Skip to content

v0.7.0 — audited acceptance and a default coding Skill for workers

Choose a tag to compare

@xenups xenups released this 14 Nov 07:52
· 21 commits to main since this release
e963a27

Headline

asha run --spec can no longer seal a worker whose acceptance never ran, and a
coding worker now receives the mandated clean-code Skill because Asha put it
there — not because the Spec author remembered a flag.

Fixed

  • Acceptance was recorded as not_executed for every governed worker. The
    DAG coordinator bound its execution hook to default_execute, which discards
    the runner's audit record, so build_acceptance always saw
    verify_ran=False. Any Spec with verify_command failed acceptance_failed
    regardless of the command's real result. The hook is now
    execute_with_audit, and the audit reaches evidence collection. Injected
    2-tuple hooks keep working (they simply carry no audit).
  • A non-passing acceptance could still be sealed. Acceptance was consulted
    only when the worker exited nonzero; a zero exit with not_executed or
    inconclusive acceptance stayed DONE. refuse_unproven_acceptance() now
    turns that into FAILED / acceptance_failed — in the DAG and in the
    single-worker facade the MCP tools use. The sealed evidence still keeps the
    honest record, including the real exit code.
  • The acceptance record now appears in the run report, so a caller can read
    what the command actually did instead of inferring it.

Added

  • Persistent coding-Skill policy — ~/.config/asha/worker-policy.json (or
    ASHA_WORKER_POLICY). When enabled, a command worker whose argv[0]
    resolves to the configured agent CLI and whose task_kind is coding (the
    default) is launched as hermes -s clean-code -z "<task>". Skill name,
    version, path, SHA-256, overlay digest and the effective argv are sealed into
    evidence as skill_preload.
  • Fail-closed Skill availability — an enabled policy whose Skill is missing
    or unreadable blocks the worker before it starts
    (BLOCKED / mandatory_skill_unavailable, searched paths reported).
  • Engineering overlay — an optional markdown file prepended to the task
    prompt, for project rules (canonical Enum for finite domains, boundary
    validation, SRP) that supplement the Skill instead of forking it.
  • task_kind on a worker — coding (default) or non_coding; the latter
    opts out explicitly instead of relying on prompt interpretation.

Compatibility

  • No CLI or MCP signature changed. A missing policy file reproduces v0.6.0
    behaviour exactly.
  • antigravity workers and arbitrary binaries keep their argv byte-for-byte and
    receive no Skill injection — Asha does not claim Skill support it lacks.
  • The acceptance state change is the one behavioural break, and it only affects
    runs that were previously reporting success without a proven acceptance.

Verification

  • 21 new regression cases (tests/test_acceptance_audit.py,
    tests/test_worker_policy.py).
  • Real governed runs on a disposable repo: acceptance executed by Asha
    ("18 passed"), COMPARED / NO_FAILURES, sealed evidence bound to the exact
    commits and trees; a failing acceptance yields acceptance_failed with no
    seal; a missing Skill yields BLOCKED with the worker never started; a
    non-coding worker gets no preload.
  • Candidate code from the Skill probe (StrEnum, boundary parsing, typed
    tables) — see ASHA-CLEAN-CODE-DEFAULT-INTEGRATION.md.
  • CI: verify pass, fail-closed governance gate pass (PR #25).

Known limitations

  • A preloaded Skill is an instruction, not an enforcement. Asha proves the
    Skill was resolved, hashed and handed to the CLI; it cannot prove the model
    obeyed it. Acceptance and review remain the gates.
  • The legacy GovernedScheduler still owns its unaudited execution path; it is
    not used by asha run, by MCP, or by the release gates.

Full report: ASHA-ENGINEERING-HARDENING-REPORT.md.