fix(ci): authenticate origin before the Claude action's branch fetch, and count its github-actions findings

.agents/skills/iterate-pr/SKILL.md

@@ -69,6 +69,8 @@ Returns JSON with feedback categorized as:
 
 Review bot feedback (from Sentry, Warden, Copilot, Cursor, Bugbot, CodeQL, etc.) appears in `high`/`medium`/`low` with `review_bot: true` — it is NOT placed in the `bot` bucket.
 
+That includes Claude posting as `github-actions`, which happens when a workflow hands claude-code-action `github_token`. That login is shared with informational bots, so it is decided by content: inline review threads from it, and top-level comments carrying a claude-code-action signature (`Claude finished @…`, a placeholder, a `[View job](…/actions/runs/N)` link), are review feedback. Other `github-actions` comments stay in `bot`.
+
 **A review bot's comment existing is not a completion signal, and neither is its `created_at` or a green check run.** The Claude review bot posts a comment the instant it is triggered and edits that same comment in place as it works — a run has been observed to report `success` while the body still read "Review in progress" with unchecked boxes. `fetch_pr_feedback.cjs` detects this by checking whether the body still **opens** with the in-progress marker (a review that happens to discuss this behaviour mid-body is not mistaken for an unfinished one) and reports the count as `summary.review_in_progress` instead of bucketing the placeholder as ordinary feedback. **Before treating `needs_attention: 0` as "reviewed, nothing found," check `summary.review_in_progress` — a nonzero count means the review hasn't finished, not that it found nothing.** When one or more are present, `action_required` says so and outranks every other bucket, since a caller must not stop just because the buckets it can already see look clean.
 
 **Self-review feedback** (from the PR author) appears in `high`/`medium`/`low` with `self_review: true`. Because you can't "Request changes" on your own PR, a self-review lands as `COMMENTED` review summaries and ordinary review threads rather than changes-requested items — these are surfaced (not dropped). A self-review **thread** (an inline comment) is bucketed by content like any other, defaulting to `medium` absent a `h:/m:/l:` prefix. A self-review **summary** is bucketed structurally, not by content (see below), landing in `review_summary` absent a marker. Treat `self_review` items the same as any other human feedback in step 3.

.agents/skills/iterate-pr/scripts/fetch_pr_feedback.cjs

@@ -33,6 +33,9 @@
  *   with a `review_bot: true` flag — they are NOT placed in the `bot` bucket.
  * - Info bots (Codecov, Dependabot, Renovate, etc.) post status reports and are
  *   placed in the `bot` bucket for silent skipping.
+ * - `github-actions` is decided by content, since claude-code-action posts as
+ *   it when given `github_token`: its inline threads, and its top-level
+ *   comments carrying a claude-code-action signature, are review feedback.
  *
  * Self-review feedback:
  * - You can't formally "Request changes" on your own PR, so a PR author's own
@@ -150,6 +153,45 @@ const isReviewBot = (username) =>
 const isInfoBot = (username) =>
   INFO_BOT_PATTERNS.some((pattern) => pattern.test(username ?? ""));
 
+/**
+ * `github-actions` is a SHARED identity, so its login alone cannot decide a
+ * bucket. claude-code-action posts as it whenever a workflow passes
+ * `github_token: ${{ github.token }}` (this repo's review and focus workflows
+ * do), and the login is spelled `github-actions[bot]` over REST and
+ * `github-actions` over GraphQL. Filing those as info-bot noise is how
+ * taskless/marketing#73 reported `needs_attention: 0` over four open findings.
+ */
+const SHARED_ACTIONS_BOT = /^github-actions(\[bot\])?$/i;
+
+/**
+ * Openings and links that only claude-code-action's comments carry. The
+ * in-progress placeholders count too: a placeholder posted as `github-actions`
+ * must still hold the wait loop open.
+ */
+const ACTION_SIGNATURES = [
+  /^[*_]*\s*claude finished @/i,
+  /\[View job(?: run)?\]\(https:\/\/[^)\s]+\/actions\/runs\/\d+/i,
+];
+
+/**
+ * Whether a comment is a review bot's, deciding `github-actions` by content.
+ *
+ * An INLINE thread from `github-actions` is review feedback by default:
+ * informational bots post top-level status comments, not line-anchored review
+ * threads. A top-level `github-actions` comment counts only when it carries a
+ * claude-code-action signature; every other one stays informational.
+ */
+const isReviewComment = (author, body, { inline = false } = {}) => {
+  if (isReviewBot(author)) return true;
+  if (!SHARED_ACTIONS_BOT.test(author ?? "")) return false;
+  if (inline) return true;
+  const opening = body.trimStart();
+  return (
+    isReviewInProgress(opening) ||
+    ACTION_SIGNATURES.some((signature) => signature.test(opening))
+  );
+};
+
 /**
  * The placeholders a review bot posts before it has anything to say.
  *
@@ -328,8 +370,17 @@ const matchesUnnegated = (patterns, body) =>
 const categorizeComment = (comment, body) => {
   const author = comment?.author?.login || comment?.user?.login || "";
 
-  if (isInfoBot(author) && !isReviewBot(author)) return "bot";
+  if (isInfoBot(author) && !isReviewComment(author, body)) return "bot";
 
+  return categorizeContent(comment, body);
+};
+
+/**
+ * The content half of `categorizeComment`, for a caller that has already
+ * decided the author is not an info bot (`bucketByAuthor`), so a review
+ * comment posted as `github-actions` is not sent back to `bot` by its login.
+ */
+const categorizeContent = (_comment, body) => {
   const logaf = detectLogaf(body);
   if (logaf) return logaf;
 
@@ -382,28 +433,30 @@ const categorizeReviewSummary = (_comment, body) =>
  * an info bot's noise (`bot`), or ordinary human feedback, from ANY of the
  * three sources.
  *
- * `categorize` defaults to the inline-comment classifier; the review-summary
- * call site passes `categorizeReviewSummary` instead, so the bot/info-bot
- * split and the in-progress check stay one rule in one place while what
- * happens to the *content* differs by source.
+ * `categorize` defaults to the inline-comment content classifier; the
+ * review-summary call site passes `categorizeReviewSummary` instead, so the
+ * bot/info-bot split and the in-progress check stay one rule in one place while
+ * what happens to the *content* differs by source. `inline` marks a review
+ * thread, which decides a `github-actions` author (see `isReviewComment`).
  */
 const bucketByAuthor = (
   feedback,
   item,
   comment,
   body,
   author,
-  categorize = categorizeComment
+  { categorize = categorizeContent, inline = false } = {}
 ) => {
+  const reviewer = isReviewComment(author, body, { inline });
   // The author gate is load-bearing, not belt and braces. A human writing
   // "Review in progress on my end, back by EOD" would otherwise be filed as an
   // unfinished review, vanish from `needs_attention`, and hang the wait loop
   // forever: a person's comment never gets edited into a finished form the way
   // the bot's placeholder does, so the count never drops. Only a review bot has
   // the lifecycle this bucket describes.
-  if (isReviewBot(author) && isReviewInProgress(body)) {
+  if (reviewer && isReviewInProgress(body)) {
     feedback.review_in_progress.push(item);
-  } else if (isReviewBot(author)) {
+  } else if (reviewer) {
     item.review_bot = true;
     feedback[categorize(comment, body)].push(item);
   } else if (isInfoBot(author)) {
@@ -628,14 +681,9 @@ const buildFeedback = (client, { owner, repo, prInfo }) => {
     ) {
       feedback.high.push(item);
     } else {
-      bucketByAuthor(
-        feedback,
-        item,
-        review,
-        body,
-        author,
-        categorizeReviewSummary
-      );
+      bucketByAuthor(feedback, item, review, body, author, {
+        categorize: categorizeReviewSummary,
+      });
     }
   }
 
@@ -671,7 +719,7 @@ const buildFeedback = (client, { owner, repo, prInfo }) => {
     if (isResolved) {
       feedback.resolved.push(item);
     } else {
-      bucketByAuthor(feedback, item, first, body, author);
+      bucketByAuthor(feedback, item, first, body, author, { inline: true });
     }
   }
 
