kam6l · kam6l · Aug 21, 2026 · Aug 21, 2026 · Aug 21, 2026 · Aug 21, 2026
@@ -44,10 +44,11 @@ jobs:
             python-version: "3.13"
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
-      - uses: astral-sh/setup-uv@v5
+      - uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
         with:
+          version: "0.12.5"
           enable-cache: true
           python-version: ${{ matrix.python-version }}
           cache-dependency-glob: uv.lock
@@ -63,7 +64,7 @@ jobs:
 
       - name: Upload reference coverage
         if: matrix.os == 'ubuntu-latest' && matrix.python-version == '3.13'
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
           name: coverage-python-3.13
           path: coverage.xml
@@ -76,10 +77,11 @@ jobs:
     timeout-minutes: 15
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
-      - uses: astral-sh/setup-uv@v5
+      - uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
         with:
+          version: "0.12.5"
           enable-cache: true
           python-version: "3.12"
           cache-dependency-glob: uv.lock
@@ -88,10 +90,10 @@ jobs:
         run: uv sync --locked --group dev --extra langchain
 
       - name: Check formatting
-        run: uv run ruff format --check src tests examples benchmarks scripts
+        run: uv run ruff format --check src tests examples benchmarks demos
 
       - name: Lint
-        run: uv run ruff check src tests examples benchmarks scripts
+        run: uv run ruff check src tests examples benchmarks demos
 
       - name: Type-check
         run: uv run mypy src/agentdiff
@@ -105,10 +107,11 @@ jobs:
     timeout-minutes: 15
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
-      - uses: astral-sh/setup-uv@v5
+      - uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
         with:
+          version: "0.12.5"
           enable-cache: true
           python-version: "3.12"
           cache-dependency-glob: uv.lock
@@ -138,79 +141,77 @@ jobs:
       contents: read
 
     steps:
-      - uses: actions/checkout@v4
-      - uses: actions/dependency-review-action@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: actions/dependency-review-action@a1d282b36b6f3519aa1f3fc636f609c47dddb294 # v5.0.0
 
-  docs:
-    name: Documentation
+  safetybench:
+    name: Local recovery regression suite
     runs-on: ubuntu-latest
     timeout-minutes: 10
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
-      - uses: astral-sh/setup-uv@v5
+      - uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
         with:
+          version: "0.12.5"
           enable-cache: true
           python-version: "3.12"
           cache-dependency-glob: uv.lock
 
-      - name: Install documentation dependencies
-        run: uv sync --locked --group docs
-
-      - name: Build strict documentation
-        env:
-          PYTHONWARNINGS: "ignore::UserWarning:mkdocs_autorefs"
-        run: uv run mkdocs build --strict
+      - name: Install benchmark dependencies
+        run: uv sync --locked --group dev
 
-      - name: Check built-site links and assets
-        run: uv run python scripts/check_docs_links.py site
+      - name: Run adversarial local regression cases
+        run: uv run python3 benchmarks/safetybench.py --output recovery-regression.json
 
-      - uses: actions/upload-artifact@v4
+      - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
-          name: mkdocs-site
-          path: site/
+          name: local-recovery-regression
+          path: recovery-regression.json
           if-no-files-found: error
           retention-days: 7
 
-  safetybench:
-    name: Local recovery regression suite
+  docker-proof:
+    name: Real Docker proof - OpenAI migration
     runs-on: ubuntu-latest
-    timeout-minutes: 10
+    timeout-minutes: 15
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
-      - uses: astral-sh/setup-uv@v5
+      - uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
         with:
+          version: "0.12.5"
           enable-cache: true
           python-version: "3.12"
           cache-dependency-glob: uv.lock
 
-      - name: Install benchmark dependencies
+      - name: Install test dependencies
         run: uv sync --locked --group dev
 
-      - name: Run adversarial local regression cases
-        run: uv run python3 benchmarks/safetybench.py --output recovery-regression.json
+      - name: Verify Docker and prefetch proof image
+        run: |
+          docker info
+          docker pull python:3.12-slim
 
-      - uses: actions/upload-artifact@v4
-        with:
-          name: local-recovery-regression
-          path: recovery-regression.json
-          if-no-files-found: error
-          retention-days: 7
+      - name: Run real container-backed proof gates
+        env:
+          AGENTDIFF_DOCKER_TESTS: "1"
+        run: uv run pytest -p no:cacheprovider tests/test_docker_integration.py -v -rs
 
   package:
     name: Package validation
     runs-on: ubuntu-latest
-    needs: [tests, quality, security, docs, safetybench]
+    needs: [tests, quality, security, safetybench, docker-proof]
     timeout-minutes: 10
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
-      - uses: astral-sh/setup-uv@v5
+      - uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
         with:
+          version: "0.12.5"
           enable-cache: true
           python-version: "3.12"
           cache-dependency-glob: uv.lock
@@ -238,9 +239,9 @@ jobs:
           /tmp/agentdiff-wheel/bin/agentdiff policy validate --policy /tmp/agentdiff.yaml
           /tmp/agentdiff-wheel/bin/agentdiff-demo --json > /tmp/agentdiff-demo.json
 
