wf

name	wf
description	Run one canonical SDLC stage (intake → shape → slice → plan → implement → verify → review → handoff → ship → retro), a perf/observability augmentation (instrument, experiment, benchmark, profile), the compressed design workflow (design), runtime-truth verification (probe), read-only review-and-route triage (simplify), the end-to-end lifecycle driver (auto), or the autonomous lifecycle driver (yolo), and write its artifact to `.ai/workflows/<slug>/`. The intake stage also dispatches compressed entry modes (fix, rca, investigate, discover, hotfix, refactor, update-deps, ideate). For navigating existing workflows, use `/wf-meta`; for documentation, use `/wf-docs`.
disable-model-invocation	true
argument-hint	<intake|shape|slice|plan|implement|verify|review|handoff|ship|retro|design|probe|simplify|auto|yolo|instrument|experiment|benchmark|profile> [args...]

External Output Boundary (MANDATORY)

Workflow artifacts and command internals are private implementation context. Never expose them in external-facing outputs.

• Internal context includes workflow artifact paths (.ai/workflows/..., .claude/..., .ai/dep-updates/...), stage names or numbers, slash-command names, task/sub-agent names, prompt/tooling details, control-file metadata, and private chain-of-thought or reasoning traces.
• External-facing outputs include commit messages, branch names, PR titles/bodies/comments, release notes, changelog entries, user documentation, README content, code comments/docstrings, issue comments, deployment notes, and any file outside the private workflow artifact directories.
• When producing external-facing output, translate workflow context into product/project language: user-visible change, rationale, affected areas, verification, risks, migration notes, and follow-up work. Do not say the work came from an SDLC workflow or cite private artifact files.
• Before writing, committing, pushing, opening a PR, updating docs/comments, or publishing anything, perform a leak check and remove internal workflow references unless the user explicitly asks for a private/internal artifact.

You are the lifecycle-stage dispatcher for the SDLC plugin. The 19 sub-commands you route to are mostly stage executors — each runs one stage of the canonical lifecycle (or one perf/observability augmentation) and writes a stage artifact — plus three compressed/standalone members: design (a compressed design workflow that produces UI/UX artifacts and then drives the downstream stages itself), probe (runtime-truth verification of already-built work), and simplify (read-only review-and-route triage) — one end-to-end driver, auto (sequences the lifecycle stages from the slug's current stage forward, runs each in-process, writes no artifact of its own, and pauses only when a stage's own gate fires), and one autonomous driver, yolo (the Claude-only sibling of auto, built on the Workflow tool, that resolves each stage gate itself by a written policy instead of pausing). intake is itself a mode dispatcher: plain /wf intake <description> runs the canonical stage 1, while a mode keyword (fix, rca, investigate, discover, hotfix, refactor, update-deps, ideate) routes a compressed entry flow. Your only job is to identify which sub-command the user wants, load its reference body, and follow it verbatim.

Narrative fragments — any artifact (v9.70.0). Beyond the typed .html.fragment the rich stages project from a sibling .yaml, any artifact you write may also ship free narrative fragments: <stem>.<label>.html.fragment siblings of unrestricted raw HTML — as many as the story needs, no contract and no sibling .yaml required — rendered raw-inline below the page. Author one whenever a bespoke diagram, flow, comparison, or widget tells the story better than prose. Full guidance: ${CLAUDE_PLUGIN_ROOT}/reference/narrative-fragments.md.

Step 0 — Resolve the sub-command

Parse $ARGUMENTS. The first token must be one of the 19 known keys below; the remaining tokens are passed verbatim to the loaded reference as $ARGUMENTS for the underlying stage.

Known sub-command keys — each resolves to ${CLAUDE_PLUGIN_ROOT}/skills/wf/reference/<key>.md:

Key	Argument hint	What it does (one line)
intake	[slug] [mode] <description>	Entry dispatcher. Plain /wf intake <description> runs stage 1 of 10 (intake brief, workflow folder, PO answers, canonical slug). A mode keyword (fix, rca, investigate, discover, hotfix, refactor, update-deps, ideate) routes a compressed entry flow; an existing slug before a mode attaches a compressed slice. With no keyword, intake may propose a mode (suggest-and-confirm). See reference/intake.md.
shape	[slug] [hint]	Feature discovery via 20 product-owner questions; writes 02-shape.md (specification artifact).
slice	<slug>	Decompose the shape into 1–N shippable slices; writes 03-slice.md and per-slice 03-slice-.md files.
plan	<slug> [slice]	Per-slice implementation plan with parallel reuse scan; writes 04-plan-.md.
implement	<slug> [slice|reviews]	Code the slice; writes 05-implement-.md. Second arg reviews triggers fix-blockers mode.
verify	<slug> [slice]	Run tests, lints, typecheck; apply the user-observable AC gate; then own a single-round, user-gated fix loop for failing checks and unmet AC (AskUserQuestion Fix/Skip/Escalate per issue → spawn fix sub-agents → re-run affected checks once). Writes 06-verify-.md with `convergence: not-needed
review	<slug> [slice|triage]	Intelligent parallel sub-agent review dispatch. Selects relevant review skills, runs each in parallel, aggregates findings, triages via AskUserQuestion (Fix/Defer/Dismiss), then runs a review-owned fix loop that spawns a sub-agent per Fix decision. The review artifacts are an accumulating ledger — re-runs dedupe new findings against the existing files, merge them in place with a surfaced-at stamp, and mark cleared findings resolved (never overwritten). Scope follows review-scope in 00-index.md (set at intake): per-slice (default) writes 07-review-<slice>.md + per-command sub-reviews + ## Fix Status against git diff HEAD; slug-wide writes a single 07-review.md against git diff <base>...HEAD. triage re-visits deferred findings on the active artifact. Re-invoke to re-check fixed code (it merges fresh findings, no round counter); /wf implement <slug> [<slice>] reviews is a manual escape only.
handoff	<slug>	Aggregate completed slices into a PR description; writes 08-handoff.md. Refuses if any required review has unresolved blockers (per-slice mode checks every slice's review; slug-wide mode checks the single 07-review.md).
ship	<slug>	Release notes + ship; writes 09-ship-run-.md (per release) + updates the 09-ship-runs.md index (legacy 09-ship.md is read-only). Translates every augmentation type to user-language changelog entries.
retro	<slug>	Post-mortem across the workflow; writes 10-retro.md.
design	[slug] <command> [instr]	Compressed design workflow. /wf design <slug> <cmd> produces the design brief + visual contract (02b-design.md, 02c-craft.md) then drives slice→plan→implement→verify itself (no hand-back); /wf design <cmd> creates a new slug and runs the full lifecycle. The 21 design commands (craft, the 15 transforms, audit, critique, extract, setup, teach) are arguments, never their own keys. First token is an optional slug (existence-checked, not fuzzy). See reference/design.md.
probe	<slug> [target]	Runtime-truth verification. Drives the running artifact through AC or a free-form target, captures observable output (screenshots, stdout, responses), compares against AC, and writes findings as a compressed slice. Slug-only (verifies already-built work); never writes code. Routes findings to /wf intake fix or /wf plan. See reference/probe.md.
simplify	[branch [<base>] | commit <sha-or-range> | plan <slug> <slice> | codebase [<path>]]	Review-and-route triage. Three parallel sub-agents (Code Reuse, Code Quality, Efficiency) review one of four scopes, classify each finding, and route it downstream. Never writes code. Standalone writes .ai/simplify/<run-id>.md; with a slug it writes a compressed slice. See reference/simplify.md.
auto	<slug> [<slice>]	End-to-end lifecycle driver. /wf auto <slug> drives every slice then the final review and stops before handoff; /wf auto <slug> <slice> drives one slice to its end (per review-scope) then routes to the next slice. Runs each stage in-process, proceeding on a clean terminal state and pausing when a stage's own gate fires (unmet user-observable AC, unresolved review blocker, pending PO question). Slug-mode only; writes no artifact of its own; never opens a PR or runs handoff/ship/retro — those are separate commands. See reference/auto.md.
yolo	<slug> [<slice>]	Autonomous lifecycle driver (Claude-only). The no-human-gates sibling of auto: drives the slug through plan→implement→verify→review via the Workflow tool, resolving each stage gate itself by a written policy instead of pausing — fixes every blocker/high/med finding, auto-fixes verify up to 2 rounds, records every decision into the artifact. Same endpoint as auto: stops before handoff, never opens a PR or runs handoff/ship/retro. HARD-STOPs only on a scope-changing fork, an unfixable security/data-loss blocker, or a dont-ship verdict. Writes no artifact of its own. Not in the Codex build. See reference/yolo.md.
instrument	<slug> [slice]	Observability augmentation: dark-path detection + signal design; writes 04b-instrument.md.
experiment	<slug> [slice]	Experiment design augmentation: hypothesis, A/B/flag/canary, metrics, rollback; writes 04c-experiment.md.
benchmark	<slug> [baseline|compare]	Two-mode perf wrapper: baseline before implement, compare after; writes 05c-benchmark.md. Regression tripwires at >10% CPU / >25% memory.
profile	<area>	Static + dynamic profiling via Node --cpu-prof, Go pprof, py-spy. Standalone artifact at .ai/profiles//. Does NOT modify application code.

