interview

name	interview
description	dure deep interview — interview a vague requirement until understanding converges quantitatively, then crystallize it into .dure/specs/<slug>.md. Four traits — evidence-grounded, red-team, auto-research, quantitative convergence gate.
argument-hint	<one-line requirement>
allowed-tools	Bash Read Grep Glob Task Write AskUserQuestion
disable-model-invocation	true

/dure:interview — Deep Interview

Request: $ARGUMENTS

Goal: interview the request using the methodology in playbook.md (spec §4), and crystallize .dure/specs/<slug>.md + .dure/interview-logs/<slug>.md only after the stop conditions (§4.4) are met.

Language: conduct the interview in the user's language; write all artifacts (spec, log) in English (ADR-0007).

0. Bootstrap + context (E1.1)

1. Run ${CLAUDE_PLUGIN_ROOT}/scripts/dure-bootstrap.sh → ensure .dure/ (idempotent).
2. Run ${CLAUDE_PLUGIN_ROOT}/scripts/dure-context.sh → detect git/non-git · greenfield/brownfield.
3. Resume check (§4.6): if .dure/interview-logs/ contains a log with status: in-progress, ask the user (in their language) whether to continue it instead of starting a new interview.
4. Compute the slug for a new interview: ${CLAUDE_PLUGIN_ROOT}/scripts/dure-slug.sh "$ARGUMENTS". Initialize the log from templates/interview-log.md.

1. Convergence loop (see playbook.md, spec §4)

Decompose the request into components, then repeat until the gate passes:

1. Evidence — Task the grounding-scout subagent to read code/issues within keyword scope → evidence + candidate answers. (trait ①; bound the scope for large repos, C9)
2. Score — self-assess each component × dimension (§4.2) honestly (0–5). You MUST NOT lower scores to finish faster.
3. Ask — target the weakest dimension with 1–4 questions via AskUserQuestion (structured choices + free-form), attaching cited candidate answers. Put any recommendation first.
4. Red-team — Task the redteam-critic subagent; it MUST inject ≥1 requirement-breaking cross-examination and return a pass|fail testable sign-off per component.
5. Research — when the user defers, Task the research-scout subagent for sourced candidates.
6. Record — append the round (evidence files · scores · red-team · decisions) to the log.

2. Convergence gate (§4.4) — deterministic, MUST NOT self-declare

Build this round's payload and feed it to the gate:

echo '{ "round": N,
  "components": [ {"name":"C1",
    "scores":{"problem":..,"scope":..,"acceptance":..,"constraints":..,"edge":..,"stakeholders":..},
    "testable_signoff":"pass|fail"} ],
  "new_ambiguity_last_round": <n>, "blocking_open_questions": <n> }' \
| "${CLAUDE_PLUGIN_ROOT}/scripts/dure-gate.py"

• Proceed to §3 ONLY when the exit code is 0 and the output is gate=PASS.
• On BLOCK, take the conditions in failed[] as the next round's target.
• Record the gate input JSON and its result verbatim in the log (auditability, AC2).

3. Crystallize (only after PASS)

1. Write .dure/specs/<slug>.md from templates/spec.md (decisions, scope, testable AC) — in English.
2. Set the interview log front matter to status: converged.
3. Write <slug> into .dure/active.
4. Summarize for the user (in their language) and point them to /dure:plan.

---

Implementation status (E1.2): the deterministic gate (dure-gate.py), slug helper, playbook, and templates are in place. Subagent orchestration (grounding-scout / redteam-critic / research-scout) MUST be driven via Task as above; if a subagent is unavailable, perform that step in the main context and state so in the log.