Optimize review: small-PR and incremental fast paths by don-petry · Pull Request #4 · petry-projects/.github-private

don-petry · 2026-04-10T11:52:23Z

Summary

Small PRs (<10 lines): single Opus call instead of full 3-member council (1 invocation vs 4)
Re-reviews (prior marker exists): single Opus call with prior review context, focused on delta
First review of non-small PRs: full council (unchanged)
New config: SMALL_PR_THRESHOLD repo variable (default 10)

Cost impact

Small docs PRs (like .github#101): ~1 min instead of ~5 min, 1 invocation instead of 4
Re-reviews after Claude fixes: ~1 min focused check instead of full re-analysis
Full council: unchanged for substantive first reviews

Test plan

Let the automation review and merge this PR
Verify mode detection in logs (mode: small|incremental|full)
Verify small PR uses single Opus path
Verify re-review after new commits uses incremental path

🤖 Generated with Claude Code

PRs are now routed to the cheapest review path: - Small PRs (<10 lines): single Opus call (1 invocation vs 4) - Re-reviews (prior marker at different SHA): single Opus call with prior review context, focused on what changed since last review - First review of non-small PRs: full 3-member council (unchanged) New files: - prompts/single-review.md: combined council+synth prompt for single reviewer mode, handles both small and incremental paths New config: - SMALL_PR_THRESHOLD repo variable (default 10) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Copilot

Pull request overview

This PR optimizes the PR-review agent’s orchestration path to reduce cost/latency by using a single-reviewer “fast path” for small PRs and for re-reviews (when a prior review marker exists), while keeping the existing full 3-member council flow for first reviews of larger PRs.

Changes:

Add review-mode detection (small|incremental|full) in the per-PR orchestrator script and short-circuit to a single Opus invocation for non-full modes.
Introduce a new prompts/single-review.md prompt that combines council + synthesizer behavior into one agent for small/incremental reviews.
Document and wire up SMALL_PR_THRESHOLD as a workflow/repo variable.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 5 comments.

File	Description
scripts/review-one-pr.sh	Adds mode selection and a single-reviewer execution path before the existing council workflow.
prompts/single-review.md	New prompt defining the single-reviewer behavior, including incremental re-review behavior using prior review context.
AGENT.md	Updates documentation to describe the new optimized review modes and the new threshold variable.
.github/workflows/pr-review.yml	Exposes `SMALL_PR_THRESHOLD` to the job environment with a default.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

+#    - "small"       : PR has < SMALL_PR_THRESHOLD lines changed → single Opus call
+#    - "incremental" : prior review exists at different SHA → single Opus call with prior context
+#    - "full"        : first review of a non-small PR → 3-member council + synthesizer
+SMALL_PR_THRESHOLD="${SMALL_PR_THRESHOLD:-10}"
+PR_SIZE=$(gh pr view "$PR_URL" --json additions,deletions --jq '.additions + .deletions')
+export PR_SIZE
+
+if [ "$PR_SIZE" -lt "$SMALL_PR_THRESHOLD" ]; then
+  REVIEW_MODE="small"
+elif [ -n "${EXISTING_MARKER_SHA:-}" ]; then
+  REVIEW_MODE="incremental"


+SMALL_PR_THRESHOLD="${SMALL_PR_THRESHOLD:-10}"
+PR_SIZE=$(gh pr view "$PR_URL" --json additions,deletions --jq '.additions + .deletions')
+export PR_SIZE
+
+if [ "$PR_SIZE" -lt "$SMALL_PR_THRESHOLD" ]; then
+  REVIEW_MODE="small"
+elif [ -n "${EXISTING_MARKER_SHA:-}" ]; then


+  # Extract the full body of the review/comment containing our marker for the prior SHA.
+  PRIOR_REVIEW_BODY=$(
+    gh pr view "$PR_URL" --json reviews,comments \
+      --jq "((.reviews // []) + (.comments // [])) | .[].body | select(. != null) | select(test(\"sha=$PRIOR_REVIEW_SHA\"))" 2>/dev/null \
+    | head -1 || true


