npm - clud-bug - Versions diffs - 0.6.25 → 0.6.27 - Mend

clud-bug 0.6.25 → 0.6.27

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (5) hide show

package/lib/prompts.js +30 -0
package/package.json +1 -1
package/templates/workflow-py.yml.tmpl +61 -11
package/templates/workflow-ts.yml.tmpl +61 -11
package/templates/workflow.yml.tmpl +140 -17

package/lib/prompts.js CHANGED Viewed

@@ -201,6 +201,36 @@ Rules:
     every file's verdict so a maintainer can verify nothing was
     skipped.
+Mid-review self-check-in (v0.6.27 / §5.5 Layer 3):
+After every 5 tool_uses, write a single-line budget heartbeat as a
+free-text "thinking" message (not a tool call — these don't cost a
+turn) of the form:
+  [budget] files_reviewed=X/N, turns_used=Y/M, pace=ok|behind
+Where:
+  - X / N is the count of files you've meaningfully looked at so far
+    over the total in this PR's diff.
+  - Y / M is your current turn count over max_turns.
+  - pace = "ok" when X / N >= Y / (M - 5). The denominator subtracts the
+    5-turn emit reservation: over the (M - 5) turns available for file
+    review, your file-coverage rate must match where you actually are
+    in the budget. (Don't subtract from Y — that would be saying "I've
+    used Y minus 5 turns" which double-counts the reservation.)
+    pace = "behind" otherwise.
+When pace = "behind", immediately pivot strategy:
+  1. Stop deep-dive analysis on the current file.
+  2. Switch to one-sentence verdicts for every remaining file.
+  3. Keep going through the whole diff — silent skipping is
+     non-negotiable. Cover everything, even if some files only get
+     "no issues found in this file" as their verdict.
+The heartbeat serves two purposes: (a) forces internal pacing — you
+can't drift past budget without noticing; (b) lands in the action's
+streaming output for post-hoc calibration of the per-line cost
+coefficients used by paths-check's Layer 1 estimator.
 Incremental-diff handshake (v0.6.10+) — emit the SHA marker:
 At the very end of the summary (after the Skills-referenced footer,
 on its own line), append:

package/package.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "clud-bug",
-  "version": "0.6.25",
+  "version": "0.6.27",
   "description": "Skill-driven Claude PR review. Ship a brand-voice skill, get brand reviews. Each finding cites the skill that motivated it. CLI installs the workflow + a baseline kit; add more from skills.sh.",
   "homepage": "https://cludbug.dev",
   "bugs": "https://github.com/thrillmade/clud-bug/issues",

package/templates/workflow-py.yml.tmpl CHANGED Viewed

@@ -1,4 +1,4 @@
-# clud-bug-template-version: v11
+# clud-bug-template-version: v12
 name: Clud Bug 🐛 Crawls Your Code
 on:
@@ -53,17 +53,32 @@ jobs:
             } >> "$GITHUB_OUTPUT"
             exit 0
           fi
-          IS_WORKFLOW_ONLY=true
+          # v0.6.26 / 0.0.W² — see workflow.yml.tmpl for design notes.
+          ALL_IN_ALLOWLIST=true
+          HAS_WORKFLOW_CHANGE=false
           while IFS= read -r f; do
             case "$f" in
-              .github/workflows/clud-bug-*.yml) ;;
-              .github/actions/strict-mode-gate/*) ;;
-              *) IS_WORKFLOW_ONLY=false; break ;;
+              .github/workflows/clud-bug-*.yml) HAS_WORKFLOW_CHANGE=true ;;
+              .github/actions/strict-mode-gate/*) HAS_WORKFLOW_CHANGE=true ;;
+              AGENTS.md) ;;
+              .cursorrules|.clinerules|.windsurfrules|.continuerules) ;;
+              .github/copilot-instructions.md) ;;
+              .claude/skills/.clud-bug.json) ;;
+              .claude/skills/critical-issues-only/SKILL.md) ;;
+              .claude/skills/evidence-based-review/SKILL.md) ;;
+              .claude/skills/respect-existing-conventions/SKILL.md) ;;
+              docs/timeline.md|docs/file-structure.md|docs/decisions.md) ;;
+              docs/decisions-branches/*.md) ;;
+              *) ALL_IN_ALLOWLIST=false; break ;;
             esac
           done <<< "$CHANGED"
+          IS_WORKFLOW_ONLY=false
+          if [ "$ALL_IN_ALLOWLIST" = "true" ] && [ "$HAS_WORKFLOW_CHANGE" = "true" ]; then
+            IS_WORKFLOW_ONLY=true
+          fi
           echo "is_workflow_only=$IS_WORKFLOW_ONLY" >> "$GITHUB_OUTPUT"
           if [ "$IS_WORKFLOW_ONLY" = "true" ]; then
-            echo "::notice title=Clud Bug 🐛::Skipping LLM review — workflow-only PR."
+            echo "::notice title=Clud Bug 🐛::Skipping LLM review — clud-bug update output."
             echo "model=$MODEL" >> "$GITHUB_OUTPUT"
             exit 0
           fi
@@ -268,28 +283,63 @@ jobs:
           $CALIBRATION"
-      # Fallback when structured_output is empty (max-retries hit).
+      # v0.6.26 / §5.5 Layer 6 fallback render-from-inlines — see workflow.yml.tmpl for design notes.
       - name: Fallback summary (structured_output empty)
         if: success() && steps.clud-bug-review.outputs.structured_output == ''
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           PR_NUMBER: ${{ github.event.pull_request.number }}
+          REPO: ${{ github.repository }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+          TURNS_ESTIMATED: ${{ needs.paths-check.outputs.turns_estimated }}
+          MAX_TURNS: ${{ needs.paths-check.outputs.max_turns }}
+          FILES_COUNT: ${{ needs.paths-check.outputs.files_count }}
+          LINES_ADDED: ${{ needs.paths-check.outputs.lines_added }}
+          LINES_DELETED: ${{ needs.paths-check.outputs.lines_deleted }}
+          THREADS_COUNT: ${{ needs.paths-check.outputs.threads_count }}
         run: |
           set -euo pipefail
-          gh pr comment "$PR_NUMBER" --body "## 🐛 Clud Bug review
+          INLINES=$(gh api "repos/$REPO/pulls/$PR_NUMBER/comments?per_page=100" \
+            --jq "[.[] | select(.user.login == \"claude[bot]\" and .commit_id == \"$HEAD_SHA\")]")
+          INLINE_COUNT=$(echo "$INLINES" | jq 'length')
+          CRITICAL=$(echo "$INLINES" | jq '[.[] | select(.body | test("🔴"))] | length')
+          MINOR=$(echo "$INLINES" | jq '[.[] | select(.body | test("🟡"))] | length')
+          PREEXISTING=$(echo "$INLINES" | jq '[.[] | select(.body | test("🟣"))] | length')
+          CALIBRATION="<!-- clud-bug-calibration: turns_estimated=$TURNS_ESTIMATED, max_turns=$MAX_TURNS, files=$FILES_COUNT, lines_added=$LINES_ADDED, lines_deleted=$LINES_DELETED, threads=$THREADS_COUNT, structured_output=empty, inline_findings=$INLINE_COUNT -->"
+          if [ "$INLINE_COUNT" -gt 0 ]; then
+            STATUS=""
+            [ "$CRITICAL" -gt 0 ] && STATUS=" — critical findings"
+            gh pr comment "$PR_NUMBER" --body "## 🐛 Clud Bug review${STATUS}
+          **This round:** $CRITICAL critical · $MINOR minor · 0 resolved from prior · 0 still open
+          Found: $CRITICAL 🔴 / $MINOR 🟡 / $PREEXISTING 🟣
+          ⚠️ **Synthetic summary** (v0.6.26 §5.5 Layer 6 fallback). Inline findings above ARE the substantive review.
+          <!-- last-reviewed-sha: $HEAD_SHA -->
+          $CALIBRATION"
+          else
+            gh pr comment "$PR_NUMBER" --body "## 🐛 Clud Bug review
           **This round:** 0 critical · 0 minor · 0 resolved from prior · 0 still open
           Found: 0 🔴 / 0 🟡 / 0 🟣
-          ⚠️ Structured output (\`--json-schema\`) returned empty — likely max-retries hit on schema validation. Investigate the action logs.
+          ⚠️ Structured output (\`--json-schema\`) returned empty AND no inline findings posted. Investigate run logs.
+          Skills referenced: [none]
+          <!-- last-reviewed-sha: $HEAD_SHA -->
-          Skills referenced: [none]"
+          $CALIBRATION"
+          fi
       # Strict-mode gate — composite action; see workflow.yml.tmpl for design notes.
       - name: Strict mode — fail check on critical findings
         if: success()
-        uses: thrillmade/clud-bug/.github/actions/strict-mode-gate@v0.6.25
+        uses: thrillmade/clud-bug/.github/actions/strict-mode-gate@v0.6.27
         with:
           github-token: ${{ secrets.GITHUB_TOKEN }}
           # v0.6.22 / 0.0.O: summary now posted by github-actions[bot].

package/templates/workflow-ts.yml.tmpl CHANGED Viewed

@@ -1,4 +1,4 @@
-# clud-bug-template-version: v11
+# clud-bug-template-version: v12
 name: Clud Bug 🐛 Crawls Your Code
 on:
@@ -53,17 +53,32 @@ jobs:
             } >> "$GITHUB_OUTPUT"
             exit 0
           fi
-          IS_WORKFLOW_ONLY=true
+          # v0.6.26 / 0.0.W² — see workflow.yml.tmpl for design notes.
+          ALL_IN_ALLOWLIST=true
+          HAS_WORKFLOW_CHANGE=false
           while IFS= read -r f; do
             case "$f" in
-              .github/workflows/clud-bug-*.yml) ;;
-              .github/actions/strict-mode-gate/*) ;;
-              *) IS_WORKFLOW_ONLY=false; break ;;
+              .github/workflows/clud-bug-*.yml) HAS_WORKFLOW_CHANGE=true ;;
+              .github/actions/strict-mode-gate/*) HAS_WORKFLOW_CHANGE=true ;;
+              AGENTS.md) ;;
+              .cursorrules|.clinerules|.windsurfrules|.continuerules) ;;
+              .github/copilot-instructions.md) ;;
+              .claude/skills/.clud-bug.json) ;;
+              .claude/skills/critical-issues-only/SKILL.md) ;;
+              .claude/skills/evidence-based-review/SKILL.md) ;;
+              .claude/skills/respect-existing-conventions/SKILL.md) ;;
+              docs/timeline.md|docs/file-structure.md|docs/decisions.md) ;;
+              docs/decisions-branches/*.md) ;;
+              *) ALL_IN_ALLOWLIST=false; break ;;
             esac
           done <<< "$CHANGED"
+          IS_WORKFLOW_ONLY=false
+          if [ "$ALL_IN_ALLOWLIST" = "true" ] && [ "$HAS_WORKFLOW_CHANGE" = "true" ]; then
+            IS_WORKFLOW_ONLY=true
+          fi
           echo "is_workflow_only=$IS_WORKFLOW_ONLY" >> "$GITHUB_OUTPUT"
           if [ "$IS_WORKFLOW_ONLY" = "true" ]; then
-            echo "::notice title=Clud Bug 🐛::Skipping LLM review — workflow-only PR."
+            echo "::notice title=Clud Bug 🐛::Skipping LLM review — clud-bug update output."
             echo "model=$MODEL" >> "$GITHUB_OUTPUT"
             exit 0
           fi
@@ -268,28 +283,63 @@ jobs:
           $CALIBRATION"
-      # Fallback when structured_output is empty (max-retries hit).
+      # v0.6.26 / §5.5 Layer 6 fallback render-from-inlines — see workflow.yml.tmpl for design notes.
       - name: Fallback summary (structured_output empty)
         if: success() && steps.clud-bug-review.outputs.structured_output == ''
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           PR_NUMBER: ${{ github.event.pull_request.number }}
+          REPO: ${{ github.repository }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+          TURNS_ESTIMATED: ${{ needs.paths-check.outputs.turns_estimated }}
+          MAX_TURNS: ${{ needs.paths-check.outputs.max_turns }}
+          FILES_COUNT: ${{ needs.paths-check.outputs.files_count }}
+          LINES_ADDED: ${{ needs.paths-check.outputs.lines_added }}
+          LINES_DELETED: ${{ needs.paths-check.outputs.lines_deleted }}
+          THREADS_COUNT: ${{ needs.paths-check.outputs.threads_count }}
         run: |
           set -euo pipefail
-          gh pr comment "$PR_NUMBER" --body "## 🐛 Clud Bug review
+          INLINES=$(gh api "repos/$REPO/pulls/$PR_NUMBER/comments?per_page=100" \
+            --jq "[.[] | select(.user.login == \"claude[bot]\" and .commit_id == \"$HEAD_SHA\")]")
+          INLINE_COUNT=$(echo "$INLINES" | jq 'length')
+          CRITICAL=$(echo "$INLINES" | jq '[.[] | select(.body | test("🔴"))] | length')
+          MINOR=$(echo "$INLINES" | jq '[.[] | select(.body | test("🟡"))] | length')
+          PREEXISTING=$(echo "$INLINES" | jq '[.[] | select(.body | test("🟣"))] | length')
+          CALIBRATION="<!-- clud-bug-calibration: turns_estimated=$TURNS_ESTIMATED, max_turns=$MAX_TURNS, files=$FILES_COUNT, lines_added=$LINES_ADDED, lines_deleted=$LINES_DELETED, threads=$THREADS_COUNT, structured_output=empty, inline_findings=$INLINE_COUNT -->"
+          if [ "$INLINE_COUNT" -gt 0 ]; then
+            STATUS=""
+            [ "$CRITICAL" -gt 0 ] && STATUS=" — critical findings"
+            gh pr comment "$PR_NUMBER" --body "## 🐛 Clud Bug review${STATUS}
+          **This round:** $CRITICAL critical · $MINOR minor · 0 resolved from prior · 0 still open
+          Found: $CRITICAL 🔴 / $MINOR 🟡 / $PREEXISTING 🟣
+          ⚠️ **Synthetic summary** (v0.6.26 §5.5 Layer 6 fallback). Inline findings above ARE the substantive review.
+          <!-- last-reviewed-sha: $HEAD_SHA -->
+          $CALIBRATION"
+          else
+            gh pr comment "$PR_NUMBER" --body "## 🐛 Clud Bug review
           **This round:** 0 critical · 0 minor · 0 resolved from prior · 0 still open
           Found: 0 🔴 / 0 🟡 / 0 🟣
-          ⚠️ Structured output (\`--json-schema\`) returned empty — likely max-retries hit on schema validation. Investigate the action logs.
+          ⚠️ Structured output (\`--json-schema\`) returned empty AND no inline findings posted. Investigate run logs.
+          Skills referenced: [none]
+          <!-- last-reviewed-sha: $HEAD_SHA -->
-          Skills referenced: [none]"
+          $CALIBRATION"
+          fi
       # Strict-mode gate — composite action; see workflow.yml.tmpl for design notes.
       - name: Strict mode — fail check on critical findings
         if: success()
-        uses: thrillmade/clud-bug/.github/actions/strict-mode-gate@v0.6.25
+        uses: thrillmade/clud-bug/.github/actions/strict-mode-gate@v0.6.27
         with:
           github-token: ${{ secrets.GITHUB_TOKEN }}
           # v0.6.22 / 0.0.O: summary now posted by github-actions[bot].

package/templates/workflow.yml.tmpl CHANGED Viewed

@@ -1,4 +1,4 @@
-# clud-bug-template-version: v11
+# clud-bug-template-version: v12
 name: Clud Bug 🐛 Crawls Your Code
 on:
@@ -18,12 +18,45 @@ jobs:
   # Pre-flight: classify the PR diff to decide (a) whether to skip the
   # LLM review entirely, and (b) which model to route to.
   #
-  # (a) Workflow-only PRs (v0.6.14 / 0.0.W): if every changed file
-  # matches `.github/workflows/clud-bug-*.yml` or
-  # `.github/actions/strict-mode-gate/**`, the LLM call is skipped
-  # entirely — claude-code-action would refuse anyway (self-modification
-  # guard), and template re-renders have no useful review surface.
