/forge:eval — Evaluation

When to Use

• User says /forge:eval
• User wants to verify the implementation, check test coverage, or validate against the SRS
• Working in a Forge project at Stage 6 or 7

Pre-flight Check

Entry gate (REQ-GATE-ENTRY-001) — before adopting the persona, verify the prior stage's artifact exists:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/state-manager.py preflight --stage 7

If it exits non-zero, STOP: present its message verbatim and do not proceed — the prior stage must be completed first (or use /forge:force-advance to skip intentionally).

1. Read pipeline/state.md — confirm Forge project.
2. Confirm pipeline/01-srs/srs.md and pipeline/04-spec/technical-spec.md exist.
3. Run python3 ${CLAUDE_PLUGIN_ROOT}/scripts/check-gate.py --stage 6 to see Stage 6 status. If critical criteria fail, warn the user before proceeding.
4. Run python3 ${CLAUDE_PLUGIN_ROOT}/scripts/load-profile.py --cwd . --stage 7 to load project-type evaluation criteria. Every profile criterion in additional_criteria (e.g., ML: model accuracy, GPU memory, drift detection; API: contract tests, load test, auth bypass; fullstack: Lighthouse, WCAG AA, responsive breakpoints; CLI: --help text, exit codes; library: API surface, semver, doc coverage) must appear as a row in the eval matrix with a pass/fail verdict and evidence.

Steps

1. Read agents/evaluator.md to load the Evaluator persona.
2. Adopt that persona — you are now the Evaluator.
3. Read pipeline/01-srs/srs.md to enumerate all REQ-IDs.
4. Follow the Evaluator workflow: run tests, check each REQ-ID for evidence, assess NFRs. Then evaluate each additional_criteria row from the profile loaded in pre-flight — for an ML project this includes the drift detection criterion (G7-ML-005).
5. Write pipeline/07-evaluation/eval-report.md per the Output Contract. The eval matrix must explicitly list every profile criterion alongside REQ-IDs and NFRs.
6. Run python3 ${CLAUDE_PLUGIN_ROOT}/scripts/state-manager.py advance --to 7 to mark Stage 7 active.

Verification

After running, confirm:

• pipeline/07-evaluation/eval-report.md exists with per-REQ-ID verdicts
• Test results are documented (not just "tests pass")
• pipeline/state.md shows current_stage: 7

Next Step

Derive the hint from the canonical stage table — never hardcode it (REQ-NEXTHINT-001, single source of truth). Run the helper and present its output to the user verbatim:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/state-manager.py next-hint --stage 7

(If the eval found gaps, fix them and re-run this stage before advancing; the hint above is the handoff once all requirements pass.)