Merge pull request #5 from Coding-Autopilot-System/hardening/enterprise-audit-20260610

OgeonX-Ai · web-flow · commit 3550b6c53519 · 2026-06-11T13:51:57.000+03:00
Harden autopilot control-plane execution
diff --git a/.github/workflows/autopilot-create-issue.yml b/.github/workflows/autopilot-create-issue.yml
@@ -17,7 +17,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Create or update issue
-        uses: actions/github-script@v7
+        uses: actions/github-script@v9
         with:
           script: |
             const run = context.payload.workflow_run;
diff --git a/.github/workflows/autopilot-docs-daily.yml b/.github/workflows/autopilot-docs-daily.yml
@@ -19,7 +19,7 @@ jobs:
       GH_TOKEN: ${{ secrets.ORG_READ_TOKEN || github.token }}
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Update status docs
         run: |
@@ -41,28 +41,7 @@ jobs:
             table="${table}\n| ${ORG}/${repo} | ${queued} | ${inprogress} | ${blocked} | ${done} |"
           done
 
-          cat > docs/status.md <<EOF
-# Autopilot Status
-
-Last updated: ${ts}
-
-## Repository coverage
-
-${table}
-EOF
-
-          python - <<'PY'
-import re
-from pathlib import Path
-
-ts = """Last updated: ${ts}
-See `docs/status.md` for the full status table."""
-path = Path("README.md")
-text = path.read_text()
-block = "<!-- AUTOPILOT-STATUS:START -->\n" + ts + "\n<!-- AUTOPILOT-STATUS:END -->"
-text = re.sub(r"<!-- AUTOPILOT-STATUS:START -->(.|\n)*?<!-- AUTOPILOT-STATUS:END -->", block, text)
-path.write_text(text)
-PY
+          printf '# Autopilot Status\n\nLast updated: %s\n\n## Repository coverage\n\n%b\n' "${ts}" "${table}" > docs/status.md
 
       - name: Create PR
         run: |
@@ -73,7 +52,7 @@ PY
           fi
           branch="docs/daily-status"
           git checkout -B "${branch}"
-          git add docs/status.md README.md
+          git add docs/status.md
           git commit -m "Update autopilot status"
           git push -f origin "${branch}"
           gh pr create -t "Daily autopilot status update" -b "Automated docs update." --base main --head "${branch}" || true
diff --git a/.github/workflows/autopilot-operator.yml b/.github/workflows/autopilot-operator.yml
@@ -15,8 +15,7 @@ jobs:
     runs-on: [self-hosted, Windows]
     env:
       ORG: ${{ vars.ORG }}
-      GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      GH_TOKEN: ${{ secrets.ORG_AUTOPILOT_TOKEN }}
       OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
     steps:
       - name: Ensure Codex on PATH
@@ -28,7 +27,7 @@ jobs:
           }
 
