Alchemy answers are now checked against their own output contracts, and three routing gaps are closed. A consumer's FULL audit turned up both. This release was re-issued at the merge of #57, with the fixes from the v5.7.0 multi-agent field report (#56).
Changes
-
scripts/check_alchemy_report.py: checks the shape of a routed Alchemy answer:- trail columns, skip rationales and stage order
- scope-block fields and their allowed words, and whether
Blocking stagematches the trail - one record for every stage that reached a decision
- worth records (required fields, a revisit trigger, measured values in retrospective mode), design principles, the test obligation table, and placement records
- every unchecked item names a closing command, and an audit names functional correctness
Run it on any saved answer:
python scripts/check_alchemy_report.py answer.md. -
Scenario check
alchemy_report, plus a new scenario:audit-full-reports-unchecked-correctness. -
full auditnow walksFULLfrom the C₀ baseline. Before,ALCHEMY-PIPELINE-DESIGN.mdcounted plainauditas full traversal, which contradictedSKILL.md. -
Audit removals: in Audit mode,
DEPRECATE,DELETEandOBSOLETEadmit the removal as the candidate's remediation increment.Blocking stageis now the first stage that no candidate passed. -
Audit reports list functional correctness as unchecked, with the command that reviews it. No gate tests behavior.
-
Installers default to
v5.8.0.
Multi-agent fixes (#56, #57)
alchemy-orchestrate/scripts/extract_handoff.py: copies a hand-off byte for byte from the agent's transcript or reply, so the lead no longer retypes it. It refuses a final reply without exactly one hand-off.- Finding IDs: the ledger refuses a repeated finding ID.
gate --reviewmatches a hand-off's findings by package as well as ID, so two agents'F1can no longer merge silently. Hand-off IDs carry the package prefix (P1-E1,P1-F1). - Read-only gate: when blocking or major findings stay open, the gate's output notes that a pass is not a clean audit.
- Analyst command cap: "at most three" now applies only to packages that verify another package's work.
- Topology record rule R10:
Graph analysisnames tool, version, and input hash, or saysNot measured. - Profile: dropped
repository.base_branch, which nothing read.
Verification
npm run validatepasses.npm run validate:mutation: all 173 deliberate violations caught.- The report checker's self-test trips all 15 rules; the topology checker's self-test passes, R10 included.
- 16 scenarios re-recorded for #54; the 7 orchestrate scenarios re-recorded for #57 meet all expectations.
- The extractor, run on 71 recorded transcripts: 57 hand-offs copied, 14 replies without a valid hand-off refused.
bash scripts/test-installers.sh: 115 checks pass.scripts/test-installers.ps1: 224 checks pass under pwsh 7 and Windows PowerShell 5.1.- CI green on #54, #55 and #57.
Limits
- Known failure: in the new audit scenario, 2 of 2 runs closed some unchecked items with a question to an owner rather than a command.
- Known failure: in
dispatch-followup-redispatch, 2 of 3 runs named a planned route inCore routethat never ran. - The checker runs on recorded scenarios in CI. It does not run inside a consumer's session unless someone runs it by hand.
- Between a version bump merging and its release tag existing, the installers on
mainpoint to a tag that is not there yet. - Re-issued: the
v5.8.0tag moved from6325718to the #57 merge. An install pinned tov5.8.0before the move holds the earlier contents. - No check confirms that an agent read the Full-mode topology references; R10 checks only the record's graph field.