| name | agent-eval |
| description | Use only when the user explicitly invokes $agent-eval, or when an active workflow routes here, to convert locked Spec behavior into acceptance evidence, automated/manual checks, failure conditions, fixtures, risk coverage, and implementation evidence requirements before Plan. |
Agent Eval
Define how to prove the Spec is correct. Own Eval only: do not write the
implementation plan or execute tests.
Iron Law
NO PLAN WITHOUT A WAY TO PROVE CORRECTNESS.
Every important behavior needs an automated check, manual evidence path, or
explicit untestable/deferred entry with risk and owner.
Contract
Apply mode-gate.md and
coverage-trace.md. Produce a
complete, concrete artifact at the selected weight; omit fields that do not
change planning or verification.
Accept only implementation-eligible confirmed Spec behavior. For every
important behavior, include each applicable obligation:
Automated command, named test, or reproducible manual steps
Pass/fail oracle a stranger can apply
Failure condition that catches a plausible wrong implementation
Behavior -> check/evidence trace
Fixtures/data/setup when the check depends on them
Regression or risk checks when change could break nearby behavior
Untestable/deferred reason, risk, and owner
Evidence the Plan and implementers must attach
Block Plan handoff when the Eval contains:
"works as expected" or "manual QA" without steps and an oracle
one vague check for unrelated behaviors
no failure condition beyond the happy path
no trace from a check to Spec behavior
checks for draft/open candidates as if they were implementation scope
Before handoff, ask whether a plausible wrong implementation could still pass.
Strengthen the oracle or failure condition until the answer is no.
Process
- Read the locked Spec and use
project-context to discover relevant test
commands, fixtures, and evidence surfaces.
- Map confirmed behavior to automated, manual, or explicitly untestable
coverage, including failure conditions and required evidence.
- Route unclear behavior to agent-spec. Route any
draft/open candidate back to its upstream owner instead of creating a check
that promotes it into scope.
- Use agent-review only when one testing frame may
miss material risk.
- Hand the ready Eval to agent-plan.
Plan Readiness Gate
Move to Plan only when:
- every important behavior has an acceptance check or manual evidence path
- no draft/open candidate appears as Eval behavior or implementation evidence
- every check/evidence path traces to the behavior it proves
- failure conditions are explicit enough to catch wrong implementations
- verification commands or evidence surfaces are known, or flagged open
- accepted blocker/major review findings are fixed and rechecked, or rejected
with evidence; a true decision/dependency blocker keeps Eval paused
If persistence is needed, use the active workflow location; default:
docs/dev-flow/evals/<YYYY-MM-DD>-<slug>-eval.md.