Releases: beyond10x/metaharness
Releases · beyond10x/metaharness
Release list
metaharness 0.7.0
Added
metaharness aep driveowns concrete model execution, native hooks and live evaluation
above AEP's neutral governor. Pass the planning executable with--aep-binary; compatible
retained runs preserve their plugin selections and spend ceilings. AEP remains the owner
of command/operator driving and offline evidence ingestion.
Changed
- Publish source with standalone bot tooling, independently of Atlas checkout freshness and
organization dependency admission, while retaining repository correctness and release checks.
Fixed
- The AEP eval's
checks/read the charter documents where they actually live, and E4
judges the decomposition edge rather than every edge.run-checks.shwent from
2 pass, 67 fail, 0 brokento11 pass, 58 fail, 0 brokenagainst agentplugins0.7.0and
AEP0.51.0— the same command, the same two checkouts, measured back to back on
2026-09-04. Nine rows moved and nothing else did:T1–T8andE4. - One source of truth for the two charter documents, and it is not this repository.
check-trace-documents.sh,check-runner-verdict.shandcheck-live-evidence.shread
expectations.decomposer.trace.yamlandexpectations.plan-reviewer.trace.yamlunder
evals/aep/, where those files have never existed. They now resolve
$AEP_REPO/conformance/eval/<case>/expectations.trace.yamlthroughcharter_specin
checks/lib.sh, withEVAL_CHARTER_SPEC_DIRandEVAL_CHARTER_SPEC_DECOMPOSER/_REVIEWERas
overrides and ared_allreason that names the unreadable path and the variable to set. Migrating
a copy here was the other option and is the worse one: AEP replays that corpus in its own
task check, nothing underevals/runs in this repository's gate (invariant 5), and a copy
nothing gates is exactly whatchecks/contracts/trace-expectations.txthad become. contracts/trace-expectations.txtis re-pinned to those documents, and is now read in both
directions. It named seven ids per document that no canonical document carries, and spelled the
CLIaep artifact …where every recording underchecks/transcripts/— evidence, not rewritten —
spells itprotocol artifact …. Every R12 bound keeps its kind, tool and matcher; the file
carries the old-id → new-id mapping so the claim can be checked rather than taken.T2now also
asserts that every gating expectation the document declares is named in the contract: read one
way only, a hand-maintained list is green while the subject adds a bound nothing here reads.- R12's
tool.absentoverWriteis gone from the contract, on the eval owner's recorded
decision, not this repository's.conformance/eval/decomposer-charter/case.yamldrops it
because the charter grants[Read, Grep, Glob, Bash], so the row is true of every possible run.
That settles the conflictchecks/README.mdhad left open asT7's, which now passes on the
pairing it was written to demand. E4asserts theaep artifact new storyexample's edge isdecomposes, through an extractor
scoped to that invocation, instead of requiring that no other*:epic:token exists in either
file. agentplugins0.7.0teachesblocks:epic:on adecision-blocker
(plugins/aep-plan/agents/decomposer.md:93), which is correct and which the old grep reddened on.
E2reads through the same extractor, so both rows are claims about the same sentence.- The checks invoke
aep, not the retiredprotocolspelling —check-trace-documents.sh,
check-scratch-fixture.shandcheck-decomposes-edge-examples.sh, the last of which guarded on
protocolbeing onPATHand then calledaep.
Notes
- This source release was cut without running gates, tests, paid evaluations or binary packaging,
at the operator's request. The dated observations below are retained historical evidence. - On 2026-09-04,
run-checks.shexited 1; its 58 red rows are enumerated inevals/aep/checks/README.md
by what each waits for: 37 onrun-agents.sh, 11 on recordings of three paid live runs, 7 on
a README section for that runner — none of whichW4-1/1ever built — 2 on a pre-task revision
b83c623that no AEP checkout can reach since the plugins moved toagentplugins, and
L7on whichever subject checkout happens to be dirty. Nothing was softened, skipped or removed
to shorten that list.
Metaharness 0.4.1
Fixed
- Stage and invoke the canonical
aepexecutable throughout the live AEP evaluation path. - Align the toolchain fixture, evaluation assertions, operator guidance, and current conformance corpus while retaining historical transcripts as compatibility evidence.