github l-gevity/l-gevity-skills v5.2.0
v5.2.0 — recorded evidence, thinner sub-agents

49 minutes ago

Recorded evidence, thinner sub-agents

Orchestrated runs spent 16–26 minutes per sub-agent. Every sub-agent loaded the lead's orchestration skill before starting work, and verifiers re-ran checks the lead had already run on the same commit. Now the lead records each check once, and sub-agents load only what their package needs.

Changes

  • New alchemy-orchestrate/scripts/record_evidence.py:
    • Runs the plan's command criteria and the profile's verification commands once, at a clean committed head.
    • Writes a log and one ledger evidence item per command (produced_by: harness, log, log_sha256, ID <criterion>@<short commit>).
    • Runs only commands the plan or profile declares, and fails when a command changes the work tree.
  • Delegated agents no longer load alchemy-orchestrate. They read only the Hand-off section of contracts.md.
  • Verifiers read the recorded logs and cite their IDs. They run a command only for a criterion no record covers, or to test one finding they suspect, at most three times per package (governance rule 20).
  • When the lead receives a hand-off, it reads each changed file against the hand-off's claims. The hand-off check proves file digests, not meaning (step 6).
  • The pipeline prompts, review track, and primer follow the same rules.

Verification

  • Skill validation and 154 deliberate validator mutations, including new cases for the agent instructions and the recorder's refusal of an uncommitted work tree.
  • A functional test of the recorder in a scratch git repository: exit codes, log digests, the uncommitted-tree refusal, and a command that writes files.
  • Six orchestration scenarios re-recorded: 26 of 26 expectations met.
  • CI green on the merged branch.

Limits

  • The speed gain and the effect on defects found are not measured yet. Run .pilot/run_pilot.py --arm multi and compare it with the recorded 5.1.0-era pilot (1368 s).
  • Capping verifiers at three commands of their own may find fewer defects than open-ended probing did.
  • The Unstoppable pipeline changed in prompt text only; no full pipeline run was executed.
  • Existing installs keep their rendered agent files (alchemy-analyst, alchemy-implementer). Re-running the installer writes the new versions next to them with a .l-gevity suffix, to merge by hand. A project alchemy-workspace.json that overrides envelopes keeps its own instructions.

Includes #45.

Don't miss a new l-gevity-skills release

NewReleases is sending notifications on new releases.