chore: update native loop settings

[itsmiso-ai]

Updates the PR reviewer native-loop settings for v2.4.0 and replaces deprecated planner input names.

[its-saffron]

AI Automated Review

Full PR review.

Analysis engine: MiniMax-M3@https://litellm.jory.dev/v1 (anthropic) — escalated (fast_low_confidence)

Recommendation

Request changes. The PR updates native-loop configuration values, but two of the five renamed settings are not verified against the action's documented input schema. The action pinned in the workflow (misospace/pr-reviewer-action@acb9476d7358801d854f9c57366bbfbef401acd5 # v2.4.0) was bumped to v2.4.0 in PR PR 337 immediately before this change. The PR body claims it "replaces deprecated planner input names," but I cannot confirm that tool_turn_timeout_sec, tool_corpus_max_bytes, and tool_max_tokens_per_turn are accepted by the v2.4.0 action — there is no schema in the corpus and the prior names (tool_planning_timeout_sec, tool_planning_max_context_bytes, tool_planning_max_tokens) that were removed may or may not still be recognized. If even one of the new keys is rejected, the workflow will fail at runtime with an input-validation error on every PR.

Change-by-change findings

• tool_max_rounds: "2" → "4" — Doubled. Consistent with giving the reviewer more chances to gather evidence. No concern.
• tool_max_requests: "4" → "8" — Doubled. Consistent with the round bump. No concern.
• tool_loop_wall_clock_sec: "300" → "600" — Doubled. Matches the new ceiling implied by the round/requests bump. No concern.
• tool_planning_timeout_sec: "300" → tool_turn_timeout_sec: "300" — Rename, value unchanged. This is presented in the PR body as replacing a deprecated name. I cannot verify from the corpus that the v2.4.0 action accepts tool_turn_timeout_sec as the new name. Repository history shows the action was bumped v2.3.3 → v2.4.0 in commit 1e17e7f (PR 337) just before this PR; no action changelog or input schema is available in the corpus. Needs verification.
• tool_planning_max_context_bytes: "15000" → tool_corpus_max_bytes: "50000" — Rename and substantial bump (3.3×). The intent here is plausible (larger corpus for the escalated-review case this PR is itself triggering) but the new name and the much higher cap both depend on the v2.4.0 schema accepting tool_corpus_max_bytes. Needs verification.
• tool_planning_max_tokens: "16000" → tool_max_tokens_per_turn: "16000" — Rename, value unchanged. Same unverifiable-name concern as above. Needs verification.
• tool_max_response_bytes: "12000" — Unchanged. Good — this setting was not touched by the rename sweep, which is consistent with it not being part of the planner family.

Sources

• PR diff: .github/workflows/ai-pr-review.yaml, lines 63–73 (the six renames/bumps live in the tools: block of the Review PR with reusable AI reviewer step).
• PR body: "Updates the PR reviewer native-loop settings for v2.4.0 and replaces deprecated planner input names."
• Repository history, commit 1e17e7f: "ci(github-action): update action misospace/pr-reviewer-action (v2.3.3 → v2.4.0) (PR 337)" — confirms the action is on v2.4.0 and the new keys must be valid against that version.
• Workflow file pins the action to commit acb9476d7358801d854f9c57366bbfbef401acd5 # v2.4.0.
• AGENTS.md does not cover reviewer-action input naming; it documents the autonomous-loop issue contract, which is not the surface this PR touches.

Standards Compliance

AGENTS.md governs how issues are filed for the autonomous coding loop (one imperative sentence + named file paths). It does not govern reviewer-action configuration. Nothing in AGENTS.md is violated by this diff. The PR touches exactly one workflow file and the change is mechanically scoped to the tools: input block — there is no scope drift. Standard-compliance check: not applicable / satisfied.

Tool Harness Findings

The native tool-calling loop issued no tool calls; the harness reviewed the corpus directly. There are no provider findings to report or rebut.

Unknowns or Needs Verification

1. Action input schema for v2.4.0 — Confirm that tool_turn_timeout_sec, tool_corpus_max_bytes, and tool_max_tokens_per_turn are recognized inputs on misospace/pr-reviewer-action@acb9476d7358801d854f9c57366bbfbef401acd5 (v2.4.0), and that the old tool_planning_* names are no longer accepted. The action's action.yml or release notes are not in the corpus; the author must point at them.
2. tool_corpus_max_bytes: 50000 — The 3.3× jump from the prior cap (15000) is large. If the schema accepts the new name, confirm this cap is intentional and not a transcription error (e.g., meant to be 25000 to mirror a typical 2× bump).
3. CI green does not prove schema validity — The test and docker CI checks succeeded, but those are unrelated to the reviewer workflow's runtime input validation. A malformed tools: block typically only surfaces when the reviewer workflow actually runs against a PR.

Given that this PR is itself the escalation case the new budget is meant to handle, the right move is to land the rename on its own first (smallest possible diff to confirm the schema), and then bump the budgets in a follow-up. As written, a single bad key name will take the reviewer workflow down on every PR until reverted.