flow-next-impl-review

name	flow-next-impl-review
description	Carmack-level implementation review of changes via the configured backend. Use when asked to review code or a diff in a flow-next repo.
user-invocable	false

Implementation Review Mode

Workflow is backend-split. Read workflow-common.md for Phase 0 (backend detection + philosophy + trivial-diff triage), then read ONLY the file matching your active backend. The opt-in --deep/--validate/--interactive phase detail (including the phase-ordering matrix) lives in optional-phases.md, loaded only when a flag fires:

• BACKEND=codex → workflow-codex.md
• BACKEND=copilot → workflow-copilot.md
• BACKEND=cursor → workflow-cursor.md
• BACKEND=claude → workflow-claude.md
• BACKEND=host → workflow-host.md
• BACKEND=rp → workflow-rp.md

Do not load the others — only the active backend's file is needed. Each backend file carries its own Critical Rules and anti-patterns.

Conduct a John Carmack-level review of implementation changes on the current branch.

Role: Code Review Coordinator (NOT the reviewer) Backends (branch on the Phase 0 RP_ELIGIBLE probe):

• When RP_ELIGIBLE=1: RepoPrompt (rp), Codex CLI (codex), GitHub Copilot CLI (copilot), Cursor CLI (cursor), Claude Code CLI (claude), or host-native (host)
• When RP_ELIGIBLE=0: Codex CLI (codex), GitHub Copilot CLI (copilot), Cursor CLI (cursor), Claude Code CLI (claude), or host-native (host) — rp is macOS-only; never list it in guidance you surface (--review=rp stays accepted)

Preamble — execute Phase 0 exactly once

The executable Phase 0 lives in workflow-common.md §"Phase 0: Backend Detection" — Read it and execute it ONCE, before any other bash in this skill. It defines $FLOWCTL (bundled — NOT installed globally; which flowctl fails, expected), probes RP_ELIGIBLE, resolves $BACKEND via the single flowctl review-backend call, and handles the ASK / none cases. Every later bash block here (triage, deep-pass selection) uses the $FLOWCTL it defines. Never invoke flowctl review-backend a second time in the same run.

Exception: a --review=<backend> argument (see Backend Selection below) wins — when present, set BACKEND from the flag and skip Phase 0's review-backend call + ASK handling (still run its $FLOWCTL / RP_ELIGIBLE setup lines).

When RP_ELIGIBLE=0 (not macOS, no supported RepoPrompt CLI), never steer the user toward rp: every backend summary, recommendation, or override hint you surface presents only the runnable configured backends codex, copilot, cursor, claude, host (plus none). export is not an impl-review mode at all — a manual export review lives in /flow-next:plan-review --review=export; never present it here. Suppression is not a ban: an explicit --review=rp, FLOW_REVIEW_BACKEND=rp, or review.backend=rp still resolves to rp and errors at runtime via require_rp_cli().

Backend Selection

Priority (first match wins):

1. --review=rp|codex|copilot|cursor|claude|host|none argument
2. FLOW_REVIEW_BACKEND env var — bare backend (rp, codex, copilot, cursor, claude, host, none) OR spec form (codex:<model>:xhigh, copilot:<model>, cursor:<model>, claude:<model>:<effort>); host is bare-only (host:<model> is rejected)
3. .flow/config.json → review.backend (same bare / spec forms)
4. Error - no auto-detection

Parse from arguments first

Check $ARGUMENTS for:

• --review=rp or --review rp → use rp
• --review=codex or --review codex → use codex
• --review=copilot or --review copilot → use copilot
• --review=cursor or --review cursor → use cursor
• --review=claude or --review claude → use claude
• --review=host or --review host → use host
• --review=export or --review export → fail closed: report that export is not an impl-review backend and stop before any dispatch; the manual path is /flow-next:plan-review --review=export
• --review=none or --review none → skip review

If found, use that backend and skip all other detection.

Otherwise: Phase 0 resolves it

No --review flag → $BACKEND comes from workflow-common.md Phase 0 (executed once per the Preamble): the single flowctl review-backend "$REVIEW_ID" call with ASK handling included. Do not re-resolve here.

