[BunsDev]

Part of #281 (CF-8), tracker #273. Val authorized one short, benign live run per CLI on the signed-in accounts on 2026-10-19.

What it adds

• e2e/live-smoke.spec.ts, an opt-in live scenario (MODEX_LIVE_SMOKE=1; CI skips it, 2 skipped). For each CLI it drives the real app, main process and native adapter in a disposable repository:

  1. A real edit and check. It asserts the file's bytes and a check tool item that succeeded (exit code 0 for Codex), plus an answer and no failure card.
  2. Stop during an approved 30-second command. It asserts the call isn't succeeded and that its file is never written.
  3. Codex only: one web search, recording the result's shape (keys only) from Codex's own session record for CF-5.

  A turn the CLI refuses settles in error, and the report keeps the failure card's code, its summary and the CLI's own message.

• docs/reviews/2026-10-19-live-smoke.md with the results, and the CF-8 matrix row now points to it.

Results

Claude Code 2.1.304 (claude -p, Haiku): passed.

• cat, Edit and grep -q 'hello, live' greeting.txt all succeeded, and the file reads hello, live. Agent mode needed no cards.
• The answer's phase is unspecified. Claude reports no phase, so Modex claims none.
• Stop: idle in 2.4 s, the call interrupted, and the notice "Stopped. Claude did not end the turn within 1.5 s of the interrupt, so Modex terminated it." No failure card, and live-stop.txt was never written.

codex-cli 0.163.0 (codex app-server, "Logged in using ChatGPT", gpt-6-astra): refused by the account.

• Codex ended the turn after 3.1 s. Its session record shows codex_error_info: "usage_limit_exceeded" and credits balance: "0".
• Modex showed one failure card: rate_limited, "Codex is rate-limited or out of quota.", with the CLI's words kept: "…try again at Oct 25th, 2026 9:35 AM." Nothing was edited or run.
• Live Codex acceptance, and CF-5's web-search sample, wait for that time, or for credits. Rerun with MODEX_LIVE_BACKENDS=codex.

How it was run

env -u CLAUDECODE -u CLAUDE_CODE_* … MODEX_LIVE_SMOKE=1 MODEX_LIVE_REPORT=… \
  npx playwright test e2e/live-smoke.spec.ts --workers=1

• First run: Claude 1 passed. Codex failed with "the turn did not settle in time", because the first version of the spec waited only for idle, while a refused turn leaves the thread in error. I found the cause in Codex's session record (usage_limit_exceeded).
• Fixed spec, Codex only: 1 failed with "the CLI refused the turn (rate_limited)", which is the correct live outcome while the account has no credits.
• Typecheck passed. With the flag unset, the spec skips 2.

Coven-Harness: Claude Code / claude-opus-5-5

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
modex	Ready	Preview	Oct 19, 2026 11:40pm UTC

[github-code-quality]

Code coverage overview

Languages: TypeScript

TypeScript / code-coverage-agent

main de856e0	cf8-live-smoke 72e2105	+/-
63%	63%	0%

ℹ️ Calculations represent line coverage. Results may change as calculations are updated.