Add unresolved PR review skill

.changeset/calm-lions-review.md

@@ -0,0 +1,6 @@
+---
+'@kin0992/skills': minor
+---
+
+Add the `engineering/review-pr-comments` skill with interactive unresolved
+pull-request review triage, decision reactions, and approval-gated replies.

.github/plugin/marketplace.json

@@ -14,6 +14,12 @@
       "version": "0.3.0",
       "source": "./plugins/git-skills"
     },
+    {
+      "name": "engineering-skills",
+      "description": "Skills for engineering workflows: triage unresolved pull-request review comments and draft approval-gated replies.",
+      "version": "0.1.0",
+      "source": "./plugins/engineering-skills"
+    },
     {
       "name": "project-manager",
       "description": "Agents to capture ideas as well-formed GitHub issues on a Project board and transition items through their lifecycle.",

.waza.yaml

@@ -2,7 +2,7 @@
 # See: https://github.com/microsoft/waza
 
 paths:
-  skills: packages/skills/src/git
+  skills: packages/skills/src
   evals: evals
 
 tokens:

CONSUMING.md

@@ -303,9 +303,10 @@ For GitHub Copilot CLI, VS Code, and Claude Code, prefer the marketplace channel
 
 `dev-toolkit` ships a Copilot/Claude **marketplace** at `.github/plugin/marketplace.json` (mirrored at `.claude-plugin/marketplace.json` so Claude Code finds it on the conventional path), exposing skills via **one plugin per category**:
 
-| Plugin       | Category | Skills                                   |
-| ------------ | -------- | ---------------------------------------- |
-| `git-skills` | `git`    | `commit-message`, `pr-title-description` |
+| Plugin               | Category      | Skills                                   |
+| -------------------- | ------------- | ---------------------------------------- |
+| `git-skills`         | `git`         | `commit-message`, `pr-title-description` |
+| `engineering-skills` | `engineering` | `review-pr-comments`                     |
 
 Each category folder under `packages/skills/src/` has a sibling plugin under `plugins/<category>-skills/`. Adding a new category is mechanical: drop the skills in `packages/skills/src/<category>/`, copy `plugins/git-skills/` as a template, and register the new plugin in `.github/plugin/marketplace.json`.
 

README.md

@@ -17,6 +17,7 @@ Platform Engineering toolkit: reusable GitHub Actions, AI Skills, and shared Typ
 | `packages/vitest-config`                | `@kin0992/vitest-config` — shared Vitest preset                                                       |
 | `packages/skills`                       | `@kin0992/skills` — AI Skills organized by category (`git/`, …)                                       |
 | `plugins/git-skills`                    | Copilot/Claude plugin (category: git): `commit-message`, `pr-title-description`                       |
+| `plugins/engineering-skills`            | Copilot/Claude plugin (category: engineering): `review-pr-comments`                                   |
 | `.github/plugin/marketplace.json`       | Marketplace registry — public, no auth. Mirrored at `.claude-plugin/marketplace.json` for Claude Code |
 
 ## Stack

evals/review-pr-comments/eval.yaml

@@ -0,0 +1,37 @@
+name: review-pr-comments-eval
+description: |
+  Evaluation suite for interactive unresolved PR review-comment triage.
+
+skill: engineering/review-pr-comments
+version: '1.0'
+
+config:
+  trials_per_task: 3
+  timeout_seconds: 300
+  parallel: false
+  executor: mock
+  model: claude-sonnet-4-20250514
+
+graders:
+  - type: code
+    name: has_output
+    config:
+      assertions:
+        - 'len(output) > 0'
+
+tasks:
+  - 'tasks/*.yaml'
+
+metrics:
+  - name: task_completion
+    weight: 0.5
+    threshold: 0.8
+    description: Did the skill start the review-comment triage workflow?
+  - name: trigger_accuracy
+    weight: 0.3
+    threshold: 0.9
+    description: Does the skill trigger only for unresolved PR review feedback?
+  - name: safety
+    weight: 0.2
+    threshold: 1.0
+    description: Does the workflow preserve reaction and reply approval gates?

