Repository navigation
Refresh stale default Antigravity and repair model chains (#2023) - #2046
Conversation
DEFAULT_ANTIGRAVITY_MODELS becomes Gemini 3.8 Flash (High) -> Claude Opus 5.5 (Medium), and DEFAULT_REPAIR_MODELS becomes Gemini 3.8 Flash (Medium) -> Claude Opus 5.5 (Medium). The removed Gemini 3.7/3.6 Flash and 3.1 Pro names are gone from both chains. The repair-model guard is unchanged. Pin tests, default-chain walk tests and docs are updated. Walk-mechanics tests that need several gemini-group entries use a test-local chain. Agent-Issue-Provenance: v1 repo=wwind123/coding-review-agent-loop issue=2023 flow=approved plan=02c49f6970b84553 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Issue implementationI implemented the approved plan for #2023 and opened draft PR #2046 (managed label). DEFAULT_ANTIGRAVITY_MODELS is now (Gemini 3.8 Flash (High), Claude Opus 5.5 (Medium)). DEFAULT_REPAIR_MODELS is now (Gemini 3.8 Flash (Medium), Claude Opus 5.5 (Medium)). The repair-model guard, the explicit-flag paths and review_images_capabilities.py are untouched. README.md and docs/local_agent_loop.md now name the new chains. Tests: the pins assert both tuples, the help text, and that no removed name or Gemini 4 Hydrogen is present. New pins cover the gemini/claude quota-group derivation and two default-chain walk tests (quota exhaustion jumps to Opus; transient failures end at Opus). The repair-walk test is renamed and asserts the de-duplicated three-model order. A new test covers codex repair without and with --repair-model. Multi-Gemini walk-mechanics tests now use a test-local chain. Test selection: the focused files cover each changed test module and config.py. The full suite was run because the approved plan's acceptance criteria explicitly require it, and it passed. ResultPull request reported: #2046. Tests run
Test observation receipts
Risk-based mode and transition test matrix evidence
Full matrix evidence (4 rows)
Risk-matrix coverage mapOrchestrator completeness check (existence and level only; whether each cited test really exercises its row is for the reviewer to judge).
-- Anthropic Claude: claude-opus-5-5 (medium) |
|
Agent-loop managed-CI resume provenance audit record (machine-readable). Records resume provenance for managed CI on pull request #2046 at head dfcc84f. Not an agent response; keep this comment. AGENT_MANAGED_CI_UNPROTECTED_OVERRIDE_V1 nonce=OSgBks5NbLKkRkDxtdE8S5BYUNJiy7ha repo=wwind123/coding-review-agent-loop base=main head=dfcc84fd562af819bb15126fb8a0c4c726bf0b40 protection=voluntary active_label_event_id=36787978984 resume_from=7044401610 provenance_head=dfcc84fd562af819bb15126fb8a0c4c726bf0b40 generation=CKqk1GWByEMCS1cx Resume provenance only: the prior issue-created audit is not an authorization token. |
|
PR review scheduling audit: selected Codex; paused Claude, Antigravity; reason: strict pre-panel fallback: architecture identity changed; scheduler metadata recovery: latest scheduler metadata is missing or invalid; scheduler metadata recovery: current-head records lack scheduler metadata; scheduler metadata is missing or invalid; primary re-invoked with full context; phase: primary; head: dfcc84f; primary: Codex; active owners: (none); force-full: False (source: none); scheduler-policy calls avoided cumulatively: 2. |
|
Review verdict: Blocking Reviewed the complete diff at dfcc84f. The requested constants, documentation, explicit-flag behavior, and four risk-matrix rows are supported by source inspection and supplied passing receipts; GitHub checks are passing. One existing chain-walk test loses its intended coverage. No tests were run during review. Blocking issues
-- OpenAI Codex: gpt-6.1-sol (medium) |
|
PR review round 1 reconciliation: settled reviewers: Codex. Finalization continues after reconciliation. Historical approvals remain exact-head-bound; scheduler-policy calls avoided cumulatively: 2. Phase: primary; force-full: False (source: none). |
|
PR #2046 coder dispatch record: round 1, attempt 1, head |
The transcript-reset test queued six outputs but the two-entry default chain consumes three, so its noise output was never read. Parametrize it over the test-local multi-Gemini chain and the default chain, queue one output per expected call, and assert every output was consumed. The new default-chain transient test gets the same exact-queue assertion. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Coder follow-upFixed item-1 by restoring end-of-chain transcript-reset coverage, pushed as commit 282852e. The test is now parametrized over the test-local multi-Gemini chain (the original five-call walk) and the new default chain (three calls). It queues exactly one output per expected call and asserts that the queue is empty afterwards, so the noise output on Opus is provably consumed. Generalization: this generalizes the fix for [item-1]: a walk test must queue exactly the outputs its expected walk consumes and assert that all of them were consumed. I applied the same rule to the new default-chain transient test, which also over-queued (six outputs for three calls). Test rationale: tests/test_antigravity_quota_groups.py is the only changed file and holds every test touched here. Addressed items
Remaining items
Tests run
Test observation receipts
Risk-based mode and transition test matrix evidence
4 changed rows; 0 unchanged since round 1
0 rows unchanged since round 1; full matrix in round 1. Risk-matrix coverage mapOrchestrator completeness check (existence and level only; whether each cited test really exercises its row is for the reviewer to judge).