+  # Extract the full body of the review/comment containing our marker for the prior SHA.
+  PRIOR_REVIEW_BODY=$(
+    gh pr view "$PR_URL" --json reviews,comments \
+      --jq "((.reviews // []) + (.comments // [])) | .[].body | select(. != null) | select(test(\"sha=$PRIOR_REVIEW_SHA\"))" 2>/dev/null \
+    | head -1 || true
+  )
+  export PRIOR_REVIEW_BODY
+  echo "    prior review SHA: $PRIOR_REVIEW_SHA (body: ${#PRIOR_REVIEW_BODY} chars)"


+   - If `isDraft` → skip. Print `{"pr":"...","decision":"skip","reason":"draft"}` and exit.
+   - Verify `headRefOid == $PR_HEAD_SHA`. If not → skip with `"reason":"head-sha-changed"`.
+2. `gh pr diff "$PR_URL"` — read the diff.
+   - **Incremental mode**: also get the diff since the prior review:


Haiku triage → Sonnet deep review → Opus security audit, where each tier only fires if the previous one escalated. This replaces the 3-parallel-members + synthesizer architecture. Cost impact: - ~80% of PRs: Haiku + Opus confirm (2 calls, ~30s vs 4 calls, ~5 min) - ~15%: + Sonnet (3 calls, ~2.5 min) - ~5%: full cascade (4 calls, ~5.5 min) Haiku works by receiving pre-fetched context (no tool use needed), which sidesteps the agentic limitations found earlier. New files: - prompts/triage.md — Tier 1: Haiku fast classification - prompts/deep-review.md — Tier 2: Sonnet full review - prompts/security-audit.md — Tier 3: Opus security audit - prompts/cascade-action.md — posts review from any resolving tier Re-reviews (prior marker at different SHA) pass prior review context to the triage, enabling faster incremental checks. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

don-petry

Automated review — NEEDS HUMAN REVIEW

Risk: MEDIUM
Reviewed commit: bab010f29438d21527bddf9c045733346b603b11
Council vote: security=MEDIUM · correctness=MEDIUM · maintainability=MEDIUM

Summary

The PR replaces the parallel 3-member review council with a cascading Haiku→Sonnet→Opus tier architecture, but the two commits contain two sequential and conflicting designs — the final HEAD is entirely undocumented in the PR description and test plan. Three functional bugs exist at HEAD: the repository JSON field is invalid for gh pr view (causing silent empty metadata), REVIEW_MODE="triage-approved" is unhandled in single-review.md, and PRIOR_REVIEW_BODY/PRIOR_REVIEW_SHA extraction uses inconsistent head/tail selectors. A dangling reference to the retired synthesize.md breaks the escalation delegation path in single-review mode.

Linked issue analysis

No linked issue found in PR metadata.

Findings

Major

[major] scripts/review-one-pr.sh:608 (correctness) — repository is not a valid gh pr view --json field; valid alternatives are headRepository/headRepositoryOwner. The || echo '{}' fallback silently gives downstream agents an empty metadata object, stripping all PR context and likely causing spurious escalations. Same invalid field appears in prompts/deep-review.md and prompts/security-audit.md.
[major] scripts/review-one-pr.sh:641 (correctness) — REVIEW_MODE="triage-approved" set by the orchestrator is not handled in prompts/single-review.md, which only documents small and incremental modes. The template renders a literal <small-pr|incremental> placeholder and incremental delta analysis is skipped.
[major] prompts/single-review.md:483 (correctness) — Dangling reference: instructs agents to follow synthesize.md step 9a, but synthesize.md is being retired by this PR in favour of cascade-action.md. The escalation delegation path is undefined in single-review mode.
[major] scripts/review-one-pr.sh:591 (correctness) — PRIOR_REVIEW_BODY is selected with head -1 (oldest match) while PRIOR_REVIEW_SHA reflects the most-recent marker. Under any multi-review history these point to different review cycles, giving agents stale body context paired with the current SHA. Should use tail -1 for the body selector.

Minor

[minor] scripts/review-one-pr.sh:89 (security) — PRIOR_REVIEW_BODY export creates a prompt-injection surface: any commenter could post a fake  marker followed by crafted content ingested as prior review context. Consider validating the full HTML comment structure in the match.
[minor] prompts/council/ (maintainability) — Orphaned prompt files (correctness.md, maintainability.md, security.md) are no longer referenced by the script but were not deleted.
[minor] scripts/review-one-pr.sh (maintainability) — SONNET_RC=$? captures the Sonnet exit code but is never referenced; the script checks the output file instead. Dead variable.
[minor] (correctness) — PR title and description describe the commit-1 architecture (small-PR threshold, incremental fast path); final HEAD replaced that with the cascade tier system. The test plan items do not match the final code.
[minor] (correctness) — No CI checks ran against HEAD (statusCheckRollup is empty). There is no automated test suite verifying the script logic or prompt correctness.

