Prepare the v0.3 release evidence profile

.github/workflows/ci.yml

@@ -442,14 +442,16 @@ jobs:
             tests/test_upstream_comparison_generator.py
           python -m agentrunproof check-upstream-bundle \
             evidence/upstream-comparison/v1/bundle.json
+          python -m agentrunproof check-upstream-bundle \
+            evidence/upstream-comparison/v2/bundle.json
       - name: Reproduce committed comparison evidence when present
         id: reproduce
         shell: bash
         env:
           INPUT_UPSTREAM_TARGET_COMMIT: ${{ inputs.upstream_target_commit }}
         run: |
           set -euo pipefail
-          committed_bundle="evidence/upstream-comparison/v2/bundle.json"
+          committed_bundle="evidence/upstream-comparison/v3/bundle.json"
           if [[ ! -f "$committed_bundle" ]]; then
             target_commit="$INPUT_UPSTREAM_TARGET_COMMIT"
             if [[ -z "$target_commit" ]]; then
@@ -528,7 +530,7 @@ jobs:
           )"
           test "${#evidence_parent_fields[@]}" -eq 2
           evidence_parent="${evidence_parent_fields[1]}"
-          git merge-base --is-ancestor "$source_commit" "$evidence_parent"
+          test "$evidence_parent" = "$source_commit"
           test "$(git rev-parse "$source_commit^{tree}")" = "$source_tree"
 
           python - "$committed_bundle" > "$RUNNER_TEMP/expected-upstream-evidence-files" <<'PY'

.github/workflows/publish.yml

@@ -8,7 +8,7 @@ on:
       tag:
         description: Existing final GitHub Release tag to publish
         required: true
-        default: v0.2.0
+        default: v0.3.0
         type: string
       publish:
         description: Upload the verified distributions to PyPI
@@ -77,6 +77,12 @@ jobs:
               evidence_checker=check-upstream-bundle
               evidence_kind=upstream-comparison
               ;;
+            v0.3.0)
+              evidence_directory=evidence/upstream-comparison/v3
+              evidence_member=
+              evidence_checker=check-upstream-bundle
+              evidence_kind=upstream-comparison
+              ;;
             *)
               echo "Unsupported release evidence profile: $RELEASE_TAG" >&2
               exit 1
@@ -229,7 +235,7 @@ jobs:
           test "${#evidence_commit_fields[@]}" -eq 2
           evidence_parent="${evidence_commit_fields[1]}"
           if [[ "$EVIDENCE_KIND" = upstream-comparison ]]; then
-            git merge-base --is-ancestor "$source_commit" "$evidence_parent"
+            test "$evidence_parent" = "$source_commit"
             python - "$committed_bundle" > "$RUNNER_TEMP/expected-evidence-files" <<'PY'
           import json
           import sys
@@ -278,6 +284,10 @@ jobs:
               test "$(git rev-parse HEAD)" = "$evidence_commit"
               : > "$RUNNER_TEMP/expected-post-evidence-files"
               ;;
+            v0.3.0)
+              test "$(git rev-parse HEAD)" = "$evidence_commit"
+              : > "$RUNNER_TEMP/expected-post-evidence-files"
+              ;;
           esac
           git diff --name-only "$evidence_commit..HEAD" | sort \
             > "$RUNNER_TEMP/actual-post-evidence-files"

CHANGELOG.md

@@ -2,15 +2,15 @@
 
 All notable changes to AgentRunProof will be documented here.
 
-## 0.3.0 - Unreleased
+## 0.3.0 - 2026-08-21
 
 - Extend the declared dependency window to `openai-agents>=0.20.0,<0.23` and add exact 0.22.0
   packaged-wheel coverage to the Python 3.10–3.14 CI matrix alongside 0.20.0 and 0.21.0.
 - Treat SDK 0.22's output-guardrail replacement semantically: preserve call/output linkage and
   require the rejected raw tool result to be absent without pinning the SDK's replacement text;
   use scenario/case revision 2 while older history keeps revision 1.
-- Preserve certificate v1 plus all existing historical evidence and lock files. Publication of
-  v0.3.0 remains blocked on a fresh clean-commit comparison bundle and complete release gates.
+- Preserve certificate v1 plus all existing historical evidence and lock files, and require a
+  fresh v3 comparison bundle bound to the 0.3.0 source and release wheel.
 - Add a provider-free real-`Runner` tool example, contribution and security guidance, structured
   issue and pull-request templates, and a five-minute upstream `RunState` case study.
 - Rework the project landing page around a 30-second PyPI check, supported SDK boundaries, honest

CITATION.cff

@@ -5,8 +5,7 @@ type: software
 authors:
   - family-names: Fu
     given-names: Xiaonan
-version: 0.2.0
-date-released: 2026-08-15
+version: 0.3.0
 repository-code: "https://github.com/FU-max-boop/agentrunproof"
 url: "https://github.com/FU-max-boop/agentrunproof"
 license: MIT

PLAN.md

@@ -56,8 +56,9 @@ Deliver a standalone, installable tool that deterministically checks OpenAI Agen
 - [x] (2026-08-21) Prepared the v0.3.0 source release candidate for SDK 0.22: widened metadata to
   `<0.23`, added the exact 0.22.0 packaged-wheel cells, and made the output-guardrail history check
   require linked redacted replay without matching replacement prose. One built wheel passed the
-  full suite locally on CPython 3.12 against exact SDK 0.20.0, 0.21.0, and 0.22.0. The full
-  Python 3.10–3.14 CI matrix and a fresh release-bound v3 evidence bundle remain publication gates.
+  full suite locally on CPython 3.12 against exact SDK 0.20.0, 0.21.0, and 0.22.0. The full hosted
+  Python 3.10–3.14 matrix passed on the release-candidate PR and main; a fresh release-bound v3
+  evidence bundle remains the publication gate.
 
 ## Milestones
 
@@ -158,6 +159,12 @@ The milestone is complete when the charter's Gate 4 is satisfied.
   signed-off evidence and historical lock files remain immutable. A v0.3 release remains blocked
   until fresh clean-commit evidence and all release gates exist; documentation and workflow
   preparation must not imply that 0.3.0 has been published.
+- **2026-08-21 — v0.3 evidence publication sequence:** Freeze all package inputs, final release
+  metadata, release-profile logic, and evidence-generation workflow in one clean source commit C.
+  Generate the versioned v3 comparison on canonical Linux from C, then add only the six JSON
+  evidence members in a direct child commit E. The v0.3.0 tag must point directly to E; v1 and v2
+  evidence remain immutable and continue to be checked. This preserves the two-commit provenance
+  design and prevents a release artifact from claiming evidence generated by a different source.
 
 ## Outcomes & Retrospective
 

PROJECT_CHARTER.md

@@ -25,10 +25,9 @@ The first public contract covers the standard text `Runner` and these released p
 
 For one logical scenario, AgentRunProof may execute several variants. A variant is a real SDK run, not a simulation of the runner. The harness controls only model responses, local test tools, and the test session.
 
-The v0.3 release candidate declares the compatibility window `openai-agents>=0.20,<0.23` on
-Python 3.10 through 3.14. Its acceptance matrix installs the packaged artifact against exactly SDK
-0.20.0, 0.21.0, and 0.22.0; a Python/SDK cell is claimed as tested only after that exact CI cell
-passes. The latest published release remains v0.2.0 until the v0.3 release gates complete.
+The v0.3 source contract declares the compatibility window `openai-agents>=0.20,<0.23` on Python 3.10
+through 3.14. Its acceptance matrix installs the packaged artifact against exactly SDK 0.20.0,
+0.21.0, and 0.22.0; a Python/SDK cell is claimed as tested only after that exact CI cell passes.
 
 v0.3 retains certificate v1, the v0.1 runtime scenarios, and v0.2's explicit
 `DeterministicModel(..., emit_traces=True)` interoperability path: supported SDKs 0.21 and 0.22

README.md

@@ -14,11 +14,10 @@ across `run` and `run_streamed`, and writes content-addressed JSON records for s
 tool-linkage, and `RunState` resume invariants. A failing record carries the normalized
 counterexample observations.
 
