    LANG=C.UTF-8 \
+    LC_ALL=C.UTF-8 \
+    UNOSERVER_PORT=2003 \
+    UNOSERVER_UNO_PORT=2002 \
+    UNOSERVER_INTERFACE=0.0.0.0 \
+    UNOSERVER_CONVERSION_TIMEOUT=1800 \
+    UNOSERVER_RECYCLE_INTERVAL_SECONDS=0 \
+    UNOSERVER_PROFILE_DIR=/var/lib/unoserver/profile
+
+RUN set -eux; \
+    apt-get update; \
+    apt-get install -y --no-install-recommends \
+      software-properties-common ca-certificates curl gnupg tini; \
+    add-apt-repository -y ppa:libreoffice/ppa; \
+    apt-get update; \
+    apt-get install -y --no-install-recommends \
+      python3 python3-venv python3-uno python3-packaging \
+      libreoffice-writer-nogui libreoffice-calc-nogui \
+      libreoffice-impress-nogui libreoffice-draw-nogui \
+      libreoffice-base-nogui libreoffice-java-common \
+      xvfb x11-utils coreutils \
+      fonts-dejavu fonts-liberation fonts-liberation2 \
+      fonts-crosextra-caladea fonts-crosextra-carlito \
+      fonts-noto-core fonts-noto-mono fonts-noto-color-emoji \
+      hyphen-en-gb hyphen-en-us hyphen-fr hyphen-de hyphen-es hyphen-it \
+      hyphen-pt-pt hyphen-pt-br hyphen-nl hyphen-pl hyphen-ru \
+      fontconfig \
+      libcairo2 libpango-1.0-0 libpangoft2-1.0-0 libgdk-pixbuf-2.0-0 \
+      libfreetype6 libfontconfig1 libx11-6 libxt6 libxext6 libxrender1 \
+      libxtst6 libxi6 libxinerama1 libxkbcommon0 libsm6 libice6 \
+      libglib2.0-0; \
+    if [ "${INSTALL_CJK_FONTS}" = "true" ]; then \
+      apt-get install -y --no-install-recommends fonts-noto-cjk; \
+    fi; \
+    apt-get remove --purge -y software-properties-common; \
+    apt-get autoremove --purge -y; \
+    rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
+
+# Noble's hyphen-* packages skip LibreOffice registration; wire it up by hand
+# or the dicts never load.
+RUN set -eu; \
+    mkdir -p /usr/share/myspell/dicts /usr/share/myspell/infos/ooo; \
+    for f in /usr/share/hyphen/hyph_*_*.dic; do \
+      base=$(basename "$f" .dic); \
+      locale=${base#hyph_}; \
+      lang=${locale%_*}; \
+      country=${locale#*_}; \
+      ln -sfn "../../hyphen/${base}.dic" "/usr/share/myspell/dicts/${base}.dic"; \
+      printf "HYPH %s %s %s\n" "$lang" "$country" "$base" \
+        > "/usr/share/myspell/infos/ooo/hyphen-${lang}-${country}"; \
+    done
+
+COPY --chmod=0644 docker/unoserver/fonts.conf /etc/fonts/conf.d/100-stirling.conf
+RUN fc-cache -f
+
+# venv with --system-site-packages so it picks up distro python3-uno.
+RUN python3 -m venv /opt/unoserver-venv --system-site-packages && \
+    /opt/unoserver-venv/bin/pip install --no-cache-dir "unoserver==${UNOSERVER_VERSION}" && \
+    ln -sf /opt/unoserver-venv/bin/unoserver  /usr/local/bin/unoserver && \
+    ln -sf /opt/unoserver-venv/bin/unoconvert /usr/local/bin/unoconvert && \
+    ln -sf /opt/unoserver-venv/bin/unoping    /usr/local/bin/unoping && \
+    find /opt/unoserver-venv -type d -name __pycache__ -exec rm -rf {} + 2>/dev/null || true
+
+RUN rm -rf /usr/lib/libreoffice/share/gallery \
+           /usr/lib/libreoffice/share/template \
+           /usr/lib/libreoffice/share/wizards \
+           /usr/lib/libreoffice/help \
+           /usr/lib/libreoffice/share/extensions/wiki-publisher \
+           /usr/lib/libreoffice/share/extensions/nlpsolver \
+           /usr/share/doc/* /usr/share/man/* /usr/share/info/*
+
+ARG PUID=1001
+ARG PGID=1001
+RUN groupadd -g "${PGID}" unoserver && \
+    useradd  -u "${PUID}" -g unoserver -m -d /home/unoserver -s /bin/bash unoserver && \
+    mkdir -p "${UNOSERVER_PROFILE_DIR}" && \
+    chown -R unoserver:unoserver "${UNOSERVER_PROFILE_DIR}" /home/unoserver
+
+COPY --chmod=0755 docker/unoserver/entrypoint.sh /usr/local/bin/entrypoint.sh
+COPY --chmod=0755 docker/unoserver/healthcheck.sh /usr/local/bin/healthcheck.sh
+
+USER unoserver
+WORKDIR /home/unoserver
+
+EXPOSE 2003
+
+HEALTHCHECK --interval=15s --timeout=5s --retries=4 --start-period=45s \
+  CMD /usr/local/bin/healthcheck.sh
+
+ENTRYPOINT ["/usr/bin/tini", "--", "/usr/local/bin/entrypoint.sh"]
+
+LABEL org.opencontainers.image.title="Stirling-PDF Unoserver" \
+      org.opencontainers.image.description="Pinned unoserver build for Stirling-PDF remote UNO mode" \
+      org.opencontainers.image.source="https://github.com/Stirling-Tools/Stirling-PDF" \
+      org.opencontainers.image.licenses="MIT" \
+      org.opencontainers.image.vendor="Stirling-Tools"
@@ -0,0 +1 @@
+1.0.0
@@ -0,0 +1,104 @@
+#!/bin/bash
+set -eu
+
+PORT="${UNOSERVER_PORT:-2003}"
+UNO_PORT="${UNOSERVER_UNO_PORT:-2002}"
+INTERFACE="${UNOSERVER_INTERFACE:-0.0.0.0}"
+CONVERSION_TIMEOUT="${UNOSERVER_CONVERSION_TIMEOUT:-1800}"
+RECYCLE_INTERVAL_SECONDS="${UNOSERVER_RECYCLE_INTERVAL_SECONDS:-0}"
+RECYCLE_INTERVAL_FLOOR=60
+PROFILE_DIR="${UNOSERVER_PROFILE_DIR:-/var/lib/unoserver/profile}"
+
+log() { printf '%s %s\n' "[unoserver-entrypoint]" "$*" >&2; }
+
+case "$PORT" in ''|*[!0-9]*) log "Invalid UNOSERVER_PORT='$PORT'"; exit 64 ;; esac
+case "$UNO_PORT" in ''|*[!0-9]*) log "Invalid UNOSERVER_UNO_PORT='$UNO_PORT'"; exit 64 ;; esac
+case "$CONVERSION_TIMEOUT" in ''|*[!0-9]*) log "Invalid UNOSERVER_CONVERSION_TIMEOUT='$CONVERSION_TIMEOUT'"; exit 64 ;; esac
+case "$RECYCLE_INTERVAL_SECONDS" in ''|*[!0-9]*) log "Invalid UNOSERVER_RECYCLE_INTERVAL_SECONDS='$RECYCLE_INTERVAL_SECONDS'"; exit 64 ;; esac
+
+mkdir -p "$PROFILE_DIR"
+
+start_xvfb() {
+  if command -v Xvfb >/dev/null 2>&1 && [ -z "${DISPLAY:-}" ]; then
+    Xvfb :99 -screen 0 1024x768x24 -ac +extension GLX +render -noreset >/dev/null 2>&1 &
+    XVFB_PID=$!
+    export DISPLAY=:99
+    sleep 1
+    log "Xvfb started (pid $XVFB_PID, DISPLAY=$DISPLAY)"
+  fi
+}
+
+cleanup() {
+  trap '' TERM INT EXIT
+  if [ -n "${UNOSERVER_PID:-}" ] && kill -0 "$UNOSERVER_PID" 2>/dev/null; then
+    log "Stopping unoserver (pid $UNOSERVER_PID)"
+    pkill -TERM -P "$UNOSERVER_PID" 2>/dev/null || true
+    kill -TERM "$UNOSERVER_PID" 2>/dev/null || true
+    wait "$UNOSERVER_PID" 2>/dev/null || true
+  fi
+  if [ -n "${XVFB_PID:-}" ] && kill -0 "$XVFB_PID" 2>/dev/null; then
+    kill -TERM "$XVFB_PID" 2>/dev/null || true
+  fi
+}
+trap cleanup TERM INT EXIT
+
+start_unoserver() {
+  log "Starting unoserver on ${INTERFACE}:${PORT} (uno-port ${UNO_PORT}, timeout ${CONVERSION_TIMEOUT}s, profile ${PROFILE_DIR})"
+  # Pass --user-installation as a plain path; unoserver 3.6 wraps it itself
+  # and crashes if pre-wrapped as a file:// URI.
+  unoserver \
+    --interface "$INTERFACE" \
+    --port "$PORT" \
+    --uno-port "$UNO_PORT" \
+    --user-installation "${PROFILE_DIR}" \
+    --conversion-timeout "$CONVERSION_TIMEOUT" \
+    2> >(grep --line-buffered -v "POST /RPC2" >&2) \
+    &
+  UNOSERVER_PID=$!
+}
+
+# Wall-clock recycle to bound LibreOffice memory growth. wait -n is unreliable
+# here because the unoserver job is wrapped in a process substitution.
+recycle_supervisor() {
+  if [ "$RECYCLE_INTERVAL_SECONDS" -le 0 ]; then
+    log "Recycle disabled (UNOSERVER_RECYCLE_INTERVAL_SECONDS=0)"
+    wait "$UNOSERVER_PID"
+    return $?
+  fi
+  local interval="$RECYCLE_INTERVAL_SECONDS"
+  if [ "$interval" -lt "$RECYCLE_INTERVAL_FLOOR" ]; then
+    log "Clamping recycle interval ${interval}s up to floor ${RECYCLE_INTERVAL_FLOOR}s"
+    interval="$RECYCLE_INTERVAL_FLOOR"
+  fi
+  log "Recycle enabled: restart every ${interval}s"
+  while true; do
+    local elapsed=0
+    while [ "$elapsed" -lt "$interval" ]; do
+      if ! kill -0 "$UNOSERVER_PID" 2>/dev/null; then
+        wait "$UNOSERVER_PID"
+        local rc=$?
+        log "unoserver exited on its own (rc=$rc); not recycling"
+        return "$rc"
+      fi
+      sleep 1
+      elapsed=$((elapsed + 1))
+    done
+    log "Recycling unoserver (pid ${UNOSERVER_PID})"
+    pkill -TERM -P "$UNOSERVER_PID" 2>/dev/null || true
+    kill  -TERM "$UNOSERVER_PID" 2>/dev/null || true
+    for _ in 1 2 3 4 5; do
+      kill -0 "$UNOSERVER_PID" 2>/dev/null || break
+      sleep 1
+    done
+    pkill -KILL -P "$UNOSERVER_PID" 2>/dev/null || true
+    kill  -KILL "$UNOSERVER_PID" 2>/dev/null || true
+    wait "$UNOSERVER_PID" 2>/dev/null || true
+    rm -rf "${PROFILE_DIR:?}"/* 2>/dev/null || true
+    start_unoserver
+    log "unoserver restarted (pid ${UNOSERVER_PID})"
+  done
+}
+
+start_xvfb
+start_unoserver
+recycle_supervisor
@@ -0,0 +1,30 @@
+<?xml version='1.0'?>
+<!DOCTYPE fontconfig SYSTEM 'fonts.dtd'>
+<!-- Subpixel hinting + slight hint + LCD filter for headless LibreOffice. -->
+<fontconfig>
+  <match target="font">
+    <edit mode="assign" name="rgba">
+      <const>rgb</const>
+    </edit>
+  </match>
+  <match target="font">
+    <edit mode="assign" name="hinting">
+      <bool>true</bool>
+    </edit>
+  </match>
+  <match target="font">
+    <edit mode="assign" name="hintstyle">
+      <const>hintslight</const>
+    </edit>
+  </match>
+  <match target="font">
+    <edit mode="assign" name="antialias">
+      <bool>true</bool>
+    </edit>
+  </match>
+  <match target="font">
+    <edit mode="assign" name="lcdfilter">
+      <const>lcddefault</const>
+    </edit>
+  </match>
+</fontconfig>
@@ -0,0 +1,12 @@
+#!/bin/bash
+# Real RPC ping — a TCP socket can be bound while LibreOffice is wedged.
+set -eu
+
+PORT="${UNOSERVER_PORT:-2003}"
+
+if command -v unoping >/dev/null 2>&1; then
+  unoping --host 127.0.0.1 --port "$PORT" >/dev/null 2>&1
+  exit $?
+fi
+
+exec timeout 2 bash -c "cat < /dev/null > /dev/tcp/127.0.0.1/${PORT}"
@@ -37,6 +37,22 @@ STIRLING_RAG_TOP_K=20
 # rather than chain more searches.
 STIRLING_RAG_MAX_SEARCHES=5

