code-review-intake

name	code-review-intake
description	Deterministic GitHub review intake workflow with ledger-based judgment. Use when processing GitHub code review comments, reconciling Code-Critic findings, or running GitHub review intake mode. DO NOT USE FOR: pre-PR readiness checks (use verification-before-completion) or post-merge cleanup (use post-pr-review).

Code Review Intake

Slim entryway for GitHub review intake, proxy-prosecution guardrails, and the extracted express-lane boundary reference.

When to Use

Activate this skill when the request includes github review, review github, or cr review. It is the entryway for deterministic intake and judgment of GitHub-originated review feedback before implementation begins, while shared mechanics stay indexed in extracted references.

The GitHub-intake path consumes the same agents/Code-Review-Response.agent.md judge body as the non-GitHub review path and therefore the same structural-criteria deferral gate. Findings ingested from GitHub are classified against the canonical criterion taxonomy in skills/review-judgment/scripts/Test-DeferralCriteria.ps1 (the predicates), and any DEFERRED-SIGNIFICANT (structural) outcome is routed through the §2e Filing Approval Gate (skills/safe-operations/SKILL.md § 2e) before it is filed via skills/safe-operations/scripts/Add-FollowUpIssue.ps1 (the filing helper) — consistent with how Code-Conductor's Auto-Tracking sequence routes structural deferrals. Before calling Add-FollowUpIssue (or any bare gh issue create) for structural deferrals, apply the board-positioning decision per skills/safe-operations/SKILL.md §2b, §2b-bis, and §2b-ter. To guarantee AC7 dedup correctness across both paths, the GitHub-intake filing call canonicalizes the title with ConvertTo-CanonicalFollowupTitle (from the same helper file) and runs §2c dedup-on-create against the canonicalized title before invoking Add-FollowUpIssue — identical to the Code-Conductor non-GitHub path. The GitHub-intake path MUST also call Get-AcTermsFromIssue (ARM 2) alongside Get-AcRefsFromIssue (ARM 1) before invoking Get-StructuralVerdict, so that behavioral-term AC cross-check fires for GitHub-ingested findings — identical to the Code-Conductor path. The ac_cross_check OUT object is required for any dismiss/defer disposition entry at severity ≥ medium per the v2 schema.

GitHub Review Mode (Proxy Prosecution Pipeline)

1. Ingest all review items from GitHub (threads, top-level comments, review summaries).
2. Build a finding ledger where each item maps to its GitHub comment/review ID and, so that per-reviewer disposition evidence is compilable from judge rulings, records a per-finding reviewer_source field (canonical field name per 768-D3 — never a flat source: key). This is the canonical statement of the reviewer_source normalization rule — other sites (skills/review-judgment/SKILL.md, Documents/Design/phase-containment-ledger.md) cite this section rather than restating the mechanism. Normalize the value deterministically: lowercase(author.login) with a trailing [bot] suffix stripped, sourced from the author's .login field only (never display name), and written as a quoted scalar. The reserved values local (pipeline-native findings only) and unresolved (lookup-failure sentinel) can never be emitted by this normalization, and neither can a value that already looks escaped: the escape fires when the normalized login IS one of the reserved values OR already STARTS WITH the ext- prefix, and in either case it is escaped by prepending ext- — so local → ext-local, a real login ext-local → ext-ext-local, and ext-unresolved → ext-ext-unresolved. This single conditional prepend (not an unbounded loop) is sufficient to keep the escaped and unescaped namespaces collision-free, because it recognizes any already-escaped-looking value, not just an exact reserved word. When one review comment enumerates multiple distinct findings, each becomes its own ledger item sharing that comment/review ID — per-finding key identity for that case is defined by the discriminator in skills/review-judgment/SKILL.md's § Stable Finding Key section.
3. Proxy prosecution: Call Code-Critic with the selector line Review mode selector: "Score and represent GitHub review". Code-Critic validates and scores each GitHub comment (critical/high→10 pts, medium→5 pts, low→1 pt). Output: scored prosecution ledger.
4. Defense pass: Call Code-Critic with the selector line Review mode selector: "Use defense review perspectives", passing the prosecution ledger.
5. Judge pass: Call Code-Review-Response with both prosecution ledger and defense report. Judge rules final and emits score summary.

Hard Guardrail

In GitHub Review Mode, do not add net-new findings outside the ingested GitHub ledger. GitHub review mode is proxy prosecution, so the R6 express lane does not apply. See references/express-lane.md for the canonical scope restriction and Tier 1 re-validation rule.

Ledger-vs-Validation Boundary

Ingestion and ledger-building (steps 1–2) are mechanical: the conductor records each ingested finding verbatim and maps it to its GitHub comment/review ID. The conductor MUST NOT independently assess the technical merit of an ingested finding — neither accepting it, rejecting it, nor forming any per-finding correctness verdict — before proxy prosecution runs. Per-finding validation is the proxy prosecution pass's responsibility (step 3).

Reading an ingested finding closely enough to record it and map it to its GitHub comment/review ID is expected; forming a verdict on whether it is correct — or assigning your own severity/type classification ahead of proxy prosecution — is not. The NEW-CRITICAL safety exception below — which concerns a newly discovered critical blocker, not an ingested finding — is the only conductor-side correctness judgment permitted before proxy prosecution.

Safety Exception

A new item may be added only for a critical correctness/security blocker discovered during verification. It must be tagged NEW-CRITICAL, include concrete evidence, and be explicitly surfaced to the user.

Judgment Guardrail

Judgment is evidence-first and deterministic:

• Every accepted/rejected/deferred item cites code, test output, architecture constraints, or issue AC evidence.
• Preference-only comments without evidence are rejected by default.
• Conflicting evidence keeps an item in disputed state until resolved or escalated.
• Do not route implementation work until judgment states are explicit.

Convergence Criteria

Converged when all items are in a terminal state:

• ✅ ACCEPT (fix inline)
• 📋 DEFERRED-SIGNIFICANT (structural)
• ❌ REJECT

Plus:

• No unresolved evidence disputes remain.
• User has visibility before any authority-boundary decision gate.

The proxy prosecution pipeline is single-shot: prosecution → defense → judge, with no rebuttal rounds. Judge rules final on all items. Unresolved items at low judge confidence are surfaced for user scoring via GitHub issue comment (async, non-blocking).

Composite References

• references/express-lane.md: canonical R6 express-lane gate, its exclusion from proxy prosecution, and the Tier 1 re-validation requirement when R6 is used elsewhere
• references/response-loop-completion.md: ordered terminal-step sequence (disposition gate, R4 dispatch, post-fix prosecution, CE Gate, persist-changes, close-out record amendment hand-off, Response Summary) that completes the /review-github response loop
• ../validation-methodology/references/review-reconciliation.md: shared non-GitHub review reconciliation, prosecution-depth setup, and post-fix R2 review mechanics that pair with intake after proxy judgment completes

Gotchas

Trigger	Gotcha	Fix
Spotting a new bug while reading GitHub review comments	Adding it informally bypasses the prosecution → defense pipeline and breaks ledger integrity	Surface as NEW-CRITICAL with concrete evidence only; present to user explicitly for decision
Routing implementation work before all judgment states are explicit	Fixes applied to some findings may contradict pending rulings on others	All items must reach terminal state (ACCEPT / REJECT / DEFERRED-SIGNIFICANT (structural)) before any routing

Response Loop Completion (Terminal Step)

The ordered terminal-step sequence that makes a bare /review-github complete the full response loop (disposition gate, Batch Specialist Dispatch R4, post-fix targeted prosecution, CE Gate, persist-changes, close-out record amendment hand-off — the rule itself lives at skills/review-judgment/SKILL.md § Close-Out Record Amendment, its single home on every lane — and Response Summary) lives in references/response-loop-completion.md. Load that reference once all judgment states reach terminal; accepted fixes are applied, committed, and pushed to the existing PR branch without requiring an additional user instruction. When a sustained finding's reviewer_source (normalized above) resolves to a real external identity, the judge's phase-containment emission may owe a post-review-observer block instead of the standard code-review block — see skills/review-judgment/SKILL.md § Phase-containment emission § Observer emission variant for the novel-gating rule and exact-equality dispatch; this section does not restate that mechanism.