Validate and configure Antigravity repair models

[wwind123]

Fixes #562

Implements configurable Antigravity repair models with the current default, one-time PTY-safe agy models discovery, actionable stale-model diagnostics, deduplicated fallback into antigravity_models, and direct attempts when discovery is unavailable. Preserves legacy Gemini repair semantics and updates documentation/tests.

Tests:

• python3 -m pytest tests/test_repair.py tests/test_antigravity.py
• python3 -m pytest

[wwind123]

Implemented and opened PR #563 for issue #562.

• Default Antigravity repair model is now Gemini 3.5 Flash (Medium).
• Repair candidates are validated once via PTY-safe agy models, with actionable stale-model diagnostics and fallback behavior.
• Added regression coverage and updated documentation.

Tests: python3 -m pytest tests/test_repair.py tests/test_antigravity.py (186 passed); python3 -m pytest (1535 passed).

-- OpenAI Codex: gpt-5.6-luna

[wwind123]

Review verdict: Blocking

Solid implementation of configurable repair-model validation and Antigravity-chain fallback (catalog fetched once via agy models, stale names skipped with actionable diagnostics, direct-attempt fallback when discovery itself fails), with good new regression coverage. One doc file was missed and now contradicts the other two updated docs in this same PR, and a parallel Literal type in usage.py wasn't extended for the new outcome value.

Same-PR follow-ups

• docs/skill_mode.md:160-162 still says 'explicit repeatable repair models define the only fallback chain,' which is now false — execute_repair's models chain is repair_models followed by antigravity_models (repair.py:1114). README.md and docs/local_agent_loop.md were correctly updated in this same PR to describe the new chain; skill_mode.md needs the same fix so the docs don't contradict each other.
• usage.py:94-96 and usage.py:202-204 define outcome: Literal["succeeded", "nonzero_exit", "empty_output", "timeout", "spawn_error", "invalid_output"] for UsageCallRecord/RunUsageContext.add_record, but repair.py:1143 now passes outcome="unavailable_model" into add_record, a value outside that Literal. repair.py's own RepairOutcome Literal was correctly extended (repair.py:1007-1010); the usage.py copy was missed.
• tests/test_repair.py:861-863 — only one blank line separates the new test_antigravity_repair_attempts_directly_when_catalog_fails from test_runner_pty_timeout_is_opt_in_and_retains_combined_log; the rest of the file uses two blank lines between top-level test functions (PEP 8).

-- Anthropic Claude: claude-sonnet-5

[wwind123]

Coder follow-up

Implemented all requested follow-up fixes and pushed them to the existing PR branch; the PR remains blocked pending another review round.

Addressed items

• item-1: Same-PR follow-up from Anthropic Claude: claude-sonnet-5, round 1: docs/skill_mode.md:160-162 still says 'explicit repeatable repair models define the only fallback...
  • Resolution: Updated docs/skill_mode.md to describe explicit repair models followed by the configured Antigravity fallback chain.
• item-2: Same-PR follow-up from Anthropic Claude: claude-sonnet-5, round 1: usage.py:94-96 and usage.py:202-204 define `outcome: Literal["succeeded", "nonzero_exit", "empty_...
  • Resolution: Added unavailable_model to both usage.py outcome Literal declarations used by UsageCallRecord and RunUsageContext.add_record.
• item-3: Same-PR follow-up from Anthropic Claude: claude-sonnet-5, round 1: tests/test_repair.py:861-863 — only one blank line separates the new `test_antigravity_repair_att...
  • Resolution: Added the missing blank line between the top-level tests in tests/test_repair.py.

Remaining items

• None.

Tests run

• python3 -m pytest tests/test_repair.py tests/test_usage.py — 147 passed
• python3 -m pytest — 1535 passed

-- OpenAI Codex: gpt-5.6-luna

[wwind123]

Review verdict: Approved

All three round-1 same-PR follow-ups are resolved: docs/skill_mode.md now correctly describes the repair-models-then-antigravity-chain fallback, usage.py's outcome Literal now includes "unavailable_model" matching repair.py's RepairOutcome, and the missing blank line in tests/test_repair.py is fixed. No new issues found in this round's diff.

Prior unresolved item dispositions

• [item-1] Same-PR follow-up from Anthropic Claude: claude-sonnet-5, round 1: docs/skill_mode.md:160-162 still says 'explicit repeatable repair models define the only fallback... -> resolved
• [item-2] Same-PR follow-up from Anthropic Claude: claude-sonnet-5, round 1: usage.py:94-96 and usage.py:202-204 define `outcome: Literal["succeeded", "nonzero_exit", "empty_... -> resolved
• [item-3] Same-PR follow-up from Anthropic Claude: claude-sonnet-5, round 1: tests/test_repair.py:861-863 — only one blank line separates the new `test_antigravity_repair_att... -> resolved

-- Anthropic Claude: claude-sonnet-5

[wwind123]

Review verdict: Approved

The PR correctly addresses all requirements and fixes the three prior unresolved items. Tests and documentation are properly updated.

Prior unresolved item dispositions

• [item-1] Same-PR follow-up from Anthropic Claude: claude-sonnet-5, round 1: docs/skill_mode.md:160-162 still says 'explicit repeatable repair models define the only fallback... -> resolved
• [item-2] Same-PR follow-up from Anthropic Claude: claude-sonnet-5, round 1: usage.py:94-96 and usage.py:202-204 define `outcome: Literal["succeeded", "nonzero_exit", "empty_... -> resolved
• [item-3] Same-PR follow-up from Anthropic Claude: claude-sonnet-5, round 1: tests/test_repair.py:861-863 — only one blank line separates the new `test_antigravity_repair_att... -> resolved

-- Google Antigravity: Gemini 3.1 Pro (High)