@plainconceptsplatform/workflows 0.6.1 → 0.16.0

package/loops/workflows/agent-audit.md

@@ -1,213 +1,201 @@
----
-# Managed by @plainconceptsplatform/workflows. Source: loops/workflows/agent-audit.md. Update with `workflows update --force`; consumer edits may be overwritten.
-env:
-  REPO_RULES: "Read-only repository audit. Report only reproducible, actionable defects with evidence. Look for: architectural layer violations, missing tests, security gaps, performance issues, and documentation drift. Do not modify files, commit, push, or run write operations."
-  AUDIT_MARKER: "<!-- agent-audit -->"
-  GIT_AUTHOR_NAME: "github-actions[bot]"
-  GIT_AUTHOR_EMAIL: "github-actions[bot]@users.noreply.github.com"
-  GIT_COMMITTER_NAME: "github-actions[bot]"
-  GIT_COMMITTER_EMAIL: "github-actions[bot]@users.noreply.github.com"
-description: |
-  Read-only repository audit. Finds 5-7 problems, scores each 1-10, files a single issue
-  with all findings listed and the top 3 refined as actionable user stories. The issue is
-  labelled `audit` + `bug` + `refine`, so Refine sizes it and splits a multi-defect report
-  into one estimated issue per finding rather than one pull request that has to fix them all. Replaces .loops/recipes/guardrails-audit-loop.yaml.
-
-  Router-only worker: triggered exclusively via workflow_call from work-router.yml.
-  Contract input: trigger-kind(scheduled|manual).
-  The router owns the weekly Monday schedule and manual dispatch.
-
-  The `skip-if-match` below is backpressure, so the audit stops filing reports while three
-  are still unactioned.
-
-name: "Agent: Audit"
-
-# Shared: network policy only. This workflow owns its Safe Outputs and OpenCode configuration.
-# permissions, engine, model and runs-on cannot be shared , see shared/platform-defaults.md.
-imports:
-  - github/gh-aw/.github/workflows/shared/opencode.md@v0.87.5
-  - shared/platform-defaults.md
-  - shared/opencode-ci.md
-
-on:
-  workflow_call:
-    inputs:
-      trigger-kind:
-        description: "Audit trigger: scheduled or manual"
-        required: false
-        type: string
-        default: manual
-
-  # Rung 1. Do not pile reports on top of unactioned reports.
-  skip-if-match:
-    query: "is:issue is:open label:audit"
-    max: 3
-
-runs-on: agents-arc
-runs-on-slim: agents-arc
-
-secrets:
-  OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
-
-engine:
-  id: opencode
-  version: "1.2.14"
-  env:
-    OPENAI_BASE_URL: https://forge.plainconcepts.com/v1
-
-model: openai/glm-5-3
-
-max-turns: 300
-max-turn-cache-misses: 3000
-max-ai-credits: 5000
-
-permissions: read-all
-
-steps:
-  - name: List what is already tracked
-    uses: ./.github/actions/list-open-issues
-    with:
-      token: ${{ github.token }}
-      repo: ${{ github.repository }}
-
-jobs:
-  conclude:
-    needs: [activation, agent, safe_outputs]
-    if: >
-      needs.agent.result == 'success' &&
-      needs.safe_outputs.result == 'success' &&
-      needs.safe_outputs.outputs.process_safe_outputs_processed_count != '0'
-    runs-on: agents-arc
-    permissions:
-      contents: read
-      issues: write
-    steps:
-      - name: Checkout workflow actions
-        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
-        with:
-          persist-credentials: false
-      - name: Create bot token
-        id: app-token
-        uses: actions/create-github-app-token@bcd2ba49218906704ab6c1aa796996da409d3eb1 # v3.2.0
-        with:
-          client-id: ${{ secrets.BOT_APP_ID }}
-          private-key: ${{ secrets.BOT_PRIVATE_KEY }}
-      # No apply-agent-output here. The safe_outputs job already created the issue and
-      # exposes its number; calling the action with create-issues would file a second copy
-      # of the same report. It was only needed while staged: true suppressed the native
-      # write, and staged is gone.
