flow-next-refine

name	flow-next-refine
description	Refine a spec or task before building: deep Q&A (business, technical, both) or a read-only research pass over external docs. Use to refine or interrogate requirements or read up on a new library.

Flow refine

Refine a task/spec: conduct an extremely thorough interview and write the refined details back, or (--scope=research) resolve the spec's external-library unknowns from official docs and write them back. The interview name remains a forwarding alias for one release.

.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.

Chart boundary

Existing-spec clarification stays primary. Interview refines a valid spec with unresolved judgment questions. Do not reopen discovery as /flow-next:chart unless the answers reveal that the effort itself is not yet specifiable - only then route backward to chart. Clear work that never needed a chart stays out of chart. Unsure of the hop: $flow-next-flow --explain.

Preamble

CRITICAL: flowctl is BUNDLED — NOT installed globally. which flowctl will fail (expected). Define once; subsequent blocks 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"

Role: technical interviewer, spec refiner Goal: extract complete implementation details through deep questioning (40+ questions typical)

Input

Full request: $ARGUMENTS

Accepts a Flow spec ID, a Flow task ID, a resolvable tracker handle, a file path, or nothing — recognition rules and the fetch command per type are in "Detect Input Type" below (single copy; the write-back command per type is in references/write-back.md).

Examples:

• /flow-next:refine fn-1-add-oauth
• /flow-next:refine fn-1-add-oauth.3
• /flow-next:refine fn-1 (legacy formats fn-1, fn-1-xxx still supported)
• /flow-next:refine docs/oauth-spec.md
• /flow-next:refine fn-1-add-oauth --scope=research (external-docs pass; no questions)

If empty, ask: "What should I interview you about? Give me a Flow ID (e.g., fn-1-add-oauth) or file path (e.g., docs/spec.md)"

Setup

Parse --scope=business|technical|both|research

Token-safe parsing for --scope / --biz / --tech lives in flowctl scope resolve — never re-implement inline. The subcommand strips scope tokens, preserves every other token in order (Flow IDs, paths, --docs, --strategy, ...), and emits the resolved scope plus a defaulted flag. The resolver's fallback when no scope flag is passed is technical (1.0.2 backward-compat) — but the skill does NOT silently run it: when defaulted == true, ask the user which pass to run after Detect Input Type (see "Scope selection when no flag passed" below). technical applies only when that question cannot be asked.

# Run BEFORE the --docs / --strategy strip block. Conflict / invalid value
# → non-zero exit; SKILL propagates.
#
# `--raw "$ARGUMENTS"` tokenizes via shlex INSIDE flowctl — preserves quoted
# paths with spaces (e.g., `/flow-next:refine --biz "docs/my spec.md"`).
# Unquoted `$ARGUMENTS` would word-split into broken tokens.
RESOLVED_JSON=$("$FLOWCTL" scope resolve --json --raw "$ARGUMENTS")
SCOPE=$(printf '%s' "$RESOLVED_JSON" | jq -r '.scope')
# true when no scope flag was passed — gates the "Scope selection when no
# flag passed" question below (older flowctl without the field → false,
# preserving the silent technical default).
SCOPE_DEFAULTED=$(printf '%s' "$RESOLVED_JSON" | jq -r '.defaulted // false')
# `remaining_args` is a JSON array of strings. Re-join with single spaces
# for downstream consumption; downstream code MUST re-tokenize via the
# same safe path (shlex) if it might re-encounter quoted paths.
ARGUMENTS=$(printf '%s' "$RESOLVED_JSON" | jq -r '.remaining_args | join(" ")')

Scope parsing, write policy, and bank selection come from flowctl scope resolve / scope write-policy / scope bank. A skill that re-implements the tokenizer, the section-ownership rules, or the bank mapping inline has broken this — the two copies drift and the inline one wins silently.

SCOPE == research asks no questions. Skip the doc-aware autodetect, the scope question, the question banks, and the interview rounds: run Detect Input Type below, then read references/research-scope.md and follow it. A business, technical, or both invocation never reads it.

Parse --docs / --no-docs / --strategy / --no-strategy flags

The four doc-aware override flags must be stripped from $ARGUMENTS before input-type detection so they don't get confused for a Flow ID or path. Two force variables carry the result — "" = autodetect, "on" = forced on, "off" = forced off:

RAW_ARGS="$ARGUMENTS"
DOC_AWARE_FORCE=""        # controls glossary + decisions
STRATEGY_AWARE_FORCE=""   # controls strategy independently

When the invocation carried ANY of --docs / --no-docs / --strategy / --no-strategy, STOP and read references/doc-aware.md § Flag parsing before proceeding — it holds the strip block (both pairs mutually exclusive, negation wins on conflict), the cascade rules, the flag matrix that is the contract for each combination, and the scope × doc/strategy interaction table. A bare invocation skips it: no flag token is present, so RAW_ARGS is $ARGUMENTS unchanged (whitespace-normalized) and both force variables stay empty (autodetect).