Info

[info] .github/workflows/pr-review.yml:48 (security · maintainability) — SMALL_PR_THRESHOLD is exposed in the workflow env: block (safe, admin-controlled) but the cascade script never reads it. Dead configuration from the first commit's approach.
[info] (security) — Haiku triage runs with --permission-mode plan (no tool access) — good least-privilege practice.
[info] (security) — jq filter interpolation of $PRIOR_REVIEW_SHA is safe; value is constrained to [a-f0-9]+.
[info] prompts/deep-review.md, prompts/security-audit.md (maintainability) — These prompts instruct agents to run gh pr view --json ...repository..., an invalid field. Agents following these prompts will encounter metadata fetch errors.
[info] (maintainability) — Tier-2 Sonnet failure exits hard (exit 1) without posting any review, while Haiku failure falls back gracefully to escalation. Consider symmetric fallback behaviour.
[info] (correctness) — 5 Copilot inline comments (COMMENTED state) remain unaddressed at HEAD, including the head -1 vs tail -1 mismatch and a repos/<owner>/<repo> placeholder in single-review.md.

CI status

No CI checks ran against HEAD commit bab010f29438d21527bddf9c045733346b603b11 — statusCheckRollup is empty.

Reviewed automatically by the don-petry PR-review council (security: opus 4.6 · correctness: sonnet 4.6 · maintainability: sonnet 4.6 · synthesis: sonnet 4.6). The marker on line 1 lets the agent detect new commits and re-review. Reply with @don-petry if you need a human.

don-petry · 2026-04-10T15:58:48Z

Review council — fix requested (cycle 1/3)

The automated review council identified the following issues. Please address each one:

Findings to fix

[major] scripts/review-one-pr.sh:608 — repository is not a valid gh pr view --json field. Replace with headRepository (or headRepositoryOwner where owner is needed) in scripts/review-one-pr.sh, prompts/deep-review.md, prompts/security-audit.md, and prompts/single-review.md.
[major] scripts/review-one-pr.sh:641 — REVIEW_MODE="triage-approved" is passed to single-review.md but that prompt only handles small and incremental modes. Either rename the value to align with an existing mode, or add a triage-approved branch in prompts/single-review.md so the template renders a meaningful mode label and delta analysis activates correctly.
[major] prompts/single-review.md:483 — Dangling reference to synthesize.md step 9a. Update to reference the correct file (cascade-action.md or equivalent) so the escalation delegation path in single-review mode is valid.
[major] scripts/review-one-pr.sh:591 — PRIOR_REVIEW_BODY uses head -1 (picks oldest match) but PRIOR_REVIEW_SHA reflects the most-recent marker. Under any multi-review history these point to different cycles. Change the body selector to tail -1 (or a jq last) to keep body and SHA in sync.
[minor] scripts/review-one-pr.sh:89 — Add validation that a matched body contains the full  HTML comment before exporting PRIOR_REVIEW_BODY, to reduce prompt-injection surface area.
[minor] prompts/council/ — Delete orphaned files prompts/council/correctness.md, prompts/council/maintainability.md, and prompts/council/security.md; they are no longer referenced by the script.
[minor] scripts/review-one-pr.sh — Remove the dead variable assignment SONNET_RC=$? (or wire it into the subsequent check in place of the output-file inspection).

Additional tasks

Resolve all unresolved review thread comments from other reviewers (CodeRabbit, Copilot, etc.)
Ensure all CI checks pass after your changes
Rebase on main if the branch is behind
Do NOT modify files unrelated to the findings above

The review council will automatically re-review after new commits are pushed.

Major fixes: - Replace invalid 'repository' gh field with 'headRepository'/'headRepositoryOwner' across script and all prompts - Add 'triage-approved' as valid REVIEW_MODE in single-review.md with dedicated mode description - Fix PRIOR_REVIEW_BODY selector: head -1 → tail -1 to match most-recent SHA - Update dangling 'synthesize.md step 9a' reference to cascade-action.md Minor fixes: - Validate PRIOR_REVIEW_BODY contains marker before exporting (anti-injection) - Write prior review body to temp file to avoid E2BIG on large reviews - Remove dead SONNET_RC=$? variable; add || true for set -e safety - Delete orphaned prompts/council/ files (no longer referenced by cascade) - Fix incremental-mode compare command: document how to derive owner/repo from headRepository metadata Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Question 4 was ambiguous — users confused "Max miles?" (meant odometer) with search radius. Now asks zip and radius as separate questions (#4 and #5) and renames the variable to MAX_ODOMETER with explicit "(highest mileage on the car itself)" clarification. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Add review mode optimization: small-PR and incremental fast paths PRs are now routed to the cheapest review path: - Small PRs (<10 lines): single Opus call (1 invocation vs 4) - Re-reviews (prior marker at different SHA): single Opus call with prior review context, focused on what changed since last review - First review of non-small PRs: full 3-member council (unchanged) New files: - prompts/single-review.md: combined council+synth prompt for single reviewer mode, handles both small and incremental paths New config: - SMALL_PR_THRESHOLD repo variable (default 10) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * Replace parallel council with cascading review tiers Haiku triage → Sonnet deep review → Opus security audit, where each tier only fires if the previous one escalated. This replaces the 3-parallel-members + synthesizer architecture. Cost impact: - ~80% of PRs: Haiku + Opus confirm (2 calls, ~30s vs 4 calls, ~5 min) - ~15%: + Sonnet (3 calls, ~2.5 min) - ~5%: full cascade (4 calls, ~5.5 min) Haiku works by receiving pre-fetched context (no tool use needed), which sidesteps the agentic limitations found earlier. New files: - prompts/triage.md — Tier 1: Haiku fast classification - prompts/deep-review.md — Tier 2: Sonnet full review - prompts/security-audit.md — Tier 3: Opus security audit - prompts/cascade-action.md — posts review from any resolving tier Re-reviews (prior marker at different SHA) pass prior review context to the triage, enabling faster incremental checks. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * Fix review council findings: 4 major + 3 minor issues Major fixes: - Replace invalid 'repository' gh field with 'headRepository'/'headRepositoryOwner' across script and all prompts - Add 'triage-approved' as valid REVIEW_MODE in single-review.md with dedicated mode description - Fix PRIOR_REVIEW_BODY selector: head -1 → tail -1 to match most-recent SHA - Update dangling 'synthesize.md step 9a' reference to cascade-action.md Minor fixes: - Validate PRIOR_REVIEW_BODY contains marker before exporting (anti-injection) - Write prior review body to temp file to avoid E2BIG on large reviews - Remove dead SONNET_RC=$? variable; add || true for set -e safety - Delete orphaned prompts/council/ files (no longer referenced by cascade) - Fix incremental-mode compare command: document how to derive owner/repo from headRepository metadata Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: DJ <dj@Rachels-Air.localdomain> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Question 4 was ambiguous — users confused "Max miles?" (meant odometer) with search radius. Now asks zip and radius as separate questions (#4 and #5) and renames the variable to MAX_ODOMETER with explicit "(highest mileage on the car itself)" clarification. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Add review mode optimization: small-PR and incremental fast paths PRs are now routed to the cheapest review path: - Small PRs (<10 lines): single Opus call (1 invocation vs 4) - Re-reviews (prior marker at different SHA): single Opus call with prior review context, focused on what changed since last review - First review of non-small PRs: full 3-member council (unchanged) New files: - prompts/single-review.md: combined council+synth prompt for single reviewer mode, handles both small and incremental paths New config: - SMALL_PR_THRESHOLD repo variable (default 10) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * Replace parallel council with cascading review tiers Haiku triage → Sonnet deep review → Opus security audit, where each tier only fires if the previous one escalated. This replaces the 3-parallel-members + synthesizer architecture. Cost impact: - ~80% of PRs: Haiku + Opus confirm (2 calls, ~30s vs 4 calls, ~5 min) - ~15%: + Sonnet (3 calls, ~2.5 min) - ~5%: full cascade (4 calls, ~5.5 min) Haiku works by receiving pre-fetched context (no tool use needed), which sidesteps the agentic limitations found earlier. New files: - prompts/triage.md — Tier 1: Haiku fast classification - prompts/deep-review.md — Tier 2: Sonnet full review - prompts/security-audit.md — Tier 3: Opus security audit - prompts/cascade-action.md — posts review from any resolving tier Re-reviews (prior marker at different SHA) pass prior review context to the triage, enabling faster incremental checks. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * Fix review council findings: 4 major + 3 minor issues Major fixes: - Replace invalid 'repository' gh field with 'headRepository'/'headRepositoryOwner' across script and all prompts - Add 'triage-approved' as valid REVIEW_MODE in single-review.md with dedicated mode description - Fix PRIOR_REVIEW_BODY selector: head -1 → tail -1 to match most-recent SHA - Update dangling 'synthesize.md step 9a' reference to cascade-action.md Minor fixes: - Validate PRIOR_REVIEW_BODY contains marker before exporting (anti-injection) - Write prior review body to temp file to avoid E2BIG on large reviews - Remove dead SONNET_RC=$? variable; add || true for set -e safety - Delete orphaned prompts/council/ files (no longer referenced by cascade) - Fix incremental-mode compare command: document how to derive owner/repo from headRepository metadata Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: DJ <dj@Rachels-Air.localdomain> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Question 4 was ambiguous — users confused "Max miles?" (meant odometer) with search radius. Now asks zip and radius as separate questions (#4 and #5) and renames the variable to MAX_ODOMETER with explicit "(highest mileage on the car itself)" clarification. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Copilot AI review requested due to automatic review settings April 10, 2026 11:52

Copilot started reviewing on behalf of don-petry April 10, 2026 11:53 View session

Copilot AI reviewed Apr 10, 2026

View reviewed changes

don-petry commented Apr 10, 2026

View reviewed changes

don-petry merged commit ec774b9 into main Apr 10, 2026

don-petry deleted the review-optimizations branch April 10, 2026 16:58

This was referenced May 7, 2026

[MEDIUM] Add retry logic around branch-update API call #37

Closed

feat: add Gemini Pro support and optimize PR review fallback chain #102

Merged

github-actions Bot mentioned this pull request May 25, 2026

petry-projects — workflow failures detected 2026-05-25 #385

Open

This was referenced Jun 1, 2026

petry-projects — workflow failures detected 2026-06-01 #416

Closed

petry-projects — workflow failures detected 2026-06-04 #429

Closed

donpetry-bot mentioned this pull request Aug 2, 2026

feat: implement issue #1411 — [Phase 4] Capture the pre-rollout baseline for the three success metrics, including a new no-action-comment noise metric #1414

Merged

donpetry-bot mentioned this pull request Aug 2, 2026

feat: implement issue #1436 — [bug] seed-repo-template.sh repins caller stubs to pre-#1184 legacy channels — template-drift enforces a baseline that contradicts standards #1441

Merged

donpetry-bot mentioned this pull request Aug 3, 2026

feat: implement issue #687 — [Phase 2] Generalized CI compliance test: reject SHA / # main pins on any first-party reusable #1449

Merged

donpetry-bot mentioned this pull request Aug 14, 2026

feat: implement issue #1482 — [Phase 4] Post-conflict-resolution integrity check — catch duplicated/corrupted content before it's trusted #1496

Merged

github-actions Bot mentioned this pull request Aug 15, 2026

petry-projects — workflow failures detected 2026-08-15 #1517

Open

Navigation Menu

Uh oh!

Optimize review: small-PR and incremental fast paths - #4

Optimize review: small-PR and incremental fast paths#4
don-petry merged 3 commits into
mainfrom
review-optimizations

don-petry commented Apr 10, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

don-petry left a comment

Uh oh!

don-petry commented Apr 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

don-petry commented Apr 10, 2026

Summary

Cost impact

Test plan

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

don-petry left a comment

Choose a reason for hiding this comment

Automated review — NEEDS HUMAN REVIEW

Summary

Linked issue analysis

Findings

Major

Minor

Info

CI status

Uh oh!

don-petry commented Apr 10, 2026

Review council — fix requested (cycle 1/3)

Findings to fix

Additional tasks

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants