Skip to content

v0.12.0

Choose a tag to compare

View all tags
@onesimplecode onesimplecode released this 28 Aug 17:20
· 2 commits to main since this release
v0.12.0
fb4e317

[0.12.0] - 2026-08-28

Added

  • Frozen, generated tool contracts (TR-AGT-010) — additive-forever field semantics,
    protocol_version on every response and error, and a spec generated from live operation
    definitions rather than hand-maintained, checked two ways: the generated doc against what's
    committed, and the declared contract against what the registration code actually sets. Promoted
    from a running private-monorepo MCP tool surface with a passing drift-guard test covering both
    checks.
  • AGENTS.md — "Disposition Field (TR-AGT-003, field 6)" — an optional sixth loop-contract field
    for a triage/gate node whose output routes to more than one next step, derived from the node's
    own verdict rather than set independently, plus a named caution about a fail-closed default
    landing on the wrong side of a gate that itself fails open on ambiguous evidence.
  • Hybrid deterministic-then-agentic classification (TR-AGT-011) — a multi-stage classification
    pipeline (deterministic → optional trained-model → LLM) tags each result with which stage
    produced it, from one closed vocabulary; a deterministic stage may short-circuit a later, more
    expensive stage entirely, and an unbuilt stage is declared, not fabricated. Confidence-gated
    override of an earlier stage by a later one is a natural extension, not itself covered by this
    ID — it stays a backlog item until a real override path exists to evidence it. Reinforces
    TR-AGT-002 and TR-SEC-012 at the level of pipeline architecture. Promoted from a running
    private-monorepo classification pipeline.
  • Graded outlet confidence (TR-SEC-015) — an optional, more finely graded confidence about a
    specific untrusted outlet, alongside TR-SEC-011's fail-closed trust tier — never upgrading trust,
    graded from a small curated allowlist with fail-closed absence. Includes a named caution about
    parsing untrusted host/URL data on a hot path without assuming the failure mode is "returns
    nothing" rather than raises. Promoted from a running private-monorepo retrieval pipeline with a
    passing drift-guard test.
  • AGENTS.md — "Public Claims Require a Pinned Benchmark" section, generalizing
    tests/test_readme_claims.py's drift-guard pattern (previously scoped to this
    repo's own asset/test counts) and TR-TEST-004/TR-TEST-005 into a named
    convention: quantitative or comparative claims about an AI tool's behavior
    must cite a versioned benchmark, not assert a number. Cites
    alibaba/open-code-review's AACR-Bench as external corroboration.

Changed

  • AGENTS.md — "Deterministic Checks Before Agent Judgment" (TR-AGT-002,
    TR-AGT-006) now cites alibaba/open-code-review as independent, at-scale
    validation of the same deterministic-scoping-before-agent-judgment split.
  • registry/tr-registry.yaml — TR-AGT-003's text extended with the disposition-field
    description (field 6); no ID renumbering, three new IDs added (TR-AGT-010, TR-AGT-011,
    TR-SEC-015). Registry now 41 requirement IDs (was 38).

Fixed

  • Credential-isolated broker example (examples/credential-isolated-broker/broker.py) —
    consume_mutation popped the lease before validating the request, so a failed validation (e.g.
    a stale current_revision) permanently destroyed an otherwise-valid lease — a validation
    failure could deny the legitimate caller's own retry. Now validates first and only consumes the
    lease on a successful match. Found by code review, fixed with a regression test
    (test_failed_validation_does_not_burn_the_lease).

Full changelog: https://github.com/onesimplecode/agent-engineering-standards/blob/main/CHANGELOG.md

Assets 2
Loading