edd

name	edd
description	Eval-Driven Development router — picks the mode and dispatches. Asks two questions (mode 1/2, aggression 1/2/3), writes `.edd/session.json`, hands off to one of the phase skills. Use this when the user wants an eval workflow but hasn't named a phase. Skill-specific work belongs in `edd:scope-agent`, `edd:scope-evals`, `edd:run`, `edd:experiment`, or `edd:expand`.

edd — router

Top-level entry. Picks the phase, persists session state, hands off.

When to invoke

After changes to prompts, tool surface, skills, model routing, or memory injection. Skip for pure UI, refactors that don't touch prompts/tools, or bug fixes with a reproducing unit test.

Ask the user — two questions only

1. Mode — which mode do you want to run?

   • 1 — Quick trace analysis. Emit + tag traces; Claude reads them inline. Judges optional via edd score --evaluators "..." if the user wants a score table without a dataset → edd:run
   • 2 — Dataset + experiment. Emit + tag, build a durable dataset, optional edd:expand for coverage, then experiment fires the judges. Inner loop is emit-only — the experiment plane is the sole scoring plane (trace-plane scores don't sync — see references/opik-endpoints.md#score-storage--two-planes-no-auto-sync). A 3–5 trace smoke-judge-check runs before the experiment to verify judges land at all → edd:run → edd:experiment

2. Aggression level — how hard should the scenarios push the agent?

   • 1 — Harness validation: normal user flows, exact trigger phrases, happy paths
   • 2 — Mixed: level 1 + edge cases (empty results, adjacent intents, partial triggers)
   • 3 — Adversarial: mostly edge cases designed to surface harness failures

Persist session state

Write .edd/session.json with the keys listed in skills/CLAUDE.md. Mint topic from the user hypothesis (≤5-word kebab slug); leave dataset_name / experiment_name null — sub-skills fill those.

Dispatch table

Mode	Scope needed first?	Then run	Then
1	Light (regressions.txt). Judges only if scoring inline.	edd:run (emit + inline read; optional edd score)	done
2	Full (promises + evaluators)	edd:run emits + smoke-judge-checks 3–5 traces (harness health, not score iteration)	edd:experiment Phase D (build) → edd:expand if coverage thin → edd:experiment Phase E (judge — sole judge plane) → inspect. Expansion is a hard gate before Phase E — never after.

Always check first:

• regressions.txt exists at repo root → scope-agent already done; skip unless agent source changed
• Opik project has the evaluators you'll need (via shared.opik_client.OpikClient().get_evaluators()) → scope-evals already done; skip unless promises changed

If either is stale, run edd:scope-agent and/or edd:scope-evals before edd:run.

Discovery rule — derive, don't ask

Never ask the user for evaluator names, dataset names, or experiment names. Derive them per naming conventions; pull dimensions from edd:scope-agent + edd:scope-evals outputs.

Next

• Sub-skill index + Pipeline DAG — sitemap with pipeline overview, anti-patterns, stop-rules, session schema.