learning-loop

What the learning loop is

stellar-build ships with a local learning loop: the more you use your skills, the better they get — without anything leaving your machine. It mirrors OpenJarvis's jarvis optimize skills / jarvis bench skills cycle, adapted for markdown skills.

Three moving parts:

1. Capture — a skill use is a multi-turn span, so capture is decoupled from "skill became active". A PostToolUse:Skill hook marks which skill is resident; a Stop hook records each turn's real assistant output; and a SessionEnd hook closes the span with a summary, counting it as one activation. Everything lands in ~/.stellar-build/traces/<day>.jsonl. Fully local. Opt out with STELLAR_BUILD_NO_TRACE=1.
2. Optimize (/optimize-skills) — compiles those traces into sharper skills: tightened triggers/instructions and few-shot examples mined from your own successful runs. Every edit is backed up and reversible.
3. Bench (/bench-skills) — scores skills on held-out prompts so you can measure whether an optimization actually helped.

   use skills  ──▶  ~/.stellar-build/traces/*.jsonl
                          │
              /optimize-skills  (DSPy-style compile)
                          │
                          ▼
            sharper ~/.claude/skills/<skill>/SKILL.md   ◀── stellar-loop restore (undo)
                          │
                /bench-skills  (measure the lift)

What to do when invoked

• Overview / "how does it work" → explain the three parts above, then point to the two action skills and the CLI.
• "status" / "is tracing on" / "show my usage" → run the CLI and relay the output:
  • stellar-loop status — install + capture + hook state, trace/optimization counts
  • stellar-loop stats — per-skill usage from local traces
  • stellar-loop tail — most recent trace records The CLI is installed at ~/.stellar-build/bin/stellar-loop (add ~/.stellar-build/bin to PATH, or call it by full path).
• "optimize" / "improve my skills" → hand off to the optimize-skills skill.
• "benchmark" / "did it help" → hand off to the bench-skills skill.
• "undo" / "restore" → stellar-loop restore <skill> reverts the last optimization for that skill from its backup.

Data & privacy

• Everything lives under ~/.stellar-build/ (traces/, backups/, bench/, optimizations.jsonl). Override the root with STELLAR_BUILD_HOME.
• Nothing is transmitted anywhere — no API calls, no telemetry. The "optimizer" is your own agent session reasoning over local files.
• Result previews are truncated and examples are generalized before they ever land in a skill, so traces and learned blocks should not accumulate secrets. Wipe anytime with stellar-loop clear (--all also clears backups).

Quick reference

Goal	Do this
See if it's working	stellar-loop status
See what you use most	stellar-loop stats
Make skills sharper	/optimize-skills (or stellar-loop optimize)
Measure the lift	/bench-skills (or stellar-loop bench)
Undo an optimization	stellar-loop restore <skill>
Turn capture off	export STELLAR_BUILD_NO_TRACE=1
Delete all traces	stellar-loop clear