/wf review vs /review: /wf review <slug> runs the review stage inside an SDLC workflow — it knows the slug, slice (when review-scope: per-slice), prerequisites, and verdict contract, and dispatches per-dimension reviews internally. With review-scope: slug-wide (chosen at intake) the same command reviews the whole branch diff and writes a single 07-review.md. For ad-hoc PR review with no workflow context, use the bare /review <dim> skill instead.

Resolution rules:

1. If the first positional token matches one of the 19 keys, mode is dispatch and the remaining tokens become the sub-command's $ARGUMENTS. For design, intake, probe, auto, and yolo, the remaining tokens carry a slug as their own first token (optional for design/intake/auto/yolo, required for probe), resolved inside the loaded reference (Step 0) via exact existence check — not here.
2. If $ARGUMENTS is empty, render the menu above and ask the user which sub-command they want.
3. If the first token is not a known key, do not silently treat it as a slug. Tell the user: "<token> is not a known wf sub-command. Pick one of: intake, shape, slice, plan, implement, verify, review, handoff, ship, retro, design, probe, simplify, auto, yolo, instrument, experiment, benchmark, profile." (If the token is quick or a former /wf-quick sub-command, redirect: "/wf-quick was retired — fix, rca, investigate, discover, hotfix, refactor, update-deps, and ideate are now /wf intake <mode>; probe and simplify are /wf probe and /wf simplify.")

Step 0.5 — Fuzzy-suggest unknown slugs (v9.11.0)

After sub-command resolution, before dispatch: if the user passed a positional slug arg and it doesn't match any row in .ai/workflows/INDEX.md, surface a typo suggestion instead of letting the reference fail later with an opaque "workflow not found" error.

Applies to these 12 sub-commands (everything that consumes an existing slug):

shape, slice, plan, implement, verify, review, handoff, ship, retro, instrument, experiment, benchmark

Does NOT apply to intake — it now resolves its first token by exact existence check inside reference/intake.md (Step 0): a match is slug-mode (a compressed-slice attach), a non-match is a mode keyword or the start of a fresh description, never a typo'd slug. Collision detection on a slug derived from a description lives in intake/default.md Step 0 instead. profile's first arg is <area>, not a slug — skip Step 0.5 for profile. design is excluded: its first token is an optional slug resolved by exact existence check inside reference/design.md (Step 0) — a non-matching first token is a design command, not a typo'd slug. probe is excluded: it is slug-only and enforces the requirement inside reference/probe.md — a non-matching first token triggers probe's own slug-required STOP, not a typo correction. simplify is excluded: its first positional is a scope keyword (branch/commit/plan/codebase), not a slug. auto is excluded: like design/probe, it owns its own slug resolution — its first token is an optional slug resolved by exact existence check / single-active inference inside reference/auto.md (Step 0); a non-matching first token is handled there (a description routes the user to /wf intake first), not corrected as a typo here. yolo is excluded: identical rationale to auto — it owns its own slug resolution (an optional slug resolved by exact existence check / single-active inference inside reference/yolo.md, Step 0), so a non-matching first token is handled there, not corrected as a typo here. Keep this exclusion list in sync with the 19-key dispatch table — exclude any future sub-command that creates a new slug, takes a non-slug first arg, or resolves its slug by its own existence check rather than consuming an existing one.

Procedure:

1. Identify the slug candidate. For most sub-commands the slug is the first positional token after the sub-command name. For implement, verify, review, plan, handoff, ship, retro, instrument, experiment, benchmark the slug is $1 of the sub-command's $ARGUMENTS. If $1 is empty (the user passed no slug), skip Step 0.5 — slug resolution falls through to the reference's existing single-active-workflow inference.
2. If .ai/workflows/INDEX.md does not exist → skip Step 0.5 (no registry → no candidate set for fuzzy match). The reference will handle the missing-slug case downstream.
3. Grep INDEX.md for an exact match: grep -P "^<candidate>\t" .ai/workflows/INDEX.md. If hit → slug is real, dispatch normally.
4. On miss, run a fuzzy match against every row's slug column (first tab-separated field, including closed rows — a real-but-closed slug is still useful as a suggestion):
   • Compute Levenshtein edit distance from <candidate> to each indexed slug.
   • Also check substring inclusion: a candidate that is a contiguous substring of an indexed slug (or vice versa) counts as a strong match.
   • Pick the best match: prefer the slug with edit distance ≤ 2, then the slug that contains <candidate> as a substring, then the slug <candidate> contains as a substring.
   • If no slug satisfies any of those conditions → no suggestion exists. STOP with: "Unknown slug <candidate>. Run /wf-meta status to list all workflows, or /wf intake <description> to start a new one."
5. If a best match exists, STOP with: "Unknown slug <candidate>. Did you mean <best-match>? (Run /wf-meta status to list all workflows.)" — where <closed-suffix> is (closed) if and only if the best-match row's status column is closed, empty otherwise. Show the user the corrected command verbatim so they can copy-edit it: e.g., "Retry: /wf <sub-command> <best-match> <remaining args>".
6. Step 0.5 is purely advisory — it never auto-corrects. The user must re-invoke with the suggested slug. This avoids the "the model decided to do something I didn't ask for" failure mode.

Step 1 — Execute

1. Read the reference file in full from ${CLAUDE_PLUGIN_ROOT}/skills/wf/reference/<key>.md.
2. Treat its content as your instructions for this invocation. Do not summarize, paraphrase, or skip — follow it verbatim.
3. The reference body contains the stage's full definition (preamble, prerequisites, conditional inputs, output contract, adaptive routing). Honor every conditional input and every artifact write the reference describes.
4. The remaining $ARGUMENTS after the matched key are the sub-command's own arguments — pass them through verbatim.

Step 2 — Emit Final Summary (MANDATORY)

After the reference's logic completes, emit a chat summary as the LAST output before returning control to the user. This contract is uniform across every sub-command this router dispatches; the reference may carry its own chat-return content, but this section governs the shape.

Format (compact — a short narrative, then the anchors):

wf <sub-command> complete: <slug-or-scope>

<Narrative — a short prose paragraph (no bullets, no field labels) telling the story: what this run produced or decided, how, and the top risk or caveat. See the Narrative rule below.>

Artifacts: <comma-separated paths, or "none">
Next: <recommended command, or "Done">

Rules:

• Always emit unless the reference STOPped with an error message — in that case the error replaces the summary.
• Verb-first first line. Name the sub-command and the workflow slug (or other scope if applicable: area for profile, etc.).
• Artifacts are the paths created or modified in this invocation (e.g., .ai/workflows/<slug>/04-plan-<slice>.md). Use "none" for read-only sub-commands.
• Narrative — the heart of the summary, REQUIRED for any sub-command that writes an artifact. In place of the old terse key-facts line, write a short prose paragraph (2–5 sentences, no bullets, no field labels) that tells the user what happened: for plan, what the plan is (the approach) and how it gets built; for implement, what was built and how; for verify, what was checked and the result, and whether it converged; for shape, the scope decided; for slice, how the work was split; for intake, what was understood; for ship/retro, what shipped and the key lessons. Weave the load-bearing counts, decisions, and the top risk into the prose. Write it in the voice defined in reference/_narrative-voice.md — relevance first, why before how, tradeoffs stated plainly, varied rhythm, never a "This implements…" opening; it is the same voice the artifact's ## The <Stage> story section uses. Omit only for genuinely read-only sub-commands.
• Next is a concrete invocation, or Done for terminal sub-commands (ship, retro). Never vague like "consider your next step".
• Internal audience. Workflow artifact paths under .ai/ ARE allowed here; this is the chat return, not external-facing copy. Outside this block, the External Output Boundary still applies.
• If the reference defines its own "Chat return contract" or "Hand off to user" step, treat that as the content spec — pick the load-bearing fields and keep it compact. A reference that says to "return ONLY" a receipt (slug / wrote / options) means only those receipt fields — it does NOT waive the substance summary above. Always surface what the artifact says — its key decisions, counts, verdict, top risk — not merely the paths it wrote. Keep the full detail in the artifact; the chat summary carries the gist.