flow-next-pilot

name	flow-next-pilot
description	Single-tick autonomous build-loop conductor. One spec or the backlog, one stage per tick (pipeline.chainStages chains qa into make-pr), emits PILOT_VERDICT. Use when asked to pilot a spec or backlog.
user-invocable	false
allowed-tools	Read, Bash, Grep, Glob, Write, Edit, Skill

/flow-next:pilot — single-tick autonomous build-loop conductor

A tick is one invocation of /flow-next:pilot: select one ready spec, classify its current stage (plan, plan-review, work, the opt-in qa, make-pr), dispatch exactly one existing stage skill (with pipeline.chainStages on, make-pr after a fresh qa verdict is the only admissible second dispatch in the same tick), verify state advanced, and end with one terminal PILOT_VERDICT line. It is intentionally not a runner; /loop in Claude Code or /goal in Claude Code / Codex owns repeated invocation.

Pilot and Ralph are alternative autonomous drivers. Ralph is an external shell loop with receipt plumbing; pilot is an in-session conductor for host loop primitives. Never nest them, and never reuse Ralph harness state inside pilot.

Human judgment lives before pilot: the spec content, depends_on_epics, and the fn-58 ready gate are the consent boundary. Pilot executes the mechanical pipeline one stage at a time, with ambiguity reported as NEEDS_HUMAN.

Chart is outside the build loop (fn-135)

Chart is optional pre-capture discovery - never a pilot pipeline stage. Pilot never advances a chart tick. Optional unattended chart driving (host /loop on /flow-next:chart) stops terminally at attended decisions (CHART_VERDICT=NEEDS_HUMAN) - pilot does not absorb or continue that work. Capture, interview, and chart stay human-gated upstream of the ready consent boundary.

Preamble

CRITICAL: flowctl is BUNDLED — NOT installed globally. which flowctl will fail (expected). Define once; subsequent blocks (here and in workflow.md) use $FLOWCTL. Subagents that run in fresh context fall back to the repo-local copy:

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"

Cold session / tick start: $FLOWCTL brief first for session-scope orientation (one budgeted call).

Re-read this skill file at every tick start. A long /loop run executing from a stale in-context copy drifts from the file the repo actually ships — the file on disk is the contract, the remembered copy is not.

Probe an idle dispatched agent read-only — its side effects (commits, receipts, status fields), never a resume message: a resume restarts the agent, so "checking on" an agent that was merely slow turns one run into two.

Hard guards (before anything else)

Run these guards before selection, ledger writes, branch changes, or skill dispatch.

if [[ -n "${FLOW_RALPH:-}" || -n "${REVIEW_RECEIPT_PATH:-}" ]]; then
  echo "Ralph and pilot are alternative drivers — never nest them" >&2
  echo 'PILOT_VERDICT=NEEDS_HUMAN spec=- stage=- reason="nested under Ralph harness (FLOW_RALPH/REVIEW_RECEIPT_PATH set) — refuse to run"'
  exit 1
fi

if git status --porcelain | grep -v '^.. \.flow/' >/dev/null; then
  echo 'PILOT_VERDICT=NEEDS_HUMAN spec=- stage=- reason="dirty working tree at tick start"'
  exit 0
fi

Dirty tree means dirty outside .flow/; pilot leaves state untouched. No cleanup, no claim reset, no strike.

Mode Detection

Parse $ARGUMENTS for the scope lock, dry-run switch, and passthroughs. Unknown flags warn to stderr and are ignored. Defaults are research=grep, depth=short, and review resolved later via $FLOWCTL review-backend.

The loop handles both --flag=value and space-separated --flag value forms directly via a PREV token holder. It deliberately avoids bash positional parameters (shift-based parsing) — the host's argument interpolation rewrites positional tokens inside skill code blocks, which corrupts a case-on-positionals parse (observed live in the 1.13.0 dogfood).

RAW_ARGS="$ARGUMENTS"
PILOT_SPEC=""
PILOT_DRY_RUN=0
PILOT_REVIEW=""
PILOT_RESEARCH="grep"
PILOT_DEPTH="short"
PILOT_BACKLOG_OVERRIDE=""   # "" = use config; "1" = force backlog (--backlog/--auto)

PREV=""
for ARG in $RAW_ARGS; do
  case "$PREV" in
    --spec)     PILOT_SPEC="$ARG"; PREV=""; continue ;;
    --review)   PILOT_REVIEW="$ARG"; PREV=""; continue ;;
    --research) PILOT_RESEARCH="$ARG"; PREV=""; continue ;;
    --depth)    PILOT_DEPTH="$ARG"; PREV=""; continue ;;
  esac
  case "$ARG" in
    --spec|--review|--research|--depth) PREV="$ARG" ;;
    --spec=*)     PILOT_SPEC="${ARG#--spec=}" ;;
    --dry-run)    PILOT_DRY_RUN=1 ;;
    --backlog|--auto) PILOT_BACKLOG_OVERRIDE=1 ;;
    --review=*)   PILOT_REVIEW="${ARG#--review=}" ;;
    --research=*) PILOT_RESEARCH="${ARG#--research=}" ;;
    --depth=*)    PILOT_DEPTH="${ARG#--depth=}" ;;
    -*) echo "Unknown flag: $ARG (ignored by /flow-next:pilot)" >&2 ;;
    *)  echo "Unknown argument: $ARG (ignored by /flow-next:pilot)" >&2 ;;
  esac
done
[[ -n "$PREV" ]] && echo "Flag $PREV given without a value (ignored by /flow-next:pilot)" >&2
export PILOT_SPEC PILOT_DRY_RUN PILOT_REVIEW PILOT_RESEARCH PILOT_DEPTH PILOT_BACKLOG_OVERRIDE

No branch flag exists in v1. Branch resolution is pilot-owned from the selected spec's branch_name.

Pilot has no no-plan flag (fn-214): the no-plan decision is spec state — the no_plan field, set at capture time or via flowctl spec set-no-plan, read from SPEC_JSON at classification. A stray --no-plan argument lands in the unknown-flag branch above (one-line notice, tick proceeds; the affected zero-task specs classify plan, the safe default). Pilot never decides no-plan on its own — the field is an explicit human instruction on the item, never inferred, and pilot only forwards it to the work dispatch.

Autonomy mode resolution (R1) — gate the wide backlog behavior

Resolve PILOT_AUTONOMY once, here, so every downstream block keys off a single value. This block also captures the tick's ROOT CONFIG SNAPSHOT (fn-110) — the ONLY config get invocation across all pilot files (SKILL.md, workflow.md, references/backlog-mode.md): workflow.md's pipeline.qa probe and the pilot.gateClasses reads derive from the snapshot file via jq, never a second config call. The gate is a strict scalar string-enum — backlog mode activates only on the literal backlog (config pilot.autonomy), or when the per-run --backlog / --auto flag forced the override. Any other config value (ready, null, a coerced bool true, a typo) leaves pilot in ready mode — byte-for-byte unchanged behavior (the references/backlog-mode.md file is never even read):

# Root config snapshot: {"key":null,"value":{<merged config>}}. Persisted to a file
# because bash vars do not survive across prompt turns; later fences RECOMPUTE this
# same deterministic repo-hash-keyed path and jq it. The path lives under
# ${TMPDIR} — NEVER under repo-controlled .flow/tmp (autonomous symlink safety:
# a committed symlink must not redirect this write out of tree) — so a dry-run
# tick mutates nothing inside the repo; dry-run terminals also `rm -f` this
# snapshot (workflow.md), leaving no persistent scratch state. On capture
# FAILURE remove the file —
# downstream jq reads then error, which keeps the pipeline.qa probe's fail-open
# contract (probe error ⇒ ACTIVE) intact.
PILOT_CFG_SNAPSHOT="${TMPDIR:-/tmp}/flow-pilot-config-$(git rev-parse --show-toplevel 2>/dev/null | cksum | cut -d' ' -f1).json"
rm -f "$PILOT_CFG_SNAPSHOT" 2>/dev/null   # drop any stale/planted file (incl. a symlinked leaf) before the fresh write
$FLOWCTL config get --json > "$PILOT_CFG_SNAPSHOT" 2>/dev/null \
  || rm -f "$PILOT_CFG_SNAPSHOT"
PILOT_AUTONOMY="$(jq -r '.value.pilot.autonomy' "$PILOT_CFG_SNAPSHOT" 2>/dev/null)"
if [ "$PILOT_BACKLOG_OVERRIDE" = "1" ]; then
  PILOT_AUTONOMY="backlog"                       # --backlog / --auto forces backlog this run
elif [ "$PILOT_AUTONOMY" != "backlog" ]; then
  PILOT_AUTONOMY="ready"                         # ONLY the literal `backlog` enables — never bool true / typos / null
fi
export PILOT_AUTONOMY

