@plainconceptsplatform/workflows 0.21.0 → 0.23.0

package/loops/actions/classify-route/action.yml

@@ -1,100 +1,100 @@
-# Managed by @plainconceptsplatform/workflows. Source: loops/actions/classify-route/action.yml. Update with workflows update --force; consumer edits may be overwritten.
-name: Classify route
-description: Classify the triggering event into exactly one route, and resolve the issue a gated pull request closes.
-inputs:
-  token:
-    description: GitHub token with pull-requests read access.
-    required: true
-outputs:
-  route:
-    description: The single route this event selects, or `none`.
-    value: ${{ steps.classify.outputs.route }}
-  issue-number:
-    description: Issue number the route operates on.
-    value: ${{ steps.classify.outputs.issue-number }}
-  pr-number:
-    description: Pull request number the route operates on.
-    value: ${{ steps.classify.outputs.pr-number }}
-  linked-issue:
-    description: Issue the pull request closes. Empty for routes that do not target one.
-    value: ${{ steps.linked-issue.outputs.issue }}
-  ci-conclusion:
-    description: Conclusion reported by the CI run that triggered a merge-gate route.
-    value: ${{ steps.classify.outputs.ci-conclusion }}
-  ci-run-id:
-    description: Workflow run ID of that CI run.
-    value: ${{ steps.classify.outputs.ci-run-id }}
-  merge-gate-attempts:
-    description: Failed merge-gate attempts already made against this CI verdict.
-    value: ${{ steps.classify.outputs.merge-gate-attempts }}
-  implement-attempts:
-    description: Implement runs already made for this issue that died before producing an answer.
-    value: ${{ steps.classify.outputs.implement-attempts }}
-  refine-mode:
-    description: first or rerefine.
-    value: ${{ steps.classify.outputs.refine-mode }}
-  triage-mode:
-    description: first or retriage.
-    value: ${{ steps.classify.outputs.triage-mode }}
-  trigger-kind:
-    description: scheduled or manual.
-    value: ${{ steps.classify.outputs.trigger-kind }}
-runs:
-  using: composite
-  steps:
-    - name: Classify the event
-      id: classify
-      shell: bash
-      env:
-        EVENT: ${{ github.event_name }}
-        ACTION: ${{ github.event.action }}
-        LABEL: ${{ github.event.label.name }}
-        ISSUE_LABELS: ${{ toJSON(github.event.issue.labels.*.name) }}
-        ISSUE_STATE: ${{ github.event.issue.state }}
-        EVENT_ISSUE_NUMBER: ${{ github.event.issue.number }}
-        EVENT_PR_NUMBER: ${{ github.event.pull_request.number }}
-        COMMENT_ON_PR: ${{ github.event.issue.pull_request != '' }}
-        COMMENT_SENDER_TYPE: ${{ github.event.comment.user.type }}
-        ACTOR: ${{ github.actor }}
-        RUN_PR_NUMBER: ${{ github.event.workflow_run.pull_requests[0].number }}
-        RUN_CONCLUSION: ${{ github.event.workflow_run.conclusion }}
-        RUN_ID: ${{ github.event.workflow_run.id }}
-        SCHEDULE: ${{ github.event.schedule }}
-        OPERATION: ${{ github.event.inputs.operation }}
-        INPUT_ISSUE_NUMBER: ${{ github.event.inputs.issue-number }}
-        INPUT_PR_NUMBER: ${{ github.event.inputs.pr-number }}
-        INPUT_MODE: ${{ github.event.inputs.mode }}
-        INPUT_CI_CONCLUSION: ${{ github.event.inputs.ci-conclusion }}
-        INPUT_CI_RUN_ID: ${{ github.event.inputs.ci-run-id }}
-        INPUT_ATTEMPTS_SO_FAR: ${{ github.event.inputs.attempts_so_far }}
-        INPUT_TRIGGER_KIND: ${{ github.event.inputs.trigger-kind }}
-      run: |
-        set -euo pipefail
-
-        bash "${GITHUB_ACTION_PATH}/classify-route.sh" | tee "$RUNNER_TEMP/route.env"
-        cat "$RUNNER_TEMP/route.env" >> "$GITHUB_OUTPUT"
-
-        route="$(sed -n 's/^route=//p' "$RUNNER_TEMP/route.env")"
-        reason="$(sed -n 's/^error=//p' "$RUNNER_TEMP/route.env")"
-
-        if [ "$route" = "none" ] && [ -n "$reason" ]; then
-          echo "::notice::No route for this event: $reason"
-        fi
-
-    - name: Resolve the issue the pull request closes
-      id: linked-issue
-      if: steps.classify.outputs.route == 'merge-gate' || steps.classify.outputs.route == 'apply-review'
-      shell: bash
-      env:
-        GH_TOKEN: ${{ inputs.token }}
-        REPO: ${{ github.repository }}
-        PR_NUMBER: ${{ steps.classify.outputs.pr-number }}
-      run: |
-        set -euo pipefail
-
-        issue="$(gh pr view "$PR_NUMBER" --repo "$REPO" \
-          --json closingIssuesReferences \
-          --jq '.closingIssuesReferences[0].number // empty')"
-
-        echo "issue=${issue}" >> "$GITHUB_OUTPUT"
-        echo "Pull request #${PR_NUMBER} closes issue #${issue:-<none>}"
+# Managed by @plainconceptsplatform/workflows. Source: loops/actions/classify-route/action.yml. Update with workflows update --force; consumer edits may be overwritten.
+name: Classify route
+description: Classify the triggering event into exactly one route, and resolve the issue a gated pull request closes.
+inputs:
+  token:
+    description: GitHub token with pull-requests read access.
+    required: true
+outputs:
+  route:
+    description: The single route this event selects, or `none`.
+    value: ${{ steps.classify.outputs.route }}
+  issue-number:
+    description: Issue number the route operates on.
+    value: ${{ steps.classify.outputs.issue-number }}
+  pr-number:
+    description: Pull request number the route operates on.
+    value: ${{ steps.classify.outputs.pr-number }}
+  linked-issue:
+    description: Issue the pull request closes. Empty for routes that do not target one.
+    value: ${{ steps.linked-issue.outputs.issue }}
+  ci-conclusion:
+    description: Conclusion reported by the CI run that triggered a merge-gate route.
+    value: ${{ steps.classify.outputs.ci-conclusion }}
+  ci-run-id:
+    description: Workflow run ID of that CI run.
+    value: ${{ steps.classify.outputs.ci-run-id }}
+  merge-gate-attempts:
+    description: Failed merge-gate attempts already made against this CI verdict.
+    value: ${{ steps.classify.outputs.merge-gate-attempts }}
+  attempts:
+    description: Runs already made for this piece of work that died before producing an answer. Carried by every worker that re-dispatches itself, so the budget is bounded.
+    value: ${{ steps.classify.outputs.attempts }}
+  refine-mode:
+    description: first or rerefine.
+    value: ${{ steps.classify.outputs.refine-mode }}
+  triage-mode:
+    description: first or retriage.
+    value: ${{ steps.classify.outputs.triage-mode }}
+  trigger-kind:
+    description: scheduled or manual.
+    value: ${{ steps.classify.outputs.trigger-kind }}
+runs:
+  using: composite
+  steps:
+    - name: Classify the event
+      id: classify
+      shell: bash
+      env:
+        EVENT: ${{ github.event_name }}
+        ACTION: ${{ github.event.action }}
+        LABEL: ${{ github.event.label.name }}
+        ISSUE_LABELS: ${{ toJSON(github.event.issue.labels.*.name) }}
+        ISSUE_STATE: ${{ github.event.issue.state }}
+        EVENT_ISSUE_NUMBER: ${{ github.event.issue.number }}
+        EVENT_PR_NUMBER: ${{ github.event.pull_request.number }}
+        COMMENT_ON_PR: ${{ github.event.issue.pull_request != '' }}
+        COMMENT_SENDER_TYPE: ${{ github.event.comment.user.type }}
+        ACTOR: ${{ github.actor }}
+        RUN_PR_NUMBER: ${{ github.event.workflow_run.pull_requests[0].number }}
+        RUN_CONCLUSION: ${{ github.event.workflow_run.conclusion }}
+        RUN_ID: ${{ github.event.workflow_run.id }}
+        SCHEDULE: ${{ github.event.schedule }}
+        OPERATION: ${{ github.event.inputs.operation }}
+        INPUT_ISSUE_NUMBER: ${{ github.event.inputs.issue-number }}
+        INPUT_PR_NUMBER: ${{ github.event.inputs.pr-number }}
+        INPUT_MODE: ${{ github.event.inputs.mode }}
+        INPUT_CI_CONCLUSION: ${{ github.event.inputs.ci-conclusion }}
+        INPUT_CI_RUN_ID: ${{ github.event.inputs.ci-run-id }}
+        INPUT_ATTEMPTS_SO_FAR: ${{ github.event.inputs.attempts_so_far }}
+        INPUT_TRIGGER_KIND: ${{ github.event.inputs.trigger-kind }}
+      run: |
+        set -euo pipefail
+
+        bash "${GITHUB_ACTION_PATH}/classify-route.sh" | tee "$RUNNER_TEMP/route.env"
+        cat "$RUNNER_TEMP/route.env" >> "$GITHUB_OUTPUT"
+
+        route="$(sed -n 's/^route=//p' "$RUNNER_TEMP/route.env")"
+        reason="$(sed -n 's/^error=//p' "$RUNNER_TEMP/route.env")"
+
+        if [ "$route" = "none" ] && [ -n "$reason" ]; then
+          echo "::notice::No route for this event: $reason"
+        fi
+
+    - name: Resolve the issue the pull request closes
+      id: linked-issue
+      if: steps.classify.outputs.route == 'merge-gate' || steps.classify.outputs.route == 'apply-review'
+      shell: bash
+      env:
+        GH_TOKEN: ${{ inputs.token }}
+        REPO: ${{ github.repository }}
+        PR_NUMBER: ${{ steps.classify.outputs.pr-number }}
+      run: |
+        set -euo pipefail
+
+        issue="$(gh pr view "$PR_NUMBER" --repo "$REPO" \
+          --json closingIssuesReferences \
+          --jq '.closingIssuesReferences[0].number // empty')"
+
+        echo "issue=${issue}" >> "$GITHUB_OUTPUT"
+        echo "Pull request #${PR_NUMBER} closes issue #${issue:-<none>}"

