address-pr-review

Étoiles0

Forks0

Mis à jour14 juin 2026 à 19:16

State-aware GitHub Copilot PR review handler. Detects review state (not triggered / in progress / has unresolved threads / clean), triggers Copilot via `gh pr edit --add-reviewer` when needed, polls until completion, classifies fresh unresolved comments as fix/resolve-only/skip, applies fixes, resolves threads via GraphQL, and pushes. Only exits when the fresh review (matching HEAD SHA) has no unresolved fresh threads. Use when the user types `/address-pr-review`, asks to address PR comments, wants to clear review findings, or says "check the PR review." Disabled for auto-invocation.

Installation

Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.

Source

szymonos

szymonos/envy-nx

Ouvrir le dépôt GitHub Voir les dépôts du créateur

Téléchargement

Métiers associésSOC

Basé sur la classification professionnelle SOC

Analystes en assurance qualité des logiciels et testeursProfessions informatiques et mathématiques·SOC 15-1253

Explorateur de fichiers

2 fichiers

SKILL.md

readonly

Plus depuis ce dépôt

même dépôt

prepare-release

szymonos/envy-nx

Cut a release end-to-end - commit uncommitted work, consolidate the branch's commits by Conventional Commits prefix via soft-reset, write the CHANGELOG entry, push the branch, then create or update the release PR. Bundled extract.py reads only the CHANGELOG sections needed (avoids the 100+ KB file); cspell_words.py reconciles new vocabulary with project-words.txt before committing; test_stats.py audits stat-callout drift in docs when `tests/` changed since the last tag. Use when the user types `/prepare-release <X.Y.Z>`, asks to cut a release, prepare release notes, ship X.Y.Z, or wrap a branch up. Disabled for auto-invocation.

2026-06-140

second-opinion

szymonos/envy-nx

Heterogeneous-model code review of the current branch's changes. Invokes GitHub Copilot CLI with gpt-5.3-codex to review git diff since merge-base with main (or user-specified commit). Reads .claude/skills/second-opinion/REVIEW-BRIEF.md for focused project context. Returns structured findings that Claude reads, summarizes, and acts on. Use when the user types `/second-opinion`, asks for a second opinion on a branch, wants GPT to review the work, or wants an independent review before pushing. Disabled for auto-invocation.

2026-06-140

review

szymonos/envy-nx

Run a periodic chunked code review against one shard of the repo. Use /review <shard> to start a review, /review next to rotate to the oldest-reviewed shard, /review act <findings-path> to triage findings and apply fixes, /review status to see what's pending. Spawns the reviewer/fixer/verifier subagents from .claude/agents/.

2026-05-240

grill-with-docs

szymonos/envy-nx

Grilling session that challenges your plan against the existing domain model, sharpens terminology, and writes agent-readable decisions to design/decisions/ as they crystallise. Use when user wants to stress-test a plan against their project's language and documented decisions.

2026-05-230

slide-deck-builder

szymonos/envy-nx

Build or update a reveal.js slide deck with a dark theme, semantic accent colors, and technical density (Outfit + JetBrains Mono fonts; indigo + green/amber/red/blue accents). Operates on docs/slides/index.html if present; scaffolds a new one otherwise. Use this skill whenever the user mentions slides, the deck, a presentation, a talk, or wants to add, edit, or fix any slide-level content - even if they don't explicitly say "reveal.js".

2026-05-230

name	address-pr-review
description	State-aware GitHub Copilot PR review handler. Detects review state (not triggered / in progress / has unresolved threads / clean), triggers Copilot via `gh pr edit --add-reviewer` when needed, polls until completion, classifies fresh unresolved comments as fix/resolve-only/skip, applies fixes, resolves threads via GraphQL, and pushes. Only exits when the fresh review (matching HEAD SHA) has no unresolved fresh threads. Use when the user types `/address-pr-review`, asks to address PR comments, wants to clear review findings, or says "check the PR review." Disabled for auto-invocation.
disable-model-invocation	true

Address PR review

State-aware Copilot PR review handler. Detects the current review state, drives toward "fresh review clean," and processes only fresh unresolved threads. Works with any reviewer that posts inline PR comments; Copilot is the default trigger target.

When to use

/address-pr-review - drive current branch's PR review to a clean state
/address-pr-review 37 - same, for a specific PR number
"address the PR review comments" / "check the PR review" / "clear the review" - same as /address-pr-review

Prerequisites

gh CLI installed and authenticated.
Copilot enabled on the repository.

Important: always trigger Copilot via pr_review.py trigger, never via ad-hoc gh pr edit --add-reviewer invocations or raw GraphQL mutations from the agent. pr_review.py trigger wraps gh pr edit --add-reviewer copilot-pull-request-reviewer with the right reviewer login, idempotency handling, and uniform error reporting; one-off calls drift in subtle ways (wrong login alias, no idempotency, no JSON output for the skill's state machine to consume). If trigger fails, surface the error to the user - Copilot may need to be enabled in repo settings.

Review states

The skill operates as a state machine. Four states are possible:

State	Fresh review exists?	Copilot requested?	Unresolved fresh threads?	Action
A	No	No	N/A	Trigger + wait
B	No	Yes (in progress)	N/A	Wait
C	Yes	No	Yes	Process
D	Yes	No	No	EXIT (clean)

"Fresh review" = a Copilot review whose commit.oid matches the PR's current headRefOid. A review from a prior push is stale and ignored - the skill triggers a new one.

