Flow work

Execute a plan systematically. Focus on finishing.

Follow this skill and linked workflows exactly. Deviations cause drift, bad gates, retries, and user frustration.

.flow/ is the only task tracker. A run that recorded task state in a markdown TODO, a plan file, TodoWrite, or any other tracker has broken this — all task state is read and written via flowctl.

Preamble

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

FLOWCTL="${DROID_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT}}/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"

Hard requirements (non-negotiable):

• Every completed task passes through flowctl done and a verified done status. A task treated as finished while flowctl show <task> still reads todo or in_progress has broken this.
• Staging is git add -A, never an explicit file list — that is what pulls .flow/ and scripts/ralph/ (when present) into the commit. A commit whose diff omits the run's .flow/ writes has broken this.
• Completion is claimed only after flowctl show <task> reports status: done. A completion claim printed ahead of that read has broken this.
• /flow-next:impl-review is dispatched only on a green tree. A review sent while tests or Quick commands are red has broken this.

Role: execution lead, plan fidelity first. Goal: complete every task in order with tests.

Ralph Mode Rules (always follow)

If REVIEW_RECEIPT_PATH is set or FLOW_RALPH=1, the Hard requirements above are the receipt contract, plus:

• The verified done status precedes the commit that carries the task. A commit landing ahead of its verified flowctl done has broken this.
• Tracking stays in .flow/ via flowctl — TodoWrite is never the task record. A Ralph iteration whose task list lives in TodoWrite has broken this.

Done when: the Hard requirements hold, and every completed task's done was verified before its commit.

Autonomous Mode (questions off, no receipt obligations)

Before gates, treat this host-expanded block as literal prompt data, never shell:

$ARGUMENTS

Strip standalone whitespace token mode:autonomous into WORK_ARGS; preserve all else verbatim (spaces/quotes/globs). Set/export AUTONOMOUS=1 if found or FLOW_AUTONOMOUS=1; otherwise set/export AUTONOMOUS=0.

Continue with WORK_ARGS; carry the exported marker into later shell fragments. If AUTONOMOUS=1:

• No setup question is asked (branch + review questions below are suppressed). A run that puts either question to the user under AUTONOMOUS=1 has broken this.
• Branch defaults deterministically to --branch=new when no explicit branch option is present — under autonomy "the user's answer" never exists, and defaulting to the current branch could commit straight to main. A chained spec (flowctl spec chain names a parent) forks from the parent's remote tip instead of main (phases.md Phase 2). Name the new branch exactly the spec's branch_name field ($FLOWCTL show <spec-id> --json | jq -r '.branch_name') — the branch matrix of flow --auto, its all-done PR probe, and make-pr's branch-match spec detection all key on that name; an ad-hoc name breaks continuity across hops and invocations.
• Review = explicit --review passthrough if present, else the configured backend (none when REVIEW_BACKEND is ASK).
• Autonomy ≠ Ralph. Neither signal sets FLOW_RALPH, implies REVIEW_RECEIPT_PATH receipt obligations, or activates ralph-guard hooks. The Ralph rules above apply only under their own markers (the done/git add -A/no-TodoWrite discipline is universal anyway).
• Never hang on a question. A genuinely unanswerable ambiguity → stop cleanly with a one-line NEEDS_HUMAN: <reason> report instead of asking.

Input

Full request after mode parsing: $WORK_ARGS

Accepts:

• Flow spec ID fn-N-slug (e.g., fn-1-add-oauth) or legacy fn-N/fn-N-xxx to work through all tasks
• Flow task ID fn-N-slug.M (e.g., fn-1-add-oauth.2) or legacy fn-N.M/fn-N-xxx.M to work on single task
• Markdown spec file path (creates spec from file, then executes)
• Idea text (creates minimal spec + single task, then executes)
• Chained instructions like "then review with /flow-next:impl-review"

Examples:

• /flow-next:work fn-1-add-oauth
• /flow-next:work fn-1-add-oauth.3
• /flow-next:work fn-1 (legacy formats fn-1, fn-1-xxx still supported)
• /flow-next:work docs/my-feature-spec.md
• /flow-next:work Add rate limiting
• /flow-next:work fn-1-add-oauth then review via /flow-next:impl-review

If no input provided, ask for it.