-      - uses: actions/upload-artifact@v4
+      - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
           name: distributions
           path: dist/
           if-no-files-found: error
-          retention-days: 7
+          retention-days: 7
@@ -27,12 +27,12 @@ jobs:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Initialize CodeQL
-        uses: github/codeql-action/init@c4dd10e44af883a891fe31ced449bcb4a6728b9b # v3
+        uses: github/codeql-action/init@7188fc363630916deb702c7fdcf4e481b751f97a # v4.37.1
         with:
           languages: python
           queries: security-extended
 
       - name: Analyze
-        uses: github/codeql-action/analyze@c4dd10e44af883a891fe31ced449bcb4a6728b9b # v3
+        uses: github/codeql-action/analyze@7188fc363630916deb702c7fdcf4e481b751f97a # v4.37.1
         with:
           category: /language:python
@@ -14,16 +14,20 @@ jobs:
     runs-on: ubuntu-latest
     timeout-minutes: 10
     steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - uses: astral-sh/setup-uv@e58605a9b6da7c637471fab8847a5e5a6b8df081 # v5
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
         with:
-          version: "0.11.12"
+          version: "0.12.5"
           python-version: "3.14"
-      - name: Build from the lockfile
+      - name: Install build tools from the lockfile
+        run: uv sync --locked --group build
+      - name: Build distributions
         run: uv build
       - name: Validate metadata
-        run: uvx twine check dist/*
-      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        run: uv run twine check dist/*
+      - name: Validate wheel contents
+        run: uv run check-wheel-contents dist/*.whl
+      - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
           name: release-distributions
           path: dist/
@@ -43,9 +47,9 @@ jobs:
       contents: read
       id-token: write
     steps:
-      - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+      - uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
         with:
           name: release-distributions
           path: dist/
       - name: Publish through OIDC
-        uses: pypa/gh-action-pypi-publish@dc37677b2e1c63e2034f94d8a5b11f265b73ba33 # release/v1
+        uses: pypa/gh-action-pypi-publish@ed0c53931b1dc9bd32cbe73a98c7f6766f8a527e # v1.13.0
@@ -55,6 +55,7 @@ dmypy.json
 # AgentDiff local run artifacts
 /.agentdiff/
 agentdiff.yaml
+!demos/*/agentdiff.yaml
 snapshot_*.json
 trajectory*.json
 report*.json

@@ -2,10 +2,17 @@
 
 AgentDiff is pre-release software. APIs and artifact schemas may change before the first stable release.
 
-## 0.3.0 (unreleased)
+## 0.4.0 (unreleased)
 
 ### Added
 
+- **Verified API migration product**: read-only simulation, provider-aware usage scanning, deterministic migration planning, private generation, exact-scope policy, impact analysis, clean-room proof, integrity certificates, and optional verified GitHub pull-request delivery.
+- **OpenAI Chat Completions → Responses migration**: current text-only mapping for literal messages and simple response consumers, with explicit fail-closed review gates for tools, streaming, structured output, multimodal content, dynamic messages, wrappers, and other semantic differences.
+- **Untrusted generator boundary**: deterministic AST and exact-argv custom generators run against a private sealed source copy; no-op, missing, unexpected, and out-of-policy results cannot inherit a passing raw proof verdict.
+- **Migration certificates**: canonical SHA-256 integrity binds the source, repository base, expected and actual files, generator, policy, proof plan/results, patch digest, and evidence capsule; verification distinguishes valid, invalid, stale, and mismatched evidence.
+- **Provider intelligence**: data-only provider initialization plus bounded HTTPS fetching with redirect, DNS/IP, time, size, content-type, provenance, cache-validator, and digest controls. Executable plugin code requires trusted metadata and explicit opt-in.
+- **Reproducible OpenAI demos**: one supported success fixture and one unsafe-worker fixture that proves unexpected workflow changes force `NOT_PROVEN`.
+- **Verified Campaigns (`agentdiff fleet`)**: read-only multi-repository simulation, independent per-repository migration proof, explicit local scope, fail-closed rollup statuses, and a campaign digest bound to child certificate, patch, and proof digests.
 - **Zero-touch sidecar (`agentdiff serve` / `wrap` / `hook`)**: a small local HTTP daemon (127.0.0.1, bearer-token auth, no hosted service) that manages transactions, evidence, policy, sandbox selection, proof, retries, promotion, and notifications. `agentdiff wrap -- <agent argv>` runs any coding-agent CLI through the full pipeline; `agentdiff init` bootstraps trust configuration and can start the sidecar.
 - **Repository trust compiler (`agentdiff bootstrap` / `init`)**: deterministic inspection of languages, package managers, tests, builds, CI, CODEOWNERS, monorepo layout, agent configs, and lockfiles, compiled into one canonical `agentdiff.yaml`, `.agentdiff/trust.lock`, `.agentdiff/repo-graph.json`, `.agentdiff/proof-plan.json`, and compiled agent instructions (`agentdiff/adapters/*.md`).
 - **Impact-aware proof + content-addressed cache**: deterministic import graph (Python/JS-TS/Go/Rust) mapping changed files to affected modules, tests, and build targets; `static`/`targeted`/`full` proof planning with high-risk widening (dependencies, CI, Dockerfiles, build config, agent configs, security paths); integrity-sealed proof cache under `.agentdiff/cache/proof` keyed by base/patch/lock/image/plan digests.
@@ -15,14 +22,22 @@ AgentDiff is pre-release software. APIs and artifact schemas may change before t
 