-      - name: Apply audit labels to created issues
-        if: needs.safe_outputs.outputs.created_issue_number != ''
-        uses: ./.github/actions/add-issue-labels
-        with:
-          token: ${{ steps.app-token.outputs.token }}
-          issue-number: ${{ needs.safe_outputs.outputs.created_issue_number }}
-          labels: |
-            audit
-            bug
-            refine
-
-safe-outputs:
-  # A failed run is already a red run. An issue per failure buries the real backlog
-  # under noise nobody closes.
-  report-failure-as-issue: false
-  threat-detection: false
-  create-issue:
-    max: 1
-
-timeout-minutes: 45
----
-
-1. Call skill("pc-repo-audit"), then run `/repo-audit` as a read-only audit of this
-   repository. Do not modify any file, do not commit, and do not push.
-
-  2. Apply repository documentation and established conventions while auditing. Focus on
-     concrete defects and avoid recommendations that weaken security, tests, or checks.
-     Adhere to ${{ env.REPO_RULES }}.
-
-    From the audit report, find **5 to 7 problems**. For each finding, verify it meets ALL
-   of these criteria before keeping it:
-   - A specific, reproducible problem in a specific file or component.
-   - Has real impact: security risk, data loss, crash, or broken functionality.
-   - Something a developer could pick up and fix without further investigation.
-   Discard anything vague, stylistic, theoretical, or nice-to-have. If you cannot find 5 that
-   meet this bar, file as many as you can. If you find zero, call `noop` and stop , that is a
-   good result.
-
-3. **Score each surviving finding from 1 to 10**, based on:
-   | Factor | Weight |
-   |---|---|
-   | Severity (how bad is the impact?) | high |
-   | Likelihood (how often does it trigger?) | medium |
-   | Blast radius (how many users/components affected?) | medium |
-   | Ease of fix (can it be fixed in one PR?) | low bonus |
-
-   Write the score next to each finding in your reasoning.
-
-4. Read `/tmp/gh-aw/agent/open-issues.json`, which lists every open issue with its title and
-   labels. Discard any finding already tracked there. Exact-title duplicates are rejected for
-   you; your job is the ones worded differently that mean the same thing.
-
-   Open issues only cover what is still open, so a finding reported weeks ago and closed
-   without a fix, or deliberately rejected, would come back every single run. Before scoring,
-   call `memory_smart_search` for prior audits of this repository and drop any finding that
-   was already reported and consciously not acted on. After you have decided what to file,
-   call `memory_save` once with a compact record of this audit: the date, each finding's file
-   and one-line description, and its disposition (filed, already tracked, or previously
-   rejected). Keep it short; it is read by the next audit, not by a person.
-
-   A finding you have reported before and that is genuinely still broken IS worth filing
-   again. What this prevents is re-filing something the team looked at and chose to live
-   with, which is the noise that makes an audit backlog get ignored.
-
-5. Call `create_issue` **once** with title "Audit: <date>". Do NOT specify labels , the
-   conclude job will apply them. The body MUST have two sections:
-
-   **Section 1 , All findings:** A numbered list of every finding (5-7) with its score,
-   file path, and a one-line description. Order by score descending.
-
-   **Section 2 , Top 3 to implement:** Call skill("pc-plan-story") and refine the top 3
-   findings by score into user stories in Mike Cohn's As a / I want to / so that format
-   with Given/When/Then acceptance criteria, edge cases, and likely files to change. Mark
-   this section clearly with a heading like `## Top 3 , To Implement`.
-
-   The issue you file goes to Refine, not straight to implementation. Refine sizes it and,
-   because a report of several unrelated defects across different files is exactly the shape
-   that does not land as one pull request, splits it into one properly estimated issue per
-   finding. Write each finding so it survives that split: self-contained, naming its own
-   files and its own acceptance criteria, never "as above" or "same as finding 2".
