Run History Skill Upgrader
Přemění důkazy o běhu na návrh upgradu existující dovednosti pouze s plánem a schvalovacím procesem
Otestováno · Funguje
Co umí
Meta-skill that converts real run evidence (failures, source/platform drift, user feedback) into a plan-only, two-stage-approval upgrade proposal for an existing skill, then applies structural changes only after explicit plan approval.
Testovací report
Naklonováno repo a spuštěn the bundled validate_upgrade_artifacts.py against the skill dir: exit 0, "validation passed"; evals.json passed python -m json.tool. Frontmatter has clean name+description; spot-checked 3 referenced files (validator script + 2 references) all HTTP 200; validator itself scans for machine-specific paths and found none. Pro OUTPUT I gave both runs the same task (improve a git-commit-writer skill after ONE subject-only failure, "go ahead and edit it"): baseline immediately edited the file and hard-coded a "never" rule from n=1; the skill-following artifact stayed PLAN ONLY (correctly rejecting "go ahead" as approval), ran the generalization gate, flagged the single run as weak evidence, and proposed one conditional guideline instead of stacked negatives. Concrete, quotable behavioral difference.
Testováno: 2026-07-21 · Claude Code 2.x (agent harness)
Instalace
git clone --depth 1 https://github.com/dongshuyan/compass-skills.git /tmp/run-history-skill-upgrader-src mkdir -p ~/.claude/skills cp -R /tmp/run-history-skill-upgrader-src/skills/run-history-skill-upgrader ~/.claude/skills/run-history-skill-upgrader # No API keys or external deps. Bundled validator needs python3 (or python / py -3): # python3 ~/.claude/skills/run-history-skill-upgrader/scripts/validate_upgrade_artifacts.py \ # --skill ~/.claude/skills/run-history-skill-upgrader # Meta-skill: it upgrades OTHER existing skills; it does not do a business task itself.
Příkazy a ukázkové prompty
/run-history-skill-upgraderPřemění důkazy o běhu na návrh upgradu existující dovednosti pouze s plánem a schvalovacím procesem
Skilly se spouštějí běžnými požadavky — žádné příkazy k zapamatování. Po instalaci ho aktivují prompty jako tyto (anglicky):
Improve this skill based on its recent failuresUpdate this skill for the platform's changed behaviorTrim bloat from this skill using recent run evidence