@@ -822,6 +870,7 @@ module.exports = {
   extractFeedbackItem,
   isInfoBot,
   isReviewBot,
+  isReviewComment,
   isReviewInProgress,
   main,
 };

.agents/skills/iterate-pr/scripts/fetch_pr_feedback.test.cjs

@@ -21,6 +21,7 @@ const {
   extractFeedbackItem,
   isInfoBot,
   isReviewBot,
+  isReviewComment,
   isReviewInProgress,
   main,
 } = require("./fetch_pr_feedback.cjs");
@@ -1072,3 +1073,106 @@ test("an explicit marker in a summary still reaches its priority bucket", () =>
   assert.equal(output.summary.high, 1);
   assert.equal(output.summary.review_summaries, 0);
 });
+
+// claude-code-action posts as `github-actions` when a workflow hands it
+// `github_token`: `github-actions[bot]` over REST, `github-actions` over
+// GraphQL. Measured on taskless/marketing#73, where four unresolved findings
+// were filed as info-bot noise and the script reported needs_attention: 0.
+test("an inline thread from github-actions is review feedback, in both spellings", () => {
+  for (const login of ["github-actions", "github-actions[bot]"]) {
+    const output = build(
+      fakeClient({
+        threads: [
+          thread({
+            comments: {
+              nodes: [
+                { body: "This will break on empty input", author: { login } },
+              ],
+            },
+          }),
+        ],
+      }),
+      {}
+    );
+    assert.equal(output.summary.bot_comments, 0, login);
+    assert.equal(output.summary.high, 1, login);
+    assert.equal(output.summary.review_bot_feedback, 1, login);
+    assert.equal(output.feedback.high[0].review_bot, true, login);
+  }
+});
+
+test("a finished claude-code-action summary from github-actions is review feedback", () => {
+  for (const login of ["github-actions", "github-actions[bot]"]) {
+    const output = build(
+      fakeClient({
+        comments: [
+          {
+            id: 1,
+            body: `${LIVE_FINISHED}\n\nThis will break on empty input.`,
+            user: { login },
+          },
+        ],
+      }),
+      {}
+    );
+    assert.equal(output.summary.bot_comments, 0, login);
+    assert.equal(output.summary.high, 1, login);
+  }
+});
+
+// The wait loop must not exit while a review posted as github-actions is still
+// running: that is the silent-success shape `review_in_progress` exists for.
+test("a placeholder posted as github-actions is an unfinished review", () => {
+  for (const body of [LIVE_CREATED, LIVE_IN_PROGRESS]) {
+    const output = build(
+      fakeClient({
+        comments: [{ id: 1, body, user: { login: "github-actions[bot]" } }],
+      }),
+      {}
+    );
+    assert.equal(output.summary.review_in_progress, 1);
+    assert.equal(output.summary.bot_comments, 0);
+    assert.match(output.action_required, /still in progress/);
+  }
+});
+
+// The login is shared, so every OTHER top-level github-actions comment stays
+// informational: a deploy preview or a coverage table is not a finding.
+test("an ordinary top-level github-actions comment stays in bot", () => {
+  const output = build(
+    fakeClient({
+      comments: [
+        {
+          id: 1,
+          body: "Preview deployed. This will break nothing, see the logs.",
+          user: { login: "github-actions[bot]" },
+        },
+      ],
+    }),
+    {}
+  );
+  assert.equal(output.summary.bot_comments, 1);
+  assert.equal(output.summary.needs_attention, 0);
+});
+
+test("isReviewComment decides github-actions by signature, and only github-actions", () => {
+  assert.ok(isReviewComment("github-actions", LIVE_FINISHED));
+  assert.ok(
+    isReviewComment(
+      "github-actions[bot]",
+      "Done.\n\n[View job](https://github.com/o/r/actions/runs/123)"
+    )
+  );
+  assert.ok(!isReviewComment("github-actions[bot]", "Coverage dropped 2%"));
+  assert.ok(isReviewComment("github-actions", "anything", { inline: true }));
+  // Another info bot carrying the same signature is still not a reviewer.
+  assert.ok(!isReviewComment("codecov[bot]", LIVE_FINISHED));
+  assert.ok(!isReviewComment("codecov[bot]", "anything", { inline: true }));
+  assert.equal(
+    categorizeComment(
+      { user: { login: "github-actions[bot]" } },
+      `${LIVE_FINISHED}\n\nThis will break at runtime`
+    ),
+    "high"
+  );
+});