-
-6. If nothing met the bar, call `noop` and stop. Filing nothing is the right outcome when
-   the codebase is clean.
-
-## Diagram
-
-```mermaid
-flowchart TD
-    auditStart("Work Router<br/>audit route<br/>(Mondays or dispatch)") --> auditBackpressure
-    auditBackpressure["Backpressure (rung 1)<br/>Fewer than 3 open reports?"] -->|✓| auditTracked
-    auditBackpressure -.->|✗| auditIdle
-    auditTracked("Tracked (rung 3)<br/>Open issue titles to disk") --> auditRun
-    auditRun("Audit<br/>/repo-audit, read-only") -->|✓| auditTriage
-    auditRun -.->|✗| auditFail
-    auditTriage["Triage<br/>Find 5-7 problems, score 1-10, dedupe"] -->|✓| auditPropose
-    auditTriage -.->|nothing found| auditQuiet
-    auditPropose("Propose<br/>Single issue: all findings + top 3 refined") -->|✓| auditReport
-    auditPropose -.->|✗| auditFail
-    auditReport(("Conclude<br/>audit+bug+refine on single issue<br/>Refine sizes and splits it"))
-    auditQuiet(("Quiet<br/>Nothing actionable, nothing proposed"))
-    auditIdle(("Idle<br/>Reports still awaiting action"))
-    auditFail(("Fail<br/>Audit or proposal failed"))
-    classDef start fill:#ffffff,stroke:#172033,stroke-width:2px,color:#172033
-    classDef action fill:#eef0ff,stroke:#554cff,stroke-width:2px,color:#172033
-    classDef decision fill:#fff8e8,stroke:#c75b00,stroke-width:2px,color:#172033
-    classDef idle fill:#202c40,stroke:#738198,stroke-width:2px,color:#ffffff
-    classDef failure fill:#fff0f0,stroke:#ef2929,stroke-width:2px,color:#8b1a2a
-    classDef success fill:#e8f8ec,stroke:#18883c,stroke-width:2px,color:#145a32
-    class auditStart start
-    class auditTracked,auditRun,auditPropose action
-    class auditBackpressure,auditTriage decision
-    class auditQuiet,auditIdle idle
-    class auditFail failure
-    class auditReport success
-```
+---
+# Managed by @plainconceptsplatform/workflows. Source: loops/workflows/agent-audit.md. Update with `workflows update --force`; consumer edits may be overwritten.
+env:
+  REPO_RULES: "Read-only repository audit. Report only reproducible, actionable defects with evidence. Do not modify files, commit, push, or run write operations."
+  # Split out of REPO_RULES, which carried both the read-only discipline above and the list
+  # below. All four consuming repositories had customised the list and none could touch it
+  # without restating the discipline; they are different kinds of rule with different owners.
+  # One line: gh-aw joins a multi-line env value onto a single line when it compiles the lock.
+  AUDIT_FOCUS: "Architectural layer violations and dependencies pointing the wrong way. Missing or misleading tests around behaviour that already shipped. Security gaps: unvalidated input, missing authorization, secrets in code. Performance anti-patterns, N+1 queries in particular. Documentation that no longer matches the code it describes."
+  AUDIT_MARKER: "<!-- agent-audit -->"
+  GIT_AUTHOR_NAME: "github-actions[bot]"
+  GIT_AUTHOR_EMAIL: "github-actions[bot]@users.noreply.github.com"
+  GIT_COMMITTER_NAME: "github-actions[bot]"
+  GIT_COMMITTER_EMAIL: "github-actions[bot]@users.noreply.github.com"
+description: |
+  Read-only repository audit. Finds 5-7 problems, scores each 1-10, files a single issue
+  with all findings listed and the top 3 refined as actionable user stories. The issue is
+  labelled `audit` + `bug` + `refine`, so Refine sizes it and splits a multi-defect report
+  into one estimated issue per finding rather than one pull request that has to fix them all. Replaces .loops/recipes/guardrails-audit-loop.yaml.