When PILOT_AUTONOMY=ready (the default), pilot behaves exactly as documented in workflow.md Phases 1–6 — no backlog-mode code path runs and references/backlog-mode.md is not loaded. When PILOT_AUTONOMY=backlog, read references/backlog-mode.md top to bottom, execute its backlog-only setup, then continue with workflow.md Phase 1. The reference owns the backlog-only verdict extension plus SELECT/TRIAGE/ASK context; workflow.md keeps the enforcing guards and action sites.

The verdict contract (read this before the workflow)

The /goal validator is transcript-blind: it reads conversation output only and never runs tools. Every tick therefore echoes its verification evidence into the output: flowctl status fields, task counts, task status transitions, and the gh-confirmed PR URL for make-pr.

Every tick ends with exactly one terminal line, the last line of the response, with nothing after it. The common ready-mode grammar is:

PILOT_VERDICT=<ADVANCED|NO_WORK|DEFERRED_TO_LAND|BLOCKED|NEEDS_HUMAN> spec=<id> stage=<stage> reason="<one line>"

Use spec=- and stage=- when no spec was selected. Stage values are exactly plan, plan-review, work, qa (opt-in — only when pipeline.qa==on), make-pr, land, or -. A chained tick (pipeline.chainStages==on) names every dispatched stage in order joined by + — exactly qa+make-pr — and carries the last dispatched stage's verdict.

Dry-run snapshot cleanup. Under --dry-run (PILOT_DRY_RUN=1), at EVERY terminal PILOT_VERDICT emission — the classification stop, the diagnostic TRIAGED exit, every NO_WORK / DEFERRED_TO_LAND / hard-guard exit — remove the root config snapshot BEFORE printing the verdict, so a dry-run leaves no persistent scratch state:

rm -f "${TMPDIR:-/tmp}/flow-pilot-config-$(git rev-parse --show-toplevel 2>/dev/null | cksum | cut -d' ' -f1).json"

Recompute the path exactly as above (vars die across prompt turns). Live (non-dry-run) ticks keep the snapshot for the tick's remaining fences; it is overwritten fresh by the next tick's capture. Never blocks, fail-open (rm -f on a missing file is a no-op).

DEFERRED_TO_LAND is a distinct non-terminal-work verdict (stage land): every remaining all-done candidate has an open PR that land — not pilot — owns. It is deliberately separated from NO_WORK so a driver can route it to /flow-next:land instead of stopping; an all-done spec with an open PR is real outstanding work, never absence of work.

Driver condition examples:

/goal keep running /flow-next:pilot until it prints PILOT_VERDICT=NO_WORK, or stop after 20 turns
/goal keep running /flow-next:pilot --review=codex until PILOT_VERDICT=NO_WORK or PILOT_VERDICT=NEEDS_HUMAN

Forbidden

• Asking the user anything in the tick path. Pilot is autonomous; ambiguity maps to NEEDS_HUMAN. In backlog mode, ambiguity that needs a person is surfaced async via the ask stage (ASKED) — never an interactive plain-text numbered prompt.
• Dispatching any skill outside the stage set {plan, plan-review, work, make-pr} - plus qa only when pipeline.qa==on (fn-72: an opt-in, gate-reversed stage at the all-done juncture before make-pr; with the gate off, qa is forbidden and the stage set is byte-for-byte unchanged). Backlog mode (PILOT_AUTONOMY=backlog) additionally invokes /flow-next:tracker-sync for the reconcile / list-open / list-comments / list-relations / question ops - these are read/surface-only tracker calls (list-comments reads parked question rounds; list-relations reads dependency relations for dep-ordering), not a pipeline stage, and run only on the backlog path. Capture, interview, chart, resolve-pr, merge, and release are never pilot stages - they stay forbidden for their distinct reasons (capture/interview/chart are human authoring and discovery upstream of the consent boundary; resolve-pr/merge/release are land's territory downstream of the PR). Opening qa under its gate, or tracker-sync under backlog mode, sets no precedent for any of those — a tick that dispatched one of them by analogy has broken this.
• Dispatching a second stage in one tick — except, under pipeline.chainStages==on, make-pr after this tick's qa verified a fresh terminal verdict (the closed one-row chain table in workflow.md Phase 5, Chained stage); any other second dispatch breaks the single-tick contract.
• Re-implementing sub-skill logic. Pilot owns selection, dispatch, verification, verdicts, and the strikes ledger only. The backlog-mode SELECT/TRIAGE/ASK workflow lives in references/backlog-mode.md (loaded only when PILOT_AUTONOMY=backlog); the question-anchor authoring + answer round-trip live in tracker-sync — backlog mode invokes them, never re-implements them.
• Never merging / never invoking land (R6) — in either mode, the terminus is make-pr (draft). Merge stays human-gated. Backlog mode never calls /flow-next:land, gh pr merge, or any merge path.
• Never authoring a spec (backlog mode) — capture/interview are human-gated upstream. A missing/too-thin spec is surfaced as a "needs capture/interview" gap and parked (ASKED), never auto-written. The only writing the ask stage may do is fill an obvious blank in an existing spec — never create a spec stub from a bare ticket.
• Touching gh anywhere except the all-done classification branch's PR probe and the make-pr verification probe.
• Printing anything after the PILOT_VERDICT line.
• Running under Ralph (FLOW_RALPH / REVIEW_RECEIPT_PATH).

Workflow

Execute workflow.md in order:

1. guards — refuse Ralph nesting, refuse dirty non-.flow/ start state, resolve the .git strikes ledger (read-only at this point). Done when: both guards have passed and LEDGER_JSON is loaded without a write.
2. select — two-pass ready-spec selection with dependency, claim, and re-bless checks. Done when: exactly one spec is selected, or the pool is empty and the terminal split has been chosen.
3. classify — derive one stage from flowctl state; probe gh only in the all-done branch. Done when: exactly one stage from the allowed set is named, with the consulted status fields echoed.
4. branch — resolve the spec branch matrix before work or make-pr. Done when: the worktree sits on the branch the matrix row names, or the tick has stopped NEEDS_HUMAN on a failed checkout.
5. dispatch — invoke exactly one stage skill with mode:autonomous plus review/research/depth passthroughs (plus the gated qa→make-pr chain when pipeline.chainStages is on). Done when: the classified stage skill has been invoked and returned — and, on a chained tick, so has make-pr; no other second dispatch happened.
6. verify — re-read flowctl state, or gh for make-pr, and echo before/after evidence. Done when: every dispatched stage has its own before/after evidence block plus stage-outcome line in the transcript and each advanced is decided from observed state.
7. report — clear or record strikes, optionally unready on the second healthy no-advance tick, and print the terminal verdict. Done when: the ledger reflects this tick and the terminal PILOT_VERDICT line is the last line of the response.

Backlog mode (PILOT_AUTONOMY=backlog). When the autonomy gate resolved to backlog, the SELECT and TRIAGE/ASK behavior follows references/backlog-mode.md — the agentic floor scheduler (loaded only in this mode). It widens SELECT (pull-before-scan, union tracker-only items, dep-order, skip parked) and adds the triage/ask stages in front of CLASSIFY; a workable item flows into the existing classify → branch → dispatch → verify path unchanged. workflow.md Phase 0.5 resolves the mode and routes; workflow.md Phase 1.5/Phase 3.5 carry the backlog SELECT + TRIAGE/ASK hooks and the safety invariants. The single-tick contract is unchanged: one item, one stage (or the gated qa→make-pr chain under pipeline.chainStages) or one durable park, one terminal verdict.

Unattended runs — rp caveat

The rp review backend runs headlessly through the CE-first CLI ladder — it needs RepoPrompt CE running on the same Mac (cold start: open -ga "RepoPrompt CE"; a stopped app fails fast). Classic is only the final compatibility fallback. On remote/CI machines, use --review=codex, --review=copilot, --review=cursor, --review=claude (Claude-family verdict from any host; same-family on Claude Code, recorded in the receipt), --review=host (host-native subagent; needs a cross-family pin in AGENTS.md model-routing), or --review=none. Wall-clock limits and iteration caps belong to the driver (/goal --tokens, /goal stop clauses, or /loop cadence); a pilot tick has no timeout machinery.