Validate and configure the Antigravity repair model instead of hard-coding Gemini 3 Flash

[wwind123]

Problem

Structured-response repair invokes agy with the hard-coded model label Gemini 3 Flash. On the current agy 1.1.2 installation, that label is not in agy models; available Flash variants are Gemini 3.5 Flash (Low|Medium|High).

agy 1.1.2 changed headless --print behavior from silently falling back when a requested model cannot be resolved to exiting non-zero. This turns the stale repair-model name into a deterministic repair failure:

Error: invalid --model "Gemini 3 Flash": model Gemini 3 Flash is not recognized as a known model or custom model in settings

This blocked repair of an otherwise recoverable Claude PR review on llm-dialectic#358.

Required behavior

• Do not hard-code an obsolete Antigravity display label for repair.
• Make the repair model configurable, with a default compatible with the supported/current Antigravity model catalog.
• At startup or before repair, validate the selected model against agy models; emit an actionable error with available choices.
• On invalid/unavailable repair model, fall back through the configured Antigravity chain where possible.
• Add regression coverage for a stale model name and for headless agy non-zero model-resolution failure.

Immediate operational workaround

Use Gemini 3.5 Flash (Medium) as the local repair default until the configurable/fallback behavior lands.

[wwind123]

Plan

Make Antigravity structured-response repair use a current configurable default, validate repair candidates against the local agy catalog, and fall back safely when a model cannot be resolved.

Plan steps

1. Update the repair-model defaults and CLI/README help in src/coding_review_agent_loop/config.py, src/coding_review_agent_loop/cli.py, and README.md to replace the obsolete Gemini 3 Flash default with Gemini 3.5 Flash (Medium), while preserving repeated --repair-model as an explicit ordered override chain.
2. Add an Antigravity catalog-query and model-name validation path around agy models (using the configured agy executable and a parseable catalog result), with sanitized actionable diagnostics that identify rejected repair names and list the available choices; define behavior for a catalog command failure or unparseable output so repair is not silently treated as a successful validation.
3. Integrate catalog validation into execute_repair/Antigravity repair invocation so unavailable configured repair models are recorded as failed repair attempts and skipped, then continue through later explicit repair models and, where the explicit chain is exhausted or invalid, eligible configured antigravity_models without retrying duplicates; retain existing usage/logging metadata and distinguish validation/model-resolution failures from successful repair output.
4. Keep the legacy Gemini repair backend unchanged, and preserve repair isolation, no-tools settings injection, timeouts, output schema validation, and the existing nonzero/empty/invalid-output fallback behavior for actual agy --print executions.
5. Extend tests in tests/test_repair.py and Antigravity/config/CLI coverage to assert the new default and documented flag behavior, valid catalog selection, stale Gemini 3 Flash rejection with available-model guidance, fallback to the next repair or Antigravity-chain model, catalog-query failure handling, and an agy --print nonzero invalid-model-resolution response that advances the chain and records the correct attempt/usage outcomes. Run the focused repair/Antigravity/config tests, then the full pytest suite.

-- OpenAI Codex: gpt-5.6-terra

[wwind123]

Review verdict: Blocking

The plan correctly replaces the stale default, adds catalog validation, and extends the repair fallback chain into antigravity_models, but leaves the catalog-query integration semantics (failure handling and query frequency) unspecified, which risks regressing repair reliability, and it misses a second doc file that states the invariant the plan is reversing.

Blocking plan issues

• Steps 2 and 3 don't specify how a catalog-query ('agy models') failure or unparseable-output case interacts with the repair fallback chain. The plan only says the model must 'not be silently treated as a successful validation' on catalog failure, but doesn't say whether the candidate model is then skipped outright or still attempted directly (falling through to the existing nonzero-exit/invalid-model reactive handling in execute_repair). If a transient or auth failure querying the catalog causes the candidate to be skipped outright, a perfectly valid configured repair model could be discarded purely because the separate agy models call failed — making the repair pass, whose entire purpose is reliability, less robust than it is today (where a bad model just produces a nonzero_exit outcome and the loop moves to the next candidate). The plan must specify that catalog-query failures fall through to attempting the model directly rather than being treated as rejection.
• The plan doesn't specify how often agy models is queried: once per repair invocation/run, or once per candidate model as execute_repair iterates the explicit repair_models chain and then the antigravity_models chain. Querying the catalog per-candidate multiplies agy subprocess spawns, auth round-trips, and failure surface for what is supposed to stay a rare, lightweight recovery pass, and adds extra latency/timeout risk to every repair attempt. The plan should specify the catalog is fetched once (and cached) per repair invocation and reused to validate every candidate in the combined repair_models + antigravity_models chain.

Same-plan follow-ups

• Step 1 only lists README.md for stale 'Gemini 3 Flash' default-text updates, but docs/local_agent_loop.md:1180-1181 repeats the same stale default AND states 'the normal coder/reviewer fallback chain is not inherited' (mirrored in README.md:617-619) — an invariant that step 3 directly reverses by falling back into antigravity_models after the explicit repair chain is exhausted. Add docs/local_agent_loop.md to the doc-update scope and reconcile that sentence with the new fallback-into-antigravity_models behavior so the docs don't contradict the new code.

-- Anthropic Claude: claude-sonnet-5

[wwind123]

Revised plan

Revise Antigravity structured-response repair to use a current configurable default, validate the complete fallback chain from one catalog snapshot per repair invocation, and preserve direct execution when catalog discovery is unavailable.

Prior plan item dispositions

• [item-1] Blocking issue from Anthropic Claude: claude-sonnet-5, round 1: Steps 2 and 3 don't specify how a catalog-query ('agy models') failure or unparseable-output case... -> resolved: The revised plan explicitly treats a failed, timed-out, nonzero, or unparseable agy models query as unknown catalog state: it emits a sanitized diagnostic but attempts every combined-chain candidate directly, preserving the existing reactive nonzero/invalid-output fallback behavior rather than rejecting a potentially valid model.
• [item-2] Blocking issue from Anthropic Claude: claude-sonnet-5, round 1: The plan doesn't specify how often agy models is queried: once per repair invocation/run, or on... -> resolved: The revised plan fetches and parses agy models exactly once at the start of each Antigravity execute_repair invocation, caches that result for the invocation, and reuses it for every candidate in the de-duplicated combined chain.
• [item-3] Same-plan follow-up from Anthropic Claude: claude-sonnet-5, round 1: Step 1 only lists README.md for stale 'Gemini 3 Flash' default-text updates, but docs/local_agent... -> resolved: The documentation scope now includes both README.md and docs/local_agent_loop.md, and both descriptions will state that repair tries explicit --repair-model entries first, then remaining configured antigravity_models, rather than claiming the normal chain is never inherited.

Plan steps

1. Update src/coding_review_agent_loop/config.py and src/coding_review_agent_loop/cli.py so DEFAULT_REPAIR_MODELS is Gemini 3.5 Flash (Medium), repeated --repair-model remains an ordered explicit repair prefix, and help/config coverage derives the displayed default from that constant. Preserve the legacy gemini repair backend and its model semantics unchanged.
2. In src/coding_review_agent_loop/repair.py (with a narrowly scoped Antigravity helper if needed), construct the Antigravity repair candidate chain once per execute_repair call: configured repair_models followed by configured antigravity_models, preserving first-seen order and removing duplicate names. Before iterating it, invoke the configured antigravity_cmd models once using the repair-safe execution context and bounded timeout, normalize ANSI/plain catalog output into model display names, and retain a single successful parsed catalog snapshot for all candidates in that repair call.
3. Define catalog-validation behavior precisely in the repair loop. When the one cached catalog is available, skip a candidate absent from it without running agy --print, record a distinct unavailable-model repair attempt and usage outcome, and include a sanitized actionable diagnostic naming the rejected model and the available catalog choices; continue to the next combined-chain candidate. When the catalog command times out, exits nonzero, cannot spawn, or yields no parseable model list, log/record a sanitized catalog-discovery diagnostic but classify catalog state as unknown, not as successful validation or candidate rejection: directly attempt each candidate with the existing agy --print path and let its nonzero, empty, timeout, spawn, and schema-validation outcomes drive fallback. Do not issue another catalog query for later candidates.
4. Keep actual Antigravity repair attempts isolated in their temporary no-tools workdir with settings injection, configured executable/arguments, per-attempt timeout, bounded logs, output validation, and usage metadata intact. Ensure fallback flags and final diagnostics reflect the de-duplicated combined chain; in particular, a headless agy --print invalid-model-resolution nonzero exit remains a failed attempted invocation and advances to the next candidate, while a catalog-confirmed unavailable model is distinguishable from that reactive execution failure.
5. Update README.md and docs/local_agent_loop.md to replace Gemini 3 Flash with Gemini 3.5 Flash (Medium), document the repeatable explicit repair prefix plus subsequent eligible antigravity_models fallback, catalog validation and actionable available-model guidance, and the availability-unknown direct-attempt behavior. Retain the repair isolation and legacy-Gemini documentation, but remove or reconcile statements that say the normal Antigravity chain is never inherited.
6. Extend focused coverage in tests/test_repair.py, tests/test_antigravity.py, and applicable config/CLI tests: assert the new default/help text; successful catalog parsing is queried once and reused across a duplicate-free explicit-plus-Antigravity chain; stale Gemini 3 Flash is skipped with available-choice guidance and falls through; valid explicit candidates work; and exhausted explicit candidates reach antigravity_models. Add catalog timeout/nonzero/spawn/unparseable cases proving direct attempts still occur and no per-candidate catalog subprocesses are made. Add a scripted headless agy --print nonzero invalid-model-resolution case proving it is recorded separately from catalog rejection, retains usage/log metadata, and advances to a succeeding fallback. Run the focused repair/Antigravity/config tests, then python -m pytest.

-- OpenAI Codex: gpt-5.6-terra

[wwind123]

Review verdict: Blocking

Round 2 correctly resolves all three round-1 blocking/same-plan items (catalog-failure fallback semantics, single cached catalog query per repair invocation, and doc reconciliation across both files), but leaves the agy models invocation mechanism unspecified in a way that risks recreating this same issue's failure mode: this CLI is documented (antigravity.py:8-10, upstream antigravity-cli issue #76) to silently drop agy --print output when stdout isn't a TTY, worked around today via use_pty=True; if agy models has the same TTY-detection behavior and the catalog query is invoked as a bare, non-PTY subprocess call, it would silently return no parseable output every time in real usage, permanently and invisibly falling back to the safe 'unknown catalog state' direct-attempt path. That would quietly neuter the proactive validation and actionable available-model diagnostics this issue exists to add, without any test catching it (mocked subprocess output in unit tests would not surface a real TTY-detection quirk).

Same-plan follow-ups

• Step 2 should explicitly specify whether the agy models catalog query reuses the existing PTY-based execution path (use_pty=True, the same workaround already applied to agy --print for the documented TTY-detection quirk in antigravity-cli issue Add plan-first issue review stage #76) rather than a bare subprocess call, and step 6's tests should include a case that would fail if the catalog call were run without a PTY in an environment where agy exhibits that quirk (or at minimum assert the same execution helper/flag is used for both agy models and agy --print).

Prior unresolved plan item dispositions

• [item-1] Blocking issue from Anthropic Claude: claude-sonnet-5, round 1: Steps 2 and 3 don't specify how a catalog-query ('agy models') failure or unparseable-output case... -> resolved: Step 3 now explicitly classifies catalog-query failure/timeout/nonzero/unparseable output as 'unknown catalog state' and requires directly attempting each candidate via the existing agy --print reactive path rather than rejecting it outright.
• [item-2] Blocking issue from Anthropic Claude: claude-sonnet-5, round 1: The plan doesn't specify how often agy models is queried: once per repair invocation/run, or on... -> resolved: Step 2 now fetches and parses the catalog exactly once per execute_repair invocation, caches it, and reuses it for every candidate in the combined chain — step 6 adds a test asserting no per-candidate catalog subprocesses are made.
• [item-3] Same-plan follow-up from Anthropic Claude: claude-sonnet-5, round 1: Step 1 only lists README.md for stale 'Gemini 3 Flash' default-text updates, but docs/local_agent... -> resolved: Step 5 now updates both README.md and docs/local_agent_loop.md and explicitly reconciles the 'chain is never inherited' language with the new explicit-then-antigravity_models fallback behavior.

-- Anthropic Claude: claude-sonnet-5

[wwind123]

Revised plan

Updated the repair-model plan so its one-per-invocation Antigravity catalog query uses the same PTY-safe runner path as agy --print, preserving proactive validation in TTY-sensitive agy environments.

Prior plan item dispositions

• [item-4] Same-plan follow-up from Anthropic Claude: claude-sonnet-5, round 2: Step 2 should explicitly specify whether the agy models catalog query reuses the existing PTY-b... -> resolved: The catalog command is now explicitly required to use Runner.run_with_log(..., use_pty=True)—the same PTY-enabled helper and flag used by Antigravity agy --print—and tests will assert that invariant.

Plan steps

1. Update src/coding_review_agent_loop/config.py and src/coding_review_agent_loop/cli.py so DEFAULT_REPAIR_MODELS becomes Gemini 3.5 Flash (Medium), repeatable --repair-model values remain an ordered explicit repair prefix, and displayed defaults derive from the constant. Leave the legacy gemini repair backend and its model semantics unchanged.
2. In src/coding_review_agent_loop/repair.py and src/coding_review_agent_loop/agents/antigravity.py, build a first-seen de-duplicated Antigravity repair chain once per execute_repair call: repair_models followed by antigravity_models. Fetch and parse agy models exactly once before that loop, using the configured executable and applicable arguments through Runner.run_with_log with the Antigravity repair-safe cwd, bounded discovery timeout, logging, and use_pty=True. Reuse the same PTY-enabled runner helper/flag as agy --print, because agy’s TTY-sensitive output behavior would otherwise make catalog discovery silently empty.
3. Define catalog behavior for the cached snapshot. If a catalog is successfully parsed, skip any absent candidate without issuing agy --print, record a distinct unavailable-model attempt/outcome, and emit a sanitized actionable message with the rejected name and available choices before continuing. If catalog discovery cannot spawn, times out, exits nonzero, or is unparseable/empty, record a sanitized discovery diagnostic and treat availability as unknown: directly attempt every candidate through the existing PTY-enabled agy --print path rather than rejecting it or re-querying the catalog.
4. Preserve each actual Antigravity repair attempt’s temporary no-tools workdir, settings injection, configured executable/arguments, per-attempt timeout, bounded logs, output validation, and usage metadata. Make fallback status and final diagnostics reflect the de-duplicated combined chain. A headless agy --print invalid-model-resolution nonzero exit must remain a recorded attempted invocation, distinct from a catalog-confirmed rejection, and advance to the next candidate.
5. Update README.md and docs/local_agent_loop.md to replace Gemini 3 Flash with Gemini 3.5 Flash (Medium) and describe the explicit repair prefix followed by remaining antigravity_models, catalog validation with available-choice guidance, and direct attempts when discovery is unavailable. Reconcile the prior statement that the regular Antigravity chain is never inherited while retaining repair isolation and legacy-Gemini documentation.
6. Extend focused tests in tests/test_repair.py, tests/test_antigravity.py, and relevant config/CLI tests. Cover the new default/help text; a single catalog lookup reused across the duplicate-free combined chain; stale Gemini 3 Flash rejection with available choices; valid explicit candidates; and fallback into antigravity_models. Add timeout/nonzero/spawn/unparseable catalog cases proving direct attempts still occur with no per-candidate lookup. Add a TTY-sensitive fake runner case that emits catalog output only when use_pty=True, and assert both agy models and agy --print calls use that same flag/helper; this must fail if discovery regresses to a bare subprocess or non-PTY call. Cover a headless invalid-model --print nonzero that retains attempt usage/log metadata and falls through. Run focused repair/Antigravity/config tests, then python -m pytest.

-- OpenAI Codex: gpt-5.6-terra

[wwind123]

Review verdict: Approved

Round 3 resolves the outstanding PTY concern by requiring the agy models catalog query to use Runner.run_with_log(..., use_pty=True), the same real, existing helper/flag already used for agy --print, plus a regression test that fails if discovery regresses to a non-PTY call. All prior blocking and same-plan items are now resolved and the plan is architecturally sound, correctly scoped to a single PR, and has adequate test coverage for the required regressions.

Prior unresolved plan item dispositions

• [item-4] Same-plan follow-up from Anthropic Claude: claude-sonnet-5, round 2: Step 2 should explicitly specify whether the agy models catalog query reuses the existing PTY-b... -> resolved: Step 2 now explicitly requires Runner.run_with_log(..., use_pty=True) for the agy models catalog query, matching the same PTY-enabled path used for agy --print, and step 6 adds a TTY-sensitive fake runner test that fails if discovery regresses to a bare/non-PTY call.

-- Anthropic Claude: claude-sonnet-5

[wwind123]

Planning complete for issue #562.

Outcome: implement

Approved plan:

Updated the repair-model plan so its one-per-invocation Antigravity catalog query uses the same PTY-safe runner path as agy --print, preserving proactive validation in TTY-sensitive agy environments.

Prior plan item dispositions

• [item-4] Same-plan follow-up from Anthropic Claude: claude-sonnet-5, round 2: Step 2 should explicitly specify whether the agy models catalog query reuses the existing PTY-b... -> resolved: The catalog command is now explicitly required to use Runner.run_with_log(..., use_pty=True)—the same PTY-enabled helper and flag used by Antigravity agy --print—and tests will assert that invariant.

Plan steps

1. Update src/coding_review_agent_loop/config.py and src/coding_review_agent_loop/cli.py so DEFAULT_REPAIR_MODELS becomes Gemini 3.5 Flash (Medium), repeatable --repair-model values remain an ordered explicit repair prefix, and displayed defaults derive from the constant. Leave the legacy gemini repair backend and its model semantics unchanged.
2. In src/coding_review_agent_loop/repair.py and src/coding_review_agent_loop/agents/antigravity.py, build a first-seen de-duplicated Antigravity repair chain once per execute_repair call: repair_models followed by antigravity_models. Fetch and parse agy models exactly once before that loop, using the configured executable and applicable arguments through Runner.run_with_log with the Antigravity repair-safe cwd, bounded discovery timeout, logging, and use_pty=True. Reuse the same PTY-enabled runner helper/flag as agy --print, because agy’s TTY-sensitive output behavior would otherwise make catalog discovery silently empty.
3. Define catalog behavior for the cached snapshot. If a catalog is successfully parsed, skip any absent candidate without issuing agy --print, record a distinct unavailable-model attempt/outcome, and emit a sanitized actionable message with the rejected name and available choices before continuing. If catalog discovery cannot spawn, times out, exits nonzero, or is unparseable/empty, record a sanitized discovery diagnostic and treat availability as unknown: directly attempt every candidate through the existing PTY-enabled agy --print path rather than rejecting it or re-querying the catalog.
4. Preserve each actual Antigravity repair attempt’s temporary no-tools workdir, settings injection, configured executable/arguments, per-attempt timeout, bounded logs, output validation, and usage metadata. Make fallback status and final diagnostics reflect the de-duplicated combined chain. A headless agy --print invalid-model-resolution nonzero exit must remain a recorded attempted invocation, distinct from a catalog-confirmed rejection, and advance to the next candidate.
5. Update README.md and docs/local_agent_loop.md to replace Gemini 3 Flash with Gemini 3.5 Flash (Medium) and describe the explicit repair prefix followed by remaining antigravity_models, catalog validation with available-choice guidance, and direct attempts when discovery is unavailable. Reconcile the prior statement that the regular Antigravity chain is never inherited while retaining repair isolation and legacy-Gemini documentation.
6. Extend focused tests in tests/test_repair.py, tests/test_antigravity.py, and relevant config/CLI tests. Cover the new default/help text; a single catalog lookup reused across the duplicate-free combined chain; stale Gemini 3 Flash rejection with available choices; valid explicit candidates; and fallback into antigravity_models. Add timeout/nonzero/spawn/unparseable catalog cases proving direct attempts still occur with no per-candidate lookup. Add a TTY-sensitive fake runner case that emits catalog output only when use_pty=True, and assert both agy models and agy --print calls use that same flag/helper; this must fail if discovery regresses to a bare subprocess or non-PTY call. Cover a headless invalid-model --print nonzero that retains attempt usage/log metadata and falls through. Run focused repair/Antigravity/config tests, then python -m pytest.

-- coding-review-agent-loop

[wwind123]

Approved plan for issue #562 handed off to PR #563 for one-shot implementation.

Mode: implement-one-shot
Plan hash: dff56639e1fbbe5c
Plan subject: dff56639e1fbbe5c39582b5275d4460ea65f1fd9747fcb02701bbc2269569955

Parent reruns will resume the PR review loop for this PR instead of re-implementing.

-- coding-review-agent-loop

[wwind123]

Review verdict: Approved

The plan addresses all requirements. The catalog query logic is appropriately hardened and specifies PTY execution identical to the agy --print path, and corresponding tests are added.

Prior unresolved plan item dispositions

• [item-4] Same-plan follow-up from Anthropic Claude: claude-sonnet-5, round 2: Step 2 should explicitly specify whether the agy models catalog query reuses the existing PTY-b... -> resolved: Plan now explicitly requires use_pty=True for the catalog query and adds tests to assert this invariant for both agy models and agy --print.

-- Google Antigravity: Gemini 3.1 Pro (High)

[wwind123]

Follow-up: agy models progress output causes a false unavailable-model result

After #563, a repair run attempted the configured Gemini 3.5 Flash (Medium) default. The availability preflight rejected it even though its own diagnostic displayed the name at the end of the model-list output:

Repair model 'Gemini 3.5 Flash (Medium)' is not available in agy models.
Available choices: ... Gemini 3.5 Flash (Low),
⠋ Fetching available models...⠙ Fetching available models...
...⠏ Fetching available models...Gemini 3.5 Flash (Medium)

The animated progress prefix is likely concatenated with the final model name. An exact line-match parser therefore fails to recognize Medium, while the same raw output is echoed misleadingly as an available choice. The fallback chain then correctly proceeded to Flash High and Pro High.

Additional requirements

• Normalize terminal control sequences, carriage-return/progress frames, and transient status text before parsing agy models.
• Prefer a documented machine-readable or noninteractive model-list mode if agy provides one.
• Do not include unparsed/progress-corrupted lines in the displayed available-model choices.
• Add a regression test where spinner/progress text is adjacent to a valid model line; the model must be recognized.
• Keep the current fallback behavior when a model is genuinely unavailable.

This is distinct from the separate repair-schema failure in #558/#561: in this run, Flash High and Pro High executed successfully but both preserved Codex's fabricated signed-human-requirement IDs.