Backend detail (model / effort / spec grammar) — on demand

The per-backend "at a glance" descriptions, the backend[:model[:effort]] spec grammar, and the FLOW_REVIEW_BACKEND spec-form examples live in references/backend-specs.md. Read it only when you must surface backend guidance to the user or resolve a model/effort spec — a normal review already has $BACKEND and needs nothing from it. When RP_ELIGIBLE=0, omit the rp line from any guidance you surface (explicit --review=rp still honored).

Critical Rules

Per-backend rules for rp, codex, copilot, cursor, and claude live at the top of each workflow-<backend>.md — read the active backend's file (routing table above) and follow its Critical Rules section.

For host backend: host is bare-only. After selection, read workflow-host.md. The review must use a fresh, tool-enforced read-only reviewer from a different model family and fail closed when no cross-family pin is available.

For all backends:

• If REVIEW_RECEIPT_PATH set: write receipt after review (any verdict)
• Any failure → output <promise>RETRY</promise> and stop

Hard invariants:

• The coordinator never authors a verdict. A SHIP with no backend response behind it has broken this.
• One backend per review. A transcript that dispatches a second backend after the first answered has broken this.
• Review is never skipped without consent. A none backend that ends the run without the user's consent has broken this.

Input

Arguments: $ARGUMENTS Format: [task ID] [--base <commit>] [--validate] [--deep[=passes]] [--interactive] [focus areas]

• --base <commit> - Compare against this commit instead of main/master (for task-scoped reviews)
• --validate - After NEEDS_WORK verdict, run a validator pass that drops false-positive findings (opt-in)
• --deep / --deep=<passes> - Run additional specialized passes (adversarial / security / performance) after primary review (opt-in)
• --interactive - On NEEDS_WORK, walk through each finding with the user (Apply/Defer/Skip/Acknowledge) (opt-in, Ralph-incompatible)
• Task ID - Optional, for context and receipt tracking
• Focus areas - Optional, specific areas to examine

Scope behavior:

• With --base: Reviews only changes since that commit (task-scoped)
• Without --base: Reviews entire branch vs main/master (full branch review)

Opt-in flags:

• --validate — adds a validator pass on NEEDS_WORK that re-checks each finding for false positives. All findings dropping upgrades verdict to SHIP.
• FLOW_VALIDATE_REVIEW=1 env var — enables --validate session-wide (works in Ralph).
• --deep — adds adversarial pass always + security/performance auto-enabled per diff paths. --deep=adversarial,security restricts to listed passes.
• FLOW_REVIEW_DEEP=1 env var — enables --deep session-wide (works in Ralph).
• --interactive — per-finding walkthrough on NEEDS_WORK. No env var form — per-invocation only, always hard-errors in Ralph mode (REVIEW_RECEIPT_PATH or FLOW_RALPH=1) to prevent accidental autonomous engagement.
• Default review behavior (no flags) is unchanged.

Workflow

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"

Step 0: Parse Arguments

Parse $ARGUMENTS for:

• --base <commit> → BASE_COMMIT (if provided, use for scoped diff)
• --no-triage → set TRIAGE_DISABLED=1 (skip trivial-diff pre-check)
• --validate → set VALIDATE=true (validator pass on NEEDS_WORK)
• --deep / --deep=<passes> → set DEEP=true + optional DEEP_PASSES CSV
• --interactive → set INTERACTIVE=true (per-finding walkthrough on NEEDS_WORK; Ralph-blocked)
• First positional arg matching fn-* → TASK_ID
• Remaining args → focus areas

If --base not provided, BASE_COMMIT stays empty (will fall back to main/master).

Opt-in flags + env vars — ONE parse fence for --validate / --deep / --interactive:

VALIDATE=false
DEEP=false
DEEP_PASSES=""  # optional CSV: "adversarial,security"
INTERACTIVE=false
for arg in $(printf '%s\n' "$ARGUMENTS"); do   # command substitution word-splits under bash AND zsh; an unquoted $ARGUMENTS does not split under zsh (dogfood E1: --validate silently dropped)
  case "$arg" in
    --validate) VALIDATE=true ;;
    --deep) DEEP=true ;;
    --deep=*) DEEP=true; DEEP_PASSES="${arg#--deep=}" ;;
    --interactive) INTERACTIVE=true ;;
  esac
done

