@plainconceptsplatform/workflows 1.6.0 → 1.6.11

package/loops/workflows/work-router.yml

@@ -69,7 +69,7 @@ run-name: >-
   && format('Agentics: Merge Train - CI on PR #{0}', inputs.pr-number)
   || github.event_name == 'workflow_dispatch'
   && inputs.operation == 'merge-gate'
-  && format('Agent: Merge Gate - PR #{0}', inputs.pr-number)
+  && format('Agent: Merge Gate - PR #{0}{1}', inputs.pr-number, inputs.ci-conclusion == 'failure' && ' (repair)' || '')
   || github.event_name == 'workflow_dispatch'
   && inputs.operation == 'merge-train'
   && format('Agentics: Merge Train{0}', '')
@@ -160,6 +160,15 @@ env:
   # "true" logs the move the train would make and makes none. For trying the train on a
   # repository before letting it act.
   MERGE_TRAIN_DRY_RUN: "false"
+  # How many pull requests the merge train may have in repair at once. A repair moves nothing on
+  # the default branch, so repairs of different pull requests do not race; only merges are serial.
+  # 1 repairs one pull request at a time.
+  MERGE_TRAIN_REPAIRS: "1"
+  # How many implement runs may work at once. Each queue is FIFO and serial. An issue joins the
+  # queue of its split parent when it has one, so the children of one refine never run side by
+  # side on the code they share; independent issues spread across the queues. 1 is one queue
+  # for everything, as before this existed. Raise it only with a fleet that has the VMs.
+  IMPLEMENT_PARALLELISM: "1"
   # Whether a merged pull request's verification plan is replayed and its screenshots archived on
   # the pull request (evidence.yml). It needs nothing configured: the app starts with the root
   # `pnpm run dev`, and a pull request with no plan costs one API call. Set it to "false" for a
@@ -522,7 +531,29 @@ jobs:
       stack-branch: ${{ steps.stack.outputs.stack-branch }}
       stack-in-flight: ${{ steps.inflight.outputs.stack-in-flight }}
       stack-blocked: ${{ steps.blocked.outputs.stack-blocked }}
+      implement-queue: ${{ steps.queue.outputs.queue }}
     steps:
+      # Which implement queue this issue waits in. The children of one refine share a queue,
+      # keyed on their split parent, because they build on each other's code and two of them at
+      # once would each cut from a main the other is about to change. Anything else is keyed on
+      # its own number. A body that cannot be read keys on the issue: a wrong queue only costs
+      # parallelism, never order within a feature.
+      - name: Pick the implement queue
+        id: queue
+        env:
+          GH_TOKEN: ${{ github.token }}
+          REPO: ${{ github.repository }}
+          ISSUE_NUMBER: ${{ needs.classify.outputs.issue-number }}
+        run: |
+          set -euo pipefail
+          queues="${IMPLEMENT_PARALLELISM:-1}"
+          case "$queues" in ''|*[!0-9]*|0) queues=1 ;; esac
+          body=$(gh issue view "$ISSUE_NUMBER" --repo "$REPO" --json body --jq '.body // ""' 2>/dev/null || echo '')
+          parent=$(printf '%s' "$body" | grep -oE '<!-- split-parent: [0-9]+ -->' | head -1 | grep -oE '[0-9]+' || true)
+          key="${parent:-$ISSUE_NUMBER}"
+          queue=$(( key % queues ))
+          echo "queue=$queue" >> "$GITHUB_OUTPUT"
+          echo "#$ISSUE_NUMBER waits in implement queue $queue of $queues (keyed on #$key)."
       - name: Resolve the feature branch from the stack markers
         id: stack
         env:
@@ -645,8 +676,10 @@ jobs:
     needs: [classify, authorize, check-implement-pr, check-stack-branch]
     if: needs.classify.outputs.route == 'implement' && needs.authorize.outputs.trusted == 'true' && needs.check-implement-pr.outputs.has-open-pr != 'true' && needs.check-stack-branch.outputs.is-stack != 'true'
     uses: ./.github/workflows/agent-implement.lock.yml
+    # One FIFO queue per IMPLEMENT_PARALLELISM slot, chosen by check-stack-branch: the children of
+    # one refine always share a queue, independent issues spread across them.
     concurrency:
-      group: implement-global
+      group: implement-queue-${{ needs.check-stack-branch.outputs.implement-queue || '0' }}
       cancel-in-progress: false
       # FIFO, up to 100 waiting. Without this the default `queue: single` cancels whatever
       # was pending, so labelling six issues at once implements one and drops the rest.
@@ -793,8 +826,12 @@ jobs:
     needs: classify
     if: needs.classify.outputs.route == 'merge-gate'
     uses: ./.github/workflows/agent-merge-gate.lock.yml
+    # A gate that was handed red CI repairs and pushes: it can never merge (the validator blocks a
+    # non-success conclusion), so it moves nothing on the default branch and cannot race another
+    # gate. Those run in a group of their own per pull request, in parallel, while everything that
+    # may merge still waits its turn on the one belt.
     concurrency:
-      group: merge-belt
+      group: ${{ needs.classify.outputs.ci-conclusion == 'failure' && format('merge-repair-{0}', needs.classify.outputs.pr-number) || 'merge-belt' }}
       cancel-in-progress: false
       queue: max
     permissions: write-all
@@ -865,6 +902,7 @@ jobs:
           hold-label: ${{ env.MERGE_TRAIN_HOLD_LABEL }}
           max-gate-attempts: ${{ env.MAX_GATE_ATTEMPTS }}
           max-retries: ${{ env.MERGE_TRAIN_MAX_RETRIES }}
+          max-repairs: ${{ env.MERGE_TRAIN_REPAIRS }}
           dry-run: ${{ env.MERGE_TRAIN_DRY_RUN }}
           ci-run-id: ${{ needs.classify.outputs.ci-run-id }}
           ci-conclusion: ${{ needs.classify.outputs.ci-conclusion }}
@@ -1254,19 +1292,28 @@ jobs:
           set -euo pipefail
 
           # Stale reservation sweep. Labelling ten issues at once labels all ten immediately,
-          # because authorize-bot-work is keyed per issue, while implement-global lets exactly
+          # because authorize-bot-work is keyed per issue, while each implement queue lets exactly
           # one run at a time. A queued run that is cancelled therefore leaves its issue
           # carrying bot-working with no run and no pull request: parked, invisible, and not
           # re-triggerable, because classify refuses an issue that is already reserved.
+          #
+          # The live runs are read by status, not as the repository's latest hundred runs. Every
+          # comment and label is a router run, so a busy repository pushes an implement that has
+          # waited two hours in the queue out of the latest hundred, and the sweep then cleared
+          # bot-working from issues whose runs were alive and about to start (Pliny-Bot #693 and
+          # #707, both still queued from 07:32). Runs that are not finished are few, whatever
+          # their age.
+          active_runs=$(for status in queued in_progress pending waiting requested; do
+              gh api "repos/$REPO/actions/runs?status=$status&per_page=100" --jq '.workflow_runs[]' 2>/dev/null || true
+            done | jq -s '{workflow_runs: .}')
           gh issue list --repo "$REPO" --label bot-working --state open --limit 100 \
             --json number --jq '.[].number' |
             while IFS= read -r issue; do
               [ -n "$issue" ] || continue
 
-              live=$(gh api "repos/$REPO/actions/runs?per_page=100" \
-                --jq "[.workflow_runs[]
+              live=$(jq "[.workflow_runs[]
                   | select(.status == \"queued\" or .status == \"in_progress\" or .status == \"pending\" or .status == \"waiting\")
-                  | select((.display_title + \" \") | test(\"#${issue}[^0-9]\"))] | length")
+                  | select((.display_title + \" \") | test(\"#${issue}[^0-9]\"))] | length" <<<"$active_runs")
               [ "${live:-0}" -eq 0 ] || continue
 
               has_pr=$(gh pr list --repo "$REPO" --state open --json body \

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@plainconceptsplatform/workflows",
-  "version": "1.6.0",
+  "version": "1.6.11",
   "description": "Install and update Platform GitHub agentic workflows.",
   "keywords": [
     "github-actions",
@@ -25,6 +25,13 @@
     "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"
   },
@@ -35,11 +42,5 @@
     "@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"
   }
-}
+}