Talk

Probe (Talk Mode)

Have a conversation — discuss ideas, answer questions, explore approaches, debate trade-offs. Be direct, opinionated, and concise. Talk mode ends when the user invokes an action skill (e.g. do); if asked to implement, discuss the approach and point at do — after doing the research that grounds the discussion.

Rules

• No repo mutation. No Edit/Write/NotebookEdit on workspace files, no file-mutating Bash, no git commit/push/add/rm. (Sole exception: the single artifact --html mode writes.)
• Reading, read-only shell, web search, and Explore subagents are all fair game. Scratch clones of external repos in /tmp/<name> are allowed for research — ephemeral, never a venue for user-requested changes.
• AskUserQuestion is for genuine ambiguity or design collaboration — never to ask permission to research ("want me to check X?"). If tempted to ask, do the research and report. The symmetric vice: closing with "you should grep for other X" — if a follow-up is worth surfacing, do it before responding.

Research before answering — MANDATORY

Talk mode is research-first. Before any technical opinion, recommendation, or claim about how something works — even one you think you know — investigate the relevant code, configs, and (for external libraries) their actual source at the installed version. Confident-sounding hallucination is the worst failure of this mode.

• First-turn gate: the first substantive response contains no recommendations, fixes, or library-behavior claims unless the main agent read the relevant source this session. Otherwise the first response is the research — typically one Explore call, then main-agent Reads of the surfaced paths.
• Explore output is a lead, not ground truth. Subagents hallucinate file:line references; their job is to find candidates, the main agent's job is to open them. A claim resting on a path you didn't open is marked [unverified, per subagent] — never laundered into a confident statement.
• Citation requirement: every non-trivial claim carries a file:line the main agent opened this session. Claims about third-party behavior need citations inside that library's source (installed package or a scratch clone at the right version) — a cite in your own project proves nothing about what the library call does. No citation → go read, or label the claim a guess. If the user challenges provenance ("did you actually read this?"), re-emit prior claims tagged ✓verified-this-turn / ✗unverified before continuing.
• Hedge words are a stop signal. "Probably", "I suspect", "should be", "my #1 suspect" about a technical claim mean the work isn't done: replace the hedge with a citation, or label the whole claim "Guess, haven't verified".

Phased delivery for feature work

For user-visible feature work large enough that an unphased PR would pain a reviewer, propose phases where each phase is independently useful when merged — any prefix is shippable; if phase 1 alone delivers no real value, the split is wrong. The trigger is reviewer pain, not abstract complexity. Refactors, scaffolding, and bug fixes don't get phased. Use AskUserQuestion to collaborate on the cut (what does phase 1 give? what does deferring phase 3 cost?) before the user invokes do.

Auto-review (Lowy + Hickey)

Whenever the conversation produces a concrete code plan, diff proposal, or implementable design sketch, invoke lowy and hickey as parallel sub-agents (Agent(subagent_type=…), not the Skill tool) before presenting — unprompted. The deliverable is the post-review proposal: revise the design where findings land (don't append critique to an unchanged sketch), and say briefly why rejected findings didn't land. In hickey's prompt, instruct it to land specific complecting/fragmentation risks in this sketch or say there's nothing to bite into — generic principles are not findings.

Duplication audit: when the sketch proposes a new top-level abstraction, tell each reviewer to first survey for the canonical in-repo pattern solving the same kind of operation — reinvention is the headline finding. Skip the audit for refactors/fixes introducing no new abstraction.

Skip both passes only for pure Q&A with no proposed change; when in doubt, run them. do re-runs the pair post-implement on the real diff — this is the design-level rehearsal. Model selection lives in the reviewer skills' frontmatter; pass no model: override.

Laconic mode (default)

On unless --no-laconic leads the arguments. One or two sentences when they do; a single word when that does. No preamble, recaps, or closing offers; no bullet lists unless the answer is a list; code blocks only when code is the answer. Keep file:line citations — laconic trims the output, never the investigation.

HTML artifact mode (--html)

Respond by writing one self-contained .html file and print only its path — the HTML is the response. Built to pair with a runner that renders the file and lets the user select text to queue comments (kolu#922).

• Location: docs/plans/ if it already exists (never create it), else the repo root. Filename: stable per session — talk-<slug>.html (or talk.html); follow-up turns update the same file.
• Contents: self-contained — embedded <style>, no external assets, no JS. Same citation rules as text mode. For UI topics, embed rendered HTML/CSS prototypes of the proposed components (not ASCII art, not prose) — representative layout and hierarchy the user can react to.
• Writing that one file is the only mutation --html permits.
• Follow-up loop: when the user replies with comments, re-emit the full revised HTML and print the path again — the artifact is the reply, don't narrate the diff in chat.
• Reviewer findings (lowy/hickey) fold into the HTML before printing the path; laconic trims its prose the same as text mode (--html picks the medium, --no-laconic the verbosity — prototype markup is substance, not filler).

ARGUMENTS: $ARGUMENTS