Edd
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`.From its SKILL.md
npx -y skills add aboyalejandro/eval-driven-development --skill eddAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 6 stars6 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
3.4 KB, as published. Nobody here has run it
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
-
Mode — which mode do you want to run?
1— Quick trace analysis. Emit + tag traces; Claude reads them inline. Judges optional viaedd score --evaluators "..."if the user wants a score table without a dataset →edd:run2— Dataset + experiment. Emit + tag, build a durable dataset, optionaledd:expandfor 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 — seereferences/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
-
Aggression level — how hard should the scenarios push the agent?
1— Harness validation: normal user flows, exact trigger phrases, happy paths2— 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.txtexists 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.