.github/workflows/claude-code-focus-on-demand.yml

@@ -129,6 +129,15 @@ jobs:
 
           echo "threads: $(jq '.data.repository.pullRequest.reviewThreads.nodes | length' "${GITHUB_WORKSPACE}/.prior-review.json")"
 
+      # The action's tag mode fetches the PR branch before it authenticates
+      # origin, which fails on a private repo. See the review workflow's step
+      # of the same name (issue #512) for why, and why not
+      # `persist-credentials: true`.
+      - name: Authenticate origin for the action's PR-branch fetch
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: git remote set-url origin "https://x-access-token:${GH_TOKEN}@${GITHUB_SERVER_URL#https://}/${GITHUB_REPOSITORY}.git"
+
       - name: Run Claude Code Focus
         id: focus
         uses: anthropics/claude-code-action@0a8d3c9443bbff909ab973b6a17a340b913f229f # v1.0.221

.github/workflows/claude-code-review-on-demand.yml

@@ -64,7 +64,9 @@ name: Claude Code Review (on demand)
 #   - that is NOT sufficient on its own. The action rewrites the origin remote
 #     URL to embed ITS token (`https://x-access-token:<token>@github.com/...`,
 #     logged as "Updated remote URL with authentication token") after checkout,
-#     whatever the checkout did. So `--disallowedTools` denies `Read(.git/**)`;
+#     whatever the checkout did, and a step just before the action writes
+#     the same URL early so private repos work (issue #512). So
+#     `--disallowedTools` denies `Read(.git/**)`;
 #   - the token also sits in the Claude process's environment, which is how
 #     `gh pr comment` authenticates, so `Read(//proc/**)` is denied as well.
 # Do not remove any of these without removing `Read`: they are safe together
@@ -231,6 +233,27 @@ jobs:
           # that had nothing to classify.
           echo "threads: $(jq '.data.repository.pullRequest.reviewThreads.nodes | length' "${GITHUB_WORKSPACE}/.prior-review.json")"
 
+      # The action fetches the PR branch BEFORE it authenticates origin.
+      # `track_progress: true` selects tag mode, whose `setupBranch` runs
+      # `git fetch origin <headRefName>` (or `pull/N/head` for a fork), and only
+      # then does `configureGitAuth` write the token into the origin URL. With
+      # `persist-credentials: false` that fetch is anonymous: fine on this
+      # public repo, `could not read Username` on a private one, which is how a
+      # vendored copy in taskless/marketing died (issue #512, checked against
+      # v1.0.221). This writes the value the action writes moments later, from
+      # the same token as `github_token`, so it exposes nothing new and
+      # `Read(.git/**)` already covers it.
+      #
+      # Do NOT fix this with `persist-credentials: true`. Checkout v7 keeps its
+      # credential in a `RUNNER_TEMP` file loaded through `includeIf.gitdir:*`;
+      # the action's `replaceCheckoutCredentials` only strips `include.path`
+      # entries, so that credential would survive, outside `.git/**`, where the
+      # `Read` deny rule does not reach.
+      - name: Authenticate origin for the action's PR-branch fetch
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: git remote set-url origin "https://x-access-token:${GH_TOKEN}@${GITHUB_SERVER_URL#https://}/${GITHUB_REPOSITORY}.git"
+
       # The 👀 on the triggering comment is not from this workflow or the
       # action. The Claude GitHub App adds it, as `claude[bot]`, to ANY comment
       # mentioning `@claude`, before this workflow has started (measured on PR