Documentation Index
Fetch the complete documentation index at: https://docs.selftune.dev/llms.txt
Use this file to discover all available pages before exploring further.
Usage
selftune search-run runs a bounded package search against your local draft
skill package. It generates routing/body variants, evaluates them through the
shared package evaluator, compares them against the accepted frontier, and
records the measured winner plus provenance.
Options
| Flag | Description |
|---|---|
--skill-path <path> | Path to a skill directory or SKILL.md file |
--skill <name> | Override the inferred skill name for lineage/reporting |
--surface routing|body|both | Limit the mutation surface; defaults to both |
--max-candidates <n> | Cap the minibatch size; defaults to 5 |
--agent <name> | Runtime agent used for shared package evaluation |
--eval-set <path> | Override the eval set used during package evaluation |
--apply-winner | Promote the winning candidate back into the draft package |
--json | Emit the full search result as JSON |
--help | Show command help |
How it works
- Generate reflective routing/body variants from measured replay, routing, and grading failures when an agent is available.
- Fill any remaining routing/body budget with eval-informed targeted variants, then deterministic fallback variants.
- Fingerprint each candidate and skip duplicates already represented in the package frontier.
- Evaluate the minibatch through the shared package evaluator.
- If routing and body both produce accepted improvements, evaluate a merged candidate that combines the complementary surfaces before final winner selection.
- Compare accepted candidates against the measured frontier parent.
- Persist the search run, winner, and provenance.
- Optionally apply the winner back into the draft package.
--surface both is used, the routing/body budget is biased toward the
weaker measured surface from the accepted frontier or canonical package
evaluation instead of always splitting evenly.
Relationship to improve
selftune improve --scope package is the lifecycle alias for this low-level
package-search command. It delegates to search-run, and unless --dry-run is
set it applies the winning candidate automatically. Plain selftune improve
also auto-selects this path when the target skill already has package evidence
or a draft package manifest.
selftune run / selftune orchestrate can also select package search
automatically for eligible skills when accepted-frontier or canonical
package-evaluation evidence exists.