Hand-offs from the agent reply, evidence for every claim
alchemy-orchestratestep 6: the lead saves each hand-off from the reply the agent tool returned, never from a runtime output file. On Claude Code for Windows, background agents'tasks/<id>.outputfiles are 0 bytes while the reply is complete; a FULL audit skipped its hand-off check because of this.- Hand-off contract: every claim cites an evidence item, findings from reading files included (
git grep -n,sed -n).commandis recorded exactly as it ran,exit_codeis the command's own (not a pipe stage's), andexpect_exitmarks a failure that is itself the finding. - The analyst envelope says every claim cites an evidence item; re-render agents with
init_workspace.pyto pick it up. adapters/README.mdnames each runtime's hand-off source and the Claude Code subagent transcript path.
Verification
validate-skills.pypassed; all 168 deliberate validator mutations were caught; the new wording is pinned.- 7 affected behavior scenarios re-recorded, all expectations met.
- On the audit that exposed the gap, the hand-off check run afterwards accepted P1 and rejected P2 and P3 for exactly the faults this release addresses.
- CI passed before publication.
Limits
No new scenario fails without this change: the hand-off checker already rejected these hand-offs; the change aligns the lead's procedure and the analyst instructions with it. Codex's hand-off source is not verified. The library is at its word budget.
Includes #49.