package/loops/actions/classify-route/classify-route.sh

@@ -28,7 +28,7 @@ is_issue_number() {
 
 classify_route() {
   local route="none" error=""
-  local issue_number="" pr_number="" ci_conclusion="" ci_run_id="" merge_gate_attempts="0" implement_attempts="0"
+  local issue_number="" pr_number="" ci_conclusion="" ci_run_id="" merge_gate_attempts="0" attempts="0"
   local refine_mode="" triage_mode="" trigger_kind=""
 
   case "${EVENT:-}" in
@@ -190,13 +190,12 @@ classify_route() {
           if is_issue_number "${INPUT_ISSUE_NUMBER:-}"; then
             route="${OPERATION}"
             issue_number="${INPUT_ISSUE_NUMBER}"
-            if [ "$OPERATION" = "refine" ]; then
-              refine_mode="${INPUT_MODE:-first}"
-            else
-              # The implement worker re-dispatches itself when a run dies before doing any work,
-              # and carries the count so the budget is bounded.
-              implement_attempts="${INPUT_ATTEMPTS_SO_FAR:-0}"
-            fi
+            # Both, not one or the other. A worker re-dispatches itself when a run dies before
+            # doing any work and carries the count so the budget is bounded; refine also carries
+            # the mode it was started in. Written as an if/else, a refine retry arrived as attempt
+            # zero every time and could never reach the park.
+            [ "$OPERATION" = "refine" ] && refine_mode="${INPUT_MODE:-first}"
+            attempts="${INPUT_ATTEMPTS_SO_FAR:-0}"
           else
             error="operation '${OPERATION}' needs a positive issue-number, got '${INPUT_ISSUE_NUMBER:-}'"
           fi
@@ -206,6 +205,7 @@ classify_route() {
             route="triage"
             issue_number="${INPUT_ISSUE_NUMBER}"
             triage_mode="${INPUT_MODE:-first}"
+            attempts="${INPUT_ATTEMPTS_SO_FAR:-0}"
           else
             error="operation 'triage' needs a positive issue-number, got '${INPUT_ISSUE_NUMBER:-}'"
           fi
@@ -214,6 +214,7 @@ classify_route() {
           if is_issue_number "${INPUT_PR_NUMBER:-}"; then
             route="apply-review"
             pr_number="${INPUT_PR_NUMBER}"
+            attempts="${INPUT_ATTEMPTS_SO_FAR:-0}"
           else
             error="operation 'apply-review' needs a positive pr-number, got '${INPUT_PR_NUMBER:-}'"
           fi
@@ -257,7 +258,7 @@ pr-number=${pr_number}
 ci-conclusion=${ci_conclusion}
 ci-run-id=${ci_run_id}
 merge-gate-attempts=${merge_gate_attempts}
-implement-attempts=${implement_attempts}
+attempts=${attempts}
 refine-mode=${refine_mode}
 triage-mode=${triage_mode}
 trigger-kind=${trigger_kind}