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           clean: false
 
@@ -44,6 +43,14 @@ jobs:
             exit 1
           }
 
+      - name: Validate org mutation token
+        shell: pwsh
+        run: |
+          if (-not $env:GH_TOKEN) {
+            Write-Host "ORG_AUTOPILOT_TOKEN is not set."
+            exit 1
+          }
+
       - name: Validate Codex auth
         shell: pwsh
         run: |
diff --git a/.github/workflows/autopilot-org-installer.yml b/.github/workflows/autopilot-org-installer.yml
@@ -18,7 +18,7 @@ jobs:
       GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
     steps:
       - name: Checkout installer
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Sync intake workflow
         run: |
@@ -60,24 +60,11 @@ jobs:
             if [ "${opt_in}" != "true" ]; then
               marker="Autopilot Installer"
               existing=$(gh issue list -R ${ORG}/${repo} -s open -S "\"${marker}\"" --json number --jq '.[0].number')
-              body=$(cat <<EOF
-              ${marker}
-
-              This repo does not have the Autopilot intake workflow installed.
-              To opt in, add a file at:
-
-              .autopilot/opt-in
-
-              Optional: specify monitored workflows in:
-              .autopilot/workflows.txt
-
-              Once opt-in is present, Autopilot will open a PR to install the workflow.
-              EOF
-              )
+              body=$(printf '%s\n' "${marker}" "" "This repo does not have the Autopilot intake workflow installed." "To opt in, add a file at:" "" ".autopilot/opt-in" "" "Optional: specify monitored workflows in:" ".autopilot/workflows.txt" "" "Once opt-in is present, Autopilot will open a PR to install the workflow.")
               if [ -n "${existing}" ]; then
                 gh issue comment -R ${ORG}/${repo} "${existing}" -b "${body}"
               else
-                gh issue create -R ${ORG}/${repo} -t "Autopilot intake not installed" -b "${body}" -l "autofix,queued,docs"
+                gh issue create -R ${ORG}/${repo} -t "Autopilot opt-in available" -b "${body}"
               fi
               continue
             fi
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -6,27 +6,22 @@ on:
   pull_request:
     branches: [main]
 
+permissions:
+  contents: read
+
 jobs:
   validate:
     runs-on: ubuntu-latest
+    timeout-minutes: 10
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
+
+      - name: Install workflow validator dependency
+        run: python -m pip install --disable-pip-version-check PyYAML==6.0.2
 