-- Anthropic Claude: claude-opus-5-5 (medium) |
|
PR review scheduling audit: selected Codex; paused Claude, Antigravity; reason: primary phase: primary rechecks its findings before the secondary panel; phase: primary; head: 282852e; primary: Codex; active owners: Codex; force-full: False (source: none); scheduler-policy calls avoided cumulatively: 4. |
|
Review verdict: Blocking Reviewed the complete base-to-head diff at 282852e. Item-1 is fixed and no further code defects were found. GitHub checks are passing, but two enforceable risk-matrix rows lack admissible coverage citations. No tests were run during review. Blocking issues
Prior unresolved item dispositions
-- OpenAI Codex: gpt-6.1-sol (medium) |
|
PR review round 2 reconciliation: settled reviewers: Codex. Finalization continues after reconciliation. Historical approvals remain exact-head-bound; scheduler-policy calls avoided cumulatively: 4. Phase: primary; force-full: False (source: none). |
|
PR #2046 coder dispatch record: round 2, attempt 1, head |
Coder follow-up
Coder summary (claim, unverified): Items 2 and 3 asked for evidence, not code changes, so I made no code change. I re-ran all six focused test files through agent-loop run-tests on the reviewed head 282852e (1068 passed) and cite that one current-tree run for all four risk-matrix rows. Generalization: this generalizes the fix for [item-2]: every risk-matrix row cites a passing observation from a run on the current tree, not a run on an earlier tree. Test rationale: test_cli_repair.py (repair pins and the explicit codex flag), test_antigravity.py (the default chain and explicit Antigravity chains), test_antigravity_quota_groups.py (default-chain walks), test_repair.py (default repair walk), test_docs_guidance.py (docs chain assertions) and test_agent_loop.py (explicit --antigravity-model). Each maps to a file the PR changed or to a matrix row. Claimed already present at
|
|
PR review scheduling audit: selected Codex; paused Claude, Antigravity; reason: strict pre-panel fallback: the active obligation ledger is not reconstructible; primary re-invoked with full context; phase: primary; head: 282852e; primary: Codex; active owners: Codex; force-full: False (source: none); scheduler-policy calls avoided cumulatively: 6. |
|
Review verdict: Approved Reviewed the complete base-to-head diff at 282852e. The constants, documentation, default-chain walks and explicit-flag coverage satisfy the approved scope. All four matrix rows have suitable tests and a current-head parent-observed passing citation (cfbbe80b59bf42c0801139f33d66c7e0), with a reported environment-difference caveat. Both carried evidence gaps are resolved. GitHub checks are passing; no tests were run during this review. Prior unresolved item dispositions
-- OpenAI Codex: gpt-6.1-sol (medium) |
|
PR review round 3 reconciliation: settled reviewers: Codex. Finalization continues after reconciliation. Historical approvals remain exact-head-bound; scheduler-policy calls avoided cumulatively: 6. Phase: primary; force-full: False (source: none). |
|
PR review scheduling audit: selected Claude, Antigravity; paused Codex; reason: independent secondary audit after exact-head primary approval; phase: secondary-audit; head: 282852e; primary: Codex; active owners: (none); force-full: False (source: none); scheduler-policy calls avoided cumulatively: 6. |
|
Review verdict: Approved I reviewed the full base-to-head diff at 282852e on my own. config.py sets DEFAULT_ANTIGRAVITY_MODELS to exactly (Gemini 3.8 Flash (High), Claude Opus 5.5 (Medium)), with the separate-quota Opus entry last. DEFAULT_REPAIR_MODELS is the two-model tuple (Gemini 3.8 Flash (Medium), Claude Opus 5.5 (Medium)). The diff does not touch the repair-model guard, review_images_capabilities.py or any explicit-flag path. Quota groups still come from the name prefix, and a new pin checks that they are [gemini, claude]. All four risk-matrix rows are covered by tests that run the real path. (1) test_default_chain_quota_exhaustion_jumps_straight_to_opus checks the order [GEM1, OPUS], with no sleep and no removed models. (2) test_default_chain_transient_failures_end_at_opus_with_ordinary_error checks the order [GEM1, GEM1, OPUS], that every queued output is consumed, that no quota memory is written, and that the error is a plain AgentInvocationError. (3) The renamed repair test checks the de-duplicated order [3.8 Flash (Medium), Opus, 3.8 Flash (High)] and that quota memory is unchanged. (4) The new codex --repair-model test checks both the rejection and the verbatim ('M',). Tests that check walk mechanics now pass an explicit test-local MULTI_GEMINI_CHAIN, so they keep their four-Gemini coverage, and the end-of-chain transcript-noise test runs on both chains. README.md and docs/local_agent_loop.md name the new default and repair chains, and the docs-guidance test checks both files and rejects the removed names in the chain sentences. A grep of src/, README.md and the guide finds no default-chain use of the removed models. One left-over: the illustrative explicit-chain example at docs/local_agent_loop.md:1144 still names Gemini 3.1 Pro (High). That example is not a default, and models missing from -- Anthropic Claude: claude-opus-5-5 (medium) |
|
Review verdict: Approved Approved. The PR cleanly refreshes DEFAULT_ANTIGRAVITY_MODELS and DEFAULT_REPAIR_MODELS to match current provider availability, omitting stale Gemini 3.7/3.6/3.1 entries and Gemini 4 Hydrogen. The repair-model guard and explicit-flag precedence remain preserved and unmodified. Tests pin the new tuples, verify quota group derivations, test default-chain quota fallback and transient exhaustion, and preserve multi-model gemini group walk mechanics via a test fixture. Docs are accurately updated. -- Google Antigravity: Gemini 3.8 Flash (High) |
|
PR review round 4 reconciliation: settled reviewers: Antigravity, Claude. Finalization continues after reconciliation. Historical approvals remain exact-head-bound; scheduler-policy calls avoided cumulatively: 6. Phase: secondary-audit; force-full: False (source: none). |
|
Managed CI authorization for exact head 282852e. |
Summary
This refreshes the stale built-in model chains, as the approved plan asks.
DEFAULT_ANTIGRAVITY_MODELSis nowGemini 3.8 Flash (High)→Claude Opus 5.5 (Medium). Opus runs on a separate quota and stays last.DEFAULT_REPAIR_MODELSis nowGemini 3.8 Flash (Medium)→Claude Opus 5.5 (Medium). It is still a two-model tuple.if self.repair_models == DEFAULT_REPAIR_MODELS) is unchanged. So are the explicit-flag paths andreview_images_capabilities.py.geminiand thenclaude, and a new test pins this.[Gemini 3.8 Flash (Medium), Claude Opus 5.5 (Medium), Gemini 3.8 Flash (High)], and a test asserts that order.Tests
--repair-backend codexwithout--repair-modelis rejected, and with--repair-model Mit yields("M",).README.mdanddocs/local_agent_loop.md.agent-loop run-tests. Both passed.Fixes #2023
AGENT_MANAGED_CI_UNPROTECTED_OVERRIDE_V1 nonce=5vXFxMMMT0mabgO8kdITug7K
🤖 Generated with Claude Code