+
+  Router-only worker: triggered exclusively via workflow_call from work-router.yml.
+  Contract input: trigger-kind(scheduled|manual).
+  The router owns the weekly Monday schedule and manual dispatch.
+
+  The `skip-if-match` below is backpressure, so the audit stops filing reports while three
+  are still unactioned.
+
+name: "Agent: Audit"
+
+# Shared: network policy only. This workflow owns its Safe Outputs and OpenCode configuration.
+# permissions, engine, model and runs-on cannot be shared , see shared/platform-defaults.md.
+imports:
+  - github/gh-aw/.github/workflows/shared/opencode.md@v0.87.5
+  - shared/platform-defaults.md
+  - shared/opencode-ci.md
+
+on:
+  workflow_call:
+    inputs:
+      trigger-kind:
+        description: "Audit trigger: scheduled or manual"
+        required: false
+        type: string
+        default: manual
+
+  # Rung 1. Do not pile reports on top of unactioned reports.
+  #
+  # `-label:stale-audit` is what keeps this backpressure from becoming a stop. Three reports
+  # nobody ever actioned used to disable the weekly audit for good: the query counted them
+  # forever, the run skipped with a green tick, and no report was ever filed again. The
+  # janitor labels a report `stale-audit` once it has sat open past its budget, which both
+  # frees the slot and lists the report in the "Needs a human" digest. Backpressure now
+  # means "three live reports", not "three reports, ever".
+  skip-if-match:
+    query: "is:issue is:open label:audit -label:stale-audit"
+    max: 3
+
+runs-on: agents-arc
+runs-on-slim: agents-arc
+
+secrets:
+  OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+
+engine:
+  id: opencode
+  version: "1.2.14"
+  env:
+    OPENAI_BASE_URL: https://forge.plainconcepts.com/v1
+
+model: openai/glm-5-3
+
+max-turns: 300
+max-turn-cache-misses: 3000
+max-ai-credits: 5000
+
+permissions: read-all
+
+steps:
+  - name: List what is already tracked
+    uses: ./.github/actions/list-open-issues
+    with:
+      token: ${{ github.token }}
+      repo: ${{ github.repository }}
+
+jobs:
+  conclude:
+    needs: [activation, agent, safe_outputs]
+    if: >
+      needs.agent.result == 'success' &&
+      needs.safe_outputs.result == 'success' &&
+      needs.safe_outputs.outputs.process_safe_outputs_processed_count != '0'
+    runs-on: agents-arc
+    permissions:
+      contents: read
+      issues: write
+    steps:
+      - name: Checkout workflow actions
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          persist-credentials: false
+      - name: Create bot token
+        id: app-token
+        uses: actions/create-github-app-token@bcd2ba49218906704ab6c1aa796996da409d3eb1 # v3.2.0
+        with:
+          client-id: ${{ secrets.BOT_APP_ID }}
+          private-key: ${{ secrets.BOT_PRIVATE_KEY }}
+      # No apply-agent-output here. The safe_outputs job already created the issue and
+      # exposes its number; calling the action with create-issues would file a second copy
+      # of the same report. It was only needed while staged: true suppressed the native
+      # write, and staged is gone.
+      - name: Apply audit labels to created issues
+        if: needs.safe_outputs.outputs.created_issue_number != ''
+        uses: ./.github/actions/add-issue-labels
+        with:
+          token: ${{ steps.app-token.outputs.token }}
+          issue-number: ${{ needs.safe_outputs.outputs.created_issue_number }}
+          labels: |
+            audit
+            bug
+            refine
+
+safe-outputs:
+  # A failed run is already a red run. An issue per failure buries the real backlog
+  # under noise nobody closes.
+  report-failure-as-issue: false
+  threat-detection: false
+  create-issue:
+    max: 1
+
+# The fleet is two machines, so this clock is also how long a stuck run can hold half of it.
+# 240 went on to every worker at once when the provider was slow, which fixed the deaths and
+# made every worker equally expensive to hang. These numbers are per worker: enough headroom
+# for a slow gateway on the work it actually does, and not four hours for a run that reads one
+# issue. Turns remain the guard against a confused agent looping; for a custom model the credit
+# ceiling is models.dev fallback pricing and guards nothing.
+#
+# Sweeps the repository read-only and writes one findings issue. Observed around 45 minutes.
+timeout-minutes: 90
+---
+
+1. Call skill("pc-repo-audit"), then run `/repo-audit` as a read-only audit of this
+   repository. Do not modify any file, do not commit, and do not push.
+
+2. Apply repository documentation and established conventions while auditing. Focus on
+   concrete defects and avoid recommendations that weaken security, tests, or checks.
+   Adhere to ${{ env.REPO_RULES }}. Look for: ${{ env.AUDIT_FOCUS }}
+
+   From the audit report, find **5 to 7 problems**. For each finding, verify it meets ALL
+   of these criteria before keeping it:
+   - A specific, reproducible problem in a specific file or component.
+   - Has real impact: security risk, data loss, crash, or broken functionality.
+   - Something a developer could pick up and fix without further investigation.
+   Discard anything vague, stylistic, theoretical, or nice-to-have. If you cannot find 5 that
+   meet this bar, file as many as you can. If you find zero, call `noop` and stop , that is a
+   good result.
+
+3. **Score each surviving finding from 1 to 10**, based on:
+   | Factor | Weight |
+   |---|---|
+   | Severity (how bad is the impact?) | high |
+   | Likelihood (how often does it trigger?) | medium |
+   | Blast radius (how many users/components affected?) | medium |
+   | Ease of fix (can it be fixed in one PR?) | low bonus |
+
+   Write the score next to each finding in your reasoning.
+
+4. Read `/tmp/gh-aw/agent/open-issues.json`, which lists every open issue with its title and
+   labels. Discard any finding already tracked there. Exact-title duplicates are rejected for
+   you; your job is the ones worded differently that mean the same thing.
+
+   Open issues only cover what is still open, so a finding reported weeks ago and closed
+   without a fix, or deliberately rejected, would come back every single run. Before scoring,
+   call `memory_smart_search` for prior audits of this repository and drop any finding that
+   was already reported and consciously not acted on. After you have decided what to file,
+   call `memory_save` once with a compact record of this audit: the date, each finding's file
+   and one-line description, and its disposition (filed, already tracked, or previously
+   rejected). Keep it short; it is read by the next audit, not by a person.
+
+   A finding you have reported before and that is genuinely still broken IS worth filing
+   again. What this prevents is re-filing something the team looked at and chose to live
+   with, which is the noise that makes an audit backlog get ignored.
+
+5. Call `create_issue` **once** with title "Audit: <date>". Do NOT specify labels , the
+   conclude job will apply them. The body MUST have two sections:
+
+   **Section 1 , All findings:** A numbered list of every finding (5-7) with its score,
+   file path, and a one-line description. Order by score descending.
+
+   **Section 2 , Top 3 to implement:** Call skill("pc-plan-story") and refine the top 3
+   findings by score into user stories in Mike Cohn's As a / I want to / so that format
+   with Given/When/Then acceptance criteria, edge cases, and likely files to change. Mark
+   this section clearly with a heading like `## Top 3 , To Implement`.
+
+   The issue you file goes to Refine, not straight to implementation. Refine sizes it and,
+   because a report of several unrelated defects across different files is exactly the shape
+   that does not land as one pull request, splits it into one properly estimated issue per
+   finding. Write each finding so it survives that split: self-contained, naming its own
+   files and its own acceptance criteria, never "as above" or "same as finding 2".
+
+6. If nothing met the bar, call `noop` and stop. Filing nothing is the right outcome when
+   the codebase is clean.