ci(github-action)!: Update action misospace/pr-reviewer-action (v1.3.0 → v2.0.0)

[its-miso]

This PR contains the following updates:

Package	Type	Update	Change
misospace/pr-reviewer-action	action	major	v1.3.0 → v2.0.0

---

Release Notes

misospace/pr-reviewer-action (misospace/pr-reviewer-action)

v2.0.0: — Consolidation, prefill diet, Forgejo complete

Compare Source

A breaking, semver-honest release. This is the post-1.3 tech-debt audit (#​258): the high-value cleanups change the input and tool-mode surface, so they land as 2.0 rather than 1.4. There are no new review capabilities over v1.3.1 — this release is consolidation, a real prefill-token diet, and finished Forgejo support.

Breaking changes

1. plan_execute tool modes removed (#​326). tool_mode now accepts only off or native_loop. A stale plan_execute_once / plan_execute_loop value degrades to off with a warning — the review still runs (corpus-only), it just won't gather evidence. native_loop has been the only mode any consumer runs since 1.3.0, so this breaks no live deployment. Migration: set tool_mode: native_loop.

2. ai_fast_* inputs removed (#​329). The deprecated ai_fast_model / ai_fast_base_url / ai_fast_api_format / ai_fast_api_key aliases — superseded by ai_primary_* in 1.3.1 — are gone. Migration: rename them to ai_primary_*, or just let the primary resolve from ai_model.

3. Publish steps collapsed into one dispatcher (#​328). Internal restructure; inputs and outputs are unchanged, so most consumers need to do nothing. Flagged breaking out of caution because the action's internal step layout changed. As a bonus it structurally kills the duplicate-env:-key drift that broke v1.2.10 — there's only one publish env: block left to drift.

Performance — prefill-token diet

native_loop is prefill-dominated with low cache reuse, so every token trimmed from the per-round corpus compounds across rounds:

• Corpus-bound JSON compacted (#​337)
• gh_api + web_search tool-result JSON compacted (#​338)
• Conditional system-prompt assembly by PR type — host-platform, image-digest, and release-notes guidance now prefill only when the PR's classification needs them, instead of every round on every PR (#​339, #​334)
• system_prompt_mode: append — a repo can add its conventions on top of the live bundled default instead of copying and re-syncing the whole prompt (#​340)

Forgejo — support complete

• compare backend (#​280), native review publishing (#​291), gh_api routed through the platform seam (#​294), and an end-to-end smoke harness (#​293)
• Golden fixtures recorded from a real forge for every seam normalizer (#​341, finishing #​252) — these guard the field-shape bug class that bit before (e.g. the per-entry commit-status status vs state regression, #​244)

Internals and quality

• run_review.sh split into sourced section modules (#​324); transport and tool executors extracted from the tool harness into their own modules (#​325); dead code pruned (#​331)
• Cache-hit-ratio surfaced in the step summary, outputs, and metadata marker (#​335)
• Per-file attribution for classifier risk flags (#​322); a warning when an evidence provider is configured as a shell string (#​320)
• Test infra: shared bash assertion helpers (#​323), an expanded agentic eval corpus beyond the single Talos scenario (#​321), a mid-flight wall-clock assertion (#​318)
• CI: actions/checkout bumped to v7 (#​316)

Upgrading

Pin to v2.0.0 or the floating v2 tag. The v1 tag is not moved to this release — it stays on the 1.3.x line, so anything pinned to @v1 is unaffected and adopts v2 only when you bump it deliberately. Before upgrading, set tool_mode: native_loop and rename any ai_fast_* inputs.

Full changelog: misospace/pr-reviewer-action@v1.3.1...v2.0.0

v2

Compare Source

v1.3.1: — Routing fix: the primary model gets its full tool budget

Compare Source

v1.3.1 — Routing fix: the primary model gets its full tool budget

A bug-fix release for review_routing_mode: auto. The auto router was sending low-risk PRs (e.g. a renovate dependency bump) to a shallow lane that capped the loop's tool budget — starving exactly the multi-hop chains (read the changed config → search → fetch the compatibility matrix → cite) that the v1.3.0 native loop was built to run. All changes are backward-compatible via deprecated aliases.

Fixes

• The primary route gets the full configured tool budget. Removed the route-based shallow cap — the loop self-limits, so the cap only ever starved chain-needing PRs. The route now selects the model, not the budget (#​278).
• The smart tier only escalates when explicitly configured. ai_smart_* no longer falls back to ai_fallback_*: with no ai_smart_model set, a review stays on the primary model instead of silently escalating to the fallback. The fallback model is for availability, not quality escalation (#​278).

Rename (backward-compatible)

• The fast route is now primary — a capable primary model shouldn't be rationed like a "fast/dumb" lane. Inputs ai_fast_* → ai_primary_*, with ai_fast_* retained as deprecated aliases, so no workflow changes are needed to upgrade. The review_route output and engine annotation now read primary; the per-route artifact is ai-output.primary.json (#​278).

---

Configuration

📅 Schedule: (in timezone America/Edmonton)

• Branch creation
  • At any time (no schedule defined)
• Automerge
  • At any time (no schedule defined)

🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.

♻ Rebasing: Whenever PR is behind base branch, or you tick the rebase/retry checkbox.

🔕 Ignore: Close this PR and you won't be reminded about this update again.

---

• If you want to rebase/retry this PR, check this box

---

This PR has been generated by Mend Renovate.

[its-saffron]

AI Automated Review

Full PR review.

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

Recommendation: Approve

This is a straightforward dependency update for the AI PR review GitHub Action, moving from v1.3.0 to v2.0.0. The change is low-risk and already validated by CI.

Change Summary

.github/workflows/ai-pr-review.yaml (1 line changed):

• Action pin updated from commit 3479bf51ab9133414ae07af783bff985d3929884 (v1.3.0) to 7e25f7d0d8651e936f3158477b0a364f27ed7383 (v2.0.0)

Breaking Change Compatibility Check

The v2.0.0 release has three breaking changes:

Breaking Change	Status in This Repo
plan_execute tool modes removed → only off or native_loop	✅ Workflow uses tool_mode: native_loop — already compliant
ai_fast_* inputs removed → use ai_primary_*	✅ Workflow uses ai_primary_* inputs — no deprecated aliases
Publish steps internal restructure	✅ No action consumer changes needed (inputs/outputs unchanged)

The repository adopted native_loop and ai_primary_* in a previous update (commit e259008, PR PR 211), so this upgrade is fully backward-compatible.

Standards Compliance

• Collaboration: Push to main is permitted per AGENTS.md. No release needed for dependency updates.
• Repository conventions: This is a CI infrastructure change; standard GitHub Actions workflow file. No deviation from repository patterns.

Linked Issue Fit

No linked issues. This is a routine dependency maintenance PR.

Evidence Provider Findings

No evidence providers configured for this workflow.

Tool Harness Findings

The tool harness completed successfully (3 rounds, tool-call budget exhausted). The gh_api calls to the misospace/pr-reviewer-action repo were blocked by repo permissions, but this is expected for external repos and does not affect the review quality — release notes and compare metadata were fetched successfully.

CI Check Results

All CI checks passed (7 checks, all completed with success):

• review (x3) — the v2.0.0 action ran on this PR itself
• macOS validation
• Headless smoke test
• Script test suite

Unknowns / Needs Verification

None. The diff is minimal (1 line), CI passed, and breaking change compatibility is confirmed from the workflow file content.

Findings

null