-  # Skipping converts what were admin-bypass merges into normal ones.
+  # (a) Workflow-only / clud-bug-update PRs (v0.6.14 / 0.0.W;
+  # v0.6.26 / 0.0.W²): the LLM call is skipped when EITHER condition:
+  #
+  #   - Every changed file matches `.github/workflows/clud-bug-*.yml`
+  #     or `.github/actions/strict-mode-gate/**` (original 0.0.W —
+  #     pure workflow PRs that claude-code-action would refuse anyway
+  #     via its self-modification guard).
+  #
+  #   - The PR is a recognized `clud-bug update` propagation: every
+  #     changed file is in the clud-bug-update output allowlist AND a
+  #     workflow file (or strict-mode-gate composite) is part of the
+  #     change. v0.6.26 / 0.0.W² added this branch so the propagation
+  #     dance doesn't require admin-bypass every cycle.
+  #
+  # Allowlist (files `clud-bug update` legitimately touches):
+  #   .github/workflows/clud-bug-*.yml          (workflow re-render)
+  #   .github/actions/strict-mode-gate/*        (composite re-render)
+  #   AGENTS.md                                 (logmind block + clud-bug stanza)
+  #   .cursorrules / .clinerules / .windsurfrules / .continuerules
+  #   .github/copilot-instructions.md
+  #   .claude/skills/.clud-bug.json             (manifest)
+  #   .claude/skills/{critical-issues-only,evidence-based-review,
+  #                   respect-existing-conventions}/SKILL.md (baselines)
+  #   docs/timeline.md / docs/file-structure.md / docs/decisions.md
+  #   docs/decisions-branches/*.md              (logmind side-effects)
+  #
+  # The workflow-change requirement is the SIGNATURE that distinguishes
+  # "agent ran clud-bug update" from "user is editing AGENTS.md by hand."
+  # An AGENTS.md-only PR (no workflow change) still goes through normal
+  # review — this catches the prompt-injection-via-AGENTS.md attack
+  # surface that a naive allowlist would open.
+  #
+  # Safety envelope:
+  #   - Workflow file is still protected by the App-side guard (it just
+  #     never runs in this branch).
+  #   - Files in the allowlist are non-executable; modifying them can't
+  #     grant code execution.
+  #   - strictMode toggle is read from BASE ref so a PR can't disable
+  #     strict-mode on itself.
   #
   # (b) Trivial PRs (v0.6.15 / 0.0.R): if the PR author is a dep-bumping
   # bot (dependabot, renovate) OR the diff is small (<2KB) AND only
@@ -82,18 +115,49 @@ jobs:
             exit 0
           fi
-          # --- (a) workflow-only classifier ---
-          IS_WORKFLOW_ONLY=true
+          # --- (a) workflow-only / clud-bug-update classifier
+          #         (v0.6.14 / 0.0.W; widened in v0.6.26 / 0.0.W²) ---
+          # Two-track check:
+          #   ALL_IN_ALLOWLIST = every changed file is in the
+          #     clud-bug-update output allowlist (see header design notes
+          #     for the full list + rationale).
+          #   HAS_WORKFLOW_CHANGE = at least one workflow-file or
+          #     strict-mode-gate change is present. This is the signature
+          #     that distinguishes "clud-bug update output" from "user
+          #     edited AGENTS.md by hand" — naked AGENTS.md edits go
+          #     through normal review.
+          #
+          # Skip when both are true. Backward-compat with v0.6.14's
+          # `is_workflow_only` output name preserved (semantic widened to
+          # "skip-review-eligible"; the dependent clud-bug-review job
+          # gates on this output without needing a rename).
+          ALL_IN_ALLOWLIST=true
+          HAS_WORKFLOW_CHANGE=false
           while IFS= read -r f; do
             case "$f" in
-              .github/workflows/clud-bug-*.yml) ;;
-              .github/actions/strict-mode-gate/*) ;;
-              *) IS_WORKFLOW_ONLY=false; break ;;
+              .github/workflows/clud-bug-*.yml) HAS_WORKFLOW_CHANGE=true ;;
+              .github/actions/strict-mode-gate/*) HAS_WORKFLOW_CHANGE=true ;;
+              # v0.6.26 / 0.0.W² additions — files clud-bug update produces.
+              AGENTS.md) ;;
+              .cursorrules|.clinerules|.windsurfrules|.continuerules) ;;
+              .github/copilot-instructions.md) ;;
+              .claude/skills/.clud-bug.json) ;;
+              .claude/skills/critical-issues-only/SKILL.md) ;;
+              .claude/skills/evidence-based-review/SKILL.md) ;;
+              .claude/skills/respect-existing-conventions/SKILL.md) ;;
+              # logmind side-effects of `logmind log` on the same commit.
+              docs/timeline.md|docs/file-structure.md|docs/decisions.md) ;;
+              docs/decisions-branches/*.md) ;;
+              *) ALL_IN_ALLOWLIST=false; break ;;
             esac
           done <<< "$CHANGED"
+          IS_WORKFLOW_ONLY=false
+          if [ "$ALL_IN_ALLOWLIST" = "true" ] && [ "$HAS_WORKFLOW_CHANGE" = "true" ]; then
+            IS_WORKFLOW_ONLY=true
+          fi
           echo "is_workflow_only=$IS_WORKFLOW_ONLY" >> "$GITHUB_OUTPUT"
           if [ "$IS_WORKFLOW_ONLY" = "true" ]; then
-            echo "::notice title=Clud Bug 🐛::Skipping LLM review — PR only touches workflow files."
+            echo "::notice title=Clud Bug 🐛::Skipping LLM review — clud-bug update output (workflow + allowlist files, no review surface)."
             echo "model=$MODEL" >> "$GITHUB_OUTPUT"
             exit 0
           fi
@@ -434,22 +498,81 @@ jobs:
       # output after max retries (structured_output is empty). Keeps a
       # bare H2 header so the strict-mode gate sees a comment and falls
       # open (advisory) rather than panicking on a missing summary.
+      # v0.6.26 / §5.5 Layer 6: when structured_output is empty BUT inline
+      # findings were posted before the budget exhausted (tokenomics #21
+      # pattern), scrape the inline findings via gh api and render a
+      # synthetic summary that cites the real findings. Avoids the
+      # "0-findings-shown / 5-inline-posted" failure mode where the
+      # maintainer reading the PR sees a meaningless empty summary even
+      # though the review actually surfaced issues.
+      #
+      # If NO inline findings were posted either, falls through to the
+      # original advisory bare-H2 (legacy v0.6.22 behaviour).
       - name: Fallback summary (structured_output empty)
         if: success() && steps.clud-bug-review.outputs.structured_output == ''
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           PR_NUMBER: ${{ github.event.pull_request.number }}
+          REPO: ${{ github.repository }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+          TURNS_ESTIMATED: ${{ needs.paths-check.outputs.turns_estimated }}
+          MAX_TURNS: ${{ needs.paths-check.outputs.max_turns }}
+          FILES_COUNT: ${{ needs.paths-check.outputs.files_count }}
+          LINES_ADDED: ${{ needs.paths-check.outputs.lines_added }}
+          LINES_DELETED: ${{ needs.paths-check.outputs.lines_deleted }}
+          THREADS_COUNT: ${{ needs.paths-check.outputs.threads_count }}
         run: |
           set -euo pipefail
-          gh pr comment "$PR_NUMBER" --body "## 🐛 Clud Bug review
+          # Layer 6: count inline findings claude[bot] posted on THIS
+          # head SHA via the inline-review-thread endpoint. Filter to
+          # the current SHA so older review-pass findings don't inflate
+          # the count.
+          INLINES=$(gh api "repos/$REPO/pulls/$PR_NUMBER/comments?per_page=100" \
+            --jq "[.[] | select(.user.login == \"claude[bot]\" and .commit_id == \"$HEAD_SHA\")]")
+          INLINE_COUNT=$(echo "$INLINES" | jq 'length')
+          CRITICAL=$(echo "$INLINES" | jq '[.[] | select(.body | test("🔴"))] | length')
+          MINOR=$(echo "$INLINES" | jq '[.[] | select(.body | test("🟡"))] | length')
+          PREEXISTING=$(echo "$INLINES" | jq '[.[] | select(.body | test("🟣"))] | length')
+          # Findings with no emoji prefix (cross-cutting / un-prefixed) — count once.
+          UNPREFIXED=$(( INLINE_COUNT - CRITICAL - MINOR - PREEXISTING ))
+          [ "$UNPREFIXED" -lt 0 ] && UNPREFIXED=0
+          CALIBRATION="<!-- clud-bug-calibration: turns_estimated=$TURNS_ESTIMATED, max_turns=$MAX_TURNS, files=$FILES_COUNT, lines_added=$LINES_ADDED, lines_deleted=$LINES_DELETED, threads=$THREADS_COUNT, structured_output=empty, inline_findings=$INLINE_COUNT -->"
+          if [ "$INLINE_COUNT" -gt 0 ]; then
+            # L6 synthetic summary — cites the real inline findings the
+            # action managed to post before structured-output emit failed.
+            STATUS=""
+            [ "$CRITICAL" -gt 0 ] && STATUS=" — critical findings"
+            gh pr comment "$PR_NUMBER" --body "## 🐛 Clud Bug review${STATUS}
+          **This round:** $CRITICAL critical · $MINOR minor · 0 resolved from prior · 0 still open
+          Found: $CRITICAL 🔴 / $MINOR 🟡 / $PREEXISTING 🟣
+          ⚠️ **Synthetic summary** (v0.6.26 §5.5 Layer 6 fallback) — the action posted $INLINE_COUNT inline finding(s) before the structured-output emit step exhausted its turn budget (or hit schema-validation retries). The inline findings above ARE the substantive review; this summary is reconstructed from them. v0.6.26's Layer 5 (auto-retry on cap-hit) is the more permanent fix; this is the safety net.
+          <!-- last-reviewed-sha: $HEAD_SHA -->
+          $CALIBRATION"
+          else
+            # No inline findings either — legacy bare-H2 advisory. The
+            # action errored before it could post anything substantive.
+            gh pr comment "$PR_NUMBER" --body "## 🐛 Clud Bug review
           **This round:** 0 critical · 0 minor · 0 resolved from prior · 0 still open
           Found: 0 🔴 / 0 🟡 / 0 🟣
-          ⚠️ Structured output (\`--json-schema\`) returned empty — likely max-retries hit on schema validation. Investigate the action logs.
+          ⚠️ Structured output (\`--json-schema\`) returned empty AND no inline findings were posted. The action likely errored before producing review output — investigate the run logs.
+          Skills referenced: [none]
-          Skills referenced: [none]"
+          <!-- last-reviewed-sha: $HEAD_SHA -->
+          $CALIBRATION"
+          fi
       # Strict-mode gate. Fails the check when the BASE ref's manifest
       # has { "strictMode": true } AND the latest clud-bug review's first
@@ -466,7 +589,7 @@ jobs:
       # Letting the action's own failure fail the check is louder and right.
       - name: Strict mode — fail check on critical findings
         if: success()
-        uses: thrillmade/clud-bug/.github/actions/strict-mode-gate@v0.6.25
+        uses: thrillmade/clud-bug/.github/actions/strict-mode-gate@v0.6.27
         with:
           github-token: ${{ secrets.GITHUB_TOKEN }}
           # v0.6.22 / 0.0.O: the summary is now posted by the workflow