evo plugin
Structured experiment-driven code optimization using tree search and parallel subagents
How to install the evo plugin
Run both commands inside Claude Code. The first adds the marketplace once; the second installs the plugin.
/plugin marketplace add evo-hq/evo
/plugin install evo@evo-hq-evoWhat's inside evo
18 parts read from the plugin's own files.
Commands 2
Run as /evo:name- /evo:discover
Discover what to optimize and initialize an evo workspace.
- /evo:optimize
Run the evo autoresearch optimization loop.
Agents 3
Call as @agent-evo:name- @agent-evo:benchmark-reviewer
Reviews an evo benchmark in two modes. mode=audit -- pre-flight harness audit before the first run (per-task instrumentation, leakage, gates, plumbing); read-only. mode=review-experiment -- post-commit per-task failure analysis for a specific experiment; reads per-task traces and the eval-runner log
- @agent-evo:ideator
Generates ranked experiment proposals for the evo orchestrator. Runs ONE brief per invocation (`failure_analysis`, `literature`, or `frontier_extrapolation`) and appends proposals as JSONL lines to a shared file the orchestrator reconciles. Use `literature` for web/arXiv/HF/GitHub research (the only
- @agent-evo:verifier
Read-only audit of one evo experiment for design-time cheating (pre-phase) or result-time validity (post-phase). Catches test-set leakage in training data, subsetted eval commands, missing gates for new artifacts, generic hypotheses, cache short-circuits, fake artifacts, and score-reproducibility fa
Skills 7
Claude loads it when the task matches- discover
Initializes an evo workspace by exploring a repository, proposing optimization dimensions, building a benchmark, and running a first experiment.
- finetuning
Guides an agent through picking and diagnosing LLM fine-tuning methods, from reward-shape selection to smoke-runs, early-stop logic and failure checks.
- infra-setup
Guides an agent through switching evo experiment backends between local worktrees, pool slots, and remote providers, including prerequisite and auth checks.
- optimize
Runs a structured autoresearch optimization loop that plans rounds, spawns subagents to test candidate variants, and tracks the improving frontier.
- report
Reports an evo optimization workspace from recorded state, showing score charts, frontier candidates, and overnight improvements without running new benchmarks.
- ship
Distills an evo run's winning experiment into a minimal mergeable diff, lands it as a PR or merge, and attaches a mergeability report.
- subagent
Defines the protocol evo optimization subagents follow: read pointed traces, form a minimal edit, run experiments, and analyze results within a budget.
Hooks 6
Runs on these events- PostToolUsecommand
.*, Bash
- PreToolUsecommand
.*
- SessionStartcommand
No description in the file.
- Stopcommand
No description in the file.
- SubagentStopcommand
No description in the file.
- UserPromptSubmitcommand
No description in the file.
Using evo in Claude Code
Type a command in a Claude Code session:
/evo:discover
/evo:optimizeAsk for an agent by name:
@agent-evo:benchmark-reviewer
@agent-evo:ideatorSkills need no command. Claude reads them when your request matches what they describe.
Update, turn off or remove
- Update the plugin (shell)
- claude plugin update evo@evo-hq-evo
- Refresh the marketplace
- /plugin marketplace update evo-hq-evo
- Turn it off for now
- /plugin disable evo@evo-hq-evo
- Remove it
- /plugin uninstall evo@evo-hq-evo
Skills in evo
- discoverevo-hq/evo
Initializes an evo workspace by exploring a repository, proposing optimization dimensions, building a benchmark, and running a first experiment.
10 files1.5Kstars - finetuningevo-hq/evo
Guides an agent through picking and diagnosing LLM fine-tuning methods, from reward-shape selection to smoke-runs, early-stop logic and failure checks.
9 files1.5Kstars - infra-setupevo-hq/evo
Guides an agent through switching evo experiment backends between local worktrees, pool slots, and remote providers, including prerequisite and auth checks.
2 files1.5Kstars - optimizeevo-hq/evo
Runs a structured autoresearch optimization loop that plans rounds, spawns subagents to test candidate variants, and tracks the improving frontier.
2 files1.5Kstars - reportevo-hq/evo
Reports an evo optimization workspace from recorded state, showing score charts, frontier candidates, and overnight improvements without running new benchmarks.
1.5Kstars - shipevo-hq/evo
Distills an evo run's winning experiment into a minimal mergeable diff, lands it as a PR or merge, and attaches a mergeability report.
1.5Kstars - subagentevo-hq/evo
Defines the protocol evo optimization subagents follow: read pointed traces, form a minimal edit, run experiments, and analyze results within a budget.
1.5Kstars
evo plugin questions
How do I install the evo plugin?
In Claude Code run /plugin marketplace add evo-hq/evo, then /plugin install evo@evo-hq-evo.
What does the evo plugin include?
evo bundles 2 commands, 3 agents, 7 skills, hooks.
What commands does the evo plugin add?
It adds /evo:discover, /evo:optimize. Run them in Claude Code with the plugin prefix, for example /evo:discover.
Which agents come with evo?
benchmark-reviewer, ideator, verifier. Call one with @agent-evo:benchmark-reviewer.
How do I update or remove the evo plugin?
Update it with claude plugin update evo@evo-hq-evo in your shell, or from the Installed tab of /plugin. Remove it with /plugin uninstall evo@evo-hq-evo.