/flow-next:map — wrap clawpatch map for a semantic feature index

Read workflow.md for full phase-by-phase execution.

Wrap the upstream clawpatch CLI's map subcommand. Default invocation is provider-free (--source heuristic) — zero LLM calls, zero API spend, deterministic mapper. Output lands at .clawpatch/features/*.json (Zod-validated upstream, schemaVersion: 1). Scout enrichment and the /flow-next:prime DE7 nudge read the resulting index; this skill is the install/init/invoke surface.

Role: thin shell-out wrapper. flowctl never imports or requires clawpatch; the skill is the only flow-next surface that touches it.

Preamble

CRITICAL: flowctl is BUNDLED — NOT installed globally. which flowctl will fail (expected). Define once; subsequent blocks (here and in workflow.md) use $FLOWCTL:

FLOWCTL="${CODEX_HOME:-$HOME/.codex}/scripts/flowctl"
[ -x "$FLOWCTL" ] || FLOWCTL="<plugin-root>/scripts/flowctl"   # <plugin-root> = the directory two levels above this skill's SKILL.md file (the harness gave you that file's absolute path when the skill loaded); substitute it literally
[ -x "$FLOWCTL" ] || FLOWCTL=".flow/bin/flowctl"

Inline skill (no context: fork) — the map skill is fully non-interactive: install detection, version-range guard, clawpatch init, .clawpatch/.gitignore skeleton, and clawpatch map invocation all proceed without prompting the user. No plain-text numbered prompt is required or used.

Input

Arguments: $ARGUMENTS

Format: [--source <heuristic|auto|agent>] [-- <extra clawpatch args>]

• Default (no args) → clawpatch map --source heuristic (provider-free, deterministic). Heuristic targets conventional app/framework layouts; unconventional repos (CLI tools, plugins, markdown/docs-heavy, non-standard monorepos) may map to 0 features — Phase 5 surfaces a --source=auto|agent suggestion when that happens.
• --source auto|agent → passthrough to clawpatch; user must have CLAWPATCH_PROVIDER configured for these paths (clawpatch's own concern, not ours).
• -- → terminator; tokens after flow to clawpatch map (e.g. --since-ref origin/main, --paths src/).

Passthrough boundary. The slash-command host delivers $ARGUMENTS as a single string; the skill word-splits on whitespace. Passthrough is therefore token-level (whitespace-separated), not full shell-verbatim — tokens containing literal spaces or shell metacharacters that require shell quoting will not survive. Globs (*, ?) are protected from expansion (set -f before the parse) so they reach clawpatch untouched. Users needing complex quoting should run clawpatch map directly.

flow-next's review backend config (rp / codex / copilot / none) stays out of clawpatch. clawpatch's provider matrix (codex / acpx / claude / cursor / grok / opencode / pi) is orthogonal — a run that sets CLAWPATCH_PROVIDER from the flow-next backend has broken this. The Phase 0 echo reports the backend as informational only.

Version pin

SUPPORTED_CLAWPATCH=">=0.4.0 <0.5.0"

Single source of truth for the supported clawpatch range. Workflow Phase 1 parses clawpatch --version with a tolerant (\d+\.\d+\.\d+) regex and compares against this range. Outside-range → warn one line to stderr and degrade (continue). Never block. Re-verify on each clawpatch minor release.

clawpatch is pre-1.0 (v0.4.0, 2026-05-22; weekly minor releases). The README forecasts breaking changes between minor releases — tolerant parsing matters.

Ralph-block (R13) — runs first, before everything else

/flow-next:map requires a user at the terminal for the install-prompt and init-prompt branches. Autonomous loops cannot install global npm packages or accept interactive consent. Decline-to-run when Ralph signals are set.

if [[ -n "${REVIEW_RECEIPT_PATH:-}" || "${FLOW_RALPH:-}" == "1" ]]; then
  if [[ -n "${REVIEW_RECEIPT_PATH:-}" ]]; then
    TRIGGER="REVIEW_RECEIPT_PATH"
  else
    TRIGGER="FLOW_RALPH"
  fi
  echo "Error: /flow-next:map declines under Ralph ($TRIGGER set); rerun interactively." >&2
  exit 2
fi

Decline-to-run only — nothing is written to $REVIEW_RECEIPT_PATH. That file belongs to the upstream review caller; a run that leaves any byte there under Ralph has corrupted an unrelated receipt and broken this. The skill exits at line 1 of the workflow under Ralph; install/init paths are unreachable.

