evo-end-to-end

name: evo-end-to-end description: Run a Codex planning-to-Evo workflow for evo-hq/evo v0.5.2+. Use when the user wants to turn a vague performance, architecture, refactor, flaky-test, slow-build, fine-tuning, post-training, or code-quality problem into an Evo-ready experiment brief, using companion skills such as $grill-me, $grill-with-docs, and $improve-codebase-architecture when needed.

Evo End To End

Turn a fuzzy improvement request into an Evo-ready experiment, then run Evo only after approval.

Target Evo release line: evo-hq/evo v0.5.2 or newer.

For companion skill sources, optimize presets, and release-specific notes, see REFERENCE.md.

Companion Skills

Names written as $skill-name are agent skills. Load and follow a named skill when its trigger applies.

• Use $grill-me when the goal, constraints, non-goals, success metric, or forbidden changes are unclear after repo inspection.
• Use $grill-with-docs only for unclear terminology, ownership, CONTEXT.md, or ADR decisions.
• Use $improve-codebase-architecture when architecture or testability must be decomposed before choosing an Evo metric.
• Use $evo finetuning before writing or changing training code, reward design, model-weight update recipes, or post-training workflows.
• Use installed Evo plugin skills whose evo_version matches evo --version.

If a referenced $xxx skill is not installed, check the host's skill list and then the user's skill lockfile for sourceUrl and skillPath. Cite the source before continuing with the best available fallback.

Flow

1. Inspect the repo for manifests, tests, docs, benchmarks, and likely editable scope before asking clarifying questions.
2. Verify Evo before running plugin skills:
   • evo --version must report evo-hq-cli, not the unrelated SLAM package.
   • The CLI version must match the loaded Evo plugin skill version.
   • If CLI and plugin drift, tell the user the lockstep updater command (evo update <host> --version <version>). Do not install or upgrade unless the user asks.
3. Draft an Evo brief with: goal, metric, baseline command/data, pass gate, editable scope, read-only context, forbidden changes, host, backend, runtime/env needs, per-experiment timeout, budget, stall rule, task skills, and merge rule.
4. Stop for approval before Evo edits production behavior, APIs, persistence, auth/security, tests, packaging, dependencies, deployment, user-visible behavior, dependency manifests, or remote/cloud infrastructure.
5. Run $evo discover with the approved brief. Optimize only after discovery records a baseline and the benchmark reviewer gate has passed.
6. Configure backend/runtime explicitly for pool, remote, or non-default runtimes. Keep benchmark variables in evo env, not in committed scripts or docs.
7. Set or confirm --per-exp-timeout / evo config set per-exp-timeout <seconds> for long benchmarks or training runs.
8. Run evo run <exp_id> --check when wiring risk is material and non-mutating validation is available.
9. Resolve autonomous and subagents-only the same way $evo optimize does, then arm that state with evo autonomous on|off and evo subagents-only on|off.
10. Size subagents=<n> from benchmark/backend resources first. Use width 1 for exclusive GPUs, fixed ports, shared databases, singleton services, or timing-sensitive harnesses unless the harness isolates them. This is Evo CLI worker sizing, not Codex spawn_agent fork behavior. If this workflow also uses Codex subagents for planning, review, or triage, launch role-specific workers without full history. In the current spawn_agent tool, omit fork_context or set fork_context: false; on tool surfaces that use fork_turns, set fork_turns: "none". Put the role and needed context in the message; do not override agent_type, model, or reasoning_effort on a full-history fork.
11. Run $evo optimize subagents=<n> budget=<n> stall=<n> within the approved scope.
12. Use evo direct "<text>" only to steer an already-running Evo session. If an agent receives an [EVO DIRECTIVE id=...] banner, it must run evo ack <event_id> before proceeding.
13. Manually review Evo output before merging behavior, API, persistence, security, packaging, deployment, or user-visible changes.

Brief Template

Goal:
Metric:
Baseline:
Gate:
Editable scope:
Read-only context:
Forbidden changes:
Host: codex
Backend: worktree | pool | remote:<provider>
Runtime/env:
Per-exp timeout:
Budget:
Stall rule:
Autonomous: on | off
Subagents-only: on | off
Task skills:
Optimize preset:
Merge rule: