ALCHEMY 5.0.0 — a smaller library, and unattended runs that are checked
The library says the same rules in fewer words, and a multi-agent run that nobody watches now has to prove it can run before its first wave starts. This is a major version: two skill names are gone, and an existing unattended setup needs three additions.
Breaking changes
-
Two skill names removed.
architecture-as-code-javascriptandarchitecture-as-code-pythonare now references ofarchitecture-as-code. The library has 24 skills, not 26. One description carries the triggers of all three. -
Unattended runs need a bound preflight. A plan with
execution.unattended: truepasses the gate only with a preflight record bound to that plan. After the upgrade an existing unattended setup fails until you add:needs[]on each package that requires access, an account, an environment, or a fixture;plan: {digest, base}on the preflight record;covers[]on each preflight entry that proves a need.
The codes you will see are
NEEDS_UNDECLARED,PREREQUISITE_UNCOVERED,PREFLIGHT_STALE, andPREFLIGHT_REQUIRED. The table is inalchemy-orchestrate/references/autonomy-preflight.mdsection 4. -
Unstoppable pipeline: plan approval asks one more question. The human chooses between continuing while attending and continuing unattended. Unattended runs the preflight first. An absent or unclear choice counts as unattended.
-
Rework is recorded as plan amendments, and checks need the approved revision. A correction after the plan was approved is an entry in the plan's
amendments[]. It is part of the plan digest, so an amended plan needs a fresh preflight binding. The plan check printsapproved_digest; pass it as--approved <digest>to every later plan check and to the gate. Without it, a plan with amendments, and a gate whose ledger records rework, fail withAMENDMENT_INVALID. -
A gate run with
--rootreports writes no package owns (UNOWNED_WRITE). A lead fix-up outside every write set now fails; record it as an amendment the lead owns.
Compression (issue #33)
SKILL.mdtext: 57,593 → 49,195 words (26 → 24 skills). No decision rule, verdict value, evidence bar, or pinned phrase was removed; each family was reviewed by an agent that wrote none of it.- Measured on the behavior scenarios, old text against new: no difference in what a run gets right, and no measurable saving in what a run costs. The gain is text that is shorter to read and maintain.
- The
alchemyrouter is dispatch-only: every call loads 383 fewer words. - Every "See also" list is gone; a rule is stated once; one-path record shapes moved to references.
functionality-complexity-tradeoffnames its cost terms in words (Maintenance, Risk, Evolution, lifetime), so they no longer collide with gate letters.continuous-improvementstates its rules for any skill library.- Two library caps hold the result: total
SKILL.mdwords (49,200) and the number of skills (24). A new skill funds its words with a cut.
Unattended runs and bounded rework (issue #35, sections 1 to 3)
- Preflight coverage. Each package declares what it needs. The preflight check fails when a declared need has no entry that covers it, and when the record was made for another plan.
- The gate refuses an unattended plan without a passing preflight (
gate --preflight <file>). - Amendments cannot widen the plan during a run. An amendment made after the approved revision that adds scope, permissions, or protected paths fails the plan check (
AMENDMENT_BROADENS), whatever approval it carries. More scope needs a revision the human approves. - The rework bound holds. Rework is counted along the amendment chain and the ledger, whichever is higher, and against every chain whose files a fix touches. An extension needs a human's recorded answer with a reference and an expiry (
AUTHORIZATION_UNPROVENotherwise). A timeout grants nothing. - A run that is not converging stops. The gate escalates (
NOT_CONVERGING) when a serious finding returns in a later round, or at the second fix that introduced or reopened one. - Unstoppable pipeline: one looping correction job replaces the fixed correction steps. Each pass amends the plan, checks it against the approved digest, corrects, reviews read-only, and gates. It restarts only on
FAILwith passes left.
Checker and validator
- The orchestration checker has 67 check codes, each exercised by one of 223 adversarial cases.
- The validator checks that the counts in the README match the tree.
- 36 behavior scenarios hold what a reader does with the text.
Not in this release
- Issue #35, sections 4 to 6 (incremental review, escalation delivery apart from approval, usable acceptance measurements) are planned for 5.1.0. Until then the pipeline's
escalatejob still waits for a checkbox, and a failed preflight ends the run without a stop report. - The pipeline has not been run in the Unstoppable runtime.
pipelines validateaccepts its structure; the scheduling of skipped jobs was read from the runtime's documentation, not observed. - The validator does not parse the pipeline's YAML. It holds the wiring by text and scans for the two quoting defects it has met.
- A hand-run gate cannot see deleted amendments unless it is given
--revisionwith the digest an earlier check accepted. The pipeline passes it within a correction pass. - Across restarted correction passes the pipeline does not carry the revision binding. The pass limit still bounds the run.
- The pipeline's first gate does not compare the ledger with the reviewer's output; the gates after a correction do.
- A hand-run
--approvedaccepts any revision the plan holds. The pipeline always passes revision 0. - Scenario
orchestrate-misleading-handoffmisses one expectation in some recordings on unchanged text. - Convergence counts findings, not distinct fixes, so it can stop a run early. The extension ceiling has no upper bound, while the runtime caps passes at 50. In the pipeline
NOT_CONVERGINGstops the whole run, also the work that could continue. The correction pass re-binds the preflight without running probes again. - The checker counts one rework cycle per amendment. The pipeline asks for one amendment per origin per pass; a lead who writes one amendment per finding uses the bound up faster.
- The preflight window is not compared with the current time.
- Open items of issue #33 are listed there.