-The v0.3.0 release candidate in this source tree declares
-`openai-agents>=0.20,<0.23` on Python 3.10–3.14. Its packaged-wheel CI matrix requires the exact
-0.20.0, 0.21.0, and 0.22.0 release baselines. The latest published package remains v0.2.0, whose
-declared window ends before SDK 0.22. Built-in scenarios make no model API call and require no API
-key.
+The 0.3.0 source contract declares `openai-agents>=0.20,<0.23` on Python 3.10–3.14. Its
+packaged-wheel CI matrix requires the exact 0.20.0, 0.21.0, and 0.22.0 release baselines. Published
+availability is shown by the PyPI badge and immutable GitHub Releases. Built-in scenarios make no
+model API call and require no API key.
 
 > AgentRunProof-backed reports are referenced by two merged maintainer fixes,
 > [#4413](https://github.com/openai/openai-agents-python/pull/4413) and
@@ -36,7 +35,7 @@ or [CLI reference](https://github.com/FU-max-boop/agentrunproof/blob/main/docs/c
 ## 30-second local check
 
 ```bash
-python -m pip install "agentrunproof==0.2.0"
+python -m pip install agentrunproof
 agentrunproof probe basic-tool-session-parity --certificate proof.json
 agentrunproof check-certificate proof.json
 ```
@@ -63,17 +62,16 @@ and the
 
 Maintaining a downstream library? The
 [isolated, test-only CI guide](https://github.com/FU-max-boop/agentrunproof/blob/main/docs/ci-adoption.md)
-provides a copyable real-`Runner` contract test and an ephemeral `uv` matrix for exact SDK 0.20.0
-and 0.21.0 using the published v0.2.0 package, plus the pending v0.3.0 expansion to SDK 0.22.0. It
-keeps AgentRunProof out of runtime metadata and the project lockfile. Provider-free means the
-built-in deterministic model sends no model-provider request; it does not make arbitrary downstream
-code or dependency installation network-isolated.
-
-For artifact review, pin `agentrunproof==0.2.0` and use the immutable
-[v0.2.0 release](https://github.com/FU-max-boop/agentrunproof/releases/tag/v0.2.0), whose
+provides a copyable real-`Runner` contract test and an ephemeral `uv` matrix. Its published v0.2.0
+example covers exact SDK 0.20.0 and 0.21.0; the 0.3.0 source contract adds SDK 0.22.0 and may be
+pinned downstream only after its immutable release and PyPI pages exist. The isolated pattern keeps
+AgentRunProof out of runtime metadata and the project lockfile.
+
+For artifact review, pin an exact published version and use its matching
+[immutable GitHub Release](https://github.com/FU-max-boop/agentrunproof/releases). Each
 `SHA256SUMS` binds the wheel and sdist. The release workflow rebuilds and byte-compares those
 artifacts, smoke-tests the wheel, and publishes the same verified files to PyPI through OIDC trusted
-publishing. The CI guide records the exact hashes and the remaining third-party-code trust boundary.
+publishing. The CI guide explains the remaining third-party-code trust boundary.
 
 ## Where it fits
 

docs/ci-adoption.md

@@ -39,10 +39,9 @@ uv run --isolated --no-project --python 3.12 \
   -- python -m pytest -q examples/ci_adoption/test_runner_contract.py
 ```
 
-These commands intentionally remain on the published v0.2.0 contract. The 0.3.0 source candidate
-adds exact SDK 0.22.0 coverage, but `agentrunproof==0.3.0` must not be used in a downstream PyPI job
-until that release exists and publishes verified artifact hashes. From this repository checkout,
-the candidate can be tested without changing project metadata:
+The 0.3.0 source contract adds exact SDK 0.22.0 coverage. Pin `agentrunproof==0.3.0` downstream
+only after its immutable release and PyPI pages exist. Before then, the source contract can be
+tested without changing project metadata:
 
 ```bash
 uv run --isolated --no-project --python 3.12 \
@@ -62,7 +61,7 @@ the adapter, name the applicable extra, for example `--with-editable ".[openai]"
 