# Env opt-ins (Ralph-friendly). --interactive has NO env var form — per-invocation only.
if [[ "${FLOW_VALIDATE_REVIEW:-}" == "1" ]]; then
  VALIDATE=true
fi
if [[ "${FLOW_REVIEW_DEEP:-}" == "1" ]]; then
  DEEP=true
fi

# Optional-phase COUNT (PR #392): sizes the scope-ownership lease the backend
# workflows hold through the post-finalize phases (one exec allowance per
# pass). --deep counts one per selected pass (3 when unrestricted: adversarial
# + the auto-gated security/performance passes), --validate one,
# --interactive one. Carry this number into the finalize / host record blocks
# as a LITERAL - shell state does not survive across prompt turns.
OPTIONAL_PHASES_COUNT=0
if [[ "$DEEP" == "true" ]]; then
  if [[ -n "$DEEP_PASSES" ]]; then
    OPTIONAL_PHASES_COUNT=$((OPTIONAL_PHASES_COUNT + $(printf '%s' "$DEEP_PASSES" | tr ',' '\n' | grep -c .)))
  else
    OPTIONAL_PHASES_COUNT=$((OPTIONAL_PHASES_COUNT + 3))
  fi
fi
[[ "$VALIDATE" == "true" ]] && OPTIONAL_PHASES_COUNT=$((OPTIONAL_PHASES_COUNT + 1))
[[ "$INTERACTIVE" == "true" ]] && OPTIONAL_PHASES_COUNT=$((OPTIONAL_PHASES_COUNT + 1))
echo "OPTIONAL_PHASES_COUNT=$OPTIONAL_PHASES_COUNT"
# 1 when a held phase resumes the primary reviewer session (--deep /
# --validate); the interactive walkthrough alone never needs one.
PHASES_RESUME_SESSION=0
[[ "$DEEP" == "true" || "$VALIDATE" == "true" ]] && PHASES_RESUME_SESSION=1
echo "PHASES_RESUME_SESSION=$PHASES_RESUME_SESSION"

# Ralph-block: Ralph must never engage interactive.
if [[ "$INTERACTIVE" == "true" ]]; then
  if [[ -n "${REVIEW_RECEIPT_PATH:-}" || "${FLOW_RALPH:-}" == "1" ]]; then
    echo "Error: --interactive requires a user at the terminal; not compatible with Ralph mode (REVIEW_RECEIPT_PATH or FLOW_RALPH detected)." >&2
    exit 2
  fi
fi

if [[ "$DEEP" == "true" || "$VALIDATE" == "true" || "$INTERACTIVE" == "true" ]]; then
  echo "OPTIONAL PHASES ACTIVE — STOP. Read optional-phases.md (deep=$DEEP validate=$VALIDATE interactive=$INTERACTIVE) before continuing."
fi

When that sentinel prints, STOP and Read optional-phases.md before any further step — it owns the phase-ordering + flag-combination matrix, the deep-pass selection bash, the validator dispatch, and the walkthrough steps (per-finding loop detail in walkthrough.md, pass prompt templates in deep-passes.md). All three phases are default-OFF: when no flag fires, run the primary review only and write no validator / deep_passes / walkthrough receipt keys.

Step 0.5: Trivial-diff triage

Before invoking the configured backend, run a fast pre-check that short-circuits lockfile-only, docs-only, release-chore, and generated-file diffs. On SKIP, the receipt is written with mode: "triage_skip" / verdict: "SHIP" and the expensive backend call is skipped entirely.

Opt-out: --no-triage argument or FLOW_RALPH_NO_TRIAGE=1 env var.

if [[ -z "${TRIAGE_DISABLED:-}" && -z "${FLOW_RALPH_NO_TRIAGE:-}" ]]; then
  ROUTE="$($FLOWCTL review-route ${TASK_ID:+"$TASK_ID"} --json)"   # pure: canonical TASK_ID + receipt path (no rotation, no state change)
  TASK_ID="$(jq -r '.task_id // empty' <<<"$ROUTE")"
  RECEIPT_PATH="$(jq -r '.receipt_path' <<<"$ROUTE")"
  # Subcommand + one literal flag stay on the command line (the Ralph guard
  # blocks a variable in either of the two tokens after the launcher).
  TRIAGE_ARGS=(--receipt "$RECEIPT_PATH")
  [[ -n "$BASE_COMMIT" ]] && TRIAGE_ARGS+=(--base "$BASE_COMMIT")