Doc-aware autodetect

Decide whether doc-aware mode activates. DOC_AWARE controls glossary + decisions; STRATEGY_AWARE controls the strategy-conflict behavior independently. Each has three paths (forced-on / forced-off / autodetect) per the flag matrix.

The default-autodetect rule is: doc-aware mode activates when any of three conditions has signal — glossary.total_terms > 0 (a) OR a decision entry exists (b) OR strategy.sections_filled >= 1 (c). The two flag pairs override (a)+(b) and (c) independently. Counting populated entries (rather than [[ -f <file> ]]) is deliberate — see references/doc-aware.md § Why counts, not file presence.

# DOC_AWARE: glossary + decisions. Probes and parses fail OPEN (|| DOC_AWARE=1).
DOC_AWARE=0
if [[ "$DOC_AWARE_FORCE" == "on" ]]; then
  DOC_AWARE=1
elif [[ "$DOC_AWARE_FORCE" == "off" ]]; then
  DOC_AWARE=0
else
  # NO pipelines in the probe — capture raw first, rc-checked; parse separately.
  GLOSSARY_RAW="$("$FLOWCTL" glossary list --json 2>/dev/null)" || DOC_AWARE=1
  DECISIONS_RAW="$("$FLOWCTL" memory list --track knowledge --category decisions --json 2>/dev/null)" || DOC_AWARE=1
  if [ "$DOC_AWARE" = "0" ]; then
    TERMS="$(printf '%s' "$GLOSSARY_RAW" | jq -r '.total_terms // 0' 2>/dev/null)" || DOC_AWARE=1
    DECS="$(printf '%s' "$DECISIONS_RAW" | jq -r '.entries | length // 0' 2>/dev/null)" || DOC_AWARE=1
  fi
  if [ "$DOC_AWARE" = "0" ] && { [ "${TERMS:-0}" -gt 0 ] || [ "${DECS:-0}" -gt 0 ]; }; then
    DOC_AWARE=1
  fi
fi

# STRATEGY_AWARE: strategy (independent of DOC_AWARE — autodetects on its own signal)
STRATEGY_AWARE=0
if [[ "$STRATEGY_AWARE_FORCE" == "on" ]]; then
  STRATEGY_AWARE=1
elif [[ "$STRATEGY_AWARE_FORCE" == "off" ]]; then
  STRATEGY_AWARE=0
else
  STRATEGY_RAW="$("$FLOWCTL" strategy status --json 2>/dev/null)" || STRATEGY_AWARE=1
  if [ "$STRATEGY_AWARE" = "0" ]; then
    STRAT_FILLED="$(printf '%s' "$STRATEGY_RAW" | jq -r '.sections_filled // 0' 2>/dev/null)" || STRATEGY_AWARE=1
  fi
  if [ "$STRATEGY_AWARE" = "0" ] && [ "${STRAT_FILLED:-0}" -ge 1 ]; then
    STRATEGY_AWARE=1
  fi
fi

if [ "$DOC_AWARE" = "1" ] || [ "$STRATEGY_AWARE" = "1" ]; then
  echo "DOC-AWARE GATE ACTIVE — STOP. Read references/doc-aware.md before drafting the first question."
fi

When the sentinel prints, STOP and read references/doc-aware.md before any further step, then apply its behaviors — Phase-zero glossary scan (a), fuzzy-term sharpening (b), code-versus-assertion contradiction (c), decision-record write (d), and code-vs-strategy contradiction (e). On the default no-docs path (DOC_AWARE=0 and STRATEGY_AWARE=0) the interview proceeds exactly as today — do not read the file.

Detect Input Type

Handle-recognition rule (R16): do NOT gate on a hard "must start with fn-" check. Before treating a single-token arg as a file path or freeform, route it through $FLOWCTL show <arg> --json — flowctl's widened resolver maps a tracker key (wor-17 / wor-17.M) to its linked spec/task, so a resolvable handle is the existing spec/task, never a new idea. Patterns 1-2 below are the common case; pattern 3 generalizes them to any resolvable handle.

1. Flow spec ID pattern: matches fn-\d+(-[a-z0-9-]+)? (e.g., fn-1-add-oauth, fn-12, fn-2-fix-login-bug)

   • Fetch: $FLOWCTL show <id> --json
   • Read spec: $FLOWCTL cat <id>

2. Flow task ID pattern: matches fn-\d+(-[a-z0-9-]+)?\.\d+ (e.g., fn-1-add-oauth.3, fn-12.5)

   • Fetch: $FLOWCTL show <id> --json
   • Read spec: $FLOWCTL cat <id>
   • Also get parent spec context: $FLOWCTL cat <spec-id>

3. Resolvable tracker handle: any single-token arg (not an .md path) that $FLOWCTL show <arg> --json resolves — e.g. a Linear key wor-17 (spec) or wor-17.3 (task). Use the canonical id from the JSON; a .-containing handle is a task (fetch parent spec too), otherwise a spec. Treat exactly like patterns 1-2; never re-create.

4. File path: a path-like token / .md extension that does NOT resolve via flowctl show

   • Read file contents
   • If file doesn't exist, ask user to provide valid path

Done when: the argument is classified as exactly one of the four patterns, every non-.md single-token arg was routed through $FLOWCTL show <arg> --json before that classification, and the target's content (spec body, task + parent spec, or file) is in hand for the scope recommendation below.

Scope selection when no flag passed

Fires ONLY when SCOPE_DEFAULTED=true (no --scope / --biz / --tech in the invocation). An explicit scope flag always wins and skips this section entirely.

Runs AFTER Detect Input Type — the spec/file content is in hand, so the recommendation is informed. Ask ONE plain-text numbered prompt (same blocking primitive as every interview question; the tool-unreachable fallback under "Question Format" applies):

• header: Interview scope
• body: Which interview pass should run? business = product framing (goal, users, boundaries, outcome AC — never decides architecture, stack, or APIs); technical = implementation details (architecture, API contracts, edge cases); both = business first, then technical. Recommended: <X> — <one-sentence rationale from the target's current state>. Confidence: [judgment-call].
• options (frozen): business, technical, both

Derive the recommendation from the target's current state:

• Biz sections empty AND tech sections empty (new idea, fresh spec, bare file) → recommend both — ground the product framing before any technical decision.
• Biz sections populated, tech sections empty or placeholder-only → recommend technical — the business layer exists; fill the technical layer.
• Tech sections populated, biz sections absent (1.0.2-shape solo spec) → recommend technical — refine in place.

Set SCOPE to the answer and proceed exactly as if the flag had been passed — write-policy, question bank, and pass behavior all follow the chosen scope. If the question genuinely cannot be asked (tool unreachable and no plain-text answer), fall back to technical and say so in the interview opener.

Why this exists: a PM invoking /flow-next:refine <spec-id> bare used to get a silent technical interrogation — stack/API questions they don't own, with skipped answers at risk of becoming rails-derived defaults. The scope question makes the business pass discoverable at the exact moment it matters.

Interview Process

CRITICAL REQUIREMENT: For every question, you MUST ask via the plain-text numbered prompt described below.

Ask the user via plain text. Render the options below as a numbered list 1. … N., followed by a final option N+1. Other — type your own answer. Print the question, then the numbered list, then stop and wait for the user's next message before continuing. Parse the reply as: a bare number 1–N+1 → that option; the literal text of an option label → that option; free text after Other → custom answer.

• ONLY ask via the plain-text numbered prompt
• Ask in rounds: each round carries the whole frontier (see Question Order below), split across plain-text numbered prompt calls of up to 4 questions each
• Expect 40+ questions total for complex specs

Question Format: Lead with Recommendation

Every plain-text numbered prompt body must include the agent's recommended option AND a confidence tier. Mirrors the canonical phrasing in flow-next-audit/SKILL.md:64 ("Lead with the recommended option and a one-sentence rationale").

Pattern:

• question.body: ". . Recommended: — . Confidence: [high | judgment-call | your-call]."
• question.options: neutral labels (no "(recommended)" markers — recommendation goes in the body; neutral options reduce anchoring)

Plain-language question contract (field feedback, eval-validated)

Applies to EVERY question, both scopes. The interviewee must be able to read a question once and answer it confidently without asking what it means — field feedback showed jargon-dense questions disempower exactly the people the interview exists to hear (baseline legibility scored 4/10 for a second-language PM; this contract scores 7.5+ at ~30% fewer tokens).

• Open the body with ONE sentence of stakes: what this question decides, in the audience's words.
• Write for the audience in everyday words; prefer the common word over the term of art. A term of art you genuinely need gets a plain-word gloss in ≤1 clause at first use (e.g. "counter-metrics — things we'd hate to make worse").
• No unexplained acronyms or tool/repo shorthand. In business scope, no implementation vocabulary (no schemas, endpoints, config keys).
• Every option description states its consequence in plain words: "Choose this if…" / "This means…".
• Gloss referenced acceptance criteria. When a question cites a spec R-ID, attach a short plain-words gist at first mention — "R3 (the audit line's required fields)" — never a bare "R3" the interviewee must open the spec to decode. Gist, not quote: pasting full criterion text bloats the question body.

Required content and trim order (priorities — NOT a length cap; never trade required content for brevity):

• ALWAYS keep, in this order: the stakes sentence; the recommendation + its one-sentence rationale; the confidence tier; the gloss for any term of art used; each option's consequence.
• TRIM FIRST, until the question reads in one pass: repetition between body and options, secondary background, hedging, restated option lists.