 This job is deliberately separate from the downstream project's normal dependency installation.
 It grants read-only repository permission, pins the action and tool revisions, and tests the two
-exact SDK releases covered by AgentRunProof 0.2.0.
+exact SDK releases covered by the published AgentRunProof 0.2.0 contract.
 
 ```yaml
 name: OpenAI Agents Runner contract
@@ -85,10 +84,10 @@ jobs:
       matrix:
         sdk-version: ["0.20.0", "0.21.0"]
     steps:
-      - uses: actions/checkout@fbc6f3992d24b796d5a048ff273f7fcc4a7b6c09 # v5.1.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
-      - uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6.3.0
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:
           python-version: "3.12.13"
       - uses: astral-sh/setup-uv@20cfd1bf945f4377ade1205e4dbc17946fc9a30d # v10.0.1
@@ -126,17 +125,17 @@ the test intentionally covers an integration boundary.
 
 ## Dependency and release trust
 
-- Pin `agentrunproof==0.2.0` and an exact `openai-agents` release in the isolated job. The supported
+- Pin `agentrunproof==0.2.0` and an exact `openai-agents` release in the isolated job. Its supported
   range is `openai-agents>=0.20.0,<0.22`; 0.20.0 and 0.21.0 are the exact packaged CI baselines.
-- The 0.3.0 source candidate expands that declaration to `openai-agents>=0.20.0,<0.23` and adds an
+- The 0.3.0 source contract expands that declaration to `openai-agents>=0.20.0,<0.23` and adds an
   exact 0.22.0 packaged-wheel cell on every supported Python. Switch downstream pins and matrices
-  only after the final release, checksums, and release-bound evidence are public.
+  after the final release, checksums, and release-bound evidence are public.
 - The immutable [v0.2.0 GitHub release](https://github.com/FU-max-boop/agentrunproof/releases/tag/v0.2.0)
   publishes wheel and sdist SHA-256 values in `SHA256SUMS`. Its release workflow re-downloads those
   assets, compares them byte-for-byte with a rebuild from the tagged source, and smoke-tests the
   wheel before the same verified files are sent to PyPI through OIDC trusted publishing.
-- The released wheel SHA-256 is
-  `e393e98bf797cc10f07ea151ec6fffd3b1ebbb21307256309e08f11e97a27d51`; the sdist SHA-256 is
+- The v0.2.0 wheel SHA-256 is
+  `e393e98bf797cc10f07ea151ec6fffd3b1ebbb21307256309e08f11e97a27d51`; its sdist SHA-256 is
   `c0ad9c2aeb425cfcaff81b6daeee9c9ded069397dfc7a3319fda28f9842f2ace`.
 - A repository with a lock or hash policy should resolve this isolated closure with its normal
   dependency-review process and commit the resulting test-only lock data. The short `uv run`

docs/openai-agents.md

@@ -23,27 +23,25 @@ SDK regression test for the fix.
 
 ## Compatibility
 
-The AgentRunProof 0.3.0 release candidate declares `openai-agents>=0.20.0,<0.23` on Python 3.10
-through 3.14. Its packaged-wheel acceptance matrix covers exact SDK 0.20.0, 0.21.0, and 0.22.0
-installations. Other patch, prerelease, or source revisions are not individually verified support
-cells. The latest published AgentRunProof release is still 0.2.0 and does not declare SDK 0.22
-support.
+The AgentRunProof 0.3.0 source contract declares `openai-agents>=0.20.0,<0.23` on Python 3.10 through 3.14. Its
+packaged-wheel acceptance matrix covers exact SDK 0.20.0, 0.21.0, and 0.22.0 installations. Other
+patch, prerelease, or source revisions are not individually verified support cells.
 
 Install the released package in a fresh environment:
 
 ```bash
 python -m venv .venv
 source .venv/bin/activate
-python -m pip install "agentrunproof==0.2.0"
+python -m pip install agentrunproof
 mkdir -p build
 ```
 
 Built-in scenarios use deterministic local responses and make zero provider requests. They require
 no API key. This guarantee covers the built-in scenarios, not arbitrary user-supplied Python code;
 package installation may still contact the configured package index.
 
-To exercise the 0.3.0 candidate against its new exact baseline from a source checkout, install the
-checkout and SDK together instead of naming a nonexistent PyPI release:
+To exercise the 0.3.0 source contract against SDK 0.22 before its immutable release and PyPI pages
+exist, install the checkout and exact SDK together:
 
