/flow-next:capture — agent-native conversation → spec

A discussion (or a /flow-next:prospect survivor) often holds a complete spec that never gets written down, and the next session loses it. Capture is the synthesis step: the host agent extracts the user's turns, drafts a spec against the resolved template, tags only what it paraphrased or inferred ([paraphrase] / [inferred] / [strategy:<track>]; the user's verbatim words stay untagged), writes it through flowctl, prints a compact summary, and offers the saved file in the editor (docs/read-back.md). The capture request authorizes the write; substantive choices still get short questions. No Python synthesizer, no model subprocess.

flowctl supplies thin plumbing (spec create, spec set-plan, spec set-branch, memory search) plus the chart link-spec callback after a chart-briefing capture. Capture never writes chart files or a chart's ready flag; chart never writes .flow/specs.

Clear ideas and finished chart briefings route here; capture never manufactures a chart for clear work. Unsure: /flow-next:flow --explain.

Read workflow.md for the phases and phases.md for the source tags. Branch-specific machinery lives in references/*.md, loaded only when its gate fires.

Read working-rules.md first unless you already have this run; it holds for every step of this skill.

Preamble

CRITICAL: flowctl is BUNDLED — NOT installed globally. which flowctl will fail (expected). workflow.md's Preamble defines $FLOWCTL; later blocks (there and in phases.md) use it.

Inline skill (no context: fork) — plain-text numbered prompt must stay reachable across phases; subagents cannot call plain-text numbered prompts.

Mode Detection

Parse $ARGUMENTS for the literal tokens mode:autofix, from:flow and from:goal and the flags --rewrite <spec-id>, --from-compacted-ok, --yes, --override-strategy, --no-plan. Strip recognized tokens; the remainder is ignored (the conversation is the input), except under from:goal, where it names the goal or its source.

RAW_ARGS="$ARGUMENTS"
MODE="interactive"
REWRITE_TARGET=""
FROM_COMPACTED_OK=0
COMMIT_YES=0
OVERRIDE_STRATEGY=0

if [[ "$RAW_ARGS" == *"mode:autofix"* ]]; then
  MODE="autofix"
  RAW_ARGS="${RAW_ARGS//mode:autofix/}"
fi

if [[ "$RAW_ARGS" =~ --rewrite[[:space:]]+([^[:space:]]+) ]]; then
  REWRITE_TARGET="${BASH_REMATCH[1]}"
  RAW_ARGS="${RAW_ARGS//--rewrite ${REWRITE_TARGET}/}"
fi

if [[ "$RAW_ARGS" == *"--from-compacted-ok"* ]]; then
  FROM_COMPACTED_OK=1
  RAW_ARGS="${RAW_ARGS//--from-compacted-ok/}"
fi

# --yes (autofix write gate), --override-strategy, --no-plan, from:flow and
# from:goal authorize durable writes, so they are EXACT-token matches: lookalikes
# ("--yesterday", "--no-planning", "from:flowchart", "from:goals") stay in the remainder.
NO_PLAN_OPT=0
FROM_FLOW=0
FROM_GOAL=0
CLEANED_ARGS=""
for TOK in $RAW_ARGS; do
  if [ "$TOK" = "--yes" ]; then
    COMMIT_YES=1
  elif [ "$TOK" = "--override-strategy" ]; then
    OVERRIDE_STRATEGY=1
  elif [ "$TOK" = "--no-plan" ]; then
    NO_PLAN_OPT=1
  elif [ "$TOK" = "from:flow" ]; then
    FROM_FLOW=1
  elif [ "$TOK" = "from:goal" ]; then
    FROM_GOAL=1
  else
    CLEANED_ARGS="$CLEANED_ARGS $TOK"
  fi
done
RAW_ARGS="$CLEANED_ARGS"

# from:goal is an unattended write the --goal invocation already consented to.
if [ "$FROM_GOAL" = "1" ]; then
  MODE="autofix"
  COMMIT_YES=1
  echo "GATE ACTIVE — STOP. Read references/from-goal.md before continuing."
fi
if [ "$MODE" = "autofix" ]; then
  echo "GATE ACTIVE — STOP. Read references/autofix-mode.md before continuing."
fi   # default branch: bare no-op — NO link, NO read path

When the sentinel prints, read references/autofix-mode.md before Phase 0; it owns the per-phase autofix rules. Under from:goal, read references/from-goal.md as well; it says where a goal run's source differs. On the interactive path, read nothing.

Questions (interactive only)

Interactive capture follows the working rules' attended contract, plus:

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.

• Use plain-text numbered prompt (load it with ToolSearch select:plain-text numbered prompt if needed); plain-text numbered options only when the tool is unreachable. Never skip a question silently.
• The body leads with the recommendation, a one-sentence reason, and a confidence tier ([high] / [judgment-call] / [your-call]; before asking, read references/confidence-tiers.md). Option labels stay neutral and state their consequence.
• Plain language: one sentence of stakes, everyday words, a short gloss for any needed term (R-ID, [inferred]).
• Never ask for facts the conversation already gave. Phase 3 asks only its must-ask cases; other [inferred] content surfaces in the summary instead.

Forbidden behaviors

These protect spec trust:

• Unstated technology. "Needs persistence" is fine; "uses PostgreSQL" needs the user to have said it. Capture writes intent; plan writes implementation.
• Unmarked invention. An untagged line claims to be the user's words and must be findable in the evidence; every paraphrase or fill-in carries its tag, and [inferred] criteria surface in the summary for the user to keep, edit, or drop. Saving never upgrades a tag.
• Process fences as spec content. New-vs-rewrite decisions, ready-marking, "do not implement", and the duplicate-scan outcome are capture's lifecycle, never spec body or ## Boundaries lines, and never evidence lines. Boundaries hold only product constraints a worker could get wrong.
• Coordinates in the spec body. State contracts (types, signatures, behaviors, invariants), not code snippets, file paths, or line numbers, unless the exact location is itself the decision. Paths belong in plan's task specs.
• Silent overwrite. Replacing a spec needs --rewrite <spec-id>; otherwise Phase 0's duplicate branch decides.
• Acting without its own consent. Never auto-split (the user picks), never mark ready or add glossary terms without their separate questions (autofix writes neither; seeding an empty glossary is /flow-next:prime's job), never treat a saved spec as implementation consent, and never admit a draft/stale briefing silently or treat a forced draft chart briefing as final (the override names the D-IDs and reads back the risk).
• Committing. Capture never stages or commits; the user owns when .flow/ changes land.
• context: fork. Blocking questions must stay reachable.

Workflow

Execute workflow.md in order:

The spec at .flow/specs/<spec-id>.md is the deliverable. Autofix without --yes prints the draft summary and exits 0 with no spec allocated.