/flow-next-flow

name: flow-next-flow
description: Conductor for whatever the user has - an idea or a request for a change, a spec or task id, a tracker issue, a branch or a path, a pasted bug report or console output, a how or why question about the code, something slow to speed up, a cleanup that keeps behaviour, a design fork to settle, or "what should I do next". Use when the user states any of these without naming a skill; --auto runs unattended.
user-invocable: false
allowed-tools: AskUserQuestion, Read, Bash, Grep, Glob, Write, Edit, Task, Skill

/flow-next:flow - the conductor

Flow chooses the next step so the user does not have to. It reads what it was given, routes from the shared routing reference, runs the routed stage skill, and continues until the next decision that belongs to a human. It re-implements no stage logic: capture, refine, plan, plan-review, work, qa, make-pr, resolve-pr, and land keep their own contracts, receipts, and gates.

**Role:** conductor, inline (no `context: fork`) so `AskUserQuestion` stays reachable. On hosts without it, fall back to a plain-text numbered prompt with a final `Other - type your own answer` option. (sync-codex.sh rewrites the tool name for the Codex mirror.)

**Read [workflow.md](workflow.md) for the hop loop.** The routing rules live in `references/`, one file per rule, each opening with its decision record. Read a reference only at the step that names it; a step that does not name one reads nothing.

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="${DROID_PLUGIN_ROOT:-${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"

Mode detection

Parse `$ARGUMENTS` as exact tokens (never substrings), before any read or write: `--auto` sets `AUTO=1`; `--tick` sets `AUTO_TICK=1` (one hop, then stop; meaningful only with `--auto`); `--explain` sets `EXPLAIN=1`; `--review=<backend>` sets `REVIEW_OVERRIDE` and is passed through unchanged to every stage it dispatches. The destination parse is shared by both modes:

FLOW_UNTIL=""
FLOW_DESTINATION_ERROR=0
LAND_AUTHORIZED=0
LAND_SCOPE_SPEC=""
LAND_SCOPE_PR=""
LAND_BASE_ROOT=""
AUTO=0
for ARG in $ARGUMENTS; do
  case "$ARG" in
    --auto) AUTO=1 ;;
    --until=merge) FLOW_UNTIL=merge ;;
    --until|--until=*) FLOW_DESTINATION_ERROR=1 ;;
  esac
done
if [ "$FLOW_DESTINATION_ERROR" = 1 ]; then
  if [ "$AUTO" = 1 ]; then
    echo 'PILOT_VERDICT=NEEDS_HUMAN spec=- stage=- reason="invalid destination; use --until=merge"'
  else
    echo 'NEEDS_HUMAN: invalid destination; use --until=merge'
  fi
  exit 1
fi
export FLOW_UNTIL

`--until=merge` authorizes landing the selected item in this invocation, independently of `--auto`. Without it the pre-merge boundary remains; current explicit item-scoped user authorization can also authorize landing. Read `references/tail.md` when reaching that boundary. Consent is host context, never recovered from an environment variable, old transcript, or receipt. A fresh session needs the flag again or current explicit authorization; revocation stops subsequent mutations. Consume destination tokens rather than passing them to a build stage. Everything else is the starting point, verbatim - flow adds no input classifier. A tracker issue id or URL is read through the access the session already has (the sync bridge, an MCP, `gh`, `glab`); flow adds no input adapter.

**`AUTO=1`: read [auto.md](auto.md) and follow it.** Attended runs never load it.

Autonomy refusal - runs right after the token parse

Attended flow, `flow --auto`, and Ralph are three drivers and are never nested. Without `--auto`, flow is attended: under any autonomy marker (scan the marker namespace: `FLOW_RALPH`, `FLOW_AUTONOMOUS`, `REVIEW_RECEIPT_PATH`, `AUTONOMOUS=1`, a `mode:autonomous` token), stop before any read or write:

NEEDS_HUMAN: /flow-next:flow is attended - run /flow-next:flow --auto for unattended runs

With `--auto`, the refusal is Ralph-only (`FLOW_RALPH`, `REVIEW_RECEIPT_PATH`, in auto.md's hard guards), because `--auto` sets `FLOW_AUTONOMOUS` and `mode:autonomous` for the stages it dispatches. A run that routed, dispatched, or asked under a marker it should have refused has broken this.

Invariants (every run)

• **Route on content and context, never on input kind.** Read what was given, decide what it is, then match `references/route-matrix.md` at the route step.
• **Ask only on a fork that is material and not observable.** Before any "which approach" or "what should this do" question, classify the fork per `references/prototype-before-ask.md`: an observable answer is settled by running something; only a product or preference call becomes a question, and at most one per hop.
• **Never fabricate a review, QA, or completion verdict** to pass a gate. Every stage flow skips is recorded with its reason (`stage: <name> - skipped(<kind>: <detail>)`), never omitted.
• **Land owns merge and spec close.** The only driver-composition exception is a currently authorized, item-scoped flow invocation of land per `references/tail.md`. Never dispatch another flow, pilot, Ralph, or a loop from inside a run.
• **`--explain` writes nothing and dispatches nothing.** It prints the route, the positive signal, the safe skip and its kind, and why not the alternatives, in the recommendation shape from `references/route-matrix.md`.
• **Host command form:** print every copy-pasteable flow-next command in the spelling this host invokes - the flat `/flow-next-<name>` form when the resolved plugin root carries `.flow-next-opencode-manifest` (an OpenCode install), otherwise exactly