package/loops/actions/decide-agent-retry/action.yml

@@ -0,0 +1,72 @@
+# Managed by @plainconceptsplatform/workflows. Source: loops/actions/decide-agent-retry/action.yml. Update with `workflows update --force`; consumer edits may be overwritten.
+name: Decide agent retry
+description: Say whether an agent job that failed is worth running again, from how long it ran and how many attempts have been made.
+inputs:
+  token:
+    description: GitHub token with actions read access, used to read this run's own jobs.
+    required: true
+  attempts-so-far:
+    description: Attempts already made against this piece of work. Empty counts as none.
+    required: false
+    default: '0'
+  park-at:
+    description: Stop retrying once this many attempts have been made.
+    required: true
+  under-minutes:
+    description: Only a run shorter than this is repeated.
+    required: true
+outputs:
+  retry:
+    description: '"true" when this failure is worth repeating.'
+    value: ${{ steps.decide.outputs.retry }}
+  next:
+    description: The attempt number a retry would be.
+    value: ${{ steps.decide.outputs.next }}
+  minutes:
+    description: How long the agent job ran, or -1 when that could not be read.
+    value: ${{ steps.decide.outputs.minutes }}
+runs:
+  using: composite
+  steps:
+    # A run that died in three minutes produced no answer, so repeating it costs three minutes. A
+    # run that worked for half an hour and then failed produced an answer that was wrong, and
+    # repeating it buys the same wrong answer half an hour later. Duration is what separates them,
+    # and it needs no model and no log parsing to read.
+    #
+    # Written once here because the rule was in the implement worker alone, and refine, triage and
+    # apply-review parked instead -- an Odyssey refine died in three minutes on
+    # `Model 'glm-5-3' not found`, a gateway fault that clears in seconds, and waited on the
+    # janitor's six-hourly sweep. Five copies of sixty lines is how one rule becomes five.
+    - name: Decide whether this failure is worth repeating
+      id: decide
+      shell: bash
+      env:
+        GH_TOKEN: ${{ inputs.token }}
+        REPO: ${{ github.repository }}
+        RUN_ID: ${{ github.run_id }}
+        ATTEMPTS: ${{ inputs.attempts-so-far }}
+        PARK_AT: ${{ inputs.park-at }}
+        UNDER_MINUTES: ${{ inputs.under-minutes }}
+      run: |
+        set -euo pipefail
+        # The agent job belongs to this same run: a called workflow shares the caller's run id.
+        read -r started finished <<<"$(gh api "repos/$REPO/actions/runs/$RUN_ID/jobs?per_page=100" \
+          --jq '[.jobs[] | select(.name | endswith("agent"))] | last // empty
+                | "\(.started_at // "") \(.completed_at // "")"')"
+        minutes=-1
+        if [ -n "${started:-}" ] && [ -n "${finished:-}" ]; then
+          minutes=$(( ( $(date -u -d "$finished" +%s) - $(date -u -d "$started" +%s) ) / 60 ))
+        fi
+        attempts="${ATTEMPTS:-0}"
+        [ -n "$attempts" ] || attempts=0
+        retry=false
+        # An unknown duration is treated as a long run: never retry on a guess.
+        if [ "$minutes" -ge 0 ] && [ "$minutes" -lt "$UNDER_MINUTES" ] && [ "$attempts" -lt "$PARK_AT" ]; then
+          retry=true
+        fi
+        {
+          echo "retry=$retry"
+          echo "next=$((attempts + 1))"
+          echo "minutes=$minutes"
+        } >> "$GITHUB_OUTPUT"
+        echo "agent job ran for ${minutes}m; attempts so far ${attempts}; retry=${retry}"

package/loops/actions/verify-route-matrix/verify-route-matrix.sh

@@ -240,14 +240,14 @@ assert "merge-gate dispatch defaults its attempt count to zero" 0 \
   "$(route_field merge-gate-attempts EVENT=workflow_dispatch OPERATION=merge-gate INPUT_PR_NUMBER=7)"
 assert "merge-gate dispatch forwards the attempt count" 3 \
   "$(route_field merge-gate-attempts EVENT=workflow_dispatch OPERATION=merge-gate INPUT_PR_NUMBER=7 INPUT_ATTEMPTS_SO_FAR=3)"
-# The implement worker re-dispatches itself when a run dies before producing an answer, so the
-# count has to survive the round trip or the budget never advances and the retry never stops.
-assert "implement dispatch defaults its attempt count to zero" 0 \
-  "$(route_field implement-attempts EVENT=workflow_dispatch OPERATION=implement INPUT_ISSUE_NUMBER=42)"
+# A worker re-dispatches itself when a run dies before producing an answer, so the count has to
+# survive the round trip or the budget never advances and the retry never stops. The field is
+# `attempts` for every route now: it was `implement-attempts`, set in the else-branch of a test
+# on refine, so a refine retry arrived as attempt zero and could never reach the park.
+assert "a dispatch defaults its attempt count to zero" 0 \
+  "$(route_field attempts EVENT=workflow_dispatch OPERATION=implement INPUT_ISSUE_NUMBER=42)"
 assert "implement dispatch forwards the attempt count" 2 \
-  "$(route_field implement-attempts EVENT=workflow_dispatch OPERATION=implement INPUT_ISSUE_NUMBER=42 INPUT_ATTEMPTS_SO_FAR=2)"
-assert "a refine dispatch carries no implement attempts" 0 \
-  "$(route_field implement-attempts EVENT=workflow_dispatch OPERATION=refine INPUT_ISSUE_NUMBER=42 INPUT_ATTEMPTS_SO_FAR=2)"
+  "$(route_field attempts EVENT=workflow_dispatch OPERATION=implement INPUT_ISSUE_NUMBER=42 INPUT_ATTEMPTS_SO_FAR=2)"
 assert_route "release dispatch needs no numbers" release \
   EVENT=workflow_dispatch OPERATION=release
 assert_route "reconcile-bot-pr-runs dispatch needs no numbers" reconcile-bot-pr-runs \
@@ -1522,6 +1522,14 @@ if [ -f "$GATE_VALIDATOR" ] && worker_installed merge-gate; then
   gate_case "remediated with two pushes"         invalid    "$(gate_items "$(gate_comment remediated),${gate_push},${gate_push}")" failure low
   gate_case "an assessment carrying a push"      invalid    "$(gate_items "$(gate_comment assessed),${gate_push}")" success low
 
+  # Correctness remediation: the agent found the diff does not satisfy the acceptance
+  # criteria, pushed a fix, and emitted remediated. The validator treats this the same as a
+  # CI-failure or conflict remediation -- one push and the remediated word -- regardless of
+  # what acceptanceCriteriaMet says in the report. The next gate cycle re-evaluates the fix.
+  gate_unmet_remedied='{\"findings\":[{\"verified\":false,\"severity\":\"medium\",\"category\":\"correctness\",\"file\":\"src/app.ts\",\"line\":42,\"finding\":\"criterion X not implemented\"}],\"recoverability\":\"high\",\"acceptanceCriteriaMet\":false,\"confidence\":0.9}'
+  gate_case "remediated for unmet acceptance criteria" \
+                                                 remediated  "$(gate_items "$(gate_comment remediated "$gate_unmet_remedied"),${gate_push}")" success low
+
   # Output from a worker version that predates the disposition table. Applying its vocabulary
   # would merge on a word this validator no longer means the same thing by.
   gate_case "the old merge vocabulary is refused"  invalid  '{"items":[{"type":"add_comment","item_number":7,"body":"<!-- agent-merge-gate -->\\n**Verdict:** merge"}]}' success low
@@ -1940,6 +1948,69 @@ if worker_installed audit; then
   if [ "$AUDIT_EMPTY_OK" -eq 1 ]; then PASS=$((PASS + 1)); else FAIL=$((FAIL + 1)); fi
 fi
 
