mirror of
https://github.com/boshu2/agentops.git
synced 2026-09-14 15:08:13 +08:00
Bo
8ba77aa0e2
Wave 7, run through a full `/rpi` on itself per the operator's
directive.
**The feature:** `validator: {kind: 'spawned' | 'command', command}` on
the rpi workflow. With `command`, the fresh Validate context becomes a
**broker** for an external judge (any vendor CLI — the command is opaque
caller input): identical evidence packet to the spawned judge, raw
stdout/stderr + the judge charter persisted under `.agents/ao/` and
referenced in `evidence_refs`, ruling transcribed verbatim. Three
load-bearing honesty rules: **no verdict laundering** (ambiguous ruling
→ NOT_PROVEN), **no silent fallback** (dead command → NOT_PROVEN, never
a quiet spawned-judge downgrade), **three distinct identities** (author,
external validator, broker attester). Spawned mode proven byte-identical
to baseline via prompt extraction.
**Built by the loop it extends:** Plan snapshotted the intent
(`660684c4…`), Implement proved the freshness wall in both modes with a
sentinel harness, a fresh validator persisted PASS `verdict.v2`
`bd733619…` over the exact four-file subject.
**The smoke story is the headline.** The first live cross-vendor run —
`codex exec --sandbox read-only` as judge — returned an honest **FAIL**:
the implement stage had gathered a git-status receipt from the wrong
working directory, and the external judge refused the inconsistent
evidence rather than rubber-stamping. That FAIL surfaced a real
prompt-precision gap; cwd pinning is now load-bearing in every stage
prompt (contract comment cites the incident). The revised invocation
passed: `verdict.v2` `1551d3be…` with **the external judge's own session
UUID as `validator_context_id`**, the broker's id as freshness attester,
and both raw transcripts in evidence — verified distinct three ways in
the persisted bytes.
**Gates:** 66/67 over this range; the one FAIL is the known
`workflow.install-drift` local-estate finding (post-merge reinstall
documented on #945). CI safe via absent→skip.
2 lines
3 B
Plaintext
2 lines
3 B
Plaintext
42
|