       - name: Validate workflow YAML
-        run: |
-          python -c "
-          import yaml, glob, sys
-          files = glob.glob('.github/workflows/*.yml')
-          errors = []
-          for f in files:
-              try:
-                  with open(f) as fh:
-                      yaml.safe_load(fh)
-                  print('  OK:', f)
-              except yaml.YAMLError as e:
-                  # GitHub Actions parses some heredoc patterns that Python
-                  # yaml.safe_load rejects (heredoc terminators at column 0
-                  # inside block scalars). Warn but do not fail.
-                  print('  WARN:', f, '-', str(e).split('\n')[0])
-          print('Validated', len(files), 'workflow files')
-          "
+        run: python tests/validate_workflows.py
+
+      - name: Validate control-plane contracts
+        shell: pwsh
+        run: ./tests/contract-tests.ps1
diff --git a/.planning/phases/01-enterprise-audit/01-UAT.md b/.planning/phases/01-enterprise-audit/01-UAT.md
@@ -0,0 +1,24 @@
+# Enterprise Control-Plane Audit
+
+Audit source: `gsd-audit-fix --severity all --max 8`
+
+## Findings
+
+| ID | Severity | Classification | Description | File references |
+| --- | --- | --- | --- | --- |
+| F-01 | high | auto-fixable | Unlabeled issues bypass the documented `autofix + queued` authorization gate. | `scripts/autopilot-operator.ps1` |
+| F-02 | high | auto-fixable | Issues with exhausted `try-3` attempts can be selected repeatedly. | `scripts/autopilot-operator.ps1` |
+| F-03 | high | auto-fixable | Untrusted issue and comment text is passed to Codex without prompt-injection boundaries. | `scripts/autopilot-operator.ps1` |
+| F-04 | high | auto-fixable | The operator stages arbitrary generated files without sensitive-path or change-size policy enforcement. | `scripts/autopilot-operator.ps1` |
+| F-05 | high | auto-fixable | Changes without supported verification can still be pushed and marked done. | `scripts/autopilot-operator.ps1` |
+| F-06 | high | auto-fixable | The operator workflow uses the repository-scoped token for org-wide mutations. | `.github/workflows/autopilot-operator.yml`, `docs/runbooks/operator.md` |
+| F-07 | medium | auto-fixable | CI intentionally ignores workflow YAML errors and has no control-plane contract tests. | `.github/workflows/ci.yml`, `tests/contract-tests.ps1` |
+| F-08 | medium | auto-fixable | The installer creates actionable autofix issues for repositories that have not opted in. | `.github/workflows/autopilot-org-installer.yml` |
+| M-01 | high | manual-only | Provisioning a production GitHub App or fine-grained token and org-level RBAC requires administrator action. | Organization settings |
+| M-02 | medium | manual-only | Replacing secret-based Codex auth with workload identity depends on provider support and an architecture decision. | `.github/workflows/autopilot-operator.yml` |
+
+## Verification policy
+
+- Run `pwsh -NoProfile -File tests/contract-tests.ps1` after each applicable fix once the contract suite exists.
+- Run `git diff --check` before every commit.
+- Retain this audit and the final verification report as GSD evidence.
diff --git a/.planning/phases/01-enterprise-audit/01-VERIFICATION.md b/.planning/phases/01-enterprise-audit/01-VERIFICATION.md
@@ -0,0 +1,31 @@
+# Enterprise Control-Plane Verification
+
+Status: passed with manual deployment prerequisites
+
+## Fixed findings
+
+| ID | Status | Evidence |
+| --- | --- | --- |
+| F-01 | fixed | `0733a4b` removes unlabeled issue discovery and promotion. |
+| F-02 | fixed | `0532832` excludes and defensively skips exhausted `try-3` issues. |
+| F-03 | fixed | `35119dc` adds an explicit untrusted-content boundary to Codex prompts. |
+| F-04 | fixed | `36daf3c`, `e34d380` enforce sensitive-path and change-size policies. |
+| F-05 | fixed | `7bbcf3c`, `bc1e0e8` require supported verification by default and document exceptions. |
+| F-06 | fixed | `f15b9f8`, `bc1e0e8` require and document the org mutation token contract. |
+| F-07 | fixed | `26cf4f4`, `f2e6e43` repair workflow YAML, add contract tests, and migrate actions to Node.js 24-compatible majors. |
+| F-08 | fixed | `c7b10fc` makes pre-opt-in installer notices non-actionable. |
+
+## Validation
+
+- `python tests/validate_workflows.py`: passed, 5 workflow files.
+- `powershell -NoProfile -ExecutionPolicy Bypass -File tests/contract-tests.ps1`: passed.
+- `python -m compileall tests`: passed.
+- `yamllint` with GitHub Actions-compatible truthy rule disabled: passed.
+- `git diff --check`: passed.
+- Remote CI run `27290403822`: passed without Node.js runtime deprecation annotations.
+
+## Manual-only findings
+
+- M-01: An administrator must provision `ORG_AUTOPILOT_TOKEN` using a short-lived GitHub App installation token or a fine-grained token with access only to opted-in repositories. Required repository permissions are Contents write, Issues write, and Pull requests write.
+- M-02: Secretless Codex authentication remains an architecture decision dependent on provider workload-identity support.
+- M-03: A sandboxed staging organization and self-hosted Windows runner are required for a live end-to-end test of issue intake, mutation, verification, push, and pull-request creation.
diff --git a/README.md b/README.md
@@ -34,9 +34,10 @@ flowchart LR
 ## Quick start
 
 1. Set org variable `ORG` in GitHub Actions for this repo.
-2. Install `autopilot-create-issue.yml` into target repos, or use `autopilot-org-installer.yml`.
-3. Ensure a self-hosted Windows runner with Codex and `OPENAI_API_KEY` is online.
-4. Trigger `autopilot-operator.yml` manually to validate the setup.
+2. Configure the least-privilege `ORG_AUTOPILOT_TOKEN` secret for opted-in repository mutations.
+3. Install `autopilot-create-issue.yml` into target repos, or use `autopilot-org-installer.yml`.
+4. Ensure a self-hosted Windows runner with Codex and `OPENAI_API_KEY` is online.
+5. Trigger `autopilot-operator.yml` manually to validate the setup.
 
 ## Enterprise proof points
 
@@ -50,7 +51,7 @@ flowchart LR
 - Acts only on issues labeled `autofix + queued`.
 - Skips issues labeled `risky` or `needs-design`.
 - Minimal diffs only - no secrets, no destructive operations.
-- Best-effort verification before PR creation.
+- Required supported verification before PR creation, with explicit approved exceptions only.
 
 ## Workflows
 
diff --git a/docs/runbooks/operator.md b/docs/runbooks/operator.md
@@ -17,11 +17,17 @@ $env:DRY_RUN = "true"
 - `MAX_ISSUES` (default 5)
 - `DRY_RUN` (default false)
 - `BASE_BRANCH_OVERRIDE` (optional)
+- `MAX_CHANGED_FILES` (default 20)
+- `MAX_CHANGED_LINES` (default 1000)
+- `ALLOW_UNVERIFIED` (default false; approved exceptions only)
 
+
+## GitHub authorization
+The scheduled workflow requires an `ORG_AUTOPILOT_TOKEN` secret backed by a short-lived GitHub App installation token or fine-grained token. Grant access only to opted-in repositories with Issues and Pull requests write plus Contents write. The repository-scoped `GITHUB_TOKEN` cannot perform cross-repository control-plane mutations.
 ## Behavior
 - Only processes issues labeled `autofix` + `queued`
 - Skips `risky` and `needs-design`
-- Attempts best-effort verification
+- Requires a supported verification command unless an approved exception sets `ALLOW_UNVERIFIED=true`
 - Opens a PR if changes are safe and tests pass
 
 ## Logging
diff --git a/memory/examples/20260610_f01_label-gate.md b/memory/examples/20260610_f01_label-gate.md
@@ -0,0 +1,8 @@
+﻿# F-01 label-gated intake
+
+Issue Description: Unlabeled issues bypassed the documented automation authorization gate.
+State: The operator searched for all open unlabeled issues and promoted them into the queue.
+Action: Removed unlabeled-issue discovery and automatic promotion.
+Result: Only issues explicitly labeled `autofix` and `queued` are executable.
+Diff Patch: Removed the `no:label` query and unlabeled issue mutation block.
+Rationale: Explicit opt-in labels are a security boundary.
diff --git a/memory/examples/20260610_f02_bounded-attempts.md b/memory/examples/20260610_f02_bounded-attempts.md
@@ -0,0 +1,8 @@
+﻿# F-02 bounded attempts
+
+Issue Description: Issues at `try-3` could be selected and executed repeatedly.
+State: Search and processing logic did not exclude exhausted issues.
+Action: Excluded `try-3` from discovery and added a defensive runtime guard.
+Result: Exhausted issues remain available for human review but are not re-executed.
+Diff Patch: Added `-label:try-3` and an attempt-limit guard.
+Rationale: Bounded retries prevent runaway cost and repeated unsafe mutations.
diff --git a/memory/examples/20260610_f03_prompt-boundary.md b/memory/examples/20260610_f03_prompt-boundary.md
@@ -0,0 +1,8 @@
+﻿# F-03 prompt-injection boundary
+
+Issue Description: Untrusted issue and comment content was mixed with operator instructions.
+State: Repository users could place instruction-like text directly in the Codex prompt.
+Action: Added explicit security policy and untrusted-content delimiters around all issue data.
+Result: The model receives a clear instruction hierarchy and treats issue content as data.
+Diff Patch: Added security rules plus BEGIN/END UNTRUSTED markers.
+Rationale: Prompt boundaries reduce indirect prompt-injection risk at the AI execution boundary.
diff --git a/memory/examples/20260610_f04_change-policy.md b/memory/examples/20260610_f04_change-policy.md
@@ -0,0 +1,8 @@
+﻿# F-04 generated change-set policy
+
+Issue Description: The operator staged every generated file without policy checks.
+State: Sensitive files or unexpectedly large patches could be committed automatically.
+Action: Added changed-file discovery and sensitive-path, file-count, and line-count guards.
+Result: Unsafe change sets fail before verification, staging, push, or PR creation.
+Diff Patch: Added `Assert-SafeChangeSet` and configurable limits.
+Rationale: AI-generated output needs a deterministic enforcement boundary.
diff --git a/memory/examples/20260610_f05_verification-gate.md b/memory/examples/20260610_f05_verification-gate.md
@@ -0,0 +1,8 @@
+﻿# F-05 verification gate
+
+Issue Description: The operator could push changes and mark issues done without running tests.
+State: Unsupported repositories silently used `verification=skipped`.
+Action: Blocked unverified changes by default and documented an explicit exception flag.
+Result: PR creation now requires supported verification unless an operator approves an exception.
+Diff Patch: Added the `ALLOW_UNVERIFIED` gate and runbook configuration.
+Rationale: Verification is a release gate, not advisory metadata.
diff --git a/memory/examples/20260610_f06_org-token.md b/memory/examples/20260610_f06_org-token.md
@@ -0,0 +1,8 @@
+﻿# F-06 org mutation token
+
+Issue Description: The org-wide operator used a repository-scoped `GITHUB_TOKEN`.
+State: Cross-repository clones, pushes, issue edits, and PR creation could not work reliably.
+Action: Required a dedicated least-privilege `ORG_AUTOPILOT_TOKEN` and added fail-fast validation.
+Result: The workflow contract now matches its org-wide mutation responsibilities.
+Diff Patch: Replaced the token binding and documented required authorization.
+Rationale: Control-plane credentials must have explicit, auditable scope.
diff --git a/memory/examples/20260610_f07_action-runtimes.md b/memory/examples/20260610_f07_action-runtimes.md
@@ -0,0 +1,8 @@
+﻿# F-07 supported GitHub Action runtimes
+
+Issue Description: Green remote CI warned that checkout and github-script action majors used deprecated Node.js 20.
+State: GitHub will force Node.js 24 on June 16, 2026 and remove Node.js 20 on September 16, 2026.
+Action: Updated to the verified latest supported majors, `actions/checkout@v6` and `actions/github-script@v9`, including templates.
+Result: Workflows use Node.js 24-compatible action releases and CI prevents regression to deprecated majors.
+Diff Patch: Updated action majors and added a contract assertion.
+Rationale: Platform deprecation warnings are reliability defects with fixed deadlines.
diff --git a/memory/examples/20260610_f07_ci-contracts.md b/memory/examples/20260610_f07_ci-contracts.md
@@ -0,0 +1,8 @@
+# F-07 CI contracts
+
+Issue Description: CI ignored YAML parser errors and did not test control-plane safety contracts.
+State: Invalid workflows and authorization regressions could merge with green CI.
+Action: Repaired fragile workflow heredocs, added workflow validation and control-plane contract tests, and made CI fail on errors.
+Result: CI now enforces workflow syntax and critical operator safety boundaries.
+Diff Patch: Added tests and replaced warn-only validation.
+Rationale: Control-plane safeguards require executable regression tests.
diff --git a/memory/examples/20260610_f08_opt-in-notice.md b/memory/examples/20260610_f08_opt-in-notice.md
@@ -0,0 +1,8 @@
+﻿# F-08 non-actionable opt-in notice
+
+Issue Description: The installer queued an autofix issue in repositories that had not opted in.
+State: A notification issue itself satisfied the operator execution gate.
+Action: Changed the pre-opt-in notice to an unlabeled informational issue.
+Result: Non-opted-in repositories cannot be enrolled through an automated issue side effect.
+Diff Patch: Removed actionable labels and clarified the notice title.
+Rationale: Enrollment must be an explicit repository-owner decision.
diff --git a/scripts/autopilot-operator.ps1 b/scripts/autopilot-operator.ps1
diff --git a/templates/demo-repo/.github/workflows/autopilot-create-issue.yml b/templates/demo-repo/.github/workflows/autopilot-create-issue.yml
diff --git a/tests/contract-tests.ps1 b/tests/contract-tests.ps1
diff --git a/tests/validate_workflows.py b/tests/validate_workflows.py