 ### Changed
 
+- Consolidated API migration verification onto the existing authoritative `ProofEngine`; removed the duplicate verifier path and made `PROVEN` conditional on the combined generator, scope, policy, proof, and evidence result.
+- Replaced the MkDocs website with a separate React 19/Vite 8 product and documentation project. Removed MkDocs dependencies, sources, scripts, and deployment workflow from the Python repository.
+- Updated package metadata and the public version to 0.4.0; refreshed CI and release actions to pinned Node 24-capable versions.
 - `PromotionEngine` accepts an optional `store_root` so a proven patch living in a private workspace capsule can be promoted to the host repository.
 - `ProofEngine` accepts an optional content-addressed `cache`, a `base_preparer` (warm snapshot), and a proof `target`; proof results surface `cache_hit`/`cached_from_run`.
 - `PatchManifest` exposes a run-independent `content_digest()` so identical patches share proof-cache identity.
 
 ### Security
 
+- Verified PR creation replays only sealed patch bytes at the exact certified base, verifies before/after file digests, stages only sealed paths, and never auto-merges.
+- Certificates are documented as local integrity evidence, not cryptographic signatures; the default proof backend still requires Docker and fails closed when it is unavailable.
 - Trust decisions (policy, risk, proof, promotion, repair routing) remain fully deterministic; Cortex stays outside deterministic trust decisions.
-- The host repository is never writable inside the agent sandbox; the proof cache is content-addressed, integrity-sealed, and invalidated by any input change; warm base snapshots are immutable.
+- Corrected custom-generator runtime evidence: a private working copy is an observation boundary, not an OS sandbox. Custom commands retain the caller's host permissions and must be trusted until a sandbox-backed generator runtime is configured.
+- Campaign config/report/certificate control paths reject symlinks, reports are written atomically, aggregate verdicts are recomputed during verification, and every `PROVEN` child certificate is re-verified against sealed repository evidence.
+- Corrected Docker 28 bind-mount syntax so the real container runtime and clean-room proof environment request their writable private workspace without an invalid bare `rw` field.
+- The Docker proof backend keeps the host repository unmounted; the proof cache is content-addressed, integrity-sealed, and invalidated by any input change; warm base snapshots are immutable.
 
 ## 0.2.0
 

@@ -13,22 +13,22 @@ AgentDiff welcomes focused fixes, adversarial tests, documentation improvements,
 ```bash
 git clone https://github.com/kam6l/agentdiff.git
 cd agentdiff
-uv sync --locked --all-groups
+uv sync --locked --all-groups --extra langchain
 ```
 
 ## Change workflow
 
 1. Create a focused branch from `main`.
 2. Add a regression test that demonstrates the missing behavior.
 3. Implement the smallest complete change.
-4. Update `README.md`, `docs_src/`, examples, and security guidance when public behavior changes.
-5. Run the local quality gates documented in [the contributor guide](docs_src/contributing.md).
+4. Update `README.md`, examples, security guidance, and the separate React documentation project when public behavior changes.
+5. Run the local quality gates documented in `README.md`.
 6. Open a pull request that states behavior, trust-boundary impact, limitations, and real verification output.
 
 ## Security-sensitive changes
 
 Scanner, policy, persistence, process, redaction, and rollback changes require tests for failure and refusal paths—not only success paths. Recovery must preserve current data when identity or post-run equality is uncertain.
 
-Do not commit generated `site/`, coverage, cache, virtual-environment, distribution, or `.agentdiff/` run artifacts.
+Do not commit coverage, cache, virtual-environment, distribution, or `.agentdiff/` run artifacts. The product website and React documentation source live in the separate `agentdiff-web` project; this Python repository has no generated website tree.
 
 By participating, you agree to follow the [Code of Conduct](CODE_OF_CONDUCT.md).