"Fresh threads" = threads with isResolved: false AND isOutdated: false. Outdated unresolved threads are silently ignored - the fresh review re-evaluates the same code.

State D is the only clean exit. Every other state drives toward it via trigger/wait/process.

Workflow

Phase 1 - detect state

.claude/skills/address-pr-review/scripts/pr_review.py state --pr <N>

Returns JSON with state, headSha, freshReviewSha, copilotRequested, and unresolvedFreshThreads. Exit code: 0=D, 1=C, 2=B, 3=A.

If no PR is found on the current branch (or --pr is invalid), the script exits with a clear error. Surface it to the user and stop.

Phase 2 - drive to State D

Branch on the state from Phase 1:

State A (not triggered): run pr_review.py trigger --pr <N> to request Copilot. Then pr_review.py wait --pr <N> to poll until the review completes. The wait subcommand returns when the state resolves to C or D. Continue from that state.
State B (in progress): skip trigger, go straight to pr_review.py wait --pr <N>. Same continuation.
State C (unresolved fresh threads): proceed to Phase 3.
State D: announce "PR review clean - no unresolved fresh threads. Exit." Done.

wait polls every 30 seconds, up to 8 min total. On timeout (exit 4), surface to the user - Copilot may be queued or the service may be slow. Don't loop the wait; let the user decide whether to retry.

Phase 3 - process unresolved fresh threads (State C only)

The state JSON already contains unresolvedFreshThreads with {id, path, line, author, body} for each. Present them as a summary table:

## PR #37 - N unresolved fresh threads

| # | File:Line | Author | Summary |
|---|-----------|--------|---------|
| 1 | src/main.py:9 | copilot | Missing error handling on API call |
| 2 | docs/index.md:109 | copilot | Stale reference count |

Known false positives - auto-resolve without reading the file:

Comments flagging ubuntu-slim as an invalid or non-standard GitHub runner (e.g., "runs-on: ubuntu-slim is likely to fail"). ubuntu-slim is a valid runner label configured for this repo (used by every workflow under .github/workflows/) - not a GitHub-hosted label, but an intentional self-hosted / org-managed runner that the reviewer cannot see.

For each remaining thread, read the comment body + the referenced file at the specified line. Classify:

fix - the comment identifies a real issue (bug, missing code, inconsistency, stale reference). Read the referenced file, formulate a fix using Claude's knowledge of the codebase (not a copy-paste of the suggestion), apply via Edit.
resolve-only - the comment is already addressed by a prior fix in this session, or the issue genuinely doesn't apply (e.g., stale on a specific line but the file changed elsewhere). Resolve silently.
skip - the comment is a suggestion, design question, or judgment call that needs human input.

Resolve fix and resolve-only threads via:

.claude/skills/address-pr-review/scripts/pr_review.py resolve <thread-id>

After processing all fix and resolve-only items:

If any skip items remain: present them to the user via AskUserQuestion. For each, offer three options: "Fix it" (Claude fixes now), "Resolve without fix" (intentional choice), "Leave open" (for later / human reviewer to decide). Act on user's choices.
Report: "N fixed, M resolved (stale/intentional), K surfaced to user."

Phase 4 - commit and push

Only runs in standalone mode - when the skill is invoked directly, not as part of a larger workflow. The caller decides commit topology; the skill's job is to surface fixes and resolve threads.

When invoked by a caller that manages its own commit flow (e.g., a consolidation skill that re-cuts commits), the caller should pass context indicating Phase 4 should be skipped. Without that context, Phase 4 runs by default.

If any files were edited in Phase 3:

Run make lint to validate.
Stage the changed files explicitly (never git add -A).
Commit: git commit -m "fix: address PR review comments"
Push: git push

The push will trigger a new Copilot review automatically. The skill is one-shot - it does not re-invoke itself after pushing. The user decides whether to re-run.

If no files were edited (all threads were resolve-only or skip → leave-open), skip the commit - just report the resolution results.

Anti-patterns

Processing outdated threads. Threads with isOutdated: true reference code that may no longer exist at that line. The fresh review re-evaluates the same code; if the issue persists, it appears as a fresh thread. Silently ignoring outdated threads is correct.
Treating a stale review as fresh. A review whose commit.oid doesn't match the current headRefOid is from a prior push. Don't process its threads even if they're unresolved - trigger a new review instead.
Resolving skip items silently. The human might want to act on them - a "consider X" suggestion might actually be a good idea. Surface, don't suppress.
Posting reply comments before resolving. Silent resolve is the design choice - keeps the PR history clean. The fix is visible in the diff; the resolution is visible in the thread state.
Looping wait on timeout. If wait exits 4 (timeout), don't immediately re-call it. Surface to the user; Copilot may be queued or rate-limited.
Copying the reviewer's suggested fix verbatim. The reviewer (Copilot or human) suggests direction; Claude writes the actual fix using its knowledge of the codebase's patterns and accepted decisions.
Resolving threads for human reviewers' comments without checking. The skill processes ALL unresolved fresh threads regardless of author. If a human reviewer left a comment expecting a human reply, resolving silently would be rude. Classify these as skip unless the fix is unambiguous.

Example invocations

/address-pr-review - address comments on current branch's PR
/address-pr-review 37 - address comments on PR #37
"Check the PR review and fix what you can" - same as /address-pr-review