+echo "── Retrying a run that died early ────────────────────────────────────────"
+
+# A run that died before producing anything is worth repeating; one that worked and then failed
+# produced an answer that was wrong, and repeating it buys the same wrong answer later. Only the
+# implement worker knew that. An Odyssey refine died in three minutes on `Model 'glm-5-3' not
+# found` -- a gateway fault that clears in seconds -- parked, and waited on the janitor's
+# six-hourly sweep.
+RETRY_OK=1
+for route in refine implement triage apply-review; do
+  worker_md="${WORKFLOWS_DIR}/agent-${route}.md"
+  [ -f "$worker_md" ] || continue
+
+  # Both halves. A worker with only the park is the state this fixes; one with only the retry
+  # never stops trying.
+  if [ "$(count -cE "^ *if:.*steps\.decide\.outputs\.retry == 'true'" "$worker_md")" -eq 0 ] ||
+     [ "$(count -cE "^ *if:.*steps\.decide\.outputs\.retry != 'true'" "$worker_md")" -eq 0 ]; then
+    RETRY_OK=0
+    echo "FAIL: agent-${route} does not have both a retry and a park path, so an early failure either never repeats or never stops" >&2
+  fi
+
+  # Each worker re-dispatches its own route. `-f operation=implement` pasted into refine produces
+  # a plausible run doing the wrong work, and nothing anywhere goes red.
+  # Only the retry dispatch: it is the line that also carries the attempt count. implement
+  # dispatches reconcile-bot-pr-runs too, after it opens a pull request, and that is not a retry.
+  # Only the retry dispatch: it is the line that also carries the attempt count. implement
+  # dispatches reconcile-bot-pr-runs too, after it opens a pull request, and that is not a retry.
+  #
+  # Both greps go through `count`. A worker with no retry dispatch matches nothing, grep exits 1,
+  # and under `set -o pipefail` that killed the whole suite after printing the failure above it --
+  # the tally never ran, so a mutation that removed a retry branch looked like it had been caught
+  # when in fact nothing after it had been checked.
+  dispatched=$(count -E 'operation=[a-z-]+.*attempts_so_far' "$worker_md" \
+    | { count -oE 'operation=[a-z-]+' || true; } | sed 's/operation=//' | sort -u | tr '\n' ' ')
+  dispatched="${dispatched% }"
+  if [ -n "$dispatched" ] && [ "$dispatched" != "$route" ]; then
+    RETRY_OK=0
+    echo "FAIL: agent-${route} re-dispatches '${dispatched}' rather than its own route; that run would do the wrong work and still look healthy" >&2
+  fi
+
+  # The count has to survive the round trip or the park is never reached.
+  if ! grep -qE '^      attempts_so_far:' "$worker_md"; then
+    RETRY_OK=0
+    echo "FAIL: agent-${route} does not take attempts_so_far, so every retry arrives as the first and the budget never binds" >&2
+  fi
+  if [ "$(count -cF 'attempts_so_far: ${{ needs.classify.outputs.attempts }}' "$ROUTER_YML")" -eq 0 ]; then
+    RETRY_OK=0
+    echo "FAIL: the router does not pass classify's attempts to its workers" >&2
+  fi
+done
+
+# The classifier carries the count for every dispatched route. refine needs its mode *and* its
+# count: written as an if/else, a refine retry arrived as attempt zero every time.
+for op in refine implement triage; do
+  got=$(route_field attempts EVENT=workflow_dispatch OPERATION="$op" INPUT_ISSUE_NUMBER=42 INPUT_ATTEMPTS_SO_FAR=3)
+  assert "a dispatched ${op} carries its attempt count" 3 "$got"
+done
+assert "a dispatched apply-review carries its attempt count" 3 \
+  "$(route_field attempts EVENT=workflow_dispatch OPERATION=apply-review INPUT_PR_NUMBER=9 INPUT_ATTEMPTS_SO_FAR=3)"
+assert "a dispatched refine still carries its mode alongside the count" rerefine \
+  "$(route_field refine-mode EVENT=workflow_dispatch OPERATION=refine INPUT_ISSUE_NUMBER=42 INPUT_MODE=rerefine INPUT_ATTEMPTS_SO_FAR=3)"
+
+if [ "$RETRY_OK" -eq 1 ]; then PASS=$((PASS + 1)); else FAIL=$((FAIL + 1)); fi
+
 echo "── Ways a run ends with nothing ──────────────────────────────────────────"
 
 # A run can end without a pull request in five ways and they need different answers. One sentence

package/loops/templates/ci/app-ci-dotnet-next.yml

@@ -62,7 +62,7 @@ jobs:
   api:
     name: API (.NET)
     if: github.event_name != 'schedule'
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     timeout-minutes: 30
     services:
       sqlserver:
@@ -123,7 +123,7 @@ jobs:
   mutation-api:
     name: Mutation (.NET, diff)
     if: github.event_name == 'pull_request'
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     timeout-minutes: 25
     defaults:
       run:
@@ -169,7 +169,7 @@ jobs:
   web:
     name: Web (Next.js)
     if: github.event_name != 'schedule'
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     timeout-minutes: 15
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
@@ -200,7 +200,7 @@ jobs:
   mutation-web:
     name: Mutation (web, diff)
     if: github.event_name == 'pull_request'
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     timeout-minutes: 25
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
@@ -241,14 +241,14 @@ jobs:
 
   secret-scan:
     name: Secret scan (TruffleHog)
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - run: docker run --rm -v "${{ github.workspace }}:/repo" "$TRUFFLEHOG_IMAGE" filesystem /repo --only-verified --fail --no-update
 
   deps-and-iac-scan:
     name: Dependencies and IaC (Trivy)
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - uses: aquasecurity/trivy-action@ed142fd0673e97e23eac54620cfb913e5ce36c25 # v0.36.0
@@ -262,14 +262,14 @@ jobs:
 
   sast-scan:
     name: SAST (Semgrep)
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - run: docker run --rm -v "${{ github.workspace }}:/src" -w /src "$SEMGREP_IMAGE" semgrep scan --error --metrics=off --oss-only --disable-version-check --config p/default --config p/security-audit --config p/secrets --config p/csharp --config p/typescript --config p/react --exclude-rule yaml.github-actions.security.github-actions-mutable-action-tag.github-actions-mutable-action-tag
 
   sbom:
     name: SBOM (CycloneDX)
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - uses: anchore/sbom-action@e22c389904149dbc22b58101806040fa8d37a610 # v0
@@ -296,7 +296,7 @@ jobs:
       github.event.pull_request.head.repo.full_name == github.repository &&
       github.event.pull_request.draft == false &&
       github.event.pull_request.user.type == 'Bot'
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     timeout-minutes: 5
     permissions:
       contents: read

package/loops/templates/ci/app-ci-node-monorepo.yml

@@ -38,7 +38,7 @@ jobs:
   lint:
     name: Lint (Biome)
     if: github.event_name != 'schedule'
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - uses: pnpm/action-setup@f40ffcd9367d9f12939873eb1018b921a783ffaa # v4
@@ -51,7 +51,7 @@ jobs:
   web:
     name: Web (Next.js static export)
     if: github.event_name != 'schedule'
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - uses: pnpm/action-setup@f40ffcd9367d9f12939873eb1018b921a783ffaa # v4
@@ -73,7 +73,7 @@ jobs:
   mutation-web:
     name: Mutation (web, diff)
     if: github.event_name == 'pull_request'
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     timeout-minutes: 25
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
@@ -103,7 +103,7 @@ jobs:
     name: E2E (Playwright)
     if: github.event_name != 'schedule'
     needs: web
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - uses: pnpm/action-setup@f40ffcd9367d9f12939873eb1018b921a783ffaa # v4
@@ -128,7 +128,7 @@ jobs:
     name: Desktop (Electron)
     if: github.event_name != 'schedule'
     needs: web
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - uses: pnpm/action-setup@f40ffcd9367d9f12939873eb1018b921a783ffaa # v4
@@ -152,7 +152,7 @@ jobs:
   mobile:
     name: Mobile (Capacitor)
     if: github.event_name != 'schedule'
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - uses: pnpm/action-setup@f40ffcd9367d9f12939873eb1018b921a783ffaa # v4
@@ -171,14 +171,14 @@ jobs:
 
   secret-scan:
     name: Secret scan (TruffleHog)
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - run: docker run --rm -v "${{ github.workspace }}:/repo" "$TRUFFLEHOG_IMAGE" filesystem /repo --only-verified --fail --no-update
 
   deps-and-iac-scan:
     name: Dependencies and IaC (Trivy)
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - uses: aquasecurity/trivy-action@ed142fd0673e97e23eac54620cfb913e5ce36c25 # v0.36.0
@@ -192,14 +192,14 @@ jobs:
 
   sast-scan:
     name: SAST (Semgrep)
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - run: docker run --rm -v "${{ github.workspace }}:/src" -w /src "$SEMGREP_IMAGE" semgrep scan --error --metrics=off --oss-only --disable-version-check --config p/default --config p/security-audit --config p/secrets --config p/typescript --config p/react --exclude-rule yaml.github-actions.security.github-actions-mutable-action-tag.github-actions-mutable-action-tag
 
   sbom:
     name: SBOM (CycloneDX)
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - uses: anchore/sbom-action@e22c389904149dbc22b58101806040fa8d37a610 # v0
@@ -226,7 +226,7 @@ jobs:
       github.event.pull_request.head.repo.full_name == github.repository &&
       github.event.pull_request.draft == false &&
       github.event.pull_request.user.type == 'Bot'
-    runs-on: agents-arc
+    runs-on: RunnerLandingZone
     timeout-minutes: 5
     permissions:
       contents: read

package/loops/templates/release/github-release.yml

@@ -12,7 +12,7 @@ permissions:
 jobs:
   publish:
     name: Publish release
-    runs-on: [self-hosted, linux, agents]
+    runs-on: ubuntu-latest
     timeout-minutes: 10
     steps:
       - name: Create or update GitHub release

package/loops/workflows/agent-apply-review.md

@@ -12,6 +12,18 @@ env:
   STALLED_LABEL: stalled
   PR_PENDING_LABEL: pr-pending
   REVIEW_MARKER: "<!-- agent-apply-review -->"
+  # A run that died before it produced anything is worth repeating; one that worked and then
+  # failed produced an answer that was wrong, and repeating it buys the same wrong answer later.
+  # Duration is what separates them. Odyssey #190 died in three minutes on `Model 'glm-5-3' not
+  # found`, a gateway fault that clears in seconds, and waited on the janitor's six-hourly sweep
+  # because only the implement worker could do this.
+  APPLY_REVIEW_ATTEMPT_MARKER: "<!-- agent-apply-review-attempt -->"
+  MAX_ATTEMPTS: "5"
+  PARK_AT_ATTEMPT: "4"
+  # Ten rather than six: the failure that prompted this took three minutes, and a slower one on a
+  # worse day would fall outside a six-minute window and park for a fault that clears by itself.
+  RETRY_UNDER_MINUTES: "10"
+  RETRY_COMMENT: "That is what a provider outage looks like -- the run ended before it could produce an answer -- so this is being tried again from the start. It is a fresh run rather than a continuation: nothing is carried over from the attempt that failed."
   INCOMPLETE_COMMENT: "Applying the review feedback ended without an outcome. This worker has no retry of its own: it runs again when somebody reviews or comments on the pull request, and the issue is flagged so it is not lost until then."
   ISSUE_CONTEXT_PATH: /tmp/gh-aw/agent/issue-context.json
   GH_AW_ALLOWED_BOTS: "platform-devbox[bot],github-actions[bot]"
@@ -38,6 +50,11 @@ imports:
 on:
   workflow_call:
     inputs:
+      attempts_so_far:
+        description: Runs already made for this work that died before producing an answer. Filled by the worker when it re-dispatches itself, not by people.
+        required: false
+        type: string
+        default: '0'
       pr-number:
         description: Pull request number to apply review feedback on.
         required: true
@@ -285,6 +302,8 @@ jobs:
     permissions:
       contents: read
       issues: write
+      # the retry re-enters through the router, which is a workflow_dispatch
+      actions: write
     steps:
       - name: Checkout workflow actions
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
@@ -296,13 +315,58 @@ jobs:
         with:
           client-id: ${{ secrets.BOT_APP_ID }}
           private-key: ${{ secrets.BOT_PRIVATE_KEY }}
+      - name: Decide whether this failure is worth repeating
+        id: decide
+        uses: ./.github/actions/decide-agent-retry
+        with:
+          token: ${{ github.token }}
+          attempts-so-far: ${{ inputs.attempts_so_far }}
+          park-at: ${{ env.PARK_AT_ATTEMPT }}
+          under-minutes: ${{ env.RETRY_UNDER_MINUTES }}
+      # Recorded before any label moves, so a failure in the steps below leaves a run that can be
+      # counted rather than work released with nothing to show for it.
+      - name: Report the failed attempt
+        if: steps.decide.outputs.retry == 'true'
+        uses: ./.github/actions/create-issue-comment
+        with:
+          token: ${{ steps.app-token.outputs.token }}
+          issue-number: ${{ needs.subject.outputs.issue }}
+          body: |
+            ${{ env.APPLY_REVIEW_ATTEMPT_MARKER }}
+            Attempt ${{ steps.decide.outputs.next }} of ${{ env.MAX_ATTEMPTS }} ended after ${{ steps.decide.outputs.minutes }} minutes, before the run could produce an answer.
+            ${{ env.RETRY_COMMENT }}
+            [View this workflow run](${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }})
+      - name: Release the reservation for the retry
+        if: steps.decide.outputs.retry == 'true'
+        uses: ./.github/actions/remove-issue-labels
+        with:
+          token: ${{ steps.app-token.outputs.token }}
+          issue-number: ${{ needs.subject.outputs.issue }}
+          labels: ${{ env.WORKING_LABEL }}
+      - name: Send the work back through the router
+        if: steps.decide.outputs.retry == 'true'
+        env:
+          GH_TOKEN: ${{ github.token }}
+          REPO: ${{ github.repository }}
+          REF: ${{ github.event.repository.default_branch }}
+          SUBJECT: ${{ inputs.pr-number }}
+          NEXT: ${{ steps.decide.outputs.next }}
+        run: |
+          set -euo pipefail
+          # The provider recovers in seconds, so pause before re-entering rather than dispatching
+          # back into the same outage. The router's own classify and authorize jobs add more.
+          sleep 30
+          gh workflow run work-router.yml --repo "$REPO" --ref "$REF"             -f operation=apply-review -f pr-number="$SUBJECT" -f attempts_so_far="$NEXT"
+          echo "Re-dispatched apply-review for $SUBJECT as attempt $NEXT."
       - name: Release the issue
+        if: steps.decide.outputs.retry != 'true'
         uses: ./.github/actions/remove-issue-labels
         with:
           token: ${{ steps.app-token.outputs.token }}
           issue-number: ${{ needs.subject.outputs.issue }}
           labels: ${{ env.WORKING_LABEL }}
       - name: Flag for human review
+        if: steps.decide.outputs.retry != 'true'
         uses: ./.github/actions/add-issue-labels
         with:
           token: ${{ steps.app-token.outputs.token }}
@@ -311,6 +375,7 @@ jobs:
             ${{ env.REVIEW_LABEL }}
             ${{ env.STALLED_LABEL }}
       - name: Report missing review feedback outcome
+        if: steps.decide.outputs.retry != 'true'
         uses: ./.github/actions/create-issue-comment
         with:
           token: ${{ steps.app-token.outputs.token }}

package/loops/workflows/agent-merge-gate.md

@@ -898,9 +898,35 @@ timeout-minutes: 120
    reassurance about their absence.
 
    **5c. Check the acceptance criteria.** The issue context at `${{ env.ISSUE_CONTEXT_PATH }}`
-   says what this change was supposed to do. Confirm the diff does it. Set
-   `acceptanceCriteriaMet` to false only when you can name a criterion the diff does not
-   satisfy.
+   says what this change was supposed to do. Confirm the diff does it. The double-check the
+   rest of this step asks you to do is the gate's own: a defect an agent did not notice and CI
+   did not catch is a more expensive rollback than a wrong auto-merge.
+
+   Set `acceptanceCriteriaMet` to `false` when you can name a criterion the diff does not
+   satisfy. Name every criterion that is missing or wrong, in `reason` and in `findings` if
+   the gap is a defect (it often is). An empty `findings` array with
+   `acceptanceCriteriaMet: false` is accepted but carries little evidence: prefer one entry
+   per gap, with `category: correctness`, the file and line, and a `suggestedFix`.
+
+   **Correctness remediation.** When `acceptanceCriteriaMet` is `false`, you may fix the
+   gap and push the fix the same way you fix a failed CI run — this is the same
+   `remediated` verdict the merge-conflict and CI-failure paths use, and it re-runs CI and
+   the gate on the new head. To do it:
+
+   - Produce the patch that closes each unmet criterion. Every criterion you named must be
+     addressed by the patch, or the next gate cycle reproduces this verdict.
+   - Run the verification commands below (scoped to the files you changed) before the push.
+   - Push the fix using `push_to_pull_request_branch` (pr_number: ${{ needs.subject.outputs.pr }},
+     branch: the current PR branch), then emit the `add_comment` with
+     **Verdict:** remediated. Set `acceptanceCriteriaMet` to `false` in the report: the
+     criteria were unmet when you reviewed, and the fix is what addresses them. The next
+     cycle validates the fix.
+   - Do not rebase, reset, amend or otherwise rewrite history: the push is fast-forward only.
+
+   If you cannot fix a criterion in one pass — it needs a decision, a question, or a code path
+   you cannot trace — do not push. Report `assessed` with `acceptanceCriteriaMet: false` and no
+   push. The workflow sends the pull request to a human, which is the correct action when the
+   gap is beyond a focused repair.
 
    **5d. Answer the recoverability checklist.** How easy would this be to undo if it were
    wrong? Cite the diff for each answer, and record the ones that fired in
@@ -991,8 +1017,9 @@ timeout-minutes: 120
 
 7. Say which of two things you did, and nothing more.
 
-   - **`remediated`** — CI failed or the branch conflicted, you fixed it, you verified the fix,
-     and you are pushing it. Exactly one `push_to_pull_request_branch` goes with this word.
+   - **`remediated`** — something was wrong (CI failed, the branch conflicted, or the diff did
+     not satisfy the acceptance criteria), you fixed it, you verified the fix, and you are
+     pushing it. Exactly one `push_to_pull_request_branch` goes with this word.
    - **`assessed`** — you reviewed the change and are reporting what you found. No push.
 
    These are the only two words the workflow accepts. You do not write `merge`, `review`,
@@ -1041,6 +1068,12 @@ timeout-minutes: 120
 
    `"findings": []` on a clean change is the expected output, not a failure to do the job.
 
+   When `acceptanceCriteriaMet` is `false` and you are pushing a fix, the verdict must be
+   `remediated`, not `assessed`: the validator refuses an `assessed` verdict that carries a
+   push. Set `acceptanceCriteriaMet` to `false` in the report you push with the fix — it
+   describes the code you reviewed, not the fix you just produced. The next gate cycle
+   re-evaluates the updated diff and sets it to `true` (or finds another gap).
+
    The workflow applies comments, labels, merges, and closures with the App token. Reading the
    repository, running verification commands and delegating a finding to be checked are all part
    of the job. What is restricted is what leaves this run: the only safe outputs you may call are

package/loops/workflows/agent-refine.md

@@ -24,6 +24,18 @@ env:
   # re-running a decision produces the same decision. Created idempotently where it is applied.
   STALLED_LABEL: stalled
   REFINE_MARKER: "<!-- agent-refine -->"
+  # A run that died before it produced anything is worth repeating; one that worked and then
+  # failed produced an answer that was wrong, and repeating it buys the same wrong answer later.
+  # Duration is what separates them. Odyssey #190 died in three minutes on `Model 'glm-5-3' not
+  # found`, a gateway fault that clears in seconds, and waited on the janitor's six-hourly sweep
+  # because only the implement worker could do this.
+  REFINE_ATTEMPT_MARKER: "<!-- agent-refine-attempt -->"
+  MAX_ATTEMPTS: "5"
+  PARK_AT_ATTEMPT: "4"
+  # Ten rather than six: the failure that prompted this took three minutes, and a slower one on a
+  # worse day would fall outside a six-minute window and park for a fault that clears by itself.
+  RETRY_UNDER_MINUTES: "10"
+  RETRY_COMMENT: "That is what a provider outage looks like -- the run ended before it could produce an answer -- so this is being tried again from the start. It is a fresh run rather than a continuation: nothing is carried over from the attempt that failed."
   DRAFT_MARKER: "<!-- agent-refine-draft -->"
   INITIAL_MODE: first
   RESPONSE_MODE: rerefine
@@ -72,6 +84,11 @@ imports:
 on:
   workflow_call:
     inputs:
+      attempts_so_far:
+        description: Runs already made for this work that died before producing an answer. Filled by the worker when it re-dispatches itself, not by people.
+        required: false
+        type: string
+        default: '0'
       issue-number:
         description: Issue number to refine.
         required: true
@@ -369,6 +386,8 @@ jobs:
     permissions:
       contents: read
       issues: write
+      # the retry re-enters through the router, which is a workflow_dispatch
+      actions: write
     steps:
       - name: Checkout workflow actions
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
@@ -378,13 +397,59 @@ jobs:
         with:
           client-id: ${{ secrets.BOT_APP_ID }}
           private-key: ${{ secrets.BOT_PRIVATE_KEY }}
+      - name: Decide whether this failure is worth repeating
+        id: decide
+        uses: ./.github/actions/decide-agent-retry
+        with:
+          token: ${{ github.token }}
+          attempts-so-far: ${{ inputs.attempts_so_far }}
+          park-at: ${{ env.PARK_AT_ATTEMPT }}
+          under-minutes: ${{ env.RETRY_UNDER_MINUTES }}
+      # Recorded before any label moves, so a failure in the steps below leaves a run that can be
+      # counted rather than work released with nothing to show for it.
+      - name: Report the failed attempt
+        if: steps.decide.outputs.retry == 'true'
+        uses: ./.github/actions/create-issue-comment
+        with:
+          token: ${{ steps.app-token.outputs.token }}
+          issue-number: ${{ inputs.issue-number }}
+          body: |
+            ${{ env.REFINE_ATTEMPT_MARKER }}
+            Attempt ${{ steps.decide.outputs.next }} of ${{ env.MAX_ATTEMPTS }} ended after ${{ steps.decide.outputs.minutes }} minutes, before the run could produce an answer.
+            ${{ env.RETRY_COMMENT }}
+            [View this workflow run](${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }})
+      - name: Release the reservation for the retry
+        if: steps.decide.outputs.retry == 'true'
+        uses: ./.github/actions/remove-issue-labels
+        with:
+          token: ${{ steps.app-token.outputs.token }}
+          issue-number: ${{ inputs.issue-number }}
+          labels: ${{ env.WORKING_LABEL }}
+      - name: Send the work back through the router
+        if: steps.decide.outputs.retry == 'true'
+        env:
+          GH_TOKEN: ${{ github.token }}
+          REPO: ${{ github.repository }}
+          REF: ${{ github.event.repository.default_branch }}
+          SUBJECT: ${{ inputs.issue-number }}
+          NEXT: ${{ steps.decide.outputs.next }}
+          MODE: ${{ inputs.mode }}
+        run: |
+          set -euo pipefail
+          # The provider recovers in seconds, so pause before re-entering rather than dispatching
+          # back into the same outage. The router's own classify and authorize jobs add more.
+          sleep 30
+          gh workflow run work-router.yml --repo "$REPO" --ref "$REF"             -f operation=refine -f issue-number="$SUBJECT" -f mode="$MODE" -f attempts_so_far="$NEXT"
+          echo "Re-dispatched refine for $SUBJECT as attempt $NEXT."
       - name: Release the issue
+        if: steps.decide.outputs.retry != 'true'
         uses: ./.github/actions/remove-issue-labels
         with:
           token: ${{ steps.app-token.outputs.token }}
           issue-number: ${{ inputs.issue-number }}
           labels: ${{ env.WORKING_LABEL }}
       - name: Flag for human review
+        if: steps.decide.outputs.retry != 'true'
         uses: ./.github/actions/add-issue-labels
         with:
           token: ${{ steps.app-token.outputs.token }}
@@ -393,6 +458,7 @@ jobs:
             ${{ env.REVIEW_LABEL }}
             ${{ env.STALLED_LABEL }}
       - name: Report missing refinement outcome
+        if: steps.decide.outputs.retry != 'true'
         uses: ./.github/actions/create-issue-comment
         with:
           token: ${{ steps.app-token.outputs.token }}