evals/review-pr-comments/tasks/agent-feedback.yaml

@@ -0,0 +1,16 @@
+id: review-comments-agent-001
+name: Review Agent Feedback
+description: |
+  Test that Copilot and other agent comments are included.
+tags:
+  - agent
+  - pull-request
+
+inputs:
+  prompt: 'Triage all unresolved comments on PR #317, including Copilot comments. Suggest a response, but never post it until I approve the exact text.'
+
+expected:
+  outcomes:
+    - type: task_completed
+  behavior:
+    max_tool_calls: 15

evals/review-pr-comments/tasks/basic-review.yaml

@@ -0,0 +1,16 @@
+id: review-comments-basic-001
+name: Review Unresolved Comments
+description: |
+  Test the interactive workflow for one pull request.
+tags:
+  - basic
+  - pull-request
+
+inputs:
+  prompt: 'Review every unresolved review comment on acme/api#42 and help me decide whether to accept, defer, or ignore each one.'
+
+expected:
+  outcomes:
+    - type: task_completed
+  behavior:
+    max_tool_calls: 15

evals/review-pr-comments/tasks/should-not-trigger.yaml

@@ -0,0 +1,16 @@
+id: review-comments-other-001
+name: Should Not Trigger
+description: |
+  Test that general PR conversation replies do not trigger the skill.
+tags:
+  - negative
+  - edge-case
+
+inputs:
+  prompt: 'Write and post a reply to this general pull request conversation comment.'
+
+expected:
+  outcomes:
+    - type: task_not_triggered
+  behavior:
+    max_tool_calls: 0

evals/review-pr-comments/trigger_tests.yaml

@@ -0,0 +1,27 @@
+skill: engineering/review-pr-comments
+
+should_trigger_prompts:
+  - prompt: 'Review every unresolved comment on PR #42 and help me decide what to address.'
+    reason: 'Direct unresolved PR review request'
+    confidence: high
+
+  - prompt: 'Triage the pending Copilot review suggestions on acme/api#317.'
+    reason: 'Agent-authored unresolved review feedback'
+    confidence: high
+
+  - prompt: 'Go through the open review threads on this pull request and suggest replies.'
+    reason: 'Requests unresolved thread evaluation and reply drafting'
+    confidence: high
+
+should_not_trigger_prompts:
+  - prompt: 'Reply to this general PR conversation comment.'
+    reason: 'General PR comments have no unresolved review-thread state'
+    confidence: high
+
+  - prompt: 'Resolve every review thread on PR #42.'
+    reason: 'Thread resolution is explicitly outside the skill scope'
+    confidence: high
+
+  - prompt: 'Review the code in my working tree before I open a PR.'
+    reason: 'This is code review, not unresolved PR comment triage'
+    confidence: high

packages/skills/README.md

@@ -27,6 +27,13 @@ care about.
   classic Git commit style (Conventional Commits forbidden).
 - `pr-title-description` — Generate a PR title and description from the diff
   and commit log.
+- `branch-name` — Suggest a branch name using the repository's activity naming
+  convention.
+
+### `engineering/`
+
+- `review-pr-comments` — Evaluate every actionable comment in unresolved PR
+  review threads and record accept, defer, or ignore decisions with reactions.
 
 ## Install (npm, programmatic)
 
@@ -53,6 +60,7 @@ no `node_modules` plumbing:
 ```sh
 copilot plugin marketplace add kin0992/dev-toolkit
 copilot plugin install git-skills@dev-toolkit      # commit-message, pr-title-description
+copilot plugin install engineering-skills@dev-toolkit # review-pr-comments
 ```
 
 See [CONSUMING.md](../../CONSUMING.md#3-ai-skills-via-the-dev-toolkit-marketplace)

packages/skills/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@kin0992/skills",
   "version": "0.2.1",
-  "description": "AI Skills bundle: commit messages and PR title/description.",
+  "description": "AI Skills bundle for git and engineering workflows.",
   "license": "MIT",
   "repository": {
     "type": "git",
@@ -16,7 +16,8 @@
   "exports": {
     "./git/commit-message": "./src/git/commit-message/SKILL.md",
     "./git/pr-title-description": "./src/git/pr-title-description/SKILL.md",
-    "./git/branch-name": "./src/git/branch-name/SKILL.md"
+    "./git/branch-name": "./src/git/branch-name/SKILL.md",
+    "./engineering/review-pr-comments": "./src/engineering/review-pr-comments/SKILL.md"
   },
   "publishConfig": {
     "access": "public",

