Work Menu

$ARGUMENTS

Purpose

One read-only pass that answers "what should Truffle improve next, from evidence we already have?" It scans the scattered local surfaces, applies an optional focus lens, and proposes a short routed work menu — 5–7 candidates Brian can graduate, park, or hand off. The full frame (problem, surfaces, non-goals) lives in _temp/2026-06-24-truffle-work-selection-frame.md; defer to it.

It runs as a dynamic-workflow harness. The harness is not about scale — the corpus fits one window. It buys two things one window does badly, which are the frame's two named risks: blind reads (Phase 1 — no surface anchors the others) and an egoless graduation gate (Phase 3 — the agent that proposed a candidate is not the one that waves it through). The gate is the spine; over-promotion is the enemy.

What this is not

• Not a roadmap or auto-prioritizer. It proposes a menu; Brian orders the big rocks. Notion stays out of scope unless run in roadmap mode.
• Not a mutation. It never writes to a source file (BACKLOG, learning, packets). Read-only. Output is chat-first; persistence is opt-in (below).
• Not graduation. It surfaces and routes; Brian decides. A surfaced card is a recommendation, not a committed work item.
• Not the learning consolidate pass (/learning-review) or a single-change review (review-change). It can route to those; it doesn't do them.

Required context — defer to these, don't restate them

• work-menu.workflow.js — the harness. The reader prompts, the closed sets (buckets, authority grains, routes), and the gate rubric live there, as the single source of truth. Treat it as a template, not a verbatim script: adapt models, surface paths, or phases to the run, but keep the four-phase shape and the gate's default-to-watch posture.
• learning-review/SKILL.md — the gate borrows its posture (strip specifics, "what does it replace?", ≥2 sightings or a severe risk-miss, keep the "left" count honest and large). Read How to judge well before trusting any pass.
• the frame (linked above) — the seven subject buckets and the non-goals are its, not this skill's.

Task

1. Resolve the lens. Read $ARGUMENTS as the focus lens. roadmap (alone or appended) flips on roadmap mode. If the argument is blank, ask Brian for a lens or confirm a balanced pass — don't silently default (a shaped pass is the point).
2. Run the harness. Invoke the Workflow tool with scriptPath pointing at work-menu.workflow.js, passing args: { lens, engineRoot: <repo abs path>, roadmap }. It runs Read → Synthesize → Gate → Render and returns the menu object (brian_digest_md, cards, left, overflow_note, second_opinion, counts).
3. Trust the gate, don't second-guess it upward. If a pass surfaces very little, that is a clean result, not a failure — report the honest "left" count as the work, per learning-review. Never re-promote a watch item just to lengthen the menu.
4. Present, then offer to act. Print brian_digest_md to chat as the menu. End with the decision surface (below). Act on a card only when Brian picks one — then hand off to its route's verb (/agent-build-propose, /learning-review <target>, /deepen-offerings, etc.).
5. Persist only on request. Default is chat-first. If Brian says --persist (or asks for cross-pass memory), write brian_digest_md + the cards to a dated, disposable digest — propose the home on first use (_review/work-selection/NNN-YYYY-MM-DD.md, mirroring learning-review's passes/). A persisted digest is a regenerable lens, never authoritative.

Quick mode (skip the harness)

For a narrow, single-lens pass over one or two surfaces, the harness's blindness buys little — run it inline as a single agent: read the surface(s), apply the lens, gate each candidate against the rubric in the workflow's Phase-3 prompt (same default-to-watch), and render the same two views. Reach for the full harness whenever the pass spans many surfaces or the over-promotion risk is real.

Final Response

A short decision surface for Brian:

• lens + counts (raw → plausible → surfaced → on-menu, and how many left as watch).
• the menu — brian_digest_md: candidates grouped by subsystem bucket (the frame's seven), each with a sharp title, one-line problem, why-now, the gate's readiness verdict (ready vs watch + why), and recommended route. (The cards array is the provenance audit trail behind it.)
• left as watch — the honest count, plus the 1–2 calls (second_opinion) you'd most want a second look at. Any overflow_note.
• Read-only — no source file was touched; nothing was graduated. Pick a card and I'll route it.