Skip to content

demo.md

Latest commit

 

History

History
61 lines (43 loc) · 2.17 KB

demo.md

File metadata and controls

61 lines (43 loc) · 2.17 KB
Raw

Demo run

This package was tested from GitHub with Pi using a disposable fixture repository.

Command shape

pi -e git:github.com/saphid/build-skill \
  --session-dir /tmp/build-skill-demo-20260516-153422/sessions \
  --thinking minimal \
  -p '/skill:simple-build Demo run from the GitHub package. Build two independent utility Beads using separate git worktrees...'

The full prompt asked the agent to build two independent Beads:

  1. src/math.ts with clampScore(value, min, max) and tests/math.test.mjs.
  2. src/slug.ts with slugifyLabel(input) and tests/slug.test.mjs.

The requested workflow was: plan Beads, use separate worktrees, run focused test-first/RGR where practical, review each Bead, then act as separate merge reviewer and merge PASS branches into main without pushing.

Result

  • Exit code: 0
  • Wall clock: 130s
  • Assistant token usage: 258,041 total tokens
  • Cost reported by Pi usage metadata: $0.4511
  • Final validation:
    • npm run test: 10/10 tests passing
    • npm run typecheck: passed
    • npm run build: passed
  • Worktrees removed after merge.
  • Local Bead branches deleted after merge.
  • No push performed.

Beads and commits

Bead Branch Commit Result
B1 math utility build/b1-math 82848c3 Add clampScore utility PASS
B2 slug utility build/b2-slug 1c7e60c Add slugifyLabel utility PASS

Merge commits on disposable main:

  • e34697e Merge build/b1-math
  • 4b8a34e Merge build/b2-slug

Comparison to historical Life Agent sessions

Historical successful prompted workflows were usually shaped like:

small Bead coder → sandbox/worktree → focused validation → report merge-ready branch
then merge reviewer → rerun validation → fast-forward/verify/cleanup

Approximate historical observations:

  • focused Bead coder: ~5-7 minutes and ~0.8M-1.2M tokens;
  • merge gate: ~3 minutes and ~150k-250k tokens.

This demo is smaller than those production Beads, so the numbers are not a direct benchmark. It does show the package can reproduce the same workflow shape with clean worktrees, review, merge validation, and cleanup at very low overhead on a simple task.