feat: split triage and pr-fixer into separate GHA workflows

Ambient Code Bot · claude · Ambient Code Bot · commit 0dbdde871acf · 2026-04-02T16:12:03.000-05:00
- triage.yml: daily cron (8am UTC weekdays), discovers untriaged
  Jira/GH issues, creates Implement sessions (max 5/cycle)
- pr-fixer.yml: 30 min cron, manages all ai-managed PRs with
  session reuse, change detection, and circuit breaker
  - fix-single: triggered only by @ambient-fix comment
  - fix-batch: scheduled + manual dispatch
- Both use ambient-action@v0.0.3 with inactivity timeout (60s)
- Unified label: ai-managed (was agent-managed)
- Removed old matrix-based batch pattern
- Skip Jira gracefully if MCP tool unavailable

Co-Authored-By: Claude Opus 4.6 (1M context) &lt;noreply@anthropic.com&gt;
diff --git a/.github/workflows/pr-fixer.yml b/.github/workflows/pr-fixer.yml
@@ -1,25 +1,16 @@
 name: PR Fixer
 
 on:
-  # Immediate: when a PR gets the agent-managed label
-  pull_request:
-    types: [labeled]
-
   # Immediate: when someone comments @ambient-fix on a PR
   issue_comment:
     types: [created]
 
-  # Cadence: every 4 hours including overnight
+  # Cadence: every 30 minutes on weekdays
   schedule:
-    - cron: '0 */4 * * 1-5'
+    - cron: '*/30 * * * 1-5'
 
-  # Manual: for one-off fixes
+  # Manual: for batch run
   workflow_dispatch:
-    inputs:
-      pr_number:
-        description: 'PR number to fix'
-        required: true
-        type: number
 
 permissions:
   contents: read
@@ -29,29 +20,18 @@ jobs:
   # -- Single PR: triggered by label, comment, or manual dispatch --
   fix-single:
     if: >-
-      (github.event_name == 'pull_request'
-       && github.event.label.name == 'agent-managed'
-       && !github.event.pull_request.head.repo.fork)
-      || github.event_name == 'workflow_dispatch'
-      || (github.event_name == 'issue_comment'
-          && github.event.issue.pull_request
-          && contains(github.event.comment.body, '@ambient-fix')
-          && (github.event.comment.author_association == 'MEMBER'
-              || github.event.comment.author_association == 'OWNER'
-              || github.event.comment.author_association == 'COLLABORATOR'))
+      github.event_name == 'issue_comment'
+      && github.event.issue.pull_request
+      && contains(github.event.comment.body, '@ambient-fix')
+      && (github.event.comment.author_association == 'MEMBER'
+          || github.event.comment.author_association == 'OWNER'
+          || github.event.comment.author_association == 'COLLABORATOR')
     runs-on: ubuntu-latest
     timeout-minutes: 30
     steps:
       - name: Resolve PR number
         id: pr
-        run: |
-          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
-            echo "number=${{ inputs.pr_number }}" >> $GITHUB_OUTPUT
-          elif [ "${{ github.event_name }}" = "issue_comment" ]; then
-            echo "number=${{ github.event.issue.number }}" >> $GITHUB_OUTPUT
-          else
-            echo "number=${{ github.event.pull_request.number }}" >> $GITHUB_OUTPUT
-          fi
+        run: echo "number=${{ github.event.issue.number }}" >> $GITHUB_OUTPUT
 
       - name: Check PR is not a fork (issue_comment)
         if: github.event_name == 'issue_comment'
@@ -67,26 +47,53 @@ jobs:
             echo "skip=false" >> $GITHUB_OUTPUT
           fi
 
+      - name: Check for existing session
+        if: steps.fork_check.outputs.skip != 'true'
+        id: existing
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          # Read PR body and extract session_id from frontmatter
+          BODY=$(gh pr view ${{ steps.pr.outputs.number }} --repo "${{ github.repository }}" --json body --jq '.body')
+          SESSION_ID=$(echo "$BODY" | grep -oP 'acp:session_id=\K[^ ]+' | head -1 || echo "")
+          SOURCE=$(echo "$BODY" | grep -oP 'source=\K[^ ]+' | head -1 || echo "")
+          echo "session_id=$SESSION_ID" >> $GITHUB_OUTPUT
+          echo "source=$SOURCE" >> $GITHUB_OUTPUT
+          if [ -n "$SESSION_ID" ]; then
+            echo "Found existing session: $SESSION_ID"
+          else
+            echo "No existing session found — will create new one"
+          fi
+
       - name: Fix PR
         if: steps.fork_check.outputs.skip != 'true'
         id: session
-        uses: ambient-code/ambient-action@v0.0.2
+        uses: ambient-code/ambient-action@v0.0.3
         with:
           api-url: ${{ secrets.AMBIENT_API_URL }}
           api-token: ${{ secrets.AMBIENT_BOT_TOKEN }}
           project: ${{ secrets.AMBIENT_PROJECT }}
-          prompt: >-
-            Fix PR #${{ steps.pr.outputs.number }} in https://github.com/${{ github.repository }}.
-            Fetch all PR data, rebase to resolve conflicts, evaluate reviewer
-            comments (fix valid issues, respond to invalid ones), run lints
-            and tests, and push the fixes.
+          prompt: |
+            You are maintaining open pull request #${{ steps.pr.outputs.number }} in https://github.com/${{ github.repository }}.
+
+            ## Instructions
+
+            1. Check out the PR branch.
+            2. Assess the current state:
+               - Are there merge conflicts? Resolve them.
+               - Is CI failing? Read the logs and fix the failures.
+               - Are there review comments (human or bot like CodeRabbit)? Address each comment.
+            3. Push fixes to the PR branch.
+            4. Ensure the PR body contains this frontmatter as the first line
+               (read your session ID from the AGENTIC_SESSION_NAME environment variable):
+               <!-- acp:session_id=$AGENTIC_SESSION_NAME source=${{ steps.existing.outputs.source || 'unknown' }} last_action=<ISO8601_NOW> retry_count=0 -->
+            5. Do not merge. Do not close. Do not force-push.
+            6. If the PR is fundamentally broken beyond repair, add a comment explaining the situation and stop.
           repos: >-
             [{"url": "https://github.com/${{ github.repository }}", "branch": "main"}]
-          workflow: >-
-            {"gitUrl": "https://github.com/ambient-code/workflows", "branch": "main", "path": "internal-workflows/pr-fixer"}
-          model: claude-sonnet-4-5
+          model: claude-opus-4-6
           wait: 'true'
-          timeout: '25'
+          timeout: '60'
 
       - name: Session summary
         if: always() && steps.fork_check.outputs.skip != 'true'
@@ -105,106 +112,88 @@ jobs:
             echo "- **Status**: Failed to create session" >> $GITHUB_STEP_SUMMARY
           fi
 
-  # -- Batch: scheduled cadence for all agent-managed PRs --
+  # -- Batch: scheduled cadence for all ai-managed PRs --
   fix-batch:
-    if: github.event_name == 'schedule'
-    runs-on: ubuntu-latest
-    timeout-minutes: 10
-    outputs:
-      pr_numbers: ${{ steps.find.outputs.pr_numbers }}
-    steps:
-      - name: Find agent-managed PRs that need work
-        id: find
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          # Get all open PRs with agent-managed label, including last committer
-          PRS=$(gh pr list --repo "${{ github.repository }}" \
-            --label "agent-managed" \
-            --state open \
-            --json number,updatedAt,isCrossRepository \
-            --limit 20)
-
-          # Filter: updated in the last 4 hours, not a fork,
-          # and last update wasn't by the fixer bot itself
-          NEEDS_WORK=$(echo "$PRS" | python3 -c "
-          import json, sys
-          from datetime import datetime, timezone, timedelta
-
-          prs = json.load(sys.stdin)
-          cutoff = datetime.now(timezone.utc) - timedelta(hours=4)
-          active = []
-          for pr in prs:
-              # Skip fork PRs
-              if pr.get('isCrossRepository', False):
-                  continue
-              updated = pr.get('updatedAt', '')
-              if updated:
-                  try:
-                      dt = datetime.fromisoformat(updated.replace('Z', '+00:00'))
-                      if dt > cutoff:
-                          active.append(pr['number'])
-                  except Exception:
-                      active.append(pr['number'])
-          print(json.dumps(active))
-          ")
-
-          echo "pr_numbers=$NEEDS_WORK" >> $GITHUB_OUTPUT
-          COUNT=$(echo "$NEEDS_WORK" | python3 -c "import json,sys; print(len(json.load(sys.stdin)))")
-          echo "Found $COUNT agent-managed PRs with recent activity"
-
-  fix-each:
-    needs: fix-batch
-    if: needs.fix-batch.outputs.pr_numbers != '[]'
+    if: github.event_name == 'schedule' || github.event_name == 'workflow_dispatch'
     runs-on: ubuntu-latest
-    timeout-minutes: 30
-    strategy:
-      matrix:
-        pr_number: ${{ fromJson(needs.fix-batch.outputs.pr_numbers) }}
+    timeout-minutes: 45
     steps:
-      - name: Skip if last update was by the fixer
-        id: churn_check
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          # Check if the most recent comment is from the fixer bot
-          LAST_AUTHOR=$(gh api "repos/${{ github.repository }}/issues/${{ matrix.pr_number }}/comments" \
-            --jq 'last | .body // ""' 2>/dev/null || echo "")
-          if echo "$LAST_AUTHOR" | grep -q '<!-- pr-fixer-bot -->'; then
-            echo "Last activity was fixer bot — skipping"
-            echo "skip=true" >> $GITHUB_OUTPUT
-          else
-            echo "skip=false" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Fix PR #${{ matrix.pr_number }}
-        if: steps.churn_check.outputs.skip != 'true'
+      - name: Run PR fixer orchestrator
         id: session
-        uses: ambient-code/ambient-action@v0.0.2
+        uses: ambient-code/ambient-action@v0.0.3
         with:
           api-url: ${{ secrets.AMBIENT_API_URL }}
           api-token: ${{ secrets.AMBIENT_BOT_TOKEN }}
           project: ${{ secrets.AMBIENT_PROJECT }}
-          prompt: >-
-            Fix PR #${{ matrix.pr_number }} in https://github.com/${{ github.repository }}.
-            Fetch all PR data, rebase to resolve conflicts, evaluate reviewer
-            comments (fix valid issues, respond to invalid ones), run lints
-            and tests, and push the fixes.
+          prompt: |
+            You are a PR fixer orchestrator. Manage all open ai-managed PRs.
+
+            ## Find PRs
+
+            Run: gh pr list --repo ${{ github.repository }} --state open --label ai-managed --search "draft:false"
+
+            For each PR:
+
+            ## 1. Read frontmatter
+            Parse <!-- acp:session_id=<ID> source=<KEY> last_action=<TIMESTAMP> retry_count=<N> --> from the PR body.
+            If no frontmatter: create a new Fix PR session, update the PR body with frontmatter. Continue.
+
+            ## 2. Circuit breaker
+            If retry_count >= 3: comment "AI was unable to resolve after 3 attempts. Needs human attention.",
+            add ai-needs-human label, remove ai-managed label. Skip.
+
+            ## 3. Check for changes since last_action
+            Ignore commits authored by the bot. Only look for:
+            - New commits by someone other than the bot
+            - New or updated review comments
+            - New CI failures
+            - Merge conflicts from base branch changes
+
+            If nothing changed → skip entirely.
+
+            ## 4. Something changed — act
+            - CI failing → send message to existing session with CI logs
+            - New review comments → send message with the comments
+            - Merge conflicts → send message to rebase
+            - New external commits → send message to review and ensure CI passes
+
+            ## 5. Session management
+            Before sending a message, check session status:
+            - Running → send the message
+            - Stopped → restart (reuse), then send
+            - Not found → create new session with this prompt:
+              "You are maintaining an open pull request.
+              PR: <URL>  Source issue: <KEY> (if known)
+              1. Check out the PR branch.
+              2. Resolve merge conflicts, fix CI failures, address review comments.
+              3. Push fixes. Do not merge/close/force-push.
+              4. Write frontmatter: <!-- acp:session_id=$AGENTIC_SESSION_NAME source=<KEY> last_action=<NOW> retry_count=0 -->
+              5. If broken beyond repair, comment and stop."
+
+            After sending: increment retry_count and update last_action in frontmatter.
+
+            ## Rules
+            - No-op if nothing changed. Do not interact with the session.
+            - Use real ACP session IDs from AGENTIC_SESSION_NAME env var.
+            - Send messages to existing sessions, don't recreate.
+            - Ignore bot's own commits.
+            - Do not merge PRs.
+
+            All child sessions use model claude-opus-4-6.
           repos: >-
             [{"url": "https://github.com/${{ github.repository }}", "branch": "main"}]
-          workflow: >-
-            {"gitUrl": "https://github.com/ambient-code/workflows", "branch": "main", "path": "internal-workflows/pr-fixer"}
-          model: claude-sonnet-4-5
-          wait: 'false'
+          model: claude-opus-4-6
+          wait: 'true'
+          timeout: '60'
 
       - name: Session summary
-        if: always() && steps.churn_check.outputs.skip != 'true'
+        if: always()
         env:
           SESSION_NAME: ${{ steps.session.outputs.session-name }}
           SESSION_UID: ${{ steps.session.outputs.session-uid }}
           SESSION_PHASE: ${{ steps.session.outputs.session-phase }}
         run: |
-          echo "### PR Fixer — #${{ matrix.pr_number }}" >> $GITHUB_STEP_SUMMARY
+          echo "### PR Fixer — Batch" >> $GITHUB_STEP_SUMMARY
           echo "" >> $GITHUB_STEP_SUMMARY
           if [ -n "$SESSION_NAME" ]; then
             echo "- **Session**: \`$SESSION_NAME\`" >> $GITHUB_STEP_SUMMARY
diff --git a/.github/workflows/triage.yml b/.github/workflows/triage.yml
@@ -0,0 +1,83 @@
+name: Issue Triage
+
+on:
+  # Daily: weekdays at 8am UTC
+  schedule:
+    - cron: '0 8 * * 1-5'
+
+  # Manual: for one-off triage runs
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  issues: write
+
+jobs:
+  triage:
+    runs-on: ubuntu-latest
+    timeout-minutes: 45
+    steps:
+      - name: Run triage orchestrator
+        id: session
+        uses: ambient-code/ambient-action@v0.0.3
+        with:
+          api-url: ${{ secrets.AMBIENT_API_URL }}
+          api-token: ${{ secrets.AMBIENT_BOT_TOKEN }}
+          project: ${{ secrets.AMBIENT_PROJECT }}
+          prompt: |
+            You are a triage orchestrator. Find untriaged issues and create sessions to fix them.
+
+            ## Find untriaged items (most recent first, max 5 per cycle)
+
+            Query these sources:
+            - Jira: project = RHOAIENG AND "Team[Team]" = ec74d716-af36-4b3c-950f-f79213d08f71-1917 AND labels NOT IN (ai-triaged) AND type IN (Bug, Story, Task) AND status IN (New, Backlog) ORDER BY created DESC
+            - GitHub Issues: gh issue list --repo ${{ github.repository }} --state open --search "-label:ai-triaged sort:created-desc"
+
+            If the Jira MCP tool is not available, skip Jira and only triage GitHub issues.
+
+            Do NOT auto-triage external PRs.
+
+            ## For each item (up to 5 this cycle)
+
+            1. Check if an open non-draft PR with ai-managed label already references this issue. If so, skip.
+            2. If the issue is unclear, duplicated, or not actionable: comment why, add ai-triaged label, move on.
+            3. If actionable: create a child session (model: claude-opus-4-6) to investigate and fix it.
+               The child session prompt should be:
+               "You are investigating and fixing an issue.
+               Source: <KEY> — <TITLE>
+               URL: <URL>
+               Context: <DESCRIPTION>
+               Instructions:
+               1. Read the issue and understand the problem.
+               2. Explore the codebase to find the relevant code.
+               3. Implement a fix. Write tests if the area has existing test coverage.
+               4. Create a PR with a clear description. Include this frontmatter as the first line of the PR body
+                  (read your session ID from the AGENTIC_SESSION_NAME environment variable):
+                  <!-- acp:session_id=$AGENTIC_SESSION_NAME source=<KEY> last_action=<NOW> retry_count=0 -->
+               5. Add the ai-managed label to the PR.
+               6. Ensure CI passes. If it fails, investigate and fix.
+               7. Do not merge. Leave the PR open for human review."
+            4. Add ai-triaged label AFTER confirming the child session was created.
+            5. Comment on the issue linking to the child session.
+          repos: >-
+            [{"url": "https://github.com/${{ github.repository }}", "branch": "main"}]
+          model: claude-opus-4-6
+          wait: 'true'
+          timeout: '60'
+
+      - name: Session summary
+        if: always()
+        env:
+          SESSION_NAME: ${{ steps.session.outputs.session-name }}
+          SESSION_UID: ${{ steps.session.outputs.session-uid }}
+          SESSION_PHASE: ${{ steps.session.outputs.session-phase }}
+        run: |
+          echo "### Issue Triage" >> $GITHUB_STEP_SUMMARY
+          echo "" >> $GITHUB_STEP_SUMMARY
+          if [ -n "$SESSION_NAME" ]; then
+            echo "- **Session**: \`$SESSION_NAME\`" >> $GITHUB_STEP_SUMMARY
+            echo "- **UID**: \`$SESSION_UID\`" >> $GITHUB_STEP_SUMMARY
+            echo "- **Phase**: \`$SESSION_PHASE\`" >> $GITHUB_STEP_SUMMARY
+          else
+            echo "- **Status**: Failed to create session" >> $GITHUB_STEP_SUMMARY
+          fi
