chore(ai-review): adopt pr-reviewer v1.3.0 native_loop + org-var tokens

[joryirving]

Canary for the misospace v1.3.0 fanout.

Changes

• Pin pr-reviewer-action v1.2.9 → v1.3.0 (3479bf5).
• tool_mode plan_execute_loop → native_loop — the v1.3.0 flagship multi-hop tool-calling harness (the reviewer gathers its own evidence: read files, grep, fetch, with an in-conversation verdict). Requires v1.3.0 for streaming + the verdict turn.
• ai_max_tokens → ${{ vars.AI_MAX_TOKENS || '16000' }} — org-wide variable so token budget is tuned once across misospace (16000 fallback until the org var is set).
• Thinking-model loop budgets: tool_planning_max_tokens 1500 → 16000 (1500 truncates a reasoning model on every loop turn → empty content → spurious escalation), tool_planning_timeout_sec 45 → 300 (longer streamed turns), and new tool_loop_wall_clock_sec 300 (default 120 is short for the slow self-hosted reviewer).

Why a canary

native_loop on the self-hosted Gemma reviewer is slower and more token-hungry per review than the 2-round planner. This validates the config (no truncation/escalation, reviews land clean) on one repo before fanning out to KubeTix / miso-chat / miso-gallery / windowstead.

pull_request runs the head workflow, so this PR self-validates the new config.

🤖 Generated with Claude Code

[its-saffron]

AI Automated Review

Full PR review.

Analysis engine: MiniMax-M2.7@https://litellm.jory.dev/v1 (anthropic) — escalated (fast_low_confidence)

Review: PR PR 374 — Adopt pr-reviewer v1.3.0 native_loop + org-var tokens

Summary

This is a canary PR that upgrades the AI PR review workflow to use pr-reviewer-action v1.3.0 with the new native_loop tool mode, org-wide token variable (AI_MAX_TOKENS), and relaxed timing budgets. The change is scoped to one file (.github/workflows/ai-pr-review.yaml) and introduces no code changes.

Change-by-Change Findings

Change	Assessment
Pin bump: @4126a27 → @3479bf5 (v1.2.9 → v1.3.0)	Confirmed valid via GitHub Releases API — v1.3.0 published 2026-06-15 by PR author.
tool_mode: plan_execute_loop → native_loop	Backward-compatible per upstream release notes: "tool_mode: native_loop is opt-in, and existing plan_execute_loop / off consumers are unaffected."
ai_max_tokens: new `${{ vars.AI_MAX_TOKENS
tool_planning_timeout_sec: 45 → 300	Rationale documented in PR body: longer for streamed turns on self-hosted reviewer.
tool_planning_max_tokens: 1500 → 16000	Rationale: "1500 truncates a reasoning model on every loop turn → empty content → spurious escalation."
tool_loop_wall_clock_sec: new 300	New parameter added by v1.3.0; default 120 is short for the slow self-hosted reviewer.

Standards Compliance

• No secrets committed: Uses ${{ secrets.LITELLM_API_KEY }} and ${{ vars.* }} appropriately — no credentials in diff.
• API routes / workflow config: N/A — this is a GitHub Actions workflow file, not an API route.
• Error handling / validation: N/A — configuration-only change.
• Prisma schema: N/A — no database changes.
• Release process: This PR is itself the canary validation step before a broader fan-out. If CI passes, the pattern is validated for other repos in the org.

Linked Issue Fit

No linked issues present in the PR body.

Evidence Provider Findings

Tool harness confirmed:

• misospace/pr-reviewer-action v1.3.0 release exists and is tagged at commit 3479bf5.
• Release title: "Native tool-calling review harness" — the native_loop mode allows the model to chain tool calls natively.
• Release is additive and backward-compatible — no breaking changes to existing plan_execute_loop or off consumers.

Tool Harness Findings

Both API calls succeeded:

• GET /repos/misospace/pr-reviewer-action/releases — returned v1.3.0 as latest.
• GET /repos/misospace/pr-reviewer-action/releases/tags/v1.3.0 — confirmed tag at commit 3479bf5 matching the pinned SHA.

CI Check Results

• Docker Build: ✅ success
• Validate: ✅ success

Both required checks passed for commit 88608fb1.

Unknowns / Needs Verification

None identified. The PR self-validates by running the updated workflow on its own pull_request event, which is a valid canary pattern.

Recommendation

Approve. The version bump is verified against the upstream release tag. The native_loop mode is opt-in and backward-compatible per upstream release notes. All configuration changes are internally consistent and use the correct secret/variable interpolation syntax. CI has passed. The canary approach is sound — this PR will validate the new config before fanning out to other repos.