0.5.1: re-verified against claude 2.1.267, codex 0.154, copilot 1.0.88, grok 1.0.40 - #38
Merged
Merged
Conversation
added 5 commits
September 23, 2026 22:14
The notification is { error: { message }, threadId, turnId, willRetry }. The
crate read params.message and always got nothing, and it ended the turn as an
error even when codex said it would retry, failing runs that went on to
succeed. The message is read from error.message, and a retried error is left
alone.
--output-format json has no init record, so a run's model name is often unknown, and claude lists its Haiku helper beside the run's model when the helper ran. With two entries and no name the [1m] window came back unknown, which the live suite caught on 2.1.267. The top-level usage is the run's own model: the entry whose four counts equal it is chosen, only when exactly one does, and anything else is still reported as unknown.
codex debug models on 0.154.0 lists gpt-6-astra first (the new default), drops gpt-5.4 and gpt-5.4-mini, adds the account-dependent gpt-daybreak-blue-latest, and marks gpt-5.5 as retiring on 2026-10-14. grok models on 1.0.40 lists grok-4.7 (the new default) and grok-4.7-build-fast ahead of 4.6 and 4.5. Grok's effort list stays as verified: none and max appear only as a string in the binary.
Checked against claude 2.1.267, codex-cli 0.154.0, Copilot CLI 1.0.88 and grok 1.0.40, flag by flag against --help, then with the live suite. - codex exec resume refuses --add-dir, so a resumed run with extra directories failed before the agent started. It now carries them as -c sandbox_workspace_write.writable_roots=[...], TOML-quoted; a fresh exec still takes the flag. - Copilot documents --reasoning-effort in 1.0.88; --effort still parses but is gone from --help, so the documented flag is passed. - Claude's login hint is `claude auth login`, the subcommand 2.1.267 documents, and the quoted stream-json error matches the new wording. - The comment claiming codex fork is TUI-only is corrected: `codex exec fork` exists, but this crate does not drive it, so fork stays Unsupported. - verified_version moves to the four versions checked.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Releases 0.5.1. The crate is re-verified against the installed CLIs: claude 2.1.267, codex-cli 0.154.0, GitHub Copilot CLI 1.0.88 and grok 1.0.40. Every flag was checked against each CLI's
--help, and then the live suite was run once against all four.Fixed
codex exec resumerefuses--add-dir(unexpected argument '--add-dir'). A resumed run now carries the directories as-c sandbox_workspace_write.writable_roots=[...]; a freshexecstill uses the flag. Has a unit test.{ error: { message }, willRetry }. The crate read a top-levelmessageand failed the turn even when codex said it would retry. Has a unit test.[1m]run sometimes reported no window. This was caught by the live suite on 2.1.267.--output-format jsonhas noinitrecord, so when claude's Haiku helper ran there were twomodelUsageentries and no name to choose between them. The crate now picks the entry whose counts equal the top-levelusage, and only when exactly one does. Has a unit test.--reasoning-effort, the flag 1.0.88 documents.--effortstill parses but is gone from--help.claude auth login.Changed
codex debug models:gpt-6-astra(the new default), sol, terra, luna,gpt-daybreak-blue-latest(depends on the account), andgpt-5.5(retiring 2026-10-14).gpt-5.4andgpt-5.4-miniare removed.grok models:grok-4.7(the new default),grok-4.7-build-fast, 4.6 and 4.5.Verification
a_1m_alias_reports_the_widened_windowpassed after the fix above.codex_interrupts_an_orphaned_active_turnandcodex_reads_a_long_idle_session_without_starting_a_turnneed prepared fixtures (AGENT_ABSTRACTION_INTERRUPT_SESSIONandAGENT_ABSTRACTION_IDLE_SESSION). They were not set, so neither test exercised any code.clippy -D warningsis clean.Not changed yet (each needs a targeted live probe first)
CLAUDE_CODE_OAUTH_TOKENlooks like the variable thatsetup-tokencredentials use, andEnvPolicy::Minimaldrops it.rate_limit_eventnow declares an optionalutilizationfield.--system-prompt-snapshot(on by default) may make--resumeignore a changed--append-system-prompt, which would be a silent downgrade.codex exec forknow exists headless, and the crate could support fork. For now it staysUnsupported, with the comment corrected.noneandmax, but they were not claimed without a live check.