git:20260729.06c6528 to git:20260731.dc16fc5

- description: Internal CCF requirements-interview engine. Invoked ONLY by the CCF commands plan, fix and init (via the Skill tool, passing a mode argument) to interrogate the user one question at a time and gather the context those commands need. Not a standalone command — do not trigger it from ordinary conversation.

- Run a focused interview using the discipline below, then hand a concise **summary of the answers** back to the calling command so it can continue.

- - **One question at a time.** Never batch. Ask, wait for the answer, and let that answer shape the next question.

- - **Explore before you ask.** Before each question, try to answer it yourself from the codebase (Read / Glob / Grep). Only ask what the code cannot tell you; for what it can, confirm instead of asking blind ("I see `X` wired in `foo.ts:42` — still correct?").

- - **Recommend with every question.** Offer your recommended answer plus a one-line rationale so the user can simply confirm. If the user defers, proceed with your recommendation.

- - **Stop when you have enough** to act. Do not interrogate past the point of diminishing returns.

- - **Summarize at the end.** Give a short, structured recap of the decisions, ready for the command to fold into its next step.

- Probe, in order, only the points that are still unclear after exploring the code:

- 6. **Implementer model (default for the WHOLE plan)** — ask ONCE which model should implement this plan's tasks by default (an alias such as `sonnet`/`opus`/`haiku`, never a dated model ID). This is asked here, at plan-mode 2 (before any task exists), so it must stay a single plan-wide default, not a per-task question — the calling command's step 5 (once the actual task list is known) is where a per-task OVERRIDE can be offered for a task that is unusually hard/simple. Recommend `sonnet` (a task that is already clear and well-scoped); `opus` for a hard task or one with many constraints; `haiku` for a simple, mechanical text edit — as guidance for that later per-task override, not as a batch of questions asked now. If `AskUserQuestion` is unavailable (a non-interactive session), skip asking, use the `sonnet` default, and say explicitly that the default was used instead of a real answer — never proceed silently as if the user had chosen.

- Reconstruct the bug, probing only what you cannot already determine from the code/logs:

- Walk this decision tree to elicit project decisions; recommend an answer for each, and prefer confirming what the repo + `git log` / `git branch` already reveal over asking blind:

- - (d) Expected user scale? → based on scale, propose hosting (e.g. Supabase or Railway) and tell the user to install the corresponding MCP (`/plugin install ...`).

- - (e) Design patterns for FE & BE? **If there is a frontend, recommend React + Tailwind CSS + shadcn/ui by default** (stable, mainstream, well-supported, and has an MCP that lets Claude browse/install components) — one-line rationale, user free to choose otherwise. When chosen, tell the user that `/ccf:init` will add the shadcn MCP to THIS project's `.mcp.json` (`{"command":"npx","args":["shadcn@latest","mcp"]}`); after `shadcn init` (which creates `components.json`) they restart Claude Code and run `/mcp` to confirm it shows `Connected`.

- - (e2) **Design source (only if there is a frontend):** ask whether the user has a Claude Design handoff bundle (a URL like `https://api.anthropic.com/v1/design/h/...`). If yes → record the URL so the FE spec can follow it as the visual source of truth. If no → suggest creating one in Claude Design for a more polished UI (it exports HTML/React + a design spec to hand to Claude Code); default to none if they decline. Do NOT try to fetch the URL (it is authenticated).

- - (i) Testing strategy? Probe one sub-question at a time (recommend a default + one-line rationale each): **test framework / run command / test location / coverage target** (as today), THEN **"Adopt the test-design discipline — a contract-level EP/BVA/decision-table matrix on each public signature (recommended: yes; catches edge/boundary bugs early, kept at the contract level so tests stay robust)?"**, THEN (only if yes) **"Enforcement — prompt-only (the spec asks for it) or a Stop-hook gate (recommended: prompt-only to start; a stop-hook can `exit 2`-block a session that edited code with no matrix test, stronger but noisier)?"**. ASK ONLY: report the answers in the summary; `/ccf:init` does the template fill — grill-me has no write tool.

- - (j) Tech stack — must be the most stable, best-supported, least-buggy (mainstream); for each library pick the most popular/well-maintained option.

- - (k) Monorepo rule: work in the root folder; if fullstack create `be/` + `fe/`; git init at the root, NOT in sub-folders; the root holds CLAUDE.md, `.claude/`, docker, CI/CD.

- - (l) Git conventions: first check whether the repo already has commits (read-only `git log` / `git branch -a`). If a pattern exists, infer the commit/branch convention from it (don't invent). If history is empty or too thin, ask the user (or default to conventional commits: `feat:`/`fix:`/`refactor:`…). This fills `git-workflow.md`'s `{{COMMIT_CONVENTION}}` / `{{BRANCH_NAMING}}` / `{{PR_RULES}}`.