+# Chunked reasoner settings: how big each per-worker slice is (in characters),
+# how many workers may run in parallel against the fast model, and how long
+# any single worker is allowed to wait for a response before being abandoned.
+# Worker timeouts protect gather_notes from upstream model stalls (which
+# otherwise hang at the provider's ~10 minute HTTP default); the affected
+# slice is dropped and the rest of the document still answers.
+STIRLING_CHUNKED_REASONER_CHARS_PER_SLICE=16000
+STIRLING_CHUNKED_REASONER_CONCURRENCY=10
+STIRLING_CHUNKED_REASONER_WORKER_TIMEOUT_SECONDS=60
+
+# When the rendered slice notes would exceed this many characters, the
+# reasoner folds them hierarchically with fast-model calls until they fit.
+# This keeps the synthesis prompt under the model's context limit on long
+# documents (a 3000-page novel produces ~900k chars of raw notes).
+STIRLING_CHUNKED_REASONER_NOTES_CHAR_BUDGET=250000
+
 # Upper bounds on PDF page text the engine will request per extraction round.
 STIRLING_MAX_PAGES=200
 STIRLING_MAX_CHARACTERS=200000
@@ -51,3 +67,8 @@ STIRLING_LOG_LEVEL=INFO

 # Path to log file. Rolls daily, keeps 1 backup. Leave empty for console only.
 STIRLING_LOG_FILE=
+
+# Set true to log every outgoing httpx / Anthropic SDK request with timing.
+# Use when diagnosing worker stalls: a hung call shows a "Request" line with
+# no matching "Response" line. Noisy; leave off in normal use.
+STIRLING_HTTP_DEBUG=false
@@ -23,7 +23,7 @@ dev = [
    "anyio>=4.0.0",
    "datamodel-code-generator[ruff]>=0.26.0",
    "pytest>=8.0.0",
-    "pyright>=1.1.408",
+    "pyrefly==1.0.0",
    "referencing>=0.35.0",
    "ruff>=0.14.10",
 ]
@@ -51,21 +51,11 @@ select = [
    "W",
    "RUF100",
    "UP",
-    "PYI",    # flake8-pyi: flags deprecated typing constructs
-    "FA",     # flake8-future-annotations: flags missing future annotations imports
-    "BLE",    # flake8-blind-except: flags bare `except Exception`
+    "PYI", # flake8-pyi: flags deprecated typing constructs
+    "FA", # flake8-future-annotations: flags missing future annotations imports
+    "BLE", # flake8-blind-except: flags bare `except Exception`
 ]

-[tool.pyright]
-pythonVersion = "3.13"
-
-reportImportCycles = "warning"
-reportUnnecessaryCast = "warning"
-reportUnnecessaryTypeIgnoreComment = "warning"
-reportUnusedImport = "warning"
-reportUnknownParameterType = "warning"
-reportDeprecated = "warning"
-
 [tool.pytest.ini_options]
 testpaths = ["tests"]
 # ``tests`` is on the path so test modules can import shared helpers (e.g.
@@ -74,3 +64,23 @@ pythonpath = ["src", "tests"]
 # Use importlib import mode so test directories don't need __init__.py files
 # and duplicate basenames (e.g. multiple test_routes.py) collect cleanly.
 addopts = "--import-mode=importlib"
+
+[tool.pyrefly]
+python-version = "3.13.0"
+search-path = [
+    "src",
+    "tests",
+]
+# Disallow `type: ignore` comments to encourage accurate error codes to be used in ignore comments
+# Pyrefly won't flag `type: ignore [wrong-code]`, but it will flag `pyrefly: ignore [wrong-code]`
+enabled-ignores = ["pyrefly"]
+min-severity = "warn"  # TODO: Change settings below to `warn` - https://github.com/facebook/pyrefly/issues/3370
+
+[tool.pyrefly.errors]
+deprecated = "error"
+implicit-any = "error"
+redundant-cast = "error"
+unused-ignore = "error"
+# TODO: Add equivalents to:
+#  - reportImportCycles
+#  - reportUnusedImport
@@ -478,6 +478,7 @@ class MathAuditorAgent:

        logger.info("[math-auditor-agent] extracting figures from page %d (%d chars)", folio.page, len(text))
        prompt = f"Page {folio.page + 1} text:\n{text}"
+        figures: list[ExtractedFigure] = []
        try:
            result = await self._figure_extractor.run(prompt)
            figures = result.output.figures
@@ -487,7 +488,6 @@ class MathAuditorAgent:
                folio.page,
                exc_info=True,
            )
-            figures = []

        logger.debug(
            "TOOL (extract_figures)\nArgs: %s\nResult: %s",
@@ -2,7 +2,7 @@ from __future__ import annotations

 import logging
 from collections.abc import Iterable
-from typing import Literal, overload
+from typing import Literal, Protocol, overload

 from pydantic import Field
 from pydantic_ai import Agent
@@ -72,6 +72,16 @@ class PdfEditSelectionAgent:
        return result.output


