Discuss

---
name: discuss
description: Start a work session with the owner, or continue one. Use whenever the owner opens a conversation about work on the project.
argument-hint: "[target: a phase folder, to re-enter it]"
---
<!-- referent: EveryInc/compound-engineering-plugin skills/ce-brainstorm/SKILL.md @3449230705a5 and mattpocock/skills skills/productivity/grilling/SKILL.md @4ae2317ddb22; differs: each step loads its reference where it needs it and the skill ends in a handoff to `fwx:session`; the owner decides without recommendations; agents are launched in the background and their files are read by other agents; the tree's first lane is the order of the steps -->

Rules. Only the owner and you speak. Launch every agent in the background, `model` exactly as the start JSON gives it, with the absolute paths of the phase's `context.md` and `decisions.tsv`. A launch returns at once: read its result only after its completion notice. You write neither `decisions.tsv` nor `research/`. A dead agent is not "no findings": say so, launch it once more.

Context. Read no long file or investigation yourself: a `researcher` returns summary and path. Its `## Unproven` lines are not findings: say them as not checked. Run `fwx budget --json` now and before each playback, heavy read or generation. Context high or `week_tight` true: delegate more, ask for shorter reports, offer to split a topic too big for one session. `unknown` is not zero; quota never stops you. After a compaction, redo the playback from `context.md`, `decisions.tsv` and the summaries.

1. **Start.** Run `fwx discuss start --json`, with `--phase $ARGUMENTS` when a `target` phase path was given. Exit 1: say in one line the framework is off, go on as plain Claude. `bound.stage` after `discuss`: `fwx:session <bound.phase_dir>`, no conversation. `bound.reentry`: follow `${CLAUDE_PLUGIN_ROOT}/skills/discuss/references/reentry.md`. `bound.stage` null with a `phase_dir` is a discuss not yet frozen: continue it, no second phase. An `open_phases` entry with `stage` null: ask in round one whether the message continues it. Exit 2: give its one-line reason and stop.

2. **Gather context** before you classify the request. Read only: the code the message touches and `.fwx/` (`PROJECT.md`, `durable/`, `adr.tsv`); for UI work, also the identity and design documents and the component skill (the `.claude/skills/` one whose description says so). What the repo, the app or Linear can answer is looked up, never asked. No phase, no agent yet.

3. **What does the owner ask?** Unsure: ask in round one.
   - Work on the project: the subproject from the message and `subprojects` (unsure: ask, research waits). Then `fwx state new-phase <slug> [--subproject S]` and launch two or three `researcher` (`models.researcher.small`): code map, `.fwx/` state, for UI the design documents and component skill. Do not wait. First check: what you understood, the subproject and why, what waits in Linear (`waiting.groups` by age, each issue with its URL; "Linear: not checked (reason)" if NOT RUN).
   - Improve the system, or a new model: give `/fwx:improve`.
   - Trivial: only the owner calls a change trivial, by no threshold of yours; when it could be, offer "plain Claude, no workflow" in round one. Picked: one line, no phase; if it passes the closed list (step 9), give `/fwx:factory-issue`.
   - Continue: with no phase bound, `fwx:session`; for an issue the factory returned (a `needs-owner` or `factory-failed` one he picks), `fwx:session retake <KEY>`; a `factory-review` one, its URL only.

4. **Check** at the top of every reply to an owner message, a question too: up to three lines, about 50 words: what you understood; "Noted: ..." per decision; a bare "yes" read aloud ("I read your yes as b: ..."). His question gets an answer and no row. Under it, if needed, a numbered round per `${CLAUDE_PLUGIN_ROOT}/skills/discuss/references/rounds.md`, with no recommendation, only for his taste or direction or what is unclear.

5. **Record.** With decisions in the turn, launch `decisions` (`models.decisions.small`) with the phase folder, the decision and the exchange word for word between `<exchange>` tags. A lettered or verdict answer is `chosen` (a verdict's JSON is its `source`); a plain yes to what you noted, `approved`; what you decide unasked, `agent-assumption`, as is whether the work is UI by its nature. A row returned `unexamined`: raise it once next round, with the alternative and what each costs.

6. **Research.** A fact you can look up goes to a `researcher` without holding the round; a bigger investigation, after his yes, to one on `models.researcher.normal`.

7. **UI with demos.** For a bounded UI decision launch `demo` (`models.demo.normal`); when it returns, end your message with its `xdg-open` line and run `python3 <verdict_tool> verdict wait <dir>` in the background. His pick goes into `Approved UI` of `context.md`. No UI choice is closed in chat. Without `verdict_tool`, give each `xdg-open`; he answers by letter.

8. **Playback and sign-off.** When all is said, wait for every `decisions` to return, then give the playback per `${CLAUDE_PLUGIN_ROOT}/skills/discuss/references/playback.md`. He may correct, want more research, split into phases (`references/reentry.md`) or sign off; only that ends it.

9. **Demo in the app, and the closed list.** With UI and `verdict_command`, `demo` (`models.demo.normal`) runs it in the app; a UI change goes back to step 7. Then ask whether it is a bug or on the closed list of mechanical changes (`factory-issue`'s rule). Yes: give him `/fwx:factory-issue`; if he declines, or no: step 10.

10. **Freeze.** Rewrite `context.md` with the sections of `${CLAUDE_PLUGIN_ROOT}/templates/phase/context.md`. `Surfaces` lists `ui` when the work is UI by its nature (what a person sees or uses), not its file extension: a typo in a `.tsx` is not, a visible change in a `.ts` is. With `decisions.tsv` complete, run `fwx:session <phase>`, stage still `discuss`. On "cancel": `fwx state cancel <phase>`, reply `{status: done, reason: "cancelled", phase_dir}`.