No env-var opt-in. Ralph never installs global tools or accepts interactive consent.

Workflow

Execute the phases in workflow.md in order:

0. Pre-flight + config-state echo (R12) — one four-line block: clawpatch version + --source, CLAWPATCH_PROVIDER env (or "none"), flow-next review backend (informational only), .clawpatch/ last-mapped timestamp (or "absent").
1. Install detection (R1, R11) — command -v clawpatch + clawpatch --version. Missing → print pnpm add -g clawpatch install instructions verbatim and exit 1. When pnpm bin -g exits 0 but command -v clawpatch still empty, also print the PNPM_HOME bin/ hint (run pnpm setup, re-source shell rc). No auto-install.
2. Version-range guard (R10) — parse clawpatch --version with (\d+\.\d+\.\d+); compare against SUPPORTED_CLAWPATCH. Outside range → one-line stderr warning naming expected vs found and continue (degrade — never block).
3. Init (R2) — when .clawpatch/ absent, run clawpatch init first. After clawpatch creates .clawpatch/project.json + .clawpatch/config.json, write a self-contained .clawpatch/.gitignore skeleton (skill owns this write — STRATEGY zero-dep means flowctl never references clawpatch).
4. Map invocation (R1) — clawpatch map --source <SOURCE> [extra passthrough]. Default <SOURCE> is heuristic (always passed explicitly — never rely on clawpatch's default in case upstream changes it). clawpatch streams stdout live; the skill does not buffer.
5. Result summary — print path to .clawpatch/features/, count of feature files, last-mapped timestamp; suggest flowctl repo-map list and the /flow-next:plan / /flow-next:capture paths that consume it.

Sharing contract — local-only by design

The .clawpatch/.gitignore skeleton this skill writes is * + !.gitignore — everything under .clawpatch/ is git-ignored; only the .gitignore itself is tracked. The feature index is local-per-developer, not committed.

This is a deliberate design choice for a pre-1.0 mapper:

• No PR review noise from regenerated feature data on every clawpatch map run.
• No merge conflicts on features/*.json when two branches mapped at different SHAs.
• No coupling of PR review to mapper-output drift as clawpatch schema bumps land (weekly minor releases at the time of writing).
• Onboarding cost is low — --source heuristic is a fast filesystem walk; new contributors regenerate locally in seconds.

Teams that want a shared, in-repo feature index can edit .clawpatch/.gitignore directly (e.g. !features/, !project.json, ignore only .cache/ and *.log). The skill is idempotent and won't clobber a customized .gitignore on re-run. This is unsupported — be prepared for review noise and merge conflicts; document the choice in your team's CLAUDE.md / AGENTS.md. See flow-next.dev/skills/map for the full trade-off table.

Forbidden

• Auto-installing clawpatch. The skill detects and instructs; users install with their own permission. Global npm installs are user-consent territory.
• Proxying flow-next's review backend into CLAWPATCH_PROVIDER. Orthogonal matrices. clawpatch users configure clawpatch directly.
• Auto-upgrading --source to auto or agent when heuristic coverage looks weak. Users opt up explicitly via --source.
• Touching the repo .gitignore. The .clawpatch/.gitignore skeleton is self-contained inside .clawpatch/ so a full deletion of that directory removes both data and ignore rules in one step.
• Writing to $REVIEW_RECEIPT_PATH from the Ralph-block path — see R13 above. The receipt belongs to the upstream review caller; the Ralph branch is decline-to-run, not a receipt producer.
• Importing or requiring clawpatch from flowctl. flowctl never references this skill or clawpatch — uninstall promise (rm -rf .flow/) stays intact and zero-dep STRATEGY track is preserved.
• Network calls beyond what clawpatch itself does. The wrap is local-only; clawpatch's network behavior is upstream's concern.
• Running under Ralph — hard-blocked by R13 above.