 ```bash
 python -m pip install -e . "openai-agents==0.22.0"

tests/test_publish_workflow.py

@@ -9,13 +9,15 @@
 def test_upstream_comparison_release_profiles_are_versioned() -> None:
     workflow = WORKFLOW.read_text(encoding="utf-8")
 
-    assert "default: v0.2.0" in workflow
+    assert "default: v0.3.0" in workflow
     assert "v0.1.0)" in workflow
     assert "v0.1.1)" in workflow
     assert "v0.1.2)" in workflow
     assert "v0.2.0)" in workflow
+    assert "v0.3.0)" in workflow
     assert "evidence_directory=evidence/upstream-comparison/v1" in workflow
     assert "evidence_directory=evidence/upstream-comparison/v2" in workflow
+    assert "evidence_directory=evidence/upstream-comparison/v3" in workflow
     assert "evidence_checker=check-upstream-bundle" in workflow
     assert "evidence_kind=upstream-comparison" in workflow
 
@@ -38,12 +40,15 @@ def test_comparison_evidence_commit_is_minimal_and_is_the_tag_target() -> None:
     tag_policy = workflow.rsplit('case "$RELEASE_TAG" in', 1)[1]
     v012_policy = tag_policy.split("v0.1.2)", 1)[1].split(";;", 1)[0]
     v020_policy = tag_policy.split("v0.2.0)", 1)[1].split(";;", 1)[0]
+    v030_policy = tag_policy.split("v0.3.0)", 1)[1].split(";;", 1)[0]
 
-    assert 'git merge-base --is-ancestor "$source_commit" "$evidence_parent"' in workflow
+    assert 'git merge-base --is-ancestor "$source_commit" "$evidence_parent"' not in workflow
     assert 'git diff --name-only "$evidence_parent..$evidence_commit"' in workflow
     assert 'git diff --exit-code "$evidence_commit" HEAD -- "$EVIDENCE_DIRECTORY"' in workflow
     assert 'test "$(git rev-parse HEAD)" = "$evidence_commit"' in v012_policy
     assert 'test "$(git rev-parse HEAD)" = "$evidence_commit"' in v020_policy
+    assert 'test "$(git rev-parse HEAD)" = "$evidence_commit"' in v030_policy
+    assert 'test "$evidence_parent" = "$source_commit"' in workflow
     assert 'bundle_path.parent / run["certificate"]["path"]' in workflow
 
 

tests/test_upstream_comparison_generator.py

@@ -184,12 +184,13 @@ def test_ci_can_generate_the_first_canonical_candidate_from_an_exact_target() ->
     assert "runstate-recursive-agent-tool-approval-serialization,PASS" in workflow
     assert "candidate-top-level" not in workflow
     assert "python -m agentrunproof check-upstream-bundle" in workflow
-    assert 'committed_bundle="evidence/upstream-comparison/v2/bundle.json"' in workflow
+    assert 'committed_bundle="evidence/upstream-comparison/v3/bundle.json"' in workflow
     assert "evidence/upstream-comparison/v1/bundle.json" in workflow
+    assert "evidence/upstream-comparison/v2/bundle.json" in workflow
     assert "find evidence/upstream " not in workflow
-    assert 'git merge-base --is-ancestor "$source_commit" "$evidence_parent"' in workflow
+    assert 'git merge-base --is-ancestor "$source_commit" "$evidence_parent"' not in workflow
     assert 'git diff --name-only "$evidence_parent..$evidence_commit"' in workflow
-    assert 'test "${evidence_parent_fields[1]}" = "$source_commit"' not in workflow
+    assert 'test "$evidence_parent" = "$source_commit"' in workflow
 
 
 def test_replaceable_source_target_binds_clean_git_provenance(tmp_path: Path) -> None: