Date: 2026-09-23. Method tested: published commit
58d3cf7a76b218850410a716637629b66b3b722e on v2.
Target: separate local-only Python queue lab, commit
5051bccea1dd5d9ce9e98807f6d07ccb46dd28ba (not a public repository).
Environment: this Codex desktop task, Windows, Python 3.13.7.
| Check | Result |
|---|---|
| Remote publication | v2 matched local commit; main remained 68f5c4a4d45cf1531816b6f173332089d1bd8f08 |
| Fresh source retrieval | Separate remote clone matched the published pin; entry and helper validated |
| Hosted CI | Four jobs passed: Windows/Linux, Python 3.10/3.13 |
| Published helper unit tests | 50 passed |
| Controlled lab checks | 50 passed, zero failed in final run |
| Routing cases | 91: 72 kind/risk/project combinations plus all 19 signals; all ten modules reached |
| Helper entry points | route, check-record and validate, including all four workflow modes |
| Product regression tests | 13 passed; deliberately flawed v0 triggered 10 failures including subtests |
| Packaged service | 14 real HTTP checks passed against an extracted artifact on loopback |
| Git/instruction preservation | Simulated user edit and existing AGENTS.md retained byte-for-byte; user edit left unstaged |
| Read-only target | Content inventory unchanged; records stored outside target |
| Enterprise/team boundaries | Missing scanner and independent review remained blocked; handoff explicitly unsent |
The GitHub CI run is evidence for the published source revision, not for other IDE behavior.
- Policy/Git: existing instruction preservation, inert untrusted text/reference, explicit authorization boundaries, dirty-file preservation and precise staging.
- Quality/security/troubleshooting: tenant-access denial, atomic idempotency, invalid inputs, baseline failure reproduction and narrow corrections.
- Architecture/SDLC: migration preparation and commit only after success; injected failure preserved original state.
- Documentation: explicit contract, findings, diagnosis, limitations and handoff.
- Delivery: source -> ZIP -> extracted files -> running identity -> observed HTTP behavior, followed by shutdown of the owned test process.
- Team/enterprise: additive policy gates, hashed policy identity, blocked scanner and reviewer requirements, and a draft that was never represented as posted.
Negative record checks rejected missing gates, unsupported fields, absent/failed/ stale evidence, policy omission/content changes, unknown-policy passes, malformed JSON and fabricated completion through an unresolved gate. Positive synthetic records only prove format acceptance and are labeled synthetic in the lab.
The first HTTP facade used request as a helper method name. Python's server base
class sets request to a socket, so the first health request failed despite passing
domain tests. The traceback identified the collision. Renaming the helper to
dispatch, rebuilding and rerunning produced the passing live results above.
The failed run was retained; it was not relabeled a pass. This was a lab application
bug, not a discovered defect in exec-ctrl's helper.
The local lab retains timestamped command logs, routing matrix, JSON records,
HTTP observations, source snapshots and failed/successful reports. The successful
run is 20260923T214555889409Z; the earlier failed run is 20260923T214520665025Z.
They are local artifacts for this task, not publicly reproducible evidence bundles.
Packaged ZIP SHA-256:
003a3f92eb036d9e48b67da2b21caa9fdcf8f780a21277ab31df2325e49eeeda.
Packaged/extracted/running source SHA-256:
fb5a987ae8371cd662b6558773e8af7b446bc09ea2871728e7fd43b81d22917c.
This is a controlled trial by the agent that authored the framework, in a task with prior context. It is not a blind task-plus-link evaluation. Deterministic tests and observed local actions do not establish universal prompt adherence or prompt-injection resistance. Other IDEs, real organization-policy acceptance, approved scanners, independent review and actual team connectors remain untested.
The local test exercise is complete. A stable v2 release remains pending the independent trials in the protocol. This report does not waive the deliberately blocked organization release gates in the fixture.