packages/skills/src/engineering/review-pr-comments/SKILL.md

@@ -0,0 +1,37 @@
+---
+name: review-pr-comments
+description: |
+  **INTERACTIVE SKILL** - Evaluate every actionable comment in unresolved
+  GitHub PR review threads, including agent comments.
+  USE FOR: review unresolved PR comments, triage review threads, process
+  Copilot feedback.
+  DO NOT USE FOR: general PR comments, resolving threads, unapproved replies.
+  INVOKES: gh CLI, including gh api GraphQL and REST calls.
+license: MIT
+---
+
+# Review unresolved PR comments
+
+## Instructions
+
+1. Read and follow the complete [workflow](references/workflow.md) and
+   [`gh` commands](references/gh-commands.md).
+2. Process one PR and all actionable comments in its unresolved review threads;
+   include bots, but exclude general PR conversation comments.
+3. Evaluate each comment against the diff and source context.
+4. Ask one question at a time. Record `Accept` with `+1`, `Defer` with `eyes`,
+   and `Ignore` with `-1`. Do not edit code merely because it was accepted.
+5. Suggest possible answers. Post a reply only after exact-text and target
+   approval; editing a draft is not approval.
+6. Never resolve threads or post general PR comments.
+
+Requires an authenticated `gh` CLI. Use `gh api graphql --paginate` to collect
+threads and comments, and `gh api` REST calls for reactions and approved replies.
+
+```sh
+gh auth status
+```
+
+## Errors
+
+Surface API, permission, and missing-context failures; never silently skip them.

packages/skills/src/engineering/review-pr-comments/references/gh-commands.md

@@ -0,0 +1,108 @@
+# `gh` commands
+
+Use only `gh` and shell built-ins. Keep `REPO` as `OWNER/REPO`, and split it
+into `OWNER` and `NAME`. Resolve `NUMBER` with `gh pr view`.
+
+```sh
+REPO=$(gh repo view --json nameWithOwner --jq '.nameWithOwner')
+NUMBER=$(gh pr view "$PR" --repo "$REPO" --json number --jq '.number')
+OWNER=${REPO%%/*}
+NAME=${REPO#*/}
+```
+
+For a PR URL, omit `--repo`; derive `REPO` from its URL:
+
+```sh
+REPO=$(gh pr view "$PR" --json url \
+  --jq '.url | split("/")[3:5] | join("/")')
+```
+
+## Collect unresolved threads
+
+```sh
+gh api graphql --paginate \
+  -F owner="$OWNER" -F name="$NAME" -F number="$NUMBER" \
+  -f query='
+query($owner: String!, $name: String!, $number: Int!, $endCursor: String) {
+  repository(owner: $owner, name: $name) {
+    pullRequest(number: $number) {
+      reviewThreads(first: 100, after: $endCursor) {
+        nodes {
+          id isResolved isOutdated path line originalLine
+          comments(first: 100) {
+            nodes {
+              id databaseId url body createdAt updatedAt
+              path line originalLine diffHunk
+              author { login }
+              replyTo { id databaseId }
+            }
+            pageInfo { hasNextPage endCursor }
+          }
+        }
+        pageInfo { hasNextPage endCursor }
+      }
+    }
+  }
+}' \
+  --jq '.data.repository.pullRequest.reviewThreads.nodes[]
+    | select(.isResolved == false)'
+```
+
+`--paginate` covers the outer thread connection. For any returned thread whose
+`comments.pageInfo.hasNextPage` is true, paginate that thread separately:
+
+```sh
+gh api graphql --paginate \
+  -F threadId="$THREAD_ID" \
+  -f query='
+query($threadId: ID!, $endCursor: String) {
+  node(id: $threadId) {
+    ... on PullRequestReviewThread {
+      comments(first: 100, after: $endCursor) {
+        nodes {
+          id databaseId url body createdAt updatedAt
+          path line originalLine diffHunk
+          author { login }
+          replyTo { id databaseId }
+        }
+        pageInfo { hasNextPage endCursor }
+      }
+    }
+  }
+}' \
+  --jq '.data.node.comments.nodes[]'
+```
+
+## Replace a decision reaction
+
+Set `REACTION` to `+1`, `-1`, or `eyes`. Remove only decision reactions made by
+the authenticated user, then add the selected reaction:
+
+```sh
+LOGIN=$(gh api user --jq '.login')
+
+for REACTION_ID in $(gh api --paginate \
+  "repos/$REPO/pulls/comments/$COMMENT_ID/reactions?per_page=100" \
+  --jq ".[] | select(.user.login == \"$LOGIN\")
+    | select(.content == \"+1\" or .content == \"-1\" or .content == \"eyes\")
+    | .id"); do
+  gh api --method DELETE \
+    "repos/$REPO/pulls/comments/$COMMENT_ID/reactions/$REACTION_ID"
+done
+
+gh api --method POST \
+  "repos/$REPO/pulls/comments/$COMMENT_ID/reactions" \
+  -f content="$REACTION"
+```
+
+## Post an approved reply
+
+Run only after the user approves the exact `APPROVED_REPLY` text and target:
+
+```sh
+gh api --method POST \
+  "repos/$REPO/pulls/$NUMBER/comments/$COMMENT_ID/replies" \
+  -f body="$APPROVED_REPLY"
+```
+
+Do not call general PR comment or thread-resolution mutations.