FIRST: Parse Options or Ask Questions

Check configured backend:

REVIEW_BACKEND=$($FLOWCTL review-backend)

Returns: ASK (not configured), or rp/codex/copilot/cursor/claude/host/none (configured).

Option Parsing (skip questions if found in arguments)

Parse WORK_ARGS for these patterns. If found, use them and skip corresponding questions:

Branch mode:

• --branch=current or --current or "current branch" or "stay on this branch" → current branch
• --branch=new or --new-branch or "new branch" or "create branch" → new branch
• --branch=worktree or --worktree or "isolated worktree" or "worktree" → isolated worktree

Review mode:

• --review=codex or "review with codex" or "codex review" or "use codex" → Codex CLI
• --review=copilot or "review with copilot" or "copilot review" → GitHub Copilot CLI
• --review=cursor or "review with cursor" or "cursor review" → Cursor CLI (cursor-agent)
• --review=claude or "review with claude" or "claude review" → Claude Code CLI (claude -p; same-family on a Claude Code host, recorded in the receipt)
• --review=host or "host review" or "host-native review" → host-native fresh-context reviewer subagent (cross-family pin from the AGENTS.md model-routing section)
• --review=rp or "review with rp" or "rp chat" or "repoprompt review" → RepoPrompt chat (via flowctl rp chat-send)
• --review=none or --no-review or "no review" or "skip review" → no review
• --review=export or "export review" or "external llm" → REFUSE at parse time, before any dispatch: export is not an impl-review backend — never fall through to the configured backend and never pass it as REVIEW_MODE; stop and point at /flow-next:plan-review --review=export, where export lives

(All non-none review modes route through /flow-next:impl-review, which resolves the configured/overridden backend — codex, copilot, cursor, claude, rp, or host — itself.)

No-plan (direct spec execution):

• --no-plan or "no plan" or "skip planning" or "work directly without planning" → set NO_PLAN=1; it pre-answers Phase 1's zero-task fork so the fork's ask never fires when intent is stated
• The spec's own no_plan field (no_plan: true in $FLOWCTL show <spec-id> --json, set at capture time or via flowctl spec set-no-plan) counts the same as the flag: it is an explicit human instruction carried by the item, read at Phase 1's fork, never inferred
• Contradictory signals (the flag or field says direct, the prose asks to plan first) → the fork asks instead of guessing
• Existing intentional tasks govern despite a stale direct signal. A sole implicit_owner task under no_plan: true retains the direct route on resume; Phase 1 resolves the distinction.
• The fork's recommendation comes from the shared rule in plan-vs-no-plan.md, read only when the fork fires; implementation review, coverage, completion policy and opt-in QA remain unchanged on this route.
• The fork's semantics (ask, autonomous refusal, durable choice, implicit-task mint) live in phases.md Phase 1's gated references/no-plan-route.md, read only when the fork fires

Autonomous mode:

• AUTONOMOUS=1 → suppress all setup questions; use the defaults above.

If the options are absent from the arguments

If AUTONOMOUS=1 (autonomous mode): ask nothing — apply the autonomous defaults and continue to the workflow.

Otherwise (interactive): the branch question is answered before anything else happens. A run that reads a file or writes code before the answer arrives has broken this. Read references/setup-questions.md, ask the block it names for the current REVIEW_BACKEND (branch-only when a backend is configured; branch AND review when REVIEW_BACKEND is ASK), and wait for the response.

Defaults when empty/ambiguous:

• Branch = new
• Review = configured backend if set, else none (no auto-detect fallback)

Done when: the branch mode (and, under REVIEW_BACKEND=ASK, the review mode) is resolved from arguments, the user's answer, or the autonomous defaults — and no file has been read and no code written before that point.

Workflow

After setup questions answered, read phases.md and execute each phase in order.

Worker subagent model: Each task is implemented by a worker subagent with fresh context. This prevents context bleed between tasks and keeps re-anchor info with the implementation. The main conversation owns the ready frontier. By default it schedules on the rolling frontier (phases.md Phase 3 → references/rolling-scheduler.md): a new ready task is admitted at every worker-return event, each worker in an isolated workspace, with review and completion conductor-owned per task. It falls back to the wave loop - concurrent safe subsets joined at wave boundaries, or a single worker in the checkout - for a task-id run, when plan-sync is on, when the spec has fewer than two open tasks, or when its tasks form a sequential chain; the route and reason print once as Scheduling:. On the rolling route, and in any concurrent wave (PARALLEL_WAVE: true), a worker implements, tests, and commits in its isolated workspace, then returns task-unique handover files without completing shared Flow state; the conductor integrates before review, completion, tracker projection, plan-sync, or the next admission. The wave route's single-worker path shares the conductor's checkout and self-completes.

If user chose review, pass the resolved review mode to every worker. On the wave route's single-worker path the worker invokes /flow-next:impl-review itself after implementation and loops until SHIP. On the rolling route, and for any concurrent wave (PARALLEL_WAVE: true), the worker never reviews or completes: the conductor runs the review after integration and calls flowctl done only on SHIP.

Completion review gate: Default-on in SPEC_MODE when a review backend is configured. After all tasks are done, phases.md 3g invokes /flow-next:spec-completion-review — except it skips when the spec has exactly one task, that task's per-task impl-review reached SHIP (REVIEW_MODE was not none), and every spec R-ID is covered by that task's declared satisfies. On skip, persist completion_review_status not_required via the CAS setter (--if-current unknown) and record the Phase 5 stage line; a miss that reads not_required is an already-excused re-entry (same skip branch), while a verdict-status miss falls through to the normal status check without a skip line — policy outcome, never a SHIP. flowctl next --require-completion-review is a flowctl-level gate for driver loops; this skill does not read it. The spec-completion-review skill handles the fix loop internally until SHIP.

Tracker sync (opt-in, off by default)

The no-tracker path is the documented default and is behaviorally unchanged. A tracker touchpoint fires only when the bridge is active and its specific event is opted in (the shared gating predicate); otherwise it is a silent no-op — no new steps, no new prerequisites. A run that adds a tracker step with the bridge inactive has broken this. The bridge is active iff flowctl sync active --json reports active: true. The touchpoint mechanics — the perEvent table, the shared gating predicate, and the three dispatch payloads (phases.md 3b.1 first-claim, 3d.1 done, 3g completion-review) — live in references/tracker-touchpoints.md. That reference is read only when a phases.md tracker gate prints its active read/execute/continue sentinel (bridge active, or the gate's probe errored — fail open); a default bridge-inactive run that loaded it has broken this. Phase 5's end-of-run sync check + retro-fire + the mandatory four-state Tracker sync: summary slot stay inline in phases.md Phase 5 and run on every run (the slot reads n/a (bridge inactive) when no tracker is configured).

Handle recognition (R16): /flow-next:work wor-17 / work wor-17.1 resolve the existing linked spec/task — the Phase 1 input grammar routes any single-token arg through flowctl show (which resolves tracker handles) before treating it as idea text, so a tracker key is never re-created as a new spec.

Spec-id scheme on mint: with a tracker configured, tracker-first is the recommended team default (tracker.specIds=tracker) — it stops parallel fn-N collisions. Gate: phases.md Phase 1.

Unlink / re-link lifecycle: documented with the touchpoints in references/tracker-touchpoints.md (Unlink / re-link lifecycle) — no work-run step.

Guardrails

• The branch question is answered before the run starts. A run that began on an unresolved branch choice has broken this.
• A plan or spec exists before implementation starts. A run that began with no .flow/ spec has broken this.
• Tests run. A task marked done with its spec's Quick commands unrun has broken this.
• No task is left half-done. A run that ends with a task still in_progress and no NEEDS_HUMAN/blocked report has broken this.
• Task tracking lives in .flow/ via flowctl. A run tracking tasks in TodoWrite, or writing a plan file outside .flow/, has broken this.