reviews: default-on automated review — promote LLM lane + auto-invoke fanout (1→2 residue)

[kyle-sexton]

Problem

Review automation is present but not default-on: the LLM review lane is advisory-only and
review:fanout never auto-invokes — a human must remember to trigger the multi-dimension
review. Boris's step-2 posture: "Automated code review and security review are on by default";
the 1→2 transition names "automate code review".

Fix shape

Two decisions plus wiring, previously parked as an open question:

1. Promote the LLM review from advisory to blocking, or keep advisory with a recorded
   promotion trigger (mirrors the WP5 verification-blocking promotion discipline — earned
   flip, ratified record).
2. Auto-trigger review:fanout (or the equivalent review pass) on PR open / pre-merge instead
   of manual invocation.

Decide, then wire whichever lane is chosen through ci-workflows so the default holds
fleet-wide. Relation: #509 (no dedicated security-review pass) is the security sibling of the
same default-on posture; #618 (scope Codex connector to high-blast-radius PRs) tunes the same
lane's noise.

Source

1→2 residue item from the AI-adoption-ladder gap map; #239 directed "file via /work-items" —
this files it.

---

Reference: Boris Cherny, "Steps of AI Adoption" (Google Doc)

[kyle-sexton]

This was generated by AI during triage (historic normalization, rule 8g — closing out the 55-issue historic sweep).

Routing human: the issue's own body already frames this as "Decide, then wire" — a genuine product/policy choice (promote LLM review from advisory to blocking, or keep advisory with a recorded promotion trigger) with real consequences either way (blocking changes merge gating fleet-wide), not a single well-grounded default with a merely-vetoable alternative. Once decided, the wiring itself (auto-trigger review:fanout on PR open/pre-merge, thread through ci-workflows) is mechanical.

[kyle-sexton]

Decision brief — default-on automated review: gating posture

Human-operator brief. I decide nothing here; this frames one choice and a recommendation. The routing comment above established this needs a real decision but did not offer options — this is the first structured brief.

Heads-up: one branch of this decision is already coded and open as #722 (mergeable, not draft), which carries Closes #696. So the operator's real action is a merge call, not a from-scratch design. Brief below so you can ratify or override deliberately rather than by default.

Problem

Boris's step-2 posture is "automated code + security review on by default." The issue framed two open items — promote the LLM review lane from advisory to blocking, or keep it advisory with a recorded promotion trigger; and auto-invoke the review pass. The only genuinely open item is the gating decision (advisory vs blocking) — see verified state below.

What's verified today (machine-checked, not inferred)

• Auto-invocation is already on, independent of feat(ci): default-on security review — path-filtered caller + posture ADR #722. .github/workflows/claude-review.yml triggers on pull_request: [opened, synchronize, ready_for_review, reopened], calling the ci-workflows reusable claude-review.yml (pinned 90f1c54). Default-on invocation is satisfied; the issue-body claim "review:fanout never auto-invokes" is stale for the CI lane. The issue's own fix-shape hedges "review:fanout (or the equivalent review pass)" — the equivalent CI pass is live.
• The lane is advisory, verified via the branch ruleset — not just asserted. main's required status checks are exactly pr-title / pr-title and ci-status. No AI review lane is a required check, so none blocks merge today. (Classic branch-protection returns 404; a repository ruleset enforces the checks above.)
• The local review:fanout skill (multi-dimension, manual) is a separate tool from the CI claude-review lane and is not auto-wired. If "auto-invoke fanout" was meant literally as this skill, that sub-item stays open even after feat(ci): default-on security review — path-filtered caller + posture ADR #722 (flagged below).
• #722 already records the decision in ADR 0002: both AI review lanes default-on + advisory, promotion to blocking earned on demonstrated precision (mirrors WP5 verification-promotion discipline, no calendar flip), plus a new path-filtered claude-security-review.yml caller scoped to security-sensitive surfaces (workflows, .claude/, scripts, hooks, shell/PS/mjs/ts). Note: that path-filter is a scope lever on an advisory lane — it is not blocking-for-hot-paths.

Options

(b) RECOMMENDED — Ratify advisory + earned-promotion; merge #722 as-is.
Adopt the posture already coded: both lanes default-on + advisory, promotion to blocking earned on a proven-precision window and recorded as a reviewed change; security lane path-filtered so doc-heavy PRs stay cheap.

• Why: matches Boris step-2 ("on by default" ≠ "blocking"), matches the WP5 earned-flip discipline, and is the reversible direction — advisory→blocking later is a one-line ruleset addition of the check context; blocking-now fleet-wide is hard to walk back once merges start failing on LLM findings. Discharges reviews: default-on automated review — promote LLM lane + auto-invoke fanout (1→2 residue) #696's gating decision, its equivalent-auto-on-pass item, and the autonomy pipeline: no dedicated security-review pass — automated code review only #509 security sibling in one merge.
• Cost: advisory findings can be ignored; reviews: scope Codex connector review to high-blast-radius PRs — its advisory threads block the merge gate on every PR #618's class of noise (advisory bot threads) persists until the security path-filter tunes it.

(a) Override — Promote to blocking now, fleet-wide.
Add the claude-review check to main's ruleset required_status_checks (and roll through ci-workflows to peers).

• Tradeoff: closes the "findings ignored" gap immediately, but LLM false-positives/non-determinism now gate every merge fleet-wide with no earned-precision record — the exact risk WP5's discipline exists to avoid. Reversible only by removing the check after friction is felt. Would require holding/amending feat(ci): default-on security review — path-filtered caller + posture ADR #722.

(c) Variant — True path-filtered blocking (blocking only on high-blast-radius paths, advisory elsewhere).
Nobody has built this. #722's path-filter looks like (c) but gates nothing. A real (c) is a sensible future promotion target — blocking on CRITICAL security findings for the filtered surface only — and ADR 0002's revisit trigger already points at it. Not needed now; record as the earned-promotion landing spot.

Residue flag (not for me to resolve)

Merging #722 discharges the gating decision + equivalent auto-on pass + security sibling. But if "auto-invoke fanout" in the title meant the literal multi-dimension review:fanout skill (distinct from the claude-review CI lane), that wiring is still manual and would remain open. Worth a one-line confirmation before #696 is closed on #722's merge.

Cross-links

• #722 — open PR implementing option (b) + security sibling; Closes #696, Closes #509.
• autonomy pipeline: no dedicated security-review pass — automated code review only #509 — dedicated security-review pass (the security sibling, closed by feat(ci): default-on security review — path-filtered caller + posture ADR #722).
• reviews: scope Codex connector review to high-blast-radius PRs — its advisory threads block the merge gate on every PR #618 — scope Codex connector to high-blast-radius PRs (advisory-noise tuning the ADR cites).
• claude-ops: tune auto-mode classifier from observed team usage (step-3 guardrail, 1→2 residue) #697 — permission-prompt mining evidence the ADR draws on.