Gaps from three consumer alchemy sessions
Three real runs exposed these: a full audit followed by a fix PR, an unattended audit with 13 agents, and a runtime-queue optimisation.
- Findings cannot vanish.
- A hand-off may carry a
findingslist, each finding citing evidence. - The gate's
--reviewcan be repeated and accepts saved hand-offs, so a dropped blocking or major finding fails withSILENT_SKIP. - A read-only audit may leave its evidenced findings open and lists them under
open_serious. A finding without evidence fails with the new codeUNSUPPORTED_FINDING.
- A hand-off may carry a
- Hand-off paths:
- An absolute path or a
file@commitpath is aSCHEMAerror, no longerSTALE_CONTEXT. - A deleted file is
{"path", "deleted": true}and is checked to be gone. - The rejection messages say how to fix the hand-off.
- An absolute path or a
- Workspace checks:
- Python bytecode caches no longer cause
WORKSPACE_MISMATCH, andrecord_evidence.pyno longer writes them. preflight --rootaccepts an integrated head.
- Python bytecode caches no longer cause
- Execution: the plan check counts analyze packages as independent parts and prints
mode_overridewhen a plan's mode differs from its recommendation. - alchemy:
- An ambiguous subject names both readings and stops.
- "Fix them all" after an audit is a new dispatch.
- Optimisation words route to Y.
- The Y baseline rule is one rule everywhere: after the increment ships, or on a running system, measure the baseline first.
- alchemy-orchestrate:
- A spawn without a passing plan check is reported as a skipped check.
- Blanket approval does not cover protected paths.
- Agent instructions carry known tool quirks.
- Each pull request head is gated.
- contracts.md: the hand-off rules are written out: criterion tagging, relative paths, plan requirement IDs only, deletions.
- system-optimization:
- New Runtime Flow layer.
- If missing access blocks measurement and the user can grant it quickly, stop with the command instead of a Decision.
- Baselines go to the owning issue.
- Installers: no
.l-gevityside file when the instruction file differs only in line endings.
Re-render agent definitions with init_workspace.py to pick up the analyst instruction to report findings in the hand-off.
Verification
validate-skills.pypassed. All 168 deliberate validator mutations were caught.- The checker has 9 new adversarial cases.
- The bytecode-cache ignore and the integrated-head preflight were checked by hand in a throwaway git repository.
- All 40 behaviour scenarios were re-recorded, 3 of them new. 39 meet every expectation.
- The shell installer suite passed (83 checks), and the PowerShell suite passed under pwsh 7 and Windows PowerShell 5.1 (160 checks).
- CI passed on both pull requests before publication.
Limits
gate-trial-expiredlists one known failure. The recorded run refuses to delete the gate, but in words its pattern does not accept. gate-lifecycle is unchanged.- No runtime hook yet blocks spawning agents without a checked plan; it is a written rule.
- Report-block sentence limits are not checked automatically.
- The checker still only accepts upper-case secret names, so credential stores need an invented name.
- The library word budget rose from 49,300 to 49,370.