package/loops/workflows/agent-triage.md

@@ -17,6 +17,18 @@ env:
   STALLED_LABEL: stalled
   REFINE_LABEL: refine
   TRIAGE_MARKER: "<!-- agent-triage -->"
+  # A run that died before it produced anything is worth repeating; one that worked and then
+  # failed produced an answer that was wrong, and repeating it buys the same wrong answer later.
+  # Duration is what separates them. Odyssey #190 died in three minutes on `Model 'glm-5-3' not
+  # found`, a gateway fault that clears in seconds, and waited on the janitor's six-hourly sweep
+  # because only the implement worker could do this.
+  TRIAGE_ATTEMPT_MARKER: "<!-- agent-triage-attempt -->"
+  MAX_ATTEMPTS: "5"
+  PARK_AT_ATTEMPT: "4"
+  # Ten rather than six: the failure that prompted this took three minutes, and a slower one on a
+  # worse day would fall outside a six-minute window and park for a fault that clears by itself.
+  RETRY_UNDER_MINUTES: "10"
+  RETRY_COMMENT: "That is what a provider outage looks like -- the run ended before it could produce an answer -- so this is being tried again from the start. It is a fresh run rather than a continuation: nothing is carried over from the attempt that failed."
   MAX_TRIAGE_ROUNDS: "3"
   INCOMPLETE_COMMENT: "Automated triage ended without an outcome. The triage label remains for a retry."
   SAFE_OUTPUT_COMMENT_PREFIX: "Triage assessment"
@@ -61,6 +73,11 @@ imports:
 on:
   workflow_call:
     inputs:
+      attempts_so_far:
+        description: Runs already made for this work that died before producing an answer. Filled by the worker when it re-dispatches itself, not by people.
+        required: false
+        type: string
+        default: '0'
       issue-number:
         description: Issue number to triage.
         required: true
@@ -295,6 +312,8 @@ jobs:
     permissions:
       contents: read
       issues: write
+      # the retry re-enters through the router, which is a workflow_dispatch
+      actions: write
     steps:
       - name: Checkout workflow actions
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
@@ -304,13 +323,59 @@ jobs:
         with:
           client-id: ${{ secrets.BOT_APP_ID }}
           private-key: ${{ secrets.BOT_PRIVATE_KEY }}
+      - name: Decide whether this failure is worth repeating
+        id: decide
+        uses: ./.github/actions/decide-agent-retry
+        with:
+          token: ${{ github.token }}
+          attempts-so-far: ${{ inputs.attempts_so_far }}
+          park-at: ${{ env.PARK_AT_ATTEMPT }}
+          under-minutes: ${{ env.RETRY_UNDER_MINUTES }}
+      # Recorded before any label moves, so a failure in the steps below leaves a run that can be
+      # counted rather than work released with nothing to show for it.
+      - name: Report the failed attempt
+        if: steps.decide.outputs.retry == 'true'
+        uses: ./.github/actions/create-issue-comment
+        with:
+          token: ${{ steps.app-token.outputs.token }}
+          issue-number: ${{ inputs.issue-number }}
+          body: |
+            ${{ env.TRIAGE_ATTEMPT_MARKER }}
+            Attempt ${{ steps.decide.outputs.next }} of ${{ env.MAX_ATTEMPTS }} ended after ${{ steps.decide.outputs.minutes }} minutes, before the run could produce an answer.
+            ${{ env.RETRY_COMMENT }}
+            [View this workflow run](${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }})
+      - name: Release the reservation for the retry
+        if: steps.decide.outputs.retry == 'true'
+        uses: ./.github/actions/remove-issue-labels
+        with:
+          token: ${{ steps.app-token.outputs.token }}
+          issue-number: ${{ inputs.issue-number }}
+          labels: ${{ env.WORKING_LABEL }}
+      - name: Send the work back through the router
+        if: steps.decide.outputs.retry == 'true'
+        env:
+          GH_TOKEN: ${{ github.token }}
+          REPO: ${{ github.repository }}
+          REF: ${{ github.event.repository.default_branch }}
+          SUBJECT: ${{ inputs.issue-number }}
+          NEXT: ${{ steps.decide.outputs.next }}
+          MODE: ${{ inputs.mode }}
+        run: |
+          set -euo pipefail
+          # The provider recovers in seconds, so pause before re-entering rather than dispatching
+          # back into the same outage. The router's own classify and authorize jobs add more.
+          sleep 30
+          gh workflow run work-router.yml --repo "$REPO" --ref "$REF"             -f operation=triage -f issue-number="$SUBJECT" -f mode="$MODE" -f attempts_so_far="$NEXT"
+          echo "Re-dispatched triage for $SUBJECT as attempt $NEXT."
       - name: Release the issue
+        if: steps.decide.outputs.retry != 'true'
         uses: ./.github/actions/remove-issue-labels
         with:
           token: ${{ steps.app-token.outputs.token }}
           issue-number: ${{ inputs.issue-number }}
           labels: ${{ env.WORKING_LABEL }}
       - name: Flag for human review
+        if: steps.decide.outputs.retry != 'true'
         uses: ./.github/actions/add-issue-labels
         with:
           token: ${{ steps.app-token.outputs.token }}
@@ -319,6 +384,7 @@ jobs:
             ${{ env.REVIEW_LABEL }}
             ${{ env.STALLED_LABEL }}
       - name: Report missing triage outcome
+        if: steps.decide.outputs.retry != 'true'
         uses: ./.github/actions/create-issue-comment
         with:
           token: ${{ steps.app-token.outputs.token }}

package/loops/workflows/work-router.yml

@@ -358,7 +358,7 @@ jobs:
       ci-conclusion: ${{ steps.route.outputs.ci-conclusion }}
       ci-run-id: ${{ steps.route.outputs.ci-run-id }}
       merge-gate-attempts: ${{ steps.route.outputs.merge-gate-attempts }}
-      implement-attempts: ${{ steps.route.outputs.implement-attempts }}
+      attempts: ${{ steps.route.outputs.attempts }}
       refine-mode: ${{ steps.route.outputs.refine-mode }}
       triage-mode: ${{ steps.route.outputs.triage-mode }}
       trigger-kind: ${{ steps.route.outputs.trigger-kind }}
@@ -388,6 +388,7 @@ jobs:
     with:
       issue-number: ${{ needs.classify.outputs.issue-number }}
       mode: ${{ needs.classify.outputs.refine-mode }}
+      attempts_so_far: ${{ needs.classify.outputs.attempts }}
     secrets:
       OPENAI_API_KEY: ${{ secrets.FORGE_API_KEY }}
       CODEX_API_KEY: ${{ secrets.FORGE_API_KEY }}
@@ -481,7 +482,7 @@ jobs:
     permissions: write-all
     with:
       issue-number: ${{ needs.classify.outputs.issue-number }}
-      attempts_so_far: ${{ needs.classify.outputs.implement-attempts }}
+      attempts_so_far: ${{ needs.classify.outputs.attempts }}
     secrets:
       OPENAI_API_KEY: ${{ secrets.FORGE_API_KEY }}
       CODEX_API_KEY: ${{ secrets.FORGE_API_KEY }}
@@ -533,6 +534,7 @@ jobs:
     with:
       issue-number: ${{ needs.classify.outputs.issue-number }}
       mode: ${{ needs.classify.outputs.triage-mode }}
+      attempts_so_far: ${{ needs.classify.outputs.attempts }}
     secrets:
       OPENAI_API_KEY: ${{ secrets.FORGE_API_KEY }}
       CODEX_API_KEY: ${{ secrets.FORGE_API_KEY }}
@@ -552,6 +554,7 @@ jobs:
     with:
       pr-number: ${{ needs.classify.outputs.pr-number }}
       linked-issue: ${{ needs.classify.outputs.linked-issue }}
+      attempts_so_far: ${{ needs.classify.outputs.attempts }}
     secrets:
       OPENAI_API_KEY: ${{ secrets.FORGE_API_KEY }}
       CODEX_API_KEY: ${{ secrets.FORGE_API_KEY }}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@plainconceptsplatform/workflows",
-  "version": "0.21.0",
+  "version": "0.23.0",
   "description": "Install and update Platform GitHub agentic workflows.",
   "keywords": [
     "github-actions",
@@ -25,13 +25,6 @@
     "loops",
     "README.md"
   ],
-  "scripts": {
-    "build": "tsc --project tsconfig.build.json && node ./scripts/copy-loops.mjs",
-    "test": "vitest run",
-    "typecheck": "tsc --noEmit --project tsconfig.json",
-    "release": "pnpm build && pnpm test && npm publish --access public",
-    "prepack": "node ./scripts/copy-loops.mjs"
-  },
   "engines": {
     "node": ">=20.19"
   },
@@ -42,5 +35,11 @@
     "@types/node": "^22.0.0",
     "typescript": "^5.8.0",
     "vitest": "^3.0.0"
+  },
+  "scripts": {
+    "build": "tsc --project tsconfig.build.json && node ./scripts/copy-loops.mjs",
+    "test": "vitest run",
+    "typecheck": "tsc --noEmit --project tsconfig.json",
+    "release": "pnpm build && pnpm test && npm publish --access public"
   }
-}
+}