Codex plugin 0.1.8

Codex plugin 0.1.8 refreshes the model catalog and carries two Claude Code releases' worth of stability work (0.1.6 and 0.1.7) that Codex users have not had yet: a code-health pass over the review, write and state code, and fixes for Windows and macOS.

What's new since 0.1.4

• Model catalog: Claude Fable 5.2 replaces 5.1, Gemini 4 Carbon replaces Gemini 4 Argon, codex gains gpt-6.1-terra, kimi leads with kimi-k3.1-pro, and kilocode adds kimi-k3.1-pro, hy4, nex-n2.6 and Meta's Muse Wonder through OpenRouter. The retired arcee-ai/trinity-large-thinking row is gone, and the free OpenCode default moves from opencode/nemotron-3-ultra-free to opencode/nemotron-3.5-super-vl-free.
• Fewer failed or stuck rounds: odd or malformed model output no longer crashes a round; a damaged checkpoint or status file is set aside and the run continues; --reaggregate records HIGH findings and finishes only whole rounds.
• Safer writes and rollback: every phase that writes files holds its folders until its last write, rollback copies come from a vetted file handle, and a killed run's leftovers are cleaned up. --output-format json prints one JSON document on the apply phases.
• Provider handling: an OpenCode slot that loops on empty steps is stopped, kilocode errors and rate limits are classified from its own error lines, and a plan outside any git repository can run native apply.
• Windows and macOS: abort sends Ctrl+Break only to a console group the plugin made, copies are never written through a link planted at their name, and a moved folder is named on macOS too.
• Host runbooks: Codex now reads the whole runbook in bounded pages it can confirm.

Needs action: the base providers.yaml changed, so re-run --init --force in each project to pick up the new defaults. If you keep your own model lists, replace claude-fable-5-1, gemini-4-argon, kimi-k3, opencode/nemotron-3-ultra-free and arcee-ai/trinity-large-thinking. No config keys were renamed or removed, and the minimum Codex stays 0.159.2. One fallback changed: a provider with no configured row now gets a 1,200 s timeout instead of 600 s.

Upgrade an existing marketplace installation:

codex plugin marketplace upgrade beastlabai
codex plugin add multi-llm@beastlabai

For a new installation, first add the marketplace:

codex plugin marketplace add beastlabai/multi-llm-plugin --ref release \
  --sparse .agents --sparse packaging/codex
codex plugin add multi-llm@beastlabai

Start a new Codex session. Existing project provider configuration stays in place. Tag-pinned users should select codex-v0.1.8.

Verification

Verified source: c36cf4cccb91f3318f6d76c3b7d7d1fff2c88681, the same commit as Claude Code 0.1.8 and OpenCode 0.1.8 (shared-payload digest b7d25d18…).

• All four required PR checks passed on that exact commit: Ubuntu, macOS, Windows and package.
• The full serial suite passed: 36,485 tests.
• The full local Codex gate passed 17 rows on Codex 0.167.0 (pinned), 10 rows on the minimum 0.159.2 and 10 rows on the latest, 0.167.2, with no features diff. The upstream validator is absent; repository layout rules and actual Codex ingestion passed.

Fresh public tag and promoted-release installs passed all six criteria on Codex 0.167.2 in a stock container, with all 327 files matching the tag. A real public-release 0.1.4 → 0.1.8 marketplace upgrade ran the new MCP payload, and a same-