Aller au contenu principal

ADR-065: STOA PR Guardian — Advisory AI Review on GitHub Actions

Metadata​

FieldValue
StatusAccepted
Date2026-04-16
Decision MakersChristophe, Claude Code
Linear(to be filed)
  • ADR-012: MCP Tools Architecture — RBAC & Multi-Tenant Governance (scope enforcement the Guardian verifies)
  • ADR-041: Plugin Architecture — Community Core vs Enterprise Extensions (boundary the Guardian flags)
  • ADR-055: Portal/Console Governance (Producer/Consumer split the Guardian enforces on Portal code)
  • ADR-061: Council Multi-Stage Review (pre-ticket and pre-merge human/AI review — complementary, not overlapping)

Context​

Roughly 85% of STOA code lands via AI co-authoring. Human reviewers already carry a wide surface (3 stacks, ~70 tickets per cycle, multi-repo). Recurring review burdens keep landing on them:

  1. ADR drift — code silently crosses Portal/Console lines (ADR-055), Community/Enterprise lines (ADR-041), or bypasses declared RBAC scopes (ADR-012). These are cheap to spot but only if someone remembers to look.
  2. Security regressions — hardcoded secrets in fixtures, auth decorators forgotten on a new route, unsafe deserialization patterns (cf. webMethods JSONPath single-item array bug) that pass tests but leak at runtime.
  3. AI code smell — dead code, single-call helper wrappers, duplicated logic across files, WHAT-not-WHY comments, swallowed exceptions. Each one is trivial; ten per week is a tax.

None of these need a human on the first pass. They need a checklist that never forgets. But the review must stay advisory — the moment it blocks merge, contributors work around it, and the checklist becomes a lie.

We already run claude-review.yml (ADR-060, /review skill, Haiku, Ship/Show/Ask classification). It is useful but generic — one prompt, one verdict, no axis separation, no ADR cross-reference, no Slack escalation. The Guardian is a different shape: three explicit axes, binary GO/NO-GO, idempotent comments, fail-visible.

Why not Anthropic Routines​

Routines ship Claude workflows as a managed product: scheduling, auth, state, retries handled by Anthropic. Valid tool — we use it elsewhere. Wrong fit here:

ConcernGitHub ActionsRoutines
Audit trail for DORA/complianceNative (workflow runs, logs, artifacts, signed commits)External dashboard, separate export
Portability to self-hosted STOA installsRuns on any Actions-compatible runnerRequires Anthropic tenancy
Policy evolution velocitygit diff on docs/ai/pr-guardian-policy.mdUI-driven, no PR review on policy changes
Scale across repos (stoa, stoa-docs, stoa-web…)Drop-in workflow per repoPer-repo setup + per-run cost
Kill-switchRepo variable, instantConfig change, delayed propagation

Framing: fitness for purpose. Routines are the right answer when you want managed scheduling + managed state for a single tenant. We want an advisory gate that is versioned, auditable, forkable, and self-hostable. That is GitHub Actions.

Non-overlap with claude-review.yml​

Dimensionclaude-review.yml (existing)PR Guardian (new)
PurposeGeneric Ship/Show/Ask on diff qualityStructured checklist on ADR / SEC / SMELL
VerdictShip | Show | Ask + Low/Medium/HighBinary GO / NO-GO + confidence
ADR awarenessNoneLoads ADR-012, ADR-041, ADR-055, dynamic
EscalationPR comment onlySlack #stoa-reviews on NO-GO
IdempotenceNew comment on each runSingle edited comment via HTML marker
Policy surfaceInline promptExternal docs/ai/pr-guardian-policy.md

Both run advisory, both skip Dependabot/forks/drafts, both use Sonnet or smaller. They complement — review covers breadth (reviewer vibe), Guardian covers depth (checklist never forgets). No plan to retire claude-review.yml.

Decision​

Ship an advisory-only PR Guardian as four separated artefacts.

1. GitHub Actions workflow — .github/workflows/pr-guardian.yml​

  • Trigger: pull_request events opened, ready_for_review, synchronize.
  • Concurrency: guardian-<PR#>, cancel-in-progress: true. Resync replaces in-flight runs.
  • Kill-switch: repo variable DISABLE_PR_GUARDIAN=true (same pattern as DISABLE_CONTEXT_COMPILER, DISABLE_L1_REVIEW).
  • Skip conditions: author matches *[bot], PR is draft, label skip-guardian or wip, fork PR (secrets protection).
  • Diff guard: compute size via gh pr diff --patch | wc -l BEFORE invoking Claude. If over 1000 lines, post one comment "Diff too large — human review required" and exit 0.
  • Idempotence: HTML marker <!-- stoa-pr-guardian -->. Workflow finds and edits the previous summary comment; does not spam on every resync.
  • Failure mode: if Claude invocation fails, post "Guardian unavailable — PR not analyzed" AND alert Slack #stoa-ops (or channel behind SLACK_WEBHOOK_GUARDIAN_ALERTS). Never silent.
  • Permissions: pull-requests: write, contents: read, issues: write. Nothing else.
  • Model: Claude Sonnet 4.6 (CI policy, CLAUDE.md).

2. Versioned skill — .claude/skills/pr-guardian/SKILL.md​

  • Invocable locally as /pr-guardian <PR#> for policy tuning and historical PR replay.
  • Owns the prompt, the three-axis review structure, and the verdict computation.
  • Delegates stack detection, ADR loading, and Slack notification to sibling scripts.

3. Policy pack — docs/ai/pr-guardian-policy.md​

Separated from the skill because the rubric evolves faster than the skill logic. Contains:

  • Three review axes: [ADR-XXX], [SEC], [SMELL] with concrete examples.
  • Verdict rules: any [SEC] → NO-GO; ≥ 3 [ADR]+[SMELL] combined → NO-GO; otherwise GO.
  • Confidence rules: full analysis = high; partial ADR load or mixed stack = medium; ADR fetch failed or diff near 1000-line cap = low.

Editing the policy is a PR against one file. Any team member can propose a rubric change without touching the skill or the workflow.

4. Runbook — docs/ai/pr-guardian-runbook.md​

Operational surface: kill-switch procedure, per-PR skip label, local invocation for testing, known failure modes, log inspection, policy iteration. Not the place to explain why — that is this ADR.

Hard rules (enforced in skill prompt)​

  • Never approve a PR. Never request changes via GitHub API. Comment only.
  • Signal risk, explain why, suggest optional fixes. Do not prescribe.
  • Terse comments, max 4 lines each, no preamble, no emoji, no TODO follow-ups injected into reviewed code.
  • English in GitHub comments. French allowed in Slack alert if that matches team convention.

Alternatives Considered​

Option B — Anthropic Routines​

Managed scheduling + managed state for the Guardian prompt. Rejected for reasons above (portability, audit trail in repo, policy PR review, self-hosted parity).

Option C — Merge-blocking required check​

Configure branch protection to require Guardian check before merge. Rejected because:

  • Forces contributors to bypass on false positives (policy drift becomes policy fiction).
  • Turns an advisory signal into a hard gate without human calibration period.
  • Contradicts the "fail-visible, not fail-closed" rule from ADR-061 Stage 3.

Re-evaluate after two sprints of advisory data (false positive rate, NO-GO acceptance rate). No target date.

Option D — Extend claude-review.yml with axes​

Add ADR-awareness and three-axis output to the existing workflow. Rejected because:

  • Conflates two different feedback shapes (generic vibe check vs structured checklist).
  • Single prompt becomes a 400-line monolith that nobody edits.
  • Loses the policy/skill separation that lets non-coders propose rubric changes.

Option E — Single-artefact workflow with embedded prompt​

Ship everything in one YAML file. Rejected because the prompt is ~200 lines; inlining kills diff review on policy edits and prevents local skill invocation for replay.

Consequences​

Positive​

  • Advisory only, low social cost. Contributors see a signal, not a block. No workaround culture.
  • Binary verdict + confidence is readable in < 5 seconds. Reviewers skim, act, move on.
  • Fail-visible means Guardian outages never hide. No ghost state where PRs land unreviewed.
  • Policy pack is a PR away — rubric drift requires review like any other change.
  • Portable: forks and self-hosted installs run the same Guardian without Anthropic tenancy.
  • Auditable via git log + GitHub Actions logs — sufficient for DORA + compliance.

Negative​

  • Token cost — every PR resync hits Sonnet. Mitigated by concurrency-cancel, bot-author skip, 1000-line diff guard, draft skip.
  • False positives on early policy — expected. Mitigated by advisory-only mode: contributors ignore noise without workflow pressure.
  • Duplicate signal with claude-review.yml until we consolidate. Acceptable for the advisory phase; consolidation is a follow-up.

Neutral​

  • Does not change merge rules, branch protection, or CODEOWNERS.
  • Does not consume RBAC scopes — workflow runs with GITHUB_TOKEN, posts comments only.
  • Existing /review and /parallel-review skills unaffected.

Rollout Plan​

Phase 1 — Advisory (from merge). Workflow runs on every non-bot, non-draft, non-fork PR. Comments posted, Slack alerts on NO-GO. No merge gate. Collect 2 sprints of signal quality data.

Phase 2 — Calibration. Review false-positive rate and NO-GO acceptance rate. Tune policy pack. Decide whether to promote to required check, keep advisory, or retire.

No date on Phase 2. It starts when Phase 1 has enough signal.

Compliance​

  • No impact on RBAC (ADR-012) — Guardian reads, does not write.
  • No impact on Community/Enterprise boundary (ADR-041) — Guardian flags violations, does not create them.
  • Aligns with ADR-055 Portal/Console boundary as an enforcement surface.
  • Aligns with ADR-061 Council Stage 3 philosophy (advisory, axis-based, fail-visible).

References​

  • claude-review.yml — existing generic PR review workflow
  • docs/ai/pr-guardian-policy.md — the rubric applied by the Guardian
  • docs/ai/pr-guardian-runbook.md — operational procedures
  • .claude/skills/pr-guardian/SKILL.md — skill source and local invocation