assess — the expert panel over the whole app

Judge the application the way a seasoned leadership review would: several experts, one honest report. Read CONVENTIONS.md — especially the workspace and changelog (§1, §13.1 — the app's memory tells you why and when; only the code tells you what exists, §5.2), the role dial (§6), the resume sweep (§5), git isolation (§11), gates (§7), multi-agent rules (§8), platform adapters (§9), grounding (§14), freshness (§17), closing output (§18), and data-driven decisions (§19), and the ledger (§2), judgment (§4.1), workspace integrity (§20).

Journal every run (§2.1). The first write of the run, once the task folder is resolved (§20.1), is a START entry in the task's log.md; then set this phase in_progress in state.json. Append a DECISION entry before acting on any choice that shapes the work, and a STOP entry before every reply that ends the turn. Write only the files the workspace tree names (§1): nothing lost, nothing stray. Before asking a question or deciding, check the user's judgment (§4.1) for a confirmed rule that covers it, and journal who decided.

Boundary with discover: assess judges the features that exist; when its conclusion is "we need net-new X", hand that to discover (or straight to define if the user already chose). Don't run both from one ambiguous ask — existing-feature health leads here. If a run is already active, read state.json.mode for the multi/single answer instead of re-asking.

Step 1 — Two questions, once

1. "Run the expert panel as multiple parallel agents, or one agent playing all roles in sequence?" [multi/single] (single-agent runtimes: always inline, §9)
2. "Do you want a report only, or report + a throwaway prototype of an improvement you pick from it, so you can feel it before deciding?" [report/report+prototype]

Step 2 — Build the evidence base

Scan the repo, engineering/ (profile, decisions, feature changelog), docs. Build the feature map from the code: every feature named with where it lives (file:line, route, table), and every "missing", "dead" or "unused" claim with the searches that found nothing, run two ways (§5.2). The changelog explains a feature's history; it never proves the feature is there, wired up, or working. For usage/business numbers, measure before asking (§19): derive the questions, write read-only queries from the app's own schema, run them via a connected data tool or hand them to the user to run — and ask only for what can't be measured (§14, never invent). Map the feature inventory and how features connect (data flows, shared models, entry points).

Step 3 — The expert panel (five lenses)

Each lens reviews the same evidence and reports findings with the evidence cited:

• Domain expert (of the app's own field — detected from the product): does each feature serve the domain's real workflows? What would a practitioner miss or distrust? A claim about how the domain works cites a source checked in this run, or becomes a question for the user; never a workflow described from memory as fact (§14).
• Software engineer: architecture health, feature coupling, test coverage of core flows (from a coverage run in this run, §5.1, or said to be unmeasured), fragile or dead areas (structural signals only — line-level review belongs to inspect).
• Product (PM): completeness of each feature's journey (empty/error/edge states, onboarding, discoverability), features that exist but don't compose into outcomes.
• Market analyst: how the feature set compares to the market as of today's date — web-researched and cited (§17); table-stakes gaps; where competitors are stronger. If the market is unknown, ask the user or research — never fabricate competitors (§14).
• Improvement lead: synthesizes the others into ranked, feasible improvements (value vs. effort), separating quick wins from structural bets. Value is tied to measured usage (§19) or a goal the user stated; effort to the files and callers counted in this run. Either one that cannot be grounded is written unestimated, not guessed.

In multi-agent mode dispatch the lenses in parallel and merge; the orchestrator owns all writes (§8). In single mode, run the lenses sequentially with fresh eyes per lens.

Step 4 — The report (only what's good and what needs work)

Write assessment.md in a task folder (bootstrap per §1):

# App assessment — <app> · <date>
## Feature map            — features and how they link (a simple diagram helps)
## Strong                 — what's genuinely good, and why (evidence)
## Needs improvement      — ranked: what's weak/thin, why it matters, the concrete fix,
                            which expert flagged it, value vs. effort
## Market position        — cited comparison; table-stakes gaps
## Recommended next moves — the 1–3 to do first

No filler: every line is evidence-backed, or labelled an unverified suggestion with how to confirm it (§14). "Expert judgment" is not a label that makes a claim true: the panel is the model, and a verdict under Strong or Needs improvement always cites its evidence. If a prototype was requested, present the report first; the user picks which improvement to spike, and construct builds it on its own branch or worktree (§11), clearly disposable (§6.2). assess itself never edits the source.

Step 5 — Close the loop

Present the report. If the user wants changes: hand the chosen improvements to define → blueprint and stop at the plan-approval gate as always (§7). If no changes are needed, record the assessment date and outcome in decisions.md and stop.

Composition

• Consumes: the repo, engineering/ memory + changelog, web research (cited), the user's usage/business context.
• Produces: assessment.md and an ADR note. A prototype, if asked for, is built by construct on its own branch (§11), for the improvement the user picks.
• Receives from: the user directly, or engineer when the intent is whole-app analysis.
• Hands off to: define/blueprint for approved improvements. Distinct from inspect (a diff, code-level) and discover (net-new features/tools).

Record the phase (§2)

assessment.md on disk and non-empty (§20.2) comes first; then record an assess entry in the task's state.json (§2's optional phases) — done + validated + the artifact name. Directly invoked, no orchestrator will do it for you.

Guard — every verdict needs its evidence

These thoughts mean stop — you are about to grade a feature on impression:

Self-review (author's notes)

• Mis-routed? engineer routes here for whole-app analysis; wrong when the ask is really one diff (inspect), net-new ideation (discover), or a UI-only audit (design). Pick this over inspect when the unit of review is the product, not a change.
• Single-agent safe? Yes — the panel runs as sequential lenses inline.
• Leaks specifics? No — the domain expert is instantiated from the detected product.
• Grounded? Market claims cited per §17; unknown numbers asked, never invented.