---
name: wf-eval
description: 'WellForge command — Score a feature''s implementation against the eval rubric (LM-judge — the non-deterministic verification half). Invoke explicitly: $wf-eval [feature] — NNN-slug / slug / NNN; omit to infer the in-progress/most-recent feature'
---
> WellForge command, generated for Codex from `wellforge-plugin/commands/eval.md` — do not edit; regenerate with `adapters/codex/generate.py`.
> Invoked as `$wf-eval [feature] — NNN-slug / slug / NNN; omit to infer the in-progress/most-recent feature`. **ARGUMENTS** below means the text the user wrote after `$wf-eval` in the invoking message (empty if nothing).


Run the LM-judge evaluation for a feature: score the implementation against the central
rubric and produce a verdict. This is the eval half of verification — tests and CI gates
cover the deterministic part (`$wf-implement`'s QE step); this covers spec
fidelity, test quality, code conventions, and trajectory. Conventions: the **spec-driven**
skill (load it).

Feature: ARGUMENTS

## Step 1 — Resolve the feature

Same resolution as `$wf-implement`: leading token matched against `specs/` by
number / slug / full name; if omitted, infer the `in-progress` spec, else the most recent
one whose tasks are all checked. Ambiguous → list and ask. State which feature you
resolved. Read spec.md, plan.md, tasks.md.

## Step 2 — Gate check

- `tasks.md` must exist and the implementation be substantially done (all or nearly all
  tasks checked). Evaluating an unbuilt feature is meaningless — if most tasks are open,
  STOP and point at `$wf-implement`.
- Ensure fresh test/coverage output exists; if not, run the project's test+coverage task
  so the evaluator has real evidence (not stale or assumed results).

## Step 3 — Evaluate

Spawn the `wf-evaluator` agent with: the spec dir path, the central rubric, and any
`specs/NNN-slug/eval.md` override. The agent resolves the rubric itself, in the order its
prompt defines — project `gates/configs/eval-rubric.yml`, then CI's `.wellforge-gates/`
checkout, then the tool-bundled copy (`<plugin root>/config/eval-rubric.yml` for Claude
Code). **A scaffolded project has no `gates/` directory** — gates are referenced by CI, never
copied — so there the bundled copy is the one that resolves. If the agent reports it found no
rubric, that is a plugin-install problem, not a feature problem: stop and say so. It scores
each rubric dimension 1–5 with cited evidence, computes the weighted total, and writes
`specs/NNN-slug/eval-report.md` with a PASS/FAIL verdict (PASS = total ≥ pass_score AND
every dimension ≥ floor).

## Step 4 — Act on the verdict

- **PASS** → report the score table. The eval PASS is the gate into `done`, but this command
  doesn't flip the status — point the user at **`$wf-done <feature>`**, which verifies
  the full done gate (tasks + QE + this PASS) and records the close.
- **FAIL** → relay the failing/weak dimensions with the evaluator's evidence. Route the
  remediation: name the specific ACs/tasks to revisit and suggest
  `$wf-implement <feature> <those tasks>`. Do NOT fix anything here — the evaluator
  judges, fixes go through the dev agents. After remediation, re-run `$wf-eval`.
- Bounded: the evaluator is a judge, not a loop. One scored verdict per run; you don't
  silently re-evaluate until it passes.

## Step 5 — Record the run (observability)

Write `.forge/runs/<run_id>.json` per the **observability** skill: the `wf-evaluator` agent,
its outcome + `score`, and `verdicts.eval`. Short trace, same audit trail as implement/
orchestrate.

## Hard rules

- Read-only on code/specs, **and on status** — this command produces exactly one artifact,
  `eval-report.md` (via the evaluator). It never sets `status: done`, with or without the
  user's confirmation: an eval PASS is one input to the done gate, and
  `$wf-done` is the only place that gate is checked and the transition recorded
  (Step 4 already points there).
- Never lower a rubric floor or override a FAIL. The rubric is central
  (`gates/configs/eval-rubric.yml`); changing it is a PR to `gates/`, which also refreshes
  the plugin's byte-identical mirror (`wellforge-plugin/config/eval-rubric.yml`) — CI's
  `rubric-sync` job fails if the two drift.
- An unmet AC fails the eval regardless of the other dimensions (ac_satisfaction floor).
- For a feature with no implementation yet, this is the wrong command — implement first.