Skip to content

docs: record the actual platform verification results - #12

Merged
miter37 merged 1 commit into
masterfrom
docs/record-platform-verification
Jul 23, 2026
Merged

docs: record the actual platform verification results#12
miter37 merged 1 commit into
masterfrom
docs/record-platform-verification

Conversation

@miter37

@miter37 miter37 commented Jul 23, 2026

Copy link
Copy Markdown
Owner

The docs claimed Windows, Linux, and macOS were unverified because the project was built in a Linux container. That stopped being true today.

What changed

docs/KNOWN_LIMITATIONS.md §1 gains a verification record:

Worker Version Result Probe
claude 2.1.218 (Claude Code) healthy 18.6s
codex codex-cli 0.144.3 healthy 61.6s
antigravity 1.1.5 healthy 136.3s

All three with unattended_ok, output_ok, and artifact_ok true — they completed without human intervention and produced schema-conforming results and artifacts.

§2 is split into verified and not yet verified, rather than listing everything as unverified.

README.md gets the same correction in English, per the documentation language policy.

Deliberately still listed as unverified

This is a correction, not a promotion. The following remain open and are now stated plainly rather than buried:

  • Real provider CLIs on Linux and macOS. CI runs on all three OSes but exercises the bundled mocks. Field validation happened on Windows 11 only.
  • Job Object full process-tree termination — never exercised by a job that spawns children.
  • Dedicated RelayWorker account and NTFS ACLs.
  • Antivirus / corporate security interaction.
  • Browser helper cleanup.
  • Stall thresholds on long silent jobs.

One detail worth keeping

codex reported schema_hint: false — its --help does not advertise --output-schema — yet the deep probe passed. That is the "help output is advisory, deep probe is authoritative" design working exactly as intended, so it is recorded rather than glossed over. Trusting --help alone would have rejected a working worker.

Verification

66 tests pass; ruff check clean. Documentation only — no code changed.

The docs said Windows/Linux/macOS were unverified because the project was
built in a Linux container. That is no longer accurate, and it understated
the project to anyone evaluating it.

KNOWN_LIMITATIONS.md section 1 gains a verification record: all three
providers passed deep doctor against their real CLIs on Windows 11 --
Claude Code 2.1.218, codex-cli 0.144.3, Antigravity 1.1.5 -- with
unattended_ok, output_ok, and artifact_ok true in every case. Also notes the
codex case where --help omitted --output-schema but the deep probe passed
anyway, since that is the advisory-help design working as intended.

Section 2 is split into what is verified and what is not, rather than
listing everything as unverified. Still unverified, and now stated plainly:
real provider CLIs on Linux and macOS (CI exercises mocks only there),
Job Object process-tree termination, the dedicated account and ACL setup,
antivirus interaction, browser helper cleanup, and stall thresholds.

README gets the same correction in English, per the documentation language
policy.
@miter37
miter37 merged commit 921a596 into master Jul 23, 2026
20 checks passed
@miter37
miter37 deleted the docs/record-platform-verification branch July 23, 2026 09:40
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant