[0xAndoroid]

This PR lets the AI merge gate accept Claude Fable 5.2 as the reviewer model named in a review attestation, the PR comment line that records a clean review of the exact head commit. The accepted reasoning-effort settings, max and xhigh, and every other gate rule stay unchanged.

Why

The AI merge gate (scripts/ci/ai_merge_gate.py) rates a maintainer's pull request on a risk scale of 1 to 4. It approves one rated 1 to 3 only after an authorized PR comment attests a clean review of the exact head commit. The attestation is one fixed-format line that names the pull request, the head commit, the review procedure (a free-text "skill" name), the reviewer model and the reasoning-effort setting the review ran at. The gate accepts the line only if the model is on its list of accepted models, which until now held GPT-6 Astra and Claude Opus 5.5.

Anthropic released Claude Fable 5.2 on November 9, 2026, and the gate rejects an attestation that names it.

The gate compares the model text in the line with the list by exact string, so the new entry is the display name Claude Fable 5.2, spelled the same way as the existing entries. Other spellings stay rejected: with this change applied, the API id claude-fable-5-2 and the lowercase claude fable 5.2 still leave the gate check neutral.

Changes

• scripts/ci/ai_merge_gate.py: the tuple REVIEW_MODELS gains "Claude Fable 5.2". The check on the attestation line and the hint in the gate's check output for a missing attestation both read this tuple, so no other line of the script changes.
• .github/ai-merge-gate.md: the sentence in the gate guide that lists the accepted models names the new one.
• scripts/tests/test_ai_merge_gate.py: the accepted cases in test_attestation_requires_an_exact_line_for_this_pull_and_head gain Claude Fable 5.2 with max and with xhigh.

No other file lists the accepted models. CONTRIBUTING.md, CLAUDE.md and .claude/skills/ci-code-review/SKILL.md point to the gate guide for the list. The GPT-6 Astra attestation line in CONTRIBUTING.md and the gpt-6-astra command in the skill are examples and remain valid.

Testing

CI runs the gate tests with python3 -m unittest scripts.tests.test_ai_merge_gate -v. The same module, run locally through uv on Python 3.13.15, passes all 61 tests:

uv run --no-project python -m unittest scripts.tests.test_ai_merge_gate -v

With the REVIEW_MODELS change reverted, the two new cases fail: the gate check concludes neutral where the test expects success.

Side effect on open pull requests

Because the gate's cache of risk classifications depends on the bytes of the gate script, merging this PR makes the gate classify open pull requests again. The gate stores each classification together with a SHA-256 digest of six files, and scripts/ci/ai_merge_gate.py is one of them, so any edit to that file invalidates the stored classifications. The gate's workflows check out the default branch, never the pull request's copy, so every evaluation after the merge uses the edited script. At its next gate evaluation after the merge, a ready, maintainer-authored pull request with no standing change request and no current-head approval from another maintainer has its head classified again. That run uses one of the two paid classifier attempts the gate allows per head. A head that has already used both stays unclassified until its next push, and an unclassified head needs a current-head maintainer approval. A classification already running when this merges also loses its paid attempt: its publish job checks out the edited script, the prepared input digest no longer matches, and that run leaves the head unclassified.

[github-actions]

📏 PR diff

Note

Actual code changed: 0 lines

🟢 0 added    🔴 0 removed

Tests, docs, fixtures, and helper scripts excluded.

Category	🟢 Added	🔴 Removed
⌨️ Actual code	0	0
🧪 Tests	+2	−1
📚 Docs	+1	−1
📦 Fixtures	0	0
🔧 Helper scripts	+1	−1
Total diff	+4	−3

_{Every changed line is classified once.}

[0xandoroid-ai-agent]

This pull request #2394 has been reviewed at commit 21ebbb0 using skill reviewer and model Claude Opus 5.5 with xhigh reasoning effort, and no issues were found.

[github-actions]

Verifier classifier: risk-1

Head 21ebbb007d9476d73300696333d93624ead5d3d3

The change expands permitted review-attestation models in CI merge policy. In both revisions, this flows through attester and merge-gate decisions, not proof verification. No verifier acceptance behavior changes.

Risk: risk-1

Per-path ratings (3 files)

• .github/ai-merge-gate.md: risk-1 — Documentation adds an allowed review model; it has no production verifier effect.
• scripts/ci/ai_merge_gate.py: risk-1 — The allowlist affects GitHub review attestations and merge approvals only, with no path to production proof acceptance.
• scripts/tests/test_ai_merge_gate.py: risk-1 — Adds test cases for the new review model; no production behavior changes.

gpt-6.1-sol, reasoning effort xhigh

[github-actions]

The AI merge gate approves 21ebbb007d9476d73300696333d93624ead5d3d3. Review attested by 0xandoroid-ai-agent: An authorized comment attests a clean review of the current head 21ebbb007d9476d73300696333d93624ead5d3d3, so the gate approves it.