+class ParameterSelector(Protocol):
+    async def select(
+        self,
+        request: PdfEditRequest,
+        operation_plan: list[ToolEndpoint],
+        operation_index: int,
+        generated_steps: list[ToolOperationStep],
+    ) -> ParamToolModel: ...
+
+
 class PdfEditParameterSelector:
    def __init__(self, runtime: AppRuntime) -> None:
        self.runtime = runtime
@@ -141,7 +151,7 @@ class PdfEditParameterSelector:
 class PdfEditAgent:
    def __init__(self, runtime: AppRuntime) -> None:
        self.runtime = runtime
-        self.parameter_selector = PdfEditParameterSelector(runtime)
+        self.parameter_selector: ParameterSelector = PdfEditParameterSelector(runtime)

    async def orchestrate(self, request: OrchestratorRequest) -> PdfEditResponse:
        """Entry point for the orchestrator delegate — adapts the orchestrator's
@@ -6,6 +6,7 @@ from pydantic_ai import Agent
 from pydantic_ai.output import NativeOutput

 from stirling.agents.math_presentation import MathIntentClassifier, extract_math_verdict
+from stirling.agents.shared import ChunkedReasoner, WholeDocReaderCapability
 from stirling.contracts import (
    AiFile,
    EditPlanResponse,
@@ -24,44 +25,46 @@ from stirling.contracts import (
    format_conversation_history,
    format_file_names,
 )
+from stirling.documents import RagCapability
 from stirling.models.agent_tool_models import AgentToolId, MathAuditorAgentParams
-from stirling.rag import RagCapability
 from stirling.services import AppRuntime

 logger = logging.getLogger(__name__)


 PDF_QUESTION_SYSTEM_PROMPT = (
-    "You answer questions about PDF documents by retrieving relevant content with the "
-    "search_knowledge tool. Use it before answering. Do not guess or use outside knowledge.\n"
+    "You answer questions about PDF documents using two retrieval tools:\n"
    "\n"
-    "The search_knowledge tool has a finite call budget per run. When it is no longer "
-    "available, answer from what you have already retrieved.\n"
+    "1. search_knowledge(query) - returns the passages most semantically similar "
+    "to the query. Use it for targeted lookups: a specific fact, a named section, "
+    "a particular passage. Typically one or two calls is enough.\n"
+    "\n"
+    "2. read_full_document(query) - reads every page of the attached documents in "
+    "parallel and returns notes relevant to the query. Use it when answering "
+    "requires seeing the whole document end-to-end: summaries, aggregations "
+    "(largest, shortest, count), comparisons across sections. It is more "
+    "expensive than search_knowledge, so prefer search_knowledge when one or two "
+    "passages would suffice.\n"
+    "\n"
+    "Pick the right tool, call it, then answer from what you got back. Do not "
+    "guess or use outside knowledge.\n"
    "\n"
    "Guidelines:\n"
-    "- Make targeted search_knowledge calls. Typically one or two is enough.\n"
-    "- Answer from the retrieved text. If the retrieved content doesn't support a confident "
-    "answer, return not_found.\n"
-    "- For questions that would require reading the entire document end-to-end (e.g. "
-    "'what's the shortest chapter', 'how many X are there'), return not_found.\n"
-    "- Include a short list of evidence snippets (with page numbers where available) drawn "
-    "from what search_knowledge returned.\n"
+    "- If the retrieved content does not support a confident answer, return not_found.\n"
+    "- Include a short list of evidence snippets (with page numbers where available) "
+    "drawn from what the tools returned.\n"
    "\n"
    "Writing the not_found reason:\n"
    "- The reason is shown directly to the end user, so write it in plain, friendly "
    "language. One or two short sentences.\n"
    "- NEVER mention 'RAG', 'retrieval', 'chunks', 'search results', 'targeted search', "
-    "'search_knowledge', or other implementation details.\n"
-    "- Be honest about the actual limitation. For questions that require full-document "
-    "analysis (shortest chapter, word counts, etc.), explain that the document is too "
-    "long to analyse end-to-end: you can only look up specific passages, and that's "
-    "not enough to compare every part of the document against every other.\n"
+    "'search_knowledge', 'read_full_document', or other implementation details.\n"
    "- For questions where the answer just isn't in the document, say so directly: "
    "'I couldn't find that information in the document.'\n"
-    "- Do not make it sound like you're choosing not to answer. Be clear that it's "
-    "a genuine constraint."
+    "- Do not make it sound like you're choosing not to answer."
 )

+
 _MATH_SYNTH_SYSTEM_PROMPT = (
    "You are given a math-audit Verdict (structured JSON) and the user's "
    "original question. Answer the question in plain prose using only "
@@ -83,6 +86,9 @@ class PdfQuestionAgent:
            model_settings=runtime.fast_model_settings,
        )
        self._math_intent_classifier = MathIntentClassifier(runtime)
+        # Shared across whole-doc-reader instances so the worker agent and
+        # semaphore are constructed once and reused per request.
+        self._chunked_reasoner = ChunkedReasoner(runtime)

    async def handle(self, request: PdfQuestionRequest) -> PdfQuestionResponse:
        logger.info(
@@ -95,7 +101,7 @@ class PdfQuestionAgent:
            logger.info("[pdf-question] missing ingestions: %s", [file.name for file in missing])
            return NeedIngestResponse(
                resume_with=SupportedCapability.PDF_QUESTION,
-                reason="Some files have not been ingested into RAG yet.",
+                reason="Some files have not been ingested yet.",
                files_to_ingest=missing,
                content_types=[PdfContentType.PAGE_TEXT],
            )
@@ -145,23 +151,37 @@ class PdfQuestionAgent:
    async def _find_missing_files(self, files: list[AiFile]) -> list[AiFile]:
        missing: list[AiFile] = []
        for file in files:
-            if not await self.runtime.rag_service.has_collection(file.id):
+            if not await self.runtime.documents.has_collection(file.id):
                missing.append(file)
        return missing

    async def _run_answer_agent(self, request: PdfQuestionRequest) -> PdfQuestionTerminalResponse:
+        """Drive a single smart-model agent with both retrieval tools.
+
+        The agent picks ``search_knowledge`` for targeted lookups and
+        ``read_full_document`` for whole-document questions. Removing the
+        upstream classifier keeps that judgement in the same call that writes
+        the answer, and lets the agent mix tools when the question warrants it.
+        """
        rag = RagCapability(
-            rag_service=self.runtime.rag_service,
+            documents=self.runtime.documents,
            collections=[file.id for file in request.files],
            top_k=self.runtime.settings.rag_default_top_k,
            max_searches=self.runtime.settings.rag_max_searches,
        )
+        whole_doc = WholeDocReaderCapability(
+            runtime=self.runtime,
+            files=request.files,
+            reasoner=self._chunked_reasoner,
+        )
        agent = Agent(
            model=self.runtime.smart_model,
            output_type=NativeOutput([PdfQuestionAnswerResponse, PdfQuestionNotFoundResponse]),
            system_prompt=PDF_QUESTION_SYSTEM_PROMPT,
-            instructions=rag.instructions,
-            toolsets=[rag.toolset],
+            # pydantic-ai accepts a list of (string-or-callable) instruction sources;
+            # it resolves each at run time and concatenates them for the model.
+            instructions=[rag.instructions, whole_doc.instructions],
+            toolsets=[rag.toolset, whole_doc.toolset],
            model_settings=self.runtime.smart_model_settings,
        )
        prompt = self._build_prompt(request)
@@ -184,5 +204,5 @@ class PdfQuestionAgent:
            f"Conversation history:\n{history}\n"
            f"Files: {format_file_names(request.files)}\n"
            f"Question: {request.question}\n"
-            "Use search_knowledge to retrieve the relevant content, then answer."
+            "Pick the right retrieval tool for this question, then answer from what it returns."
        )
@@ -0,0 +1,10 @@
+"""Reasoning utilities shared across agents."""
+
+from stirling.agents.shared.chunked_reasoner import ChunkedReasoner, ChunkNotes
+from stirling.agents.shared.whole_doc_reader import WholeDocReaderCapability
+
+__all__ = [
+    "ChunkNotes",
+    "ChunkedReasoner",
+    "WholeDocReaderCapability",
+]
@@ -0,0 +1,616 @@
+"""Chunked reasoning over long documents.
+
+A reusable primitive for any agent that needs to answer a question that
+requires reading a whole document end-to-end. The document is split into
+character-budgeted chunks; each chunk is read by a parallel worker that
+extracts question-relevant notes; if the gathered notes overflow the
+synthesis context budget, the resulting notes are regrouped into fresh
+chunks and run through the same extractor again, until they fit.
+
+Pages are tracked by the wrapper, never asked of the model: keeps the model
+output schema small and the page list authoritative.
+
+Used wherever pure RAG retrieval is the wrong tool: aggregations ("largest
+number"), comparisons ("shortest chapter"), and full summaries.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import time
+from dataclasses import dataclass
+
+from pydantic import BaseModel, Field
+from pydantic_ai import Agent
+from pydantic_ai.output import NativeOutput
+
+from stirling.contracts import (
+    WholeDocCompressionRound,
+    WholeDocReadDone,
+    WholeDocReadStarted,
+    WholeDocSliceDone,
+)
+from stirling.contracts.documents import Page
+from stirling.models import ApiModel
+from stirling.services import AppRuntime, emit_progress
+
+logger = logging.getLogger(__name__)
+
+
+class ChunkNotes(ApiModel):
+    """Public-facing notes for a span of pages.
+
+    Returned to callers of :meth:`ChunkedReasoner.gather_notes` and to the
+    inside of :meth:`ChunkedReasoner.reason`. The wrapper builds these from
+    the model's :class:`_ExtractedNotes` output and a deterministic page list.
+    """
+
+    pages: list[int] = Field(description="Page numbers covered by these notes (1-indexed).")
+    summary: str = Field(description="One- to three-sentence summary of the covered range.")
+    relevant_excerpts: list[str] = Field(
+        default_factory=list,
+        description="Short verbatim quotes from the source content that bear on the user's question.",
+    )
+    facts: list[str] = Field(
+        default_factory=list,
+        description=(
+            "Concrete facts (numbers, names, dates, claims) the synthesiser may need. "
+            "Includes candidate values for aggregation questions."
+        ),
+    )
+
+
+class _ExtractedNotes(BaseModel):
+    """Model output for one extractor call.
+
+    No ``pages`` field: page numbers are mechanical aggregation the wrapper
+    computes deterministically. Keeping them out of the schema saves output
+    tokens for the bulkier excerpts/facts payload and prevents the model
+    from misreporting page coverage.
+    """
+
+    summary: str = Field(description="One- to three-sentence summary of the supplied content.")
+    relevant_excerpts: list[str] = Field(
+        default_factory=list,
+        description=(
+            "Short verbatim quotes drawn from the supplied content that bear on the question. "
+            "Deduplicate; drop ones that don't bear on the question."
+        ),
+    )
+    facts: list[str] = Field(
+        default_factory=list,
+        description=(
+            "Distinct, deduplicated facts (numbers, names, dates, claims) needed to answer "
+            "the question. For aggregation questions retain ALL candidate values across the "
+            "supplied content so a later round can still pick the global winner."
+        ),
+    )
+
+
+@dataclass(frozen=True)
+class _Chunk:
+    """A unit of work for the extractor: content + the pages it covers + a fallback.
+
+    ``content`` is the formatted text fed to the model: raw page text with
+    ``[Page N]`` markers in the first round, formatted prior-pass notes with
+    ``[Notes from pages A-B]`` markers in subsequent rounds. ``pages`` is
+    attached to the resulting :class:`ChunkNotes` deterministically.
+
+    ``fallback`` is the list of notes to keep if the extractor call fails. For
+    raw page chunks it's empty (a failed slice has no pre-extracted notes to
+    preserve). For chunks built from existing notes it's the input notes
+    themselves, so a failure doesn't lose page coverage.
+    """
+
+    content: str
+    pages: list[int]
+    fallback: list[ChunkNotes]
+    label: str
+
+
+@dataclass(frozen=True)
+class _RoundResult:
+    """Outcome of one extraction round.
+
+    ``successes`` lets the loop detect rounds that made no forward progress
+    (every chunk failed) and bail rather than spinning. ``slowest`` is the
+    chunk with the longest successful extractor call this round, used for
+    diagnostic log lines on the first round.
+    """
+
+    notes: list[ChunkNotes]
+    successes: int
+    slowest: tuple[str, float] | None
+
+
+def _page_range_label(pages: list[Page]) -> str:
+    if not pages:
+        return "pages=?"
+    elif len(pages) == 1:
+        return f"pages={pages[0].page_number}"
+    else:
+        return f"pages={pages[0].page_number}-{pages[-1].page_number}"
+
+
+def _note_range_label(notes: list[ChunkNotes]) -> str:
+    """Render a "pages=A-B" label for a group of already-extracted notes."""
+    page_numbers = sorted({p for note in notes for p in note.pages})
+    if not page_numbers:
+        return "pages=?"
+    if len(page_numbers) == 1:
+        return f"pages={page_numbers[0]}"
+    return f"pages={page_numbers[0]}-{page_numbers[-1]}"
+
+
+_EXTRACTOR_SYSTEM_PROMPT = (
+    "You are reading content from a document - either raw page text or "
+    "condensed notes from an earlier extraction pass - and your job is to "
+    "produce a tight set of notes that captures everything relevant to the "
+    "user's question. The same job runs many times in parallel across the "
+    "document and may run again to consolidate notes into smaller batches, "
+    "so be thorough: anything you skip cannot be recovered later.\n"
+    "\n"
+    "Output:\n"
+    "- summary: 1-3 sentences covering the supplied content.\n"
+    "- relevant_excerpts: short verbatim quotes from the supplied content "
+    "that bear on the question. Deduplicate; drop quotes that don't help.\n"
+    "- facts: concrete facts (numbers, names, dates, claims). Deduplicate; "
+    "drop irrelevant ones. For aggregation questions (largest, smallest, "
+    "count, total) retain ALL candidate values across the content so a "
+    "later step can still pick the global winner.\n"
+    "\n"
+    "Stay grounded in the supplied content. Do not infer or fabricate "
+    "anything that isn't already present. If nothing in the content is "
+    "relevant to the question, return empty excerpts and facts and a short "
+    "neutral summary."
+)
+
+
+class ChunkedReasoner:
+    """Run a question against a long document by chunking, mapping, and looping.
+
+    Two consumption styles:
+
+    * Tools that already have a synthesising LLM call upstream call
+      :meth:`gather_notes` to get the structured notes and format them
+      themselves with :meth:`format_notes`.
+    * Callers that just want an answer call :meth:`reason`, which runs
+      :meth:`gather_notes` and then a single synthesis call governed by the
+      caller's ``answer_prompt`` and ``answer_type``.
+
+    Lifetime:
+        Construct once per agent that uses it. The extractor agent is built
+        at construction time and reused; the synthesis agent in :meth:`reason`
+        is built per call because its output type is generic.
+    """
+
+    def __init__(
+        self,
+        runtime: AppRuntime,
+        *,
+        chars_per_slice: int | None = None,
+        concurrency: int | None = None,
+        worker_timeout_seconds: float | None = None,
+        notes_char_budget: int | None = None,
+    ) -> None:
+        chars = chars_per_slice if chars_per_slice is not None else runtime.settings.chunked_reasoner_chars_per_slice
+        conc = concurrency if concurrency is not None else runtime.settings.chunked_reasoner_concurrency
+        timeout = (
+            worker_timeout_seconds
+            if worker_timeout_seconds is not None
+            else runtime.settings.chunked_reasoner_worker_timeout_seconds
+        )
+        budget = (
+            notes_char_budget if notes_char_budget is not None else runtime.settings.chunked_reasoner_notes_char_budget
+        )
+        if chars <= 0:
+            raise ValueError("chars_per_slice must be positive")
+        if conc <= 0:
+            raise ValueError("concurrency must be positive")
+        if timeout <= 0:
+            raise ValueError("worker_timeout_seconds must be positive")
+        if budget <= 0:
+            raise ValueError("notes_char_budget must be positive")
+        self._runtime = runtime
+        self._chars_per_slice = chars
+        self._worker_timeout_seconds = timeout
+        self._notes_char_budget = budget
+        self._semaphore = asyncio.Semaphore(conc)
+        self._extractor: Agent[None, _ExtractedNotes] = Agent(
+            model=runtime.fast_model,
+            output_type=NativeOutput(_ExtractedNotes),
+            system_prompt=_EXTRACTOR_SYSTEM_PROMPT,
+            model_settings=runtime.fast_model_settings,
+        )
+
+    async def gather_notes(self, pages: list[Page], question: str) -> list[ChunkNotes]:
+        """Return notes covering every page that fit the synthesis budget.
+
+        Worker failures are tolerated: surviving notes are returned. Returns
+        an empty list only when every first-round chunk raises, which the
+        caller can treat as a hard failure.
+
+        Progress events fire as each first-round chunk finishes (in completion
+        order, not chunk order) carrying a monotonic ``completed`` counter so
+        consumers can render "Read X of Y" with X advancing by exactly one
+        per event. Subsequent compression rounds emit a single round-start
+        event each.
+        """
+        if not pages:
+            raise ValueError("ChunkedReasoner.gather_notes requires at least one page")
+
+        chunks = [self._chunk_from_pages(slice_pages) for slice_pages in self._slice_pages(pages)]
+        slice_total = len(chunks)
+        logger.info(
+            "[chunked-reasoner] question=%r pages=%d slices=%d",
+            question,
+            len(pages),
+            slice_total,
+        )
+        await emit_progress(WholeDocReadStarted(question=question, pages=len(pages), slices=slice_total))
+
+        gather_start = time.perf_counter()
+        notes = await self._run_chunks(chunks, question)
+
+        await emit_progress(
+            WholeDocReadDone(
+                completed=len(notes),
+                slices=slice_total,
+                duration_seconds=round(time.perf_counter() - gather_start, 2),
+            )
+        )
+        return notes
+
+    async def _run_chunks(self, chunks: list[_Chunk], question: str) -> list[ChunkNotes]:
+        """Run chunks through the extractor, regrouping and looping until under budget.
+
+        The first round emits per-chunk progress events for streaming UIs;
+        later rounds emit a single round-start event. Each round may produce
+        fewer notes than chunks (every chunk maps to at most one consolidated
+        note); when the rendered notes still exceed the budget, the survivors
+        are regrouped into fresh chunks and the loop runs again.
+        """
+        round_number = 0
+        while True:
+            chunks_in = len(chunks)
+            result = await self._extract_chunks(chunks, question, round_number)
+
+            if result.slowest is not None:
+                slow_label, slow_duration = result.slowest
+                logger.info(
+                    "[chunked-reasoner] round %d: %d/%d chunks succeeded; slowest %s (%.1fs)",
+                    round_number,
+                    result.successes,
+                    chunks_in,
+                    slow_label,
+                    slow_duration,
+                )
+            else:
+                logger.info(
+                    "[chunked-reasoner] round %d: 0/%d chunks succeeded",
+                    round_number,
+                    chunks_in,
+                )
+
+            rendered_size = self._rendered_notes_size(result.notes)
+            if rendered_size <= self._notes_char_budget or len(result.notes) <= 1:
+                if round_number > 0:
+                    logger.info(
+                        "[chunked-reasoner] compression done after %d round(s): %d notes, %d chars",
+                        round_number,
+                        len(result.notes),
+                        rendered_size,
+                    )
+                return result.notes
+
+            if result.successes == 0:
+                # No forward progress this round; further rounds would
+                # reproduce the same shape. Return what we have.
+                logger.warning(
+                    "[chunked-reasoner] round %d produced no successful extractions; bailing with %d notes",
+                    round_number,
+                    len(result.notes),
+                )
+                return result.notes
+
+            round_number += 1
+            groups = self._group_notes_for_compression(result.notes)
+            chunks = [self._chunk_from_notes(group) for group in groups]
+            logger.info(
+                "[chunked-reasoner] compression round %d: %d notes (%d chars) -> %d groups",
+                round_number,
+                len(result.notes),
+                rendered_size,
+                len(groups),
+            )
+            await emit_progress(
+                WholeDocCompressionRound(
+                    round_number=round_number,
+                    notes_in=len(result.notes),
+                    groups=len(groups),
+                )
+            )
+
+    async def _extract_chunks(
+        self,
+        chunks: list[_Chunk],
+        question: str,
+        round_number: int,
+    ) -> _RoundResult:
+        """Run all chunks through the extractor in parallel; collect surviving notes.
+
+        Failures fall back to ``chunk.fallback`` (empty in the first round, so
+        failures drop; populated in compression rounds, so failures preserve
+        their input notes). The first round emits a
+        :class:`WholeDocSliceDone` per successful completion in completion
+        order, with a monotonic ``completed`` counter.
+
+        Returned notes are sorted by first page so downstream grouping packs
+        document-adjacent content together regardless of which task happened
+        to finish first.
+        """
+        total = len(chunks)
+        pending: dict[asyncio.Task[tuple[ChunkNotes, float]], _Chunk] = {
+            asyncio.create_task(self._extract_chunk(chunk, question)): chunk for chunk in chunks
+        }
+
+        notes: list[ChunkNotes] = []
+        successes = 0
+        slowest: tuple[str, float] | None = None
+        completed = 0
+
+        try:
+            while pending:
+                done, _ = await asyncio.wait(pending.keys(), return_when=asyncio.FIRST_COMPLETED)
+                for task in done:
+                    chunk = pending.pop(task)
+                    exc = task.exception()
+                    if exc is not None:
+                        if chunk.fallback:
+                            logger.warning(
+                                "[chunked-reasoner] chunk %s failed: %s; preserving %d input note(s)",
+                                chunk.label,
+                                exc,
+                                len(chunk.fallback),
+                            )
+                            notes.extend(chunk.fallback)
+                        else:
+                            logger.warning("[chunked-reasoner] chunk %s failed: %s", chunk.label, exc)
+                        continue
+                    extracted, duration = task.result()
+                    notes.append(extracted)
+                    successes += 1
+                    completed += 1
+                    if slowest is None or duration > slowest[1]:
+                        slowest = (chunk.label, duration)
+                    if round_number == 0:
+                        await emit_progress(
+                            WholeDocSliceDone(
+                                completed=completed,
+                                total=total,
+                                pages=chunk.label,
+                                duration_ms=int(duration * 1000),
+                                excerpts=len(extracted.relevant_excerpts),
+                                facts=len(extracted.facts),
+                            )
+                        )
+        finally:
+            # On cancellation (typically a frontend disconnect propagating up
+            # through the streaming orchestrator) the per-chunk model calls
+            # would otherwise keep running to completion, billing tokens whose
+            # results nobody is reading. Cancel and drain so the upstream
+            # cancellation is the cancellation that matters.
+            if pending:
+                for task in pending:
+                    task.cancel()
+                await asyncio.gather(*pending.keys(), return_exceptions=True)
+
+        notes.sort(key=lambda n: n.pages[0] if n.pages else 0)
+        return _RoundResult(notes=notes, successes=successes, slowest=slowest)
+
+    async def _extract_chunk(self, chunk: _Chunk, question: str) -> tuple[ChunkNotes, float]:
+        """Run the extractor on one chunk and attach the chunk's pages to the output."""
+        try:
+            extracted, duration = await self._run_extractor(chunk.content, question, chunk.label)
+        except TimeoutError:
+            logger.warning(
+                "[chunked-reasoner] chunk %s timed out (limit %.1fs)",
+                chunk.label,
+                self._worker_timeout_seconds,
+            )
+            raise
+        logger.debug(
+            "[chunked-reasoner] chunk %s: %d excerpt(s), %d fact(s) in %dms",
+            chunk.label,
+            len(extracted.relevant_excerpts),
+            len(extracted.facts),
+            int(duration * 1000),
+        )
+        return self._build_chunk_notes(extracted, chunk.pages), duration
+
+    async def _run_extractor(
+        self,
+        content: str,
+        question: str,
+        page_label: str,
+    ) -> tuple[_ExtractedNotes, float]:
+        """Inner primitive: run the extractor agent under semaphore + timeout."""
+        prompt = self._build_extraction_prompt(content, question)
+        async with self._semaphore:
+            start = time.perf_counter()
+            try:
+                result = await asyncio.wait_for(self._extractor.run(prompt), timeout=self._worker_timeout_seconds)
+            except TimeoutError:
+                duration = time.perf_counter() - start
+                logger.debug(
+                    "[chunked-reasoner] extractor %s timed out after %dms",
+                    page_label,
+                    int(duration * 1000),
+                )
+                raise
+            duration = time.perf_counter() - start
+        return result.output, duration
+
+    def _chunk_from_pages(self, pages: list[Page]) -> _Chunk:
+        """Build a first-round chunk from a slice of raw pages."""
+        return _Chunk(
+            content="\n\n".join(f"[Page {p.page_number}]\n{p.text}" for p in pages),
+            pages=[p.page_number for p in pages],
+            fallback=[],
+            label=_page_range_label(pages),
+        )
+
+    def _chunk_from_notes(self, group: list[ChunkNotes]) -> _Chunk:
+        """Build a compression-round chunk from a group of prior-pass notes.
+
+        ``fallback`` is the input group itself: if the extractor call fails,
+        the originals stay in the working set so page coverage isn't lost.
+        """
+        return _Chunk(
+            content=self.format_notes(group),
+            pages=sorted({p for note in group for p in note.pages}),
+            fallback=group,
+            label=_note_range_label(group),
+        )
+
+    def _group_notes_for_compression(self, notes: list[ChunkNotes]) -> list[list[ChunkNotes]]:
+        """Pack consecutive notes into groups whose rendered size fits ``chars_per_slice``.
+
+        Each group becomes one compression-round chunk. Sized to match the
+        first-round slice budget so the extractor sees roughly the same input
+        footprint regardless of which round is running. Single notes that
+        exceed the budget on their own become their own group.
+        """
+        groups: list[list[ChunkNotes]] = []
+        current: list[ChunkNotes] = []
+        current_chars = 0
+        for note in notes:
+            note_chars = self._rendered_notes_size([note])
+            if current and current_chars + note_chars > self._chars_per_slice:
+                groups.append(current)
+                current = []
+                current_chars = 0
+            current.append(note)
+            current_chars += note_chars
+        if current:
+            groups.append(current)
+        return groups
+
+    @staticmethod
+    def _build_chunk_notes(extracted: _ExtractedNotes, pages: list[int]) -> ChunkNotes:
+        """Build a public ChunkNotes from the model's output and the wrapper's pages."""
+        return ChunkNotes(
+            pages=pages,
+            summary=extracted.summary,
+            relevant_excerpts=extracted.relevant_excerpts,
+            facts=extracted.facts,
+        )
+
+    @staticmethod
+    def _build_extraction_prompt(content: str, question: str) -> str:
+        """Single prompt shape used for every round.
+
+        The system prompt explains the role; the user prompt just hands over
+        the question and the content. Whether ``content`` is raw page text
+        with ``[Page N]`` markers or formatted notes with ``[Notes from
+        pages A-B]`` markers, the same instructions apply.
+        """
+        return f"User question:\n{question}\n\nContent:\n{content}"
+
+    @staticmethod
+    def _rendered_notes_size(notes: list[ChunkNotes]) -> int:
+        """Length in characters of what :meth:`format_notes` would produce."""
+        return len(ChunkedReasoner.format_notes(notes))
+
+    async def reason[T: BaseModel](
+        self,
+        *,
+        pages: list[Page],
+        question: str,
+        answer_prompt: str,
+        answer_type: type[T],
+    ) -> T:
+        """Map over pages, then synthesise a structured answer of type ``T``.
+
+        Args:
+            pages: Document pages in order.
+            question: The user's question, passed to both workers and the
+                synthesiser. Workers use it to decide what's relevant.
+            answer_prompt: System prompt for the synthesis stage. Should
+                instruct the model to answer ``question`` from the notes
+                supplied. Owned by the caller because the answer's tone,
+                format, and grounding rules are domain-specific.
+            answer_type: Pydantic model describing the structured answer.
+
+        Returns:
+            An instance of ``answer_type`` produced by the synthesis stage.
+        """
+        notes = await self.gather_notes(pages, question)
+        if not notes:
+            raise RuntimeError("All chunked-reasoning workers failed; no notes to synthesise from")
+        return await self._synthesise(question, notes, answer_prompt, answer_type)
+
+    @staticmethod
+    def format_notes(notes: list[ChunkNotes]) -> str:
+        """Render notes as readable text for inclusion in another agent's tool result.
+
+        Order is preserved. Page numbers, summary, excerpts and facts are all
+        emitted; empty sections are omitted.
+        """
+        sections: list[str] = []
+        for n in notes:
+            page_label = (
+                f"pages {n.pages[0]}-{n.pages[-1]}"
+                if len(n.pages) > 1
+                else f"page {n.pages[0]}"
+                if n.pages
+                else "unknown pages"
+            )
+            block = [f"[Notes from {page_label}]", f"Summary: {n.summary}"]
+            if n.relevant_excerpts:
+                block.append("Relevant excerpts:")
+                block.extend(f"- {e}" for e in n.relevant_excerpts)
+            if n.facts:
+                block.append("Facts:")
+                block.extend(f"- {f}" for f in n.facts)
+            sections.append("\n".join(block))
+        return "\n\n".join(sections)
+
+    def _slice_pages(self, pages: list[Page]) -> list[list[Page]]:
+        """Group consecutive pages into character-budgeted slices.
+
+        Page boundaries are preserved: a single page is never split across
+        slices. If one page exceeds the budget on its own, it becomes its
+        own slice.
+        """
+        slices: list[list[Page]] = []
+        current: list[Page] = []
+        current_chars = 0
+        for page in pages:
+            if current and current_chars + page.char_count > self._chars_per_slice:
+                slices.append(current)
+                current = []
+                current_chars = 0
+            current.append(page)
+            current_chars += page.char_count
+        if current:
+            slices.append(current)
+        return slices
+
+    async def _synthesise[T: BaseModel](
+        self,
+        question: str,
+        notes: list[ChunkNotes],
+        answer_prompt: str,
+        answer_type: type[T],
+    ) -> T:
+        agent: Agent[None, T] = Agent(
+            model=self._runtime.smart_model,
+            output_type=NativeOutput(answer_type),
+            system_prompt=answer_prompt,
+            model_settings=self._runtime.smart_model_settings,
+        )
+        prompt = f"User question:\n{question}\n\nNotes from across the document:\n\n{self.format_notes(notes)}"
+        result = await agent.run(prompt)
+        return result.output
@@ -0,0 +1,140 @@
+"""Tool capability that lets an agent read whole documents end-to-end.
+
+Companion to :class:`stirling.documents.RagCapability`. Where ``RagCapability``
+gives an agent targeted vector retrieval, this gives it map-style whole-document
+reading: every page is read in parallel by fast-model workers, and the
+question-relevant notes are returned for the agent to synthesise.
+
+Use both capabilities together when the agent should pick its strategy:
+``search_knowledge`` for specific lookups, ``read_full_document`` for
+aggregations, comparisons, and summaries.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from pydantic_ai import FunctionToolset, RunContext, ToolDefinition
+from pydantic_ai.toolsets import AbstractToolset
+
+from stirling.agents.shared.chunked_reasoner import ChunkedReasoner
+from stirling.contracts import AiFile
+from stirling.services import AppRuntime
+
+logger = logging.getLogger(__name__)
+
+
+# Cap on per-run calls. One pass already reads every page of every attached
+# document, so a second call is almost always the smart model second-guessing
+# itself on a near-identical query (and doubles wall-clock time for a sizeable
+# document). If a follow-up genuinely needs more, ``search_knowledge`` is the
+# right escape hatch. Configurable per-construction in case a future caller
+# can prove a real two-read use case; the default stays at 1.
+DEFAULT_MAX_READS = 1
+
+
+class WholeDocReaderCapability:
+    """Bundles instructions and the ``read_full_document`` toolset for agent injection.
+
+    Lifecycle: a ``WholeDocReaderCapability`` instance is intended to live for
+    the duration of a single agent run.
+
+    The agent picks between this and :class:`RagCapability` per the tool
+    descriptions: targeted retrieval vs whole-document reading.
+    """
+
+    def __init__(
+        self,
+        runtime: AppRuntime,
+        files: list[AiFile],
+        *,
+        reasoner: ChunkedReasoner | None = None,
+        max_reads: int = DEFAULT_MAX_READS,
+    ) -> None:
+        self._runtime = runtime
+        self._files = files
+        self._reasoner = reasoner if reasoner is not None else ChunkedReasoner(runtime)
+        self._max_reads = max_reads
+        self._read_count = 0
+        toolset: FunctionToolset[None] = FunctionToolset()
+        toolset.add_function(
+            self._read_full_document,
+            name="read_full_document",
+            prepare=self._prepare_read_full_document,
+        )
+        self._toolset = toolset
+
+    @property
+    def instructions(self) -> str:
+        names = ", ".join(f.name for f in self._files) if self._files else "the attached documents"
+        return (
+            "You have a 'read_full_document' tool that reads every page of "
+            f"{names} in parallel and returns notes relevant to a query. "
+            "Use it when answering requires seeing the whole document end-to-end "
+            "(summaries, aggregations, comparisons across sections). One call "
+            "already reads everything; phrase the query to cover all the angles "
+            "you need in a single pass. For follow-ups or specific lookups use "
+            "'search_knowledge', which is cheaper and targeted."
+        )
+
+    @property
+    def toolset(self) -> AbstractToolset[None]:
+        return self._toolset
+
+    async def _prepare_read_full_document(
+        self,
+        ctx: RunContext[None],
+        tool_def: ToolDefinition,
+    ) -> ToolDefinition | None:
+        """Hide the tool from the agent's toolset once the per-run budget is spent.
+        Mirrors the search_knowledge prepare callback."""
+        if self._read_count >= self._max_reads:
+            return None
+        return tool_def
+
+    async def _read_full_document(self, query: str) -> str:
+        """Read every page of the attached documents and return notes relevant to the query.
+
+        Use this when answering needs the whole document end-to-end - summaries,
+        aggregations like 'largest number' or 'shortest chapter', or comparisons
+        across sections. Slow and expensive (one fast-model call per slice per
+        document); prefer search_knowledge for targeted lookups.
+
+        Args:
+            query: A focused description of what to extract from the documents,
+                phrased so a worker reading just one slice can decide what's
+                relevant to the user's question.
+
+        Returns:
+            Per-document sections of structured notes (page numbers, summary,
+            relevant excerpts, extracted facts), already ordered by page.
+        """
+        self._read_count += 1
+        if not self._files:
+            return "No documents attached to read."
+
+        sections: list[str] = []
+        for file in self._files:
+            pages = await self._runtime.documents.read_pages(file.id)
+            if not pages:
+                logger.info(
+                    "[whole-doc-reader] no stored pages for %s (id=%s); skipping",
+                    file.name,
+                    file.id,
+                )
+                continue
+            notes = await self._reasoner.gather_notes(pages, query)
+            if not notes:
+                continue
+            sections.append(f"=== {file.name} ===\n{ChunkedReasoner.format_notes(notes)}")
+
+        if not sections:
+            return "Could not read any document content."
+
+        logger.info(
+            "[whole-doc-reader] read query=%r files=%d -> %d chars",
+            query,
+            len(self._files),
+            sum(len(s) for s in sections),
+        )
+        return "\n\n".join(sections)
@@ -19,13 +19,13 @@ from stirling.agents.pdf_comment import PdfCommentAgent
 from stirling.api.middleware import UserIdMiddleware
 from stirling.api.routes import (
    agent_draft_router,
+    document_router,
    execution_router,
    ledger_router,
    orchestrator_router,
    pdf_comments_router,
    pdf_edit_router,
    pdf_question_router,
-    rag_router,
 )
 from stirling.config import AppSettings, load_settings
 from stirling.contracts import HealthResponse
@@ -57,7 +57,7 @@ async def lifespan(fast_api: FastAPI):
    if tracer_provider:
        Agent.instrument_all(InstrumentationSettings(tracer_provider=tracer_provider))
    yield
-    await runtime.rag_service.close()
+    await runtime.documents.close()
    if tracer_provider:
        tracer_provider.shutdown()

@@ -69,7 +69,7 @@ app.include_router(pdf_edit_router)
 app.include_router(pdf_question_router)
 app.include_router(agent_draft_router)
 app.include_router(execution_router)
-app.include_router(rag_router)
+app.include_router(document_router)
 app.include_router(ledger_router)
 app.include_router(pdf_comments_router)

@@ -11,7 +11,7 @@ from stirling.agents import (
 )
 from stirling.agents.ledger import MathAuditorAgent
 from stirling.agents.pdf_comment import PdfCommentAgent
-from stirling.rag import RagService
+from stirling.documents import DocumentService
 from stirling.services import AppRuntime


@@ -39,8 +39,8 @@ def get_execution_planning_agent(request: Request) -> ExecutionPlanningAgent:
    return request.app.state.execution_planning_agent


-def get_rag_service(request: Request) -> RagService:
-    return request.app.state.runtime.rag_service
+def get_document_service(request: Request) -> DocumentService:
+    return request.app.state.runtime.documents


 def get_math_auditor_agent(request: Request) -> MathAuditorAgent:
@@ -1,19 +1,19 @@
 from .agent_drafts import router as agent_draft_router
+from .documents import router as document_router
 from .execution import router as execution_router
 from .ledger import router as ledger_router
 from .orchestrator import router as orchestrator_router
 from .pdf_comments import router as pdf_comments_router
 from .pdf_edit import router as pdf_edit_router
 from .pdf_questions import router as pdf_question_router
-from .rag import router as rag_router

 __all__ = [
    "agent_draft_router",
+    "document_router",
    "execution_router",
    "ledger_router",
    "orchestrator_router",
    "pdf_comments_router",
    "pdf_edit_router",
    "pdf_question_router",
-    "rag_router",
 ]
@@ -0,0 +1,49 @@
+from __future__ import annotations
+
+from typing import Annotated
+
+from fastapi import APIRouter, Depends
+
+from stirling.api.dependencies import get_document_service
+from stirling.contracts import (
+    DeleteDocumentResponse,
+    IngestDocumentRequest,
+    IngestDocumentResponse,
+)
+from stirling.documents import DocumentService
+from stirling.models import FileId
+
+router = APIRouter(prefix="/api/v1/documents", tags=["documents"])
+
+
+@router.post("", response_model=IngestDocumentResponse)
+async def ingest_document(
+    request: IngestDocumentRequest,
+    documents: Annotated[DocumentService, Depends(get_document_service)],
+) -> IngestDocumentResponse:
+    """Replace-ingest a document's content under ``document_id``.
+
+    Stores both representations in one shot:
+      * embedded chunks for RAG search,
+      * ordered page text for whole-document reading.
+    Any previously-stored content for this document is removed first.
+    """
+    pages = request.page_text or []
+    chunks_indexed = await documents.ingest(
+        collection=request.document_id,
+        pages=pages,
+        source=request.source,
+    )
+    return IngestDocumentResponse(document_id=request.document_id, chunks_indexed=chunks_indexed)
+
+
+@router.delete("/{document_id}", response_model=DeleteDocumentResponse)
+async def delete_document(
+    document_id: FileId,
+    documents: Annotated[DocumentService, Depends(get_document_service)],
+) -> DeleteDocumentResponse:
+    """Remove a document's content. Idempotent."""
+    existed = await documents.has_collection(document_id)
+    if existed:
+        await documents.delete_collection(document_id)
+    return DeleteDocumentResponse(document_id=document_id, deleted=existed)
@@ -1,19 +1,170 @@
 from __future__ import annotations

-from typing import Annotated
+import asyncio
+import json
+import logging
+from collections.abc import AsyncIterator
+from dataclasses import dataclass
+from typing import Annotated, assert_never

 from fastapi import APIRouter, Depends
+from fastapi.responses import StreamingResponse

 from stirling.agents import OrchestratorAgent
 from stirling.api.dependencies import get_orchestrator_agent
-from stirling.contracts import OrchestratorRequest, OrchestratorResponse
+from stirling.contracts import OrchestratorRequest, OrchestratorResponse, ProgressEvent
+from stirling.services import reset_progress_emitter, set_progress_emitter
+
+logger = logging.getLogger(__name__)
+
+# Cadence for keep-alive heartbeats on the streaming endpoint. Java forwards
+# them to the frontend as SSE comments; their job is to make every layer of
+# the connection visibly alive at this rhythm so disconnects surface within a
+# bounded window instead of waiting for the next progress event.
+HEARTBEAT_INTERVAL_SECONDS = 10.0

 router = APIRouter(prefix="/api/v1/orchestrator", tags=["orchestrator"])


-@router.post("", response_model=OrchestratorResponse)
+@router.post("")
 async def orchestrate(
    request: OrchestratorRequest,
    agent: Annotated[OrchestratorAgent, Depends(get_orchestrator_agent)],
-) -> OrchestratorResponse:
-    return await agent.handle(request)
+) -> StreamingResponse:
+    """Run the orchestrator and stream NDJSON events.
+
+    Each output line is a JSON object with an ``event`` field. ``progress``
+    events arrive whenever an inner agent reports work (e.g. each
+    chunked-reasoner slice completing); the final ``result`` event carries the
+    typed orchestrator response. ``error`` events surface failures without
+    breaking the connection. ``heartbeat`` events fire on a fixed cadence to
+    keep idle connections visibly alive so disconnects propagate.
+
+    The stream itself is the liveness signal: as long as events flow, work is
+    alive. Java consumes this with a long total timeout and treats line
+    arrival as forward progress.
+    """
+    return StreamingResponse(
+        _OrchestratorStream(
+            agent=agent,
+            request=request,
+            heartbeat_interval_seconds=HEARTBEAT_INTERVAL_SECONDS,
+        ).iterate(),
+        media_type="application/x-ndjson",
+    )
+
+
+@dataclass(frozen=True, slots=True)
+class _ProgressFrame:
+    event: ProgressEvent
+
+
+@dataclass(frozen=True, slots=True)
+class _ResultFrame:
+    response: OrchestratorResponse
+
+
+@dataclass(frozen=True, slots=True)
+class _ErrorFrame:
+    message: str
+
+
+@dataclass(frozen=True, slots=True)
+class _HeartbeatFrame:
+    """No payload: a heartbeat exists only to push bytes through the pipe.
+
+    Without periodic traffic, a slow workflow phase (e.g. all extractor
+    workers busy on long calls) leaves the engine writer, Java's SSE
+    forwarder, and the frontend's fetch all silently waiting. A closed
+    connection at any layer wouldn't surface until the next real event,
+    which could be many tens of seconds away. Heartbeats bound that window
+    to :data:`HEARTBEAT_INTERVAL_SECONDS`.
+    """
+
+
+type _StreamFrame = _ProgressFrame | _ResultFrame | _ErrorFrame | _HeartbeatFrame
+
+
+def _serialize_frame(frame: _StreamFrame) -> bytes:
+    """Render a frame as one NDJSON line."""
+    match frame:
+        case _ProgressFrame(event=event):
+            body = {"event": "progress", **event.model_dump(mode="json")}
+        case _ResultFrame(response=response):
+            body = {"event": "result", "response": response.model_dump(mode="json")}
+        case _ErrorFrame(message=message):
+            body = {"event": "error", "message": message}
+        case _HeartbeatFrame():
+            body = {"event": "heartbeat"}
+        case _:
+            assert_never(frame)
+    return (json.dumps(body) + "\n").encode("utf-8")
+
+
+class _OrchestratorStream:
+    """Drives one streaming orchestrator request.
+
+    Owns the per-request queue and pumps progress events through it; the agent
+    runs as a child task so its emissions and the streaming response interleave.
+    A heartbeat task pushes keep-alive messages onto the same queue at a fixed
+    cadence so the connection stays visibly alive between progress events.
+    """
+
+    def __init__(
+        self,
+        *,
+        agent: OrchestratorAgent,
+        request: OrchestratorRequest,
+        heartbeat_interval_seconds: float,
+    ) -> None:
+        self._agent = agent
+        self._request = request
+        self._heartbeat_interval_seconds = heartbeat_interval_seconds
+        self._queue: asyncio.Queue[_StreamFrame | None] = asyncio.Queue()
+
+    async def iterate(self) -> AsyncIterator[bytes]:
+        token = set_progress_emitter(self._emit_progress)
+        agent_task = asyncio.create_task(self._run_agent())
+        heartbeat_task = asyncio.create_task(self._emit_heartbeats())
+        try:
+            while True:
+                frame = await self._queue.get()
+                if frame is None:
+                    break
+                yield _serialize_frame(frame)
+        finally:
+            reset_progress_emitter(token)
+            await self._cancel_task(heartbeat_task)
+            await self._cancel_task(agent_task)
+
+    async def _emit_progress(self, event: ProgressEvent) -> None:
+        await self._queue.put(_ProgressFrame(event=event))
+
+    async def _emit_heartbeats(self) -> None:
+        while True:
+            await asyncio.sleep(self._heartbeat_interval_seconds)
+            await self._queue.put(_HeartbeatFrame())
+
+    async def _run_agent(self) -> None:
+        try:
+            response = await self._agent.handle(self._request)
+            await self._queue.put(_ResultFrame(response=response))
+        except asyncio.CancelledError:
+            raise
+        except Exception as exc:
+            logger.exception("orchestrator stream failed")
+            await self._queue.put(_ErrorFrame(message=str(exc)))
+        finally:
+            await self._queue.put(None)
+
+    @staticmethod
+    async def _cancel_task(task: asyncio.Task[None]) -> None:
+        if task.done():
+            return
+        task.cancel()
+        try:
+            await task
+        except asyncio.CancelledError:
+            pass
+        except Exception:
+            logger.exception("background task failed during cancellation", exc_info=True)
@@ -1,63 +0,0 @@
-from __future__ import annotations
-
-from typing import Annotated
-
-from fastapi import APIRouter, Depends
-
-from stirling.api.dependencies import get_rag_service
-from stirling.contracts import (
-    DeleteDocumentResponse,
-    IngestDocumentRequest,
-    IngestDocumentResponse,
-    PdfContentType,
-)
-from stirling.models import FileId
-from stirling.rag import Document, RagService
-
-router = APIRouter(prefix="/api/v1/rag", tags=["rag"])
-
-
-@router.post("/documents", response_model=IngestDocumentResponse)
-async def ingest_document(
-    request: IngestDocumentRequest,
-    rag: Annotated[RagService, Depends(get_rag_service)],
-) -> IngestDocumentResponse:
-    """Replace-ingest a document's content under ``document_id``.
-
-    Any previously-stored content for this document is removed and the
-    provided content replaces it wholesale. All pages are chunked up front
-    and then embedded in a single batched call so large documents (e.g. a
-    500-page book) don't fan out into hundreds of embedding requests.
-    """
-    await rag.delete_collection(request.document_id)
-
-    chunks: list[Document] = []
-    if request.page_text:
-        for page in request.page_text:
-            if not page.text.strip():
-                continue
-            chunks.extend(
-                rag.chunk_text(
-                    text=page.text,
-                    source=f"{request.source}:page:{page.page_number}",
-                    base_metadata={
-                        "page_number": str(page.page_number),
-                        "content_type": PdfContentType.PAGE_TEXT.value,
-                    },
-                )
-            )
-
-    indexed = await rag.index_documents(request.document_id, chunks) if chunks else 0
-    return IngestDocumentResponse(document_id=request.document_id, chunks_indexed=indexed)
-
-
-@router.delete("/documents/{document_id}", response_model=DeleteDocumentResponse)
-async def delete_document(
-    document_id: FileId,
-    rag: Annotated[RagService, Depends(get_rag_service)],
-) -> DeleteDocumentResponse:
-    """Remove a document's content from RAG. Idempotent."""
-    existed = await rag.has_collection(document_id)
-    if existed:
-        await rag.delete_collection(document_id)
-    return DeleteDocumentResponse(document_id=document_id, deleted=existed)
@@ -38,18 +38,38 @@ class AppSettings(BaseSettings):
    rag_default_top_k: int = Field(validation_alias="STIRLING_RAG_TOP_K")
    rag_max_searches: int = Field(validation_alias="STIRLING_RAG_MAX_SEARCHES")

+    # Chunked reasoner settings (whole-document map-reduce).
+    chunked_reasoner_chars_per_slice: int = Field(validation_alias="STIRLING_CHUNKED_REASONER_CHARS_PER_SLICE")
+    chunked_reasoner_concurrency: int = Field(validation_alias="STIRLING_CHUNKED_REASONER_CONCURRENCY")
+    chunked_reasoner_worker_timeout_seconds: float = Field(
+        validation_alias="STIRLING_CHUNKED_REASONER_WORKER_TIMEOUT_SECONDS"
+    )
+    # Maximum size, in characters, of the rendered notes block before the
+    # reasoner folds slice notes hierarchically. The Anthropic context limit
+    # is 200k tokens (~880k chars); we leave a generous margin for the
+    # downstream agent's system prompt, history, tool definitions, and
+    # response budget.
+    chunked_reasoner_notes_char_budget: int = Field(validation_alias="STIRLING_CHUNKED_REASONER_NOTES_CHAR_BUDGET")
+
    max_pages: int = Field(validation_alias="STIRLING_MAX_PAGES")
    max_characters: int = Field(validation_alias="STIRLING_MAX_CHARACTERS")

    log_level: str = Field(default="INFO", validation_alias="STIRLING_LOG_LEVEL")
    log_file: str = Field(default="", validation_alias="STIRLING_LOG_FILE")
+    # When true, raises httpx + httpcore logger levels so every outgoing
+    # model SDK call is logged with timing. Use to diagnose worker stalls:
+    # a hung request shows the "Request: POST ..." line with no matching
+    # response line, confirming the hang is transport-layer (not in our
+    # code or the Anthropic SDK itself). Off by default — DEBUG-level
+    # output is high-volume.
+    http_debug: bool = Field(default=False, validation_alias="STIRLING_HTTP_DEBUG")

    posthog_enabled: bool = Field(validation_alias="STIRLING_POSTHOG_ENABLED")
    posthog_api_key: str = Field(validation_alias="STIRLING_POSTHOG_API_KEY")
    posthog_host: str = Field(validation_alias="STIRLING_POSTHOG_HOST")


-def _configure_logging(level_name: str, log_file: str) -> None:
+def _configure_logging(level_name: str, log_file: str, http_debug: bool) -> None:
    """Configure the ``stirling`` logger hierarchy."""
    level = logging.getLevelNamesMapping().get(level_name.upper())
    if level is None:
@@ -83,11 +103,43 @@ def _configure_logging(level_name: str, log_file: str) -> None:
        fh.setLevel(level)
        root.addHandler(fh)

+    if http_debug:
+        _enable_http_debug(formatter)
+
+
+def _enable_http_debug(formatter: logging.Formatter) -> None:
+    """Surface every httpx/httpcore call against the Anthropic API.
+
+    httpx emits one INFO line per request with the URL and final status,
+    which is the most useful signal for diagnosing hung worker calls: a
+    successful call shows "Request" then "Response" within a second or two;
+    a hung one shows "Request" with no matching response until it's
+    cancelled. httpcore at DEBUG drills down to TCP / HTTP/2 stream events
+    if the user wants to see exactly where bytes stop flowing.
+
+    The ``stirling`` console handler is scoped to its own logger tree, so
+    we attach a dedicated stream handler here. Without it, httpx records
+    propagate to the root logger which has no handler in our setup and the
+    output is silently dropped.
+    """
+    handler = logging.StreamHandler()
+    handler.setFormatter(formatter)
+    handler.setLevel(logging.DEBUG)
+
+    for name, level in (("httpx", logging.INFO), ("httpcore", logging.DEBUG)):
+        lg = logging.getLogger(name)
+        lg.setLevel(level)
+        # Idempotent: avoid stacking handlers on settings reload.
+        if not any(getattr(h, "_stirling_http_debug", False) for h in lg.handlers):
+            handler._stirling_http_debug = True  # pyrefly: ignore [missing-attribute]
+            lg.addHandler(handler)
+        lg.propagate = False
+

@lru_cache(maxsize=1)
 def load_settings() -> AppSettings:
    load_dotenv(ENV_FILE)
    load_dotenv(ENV_LOCAL_FILE, override=True)
    settings = AppSettings.model_validate({})
-    _configure_logging(settings.log_level, settings.log_file)
+    _configure_logging(settings.log_level, settings.log_file, settings.http_debug)
    return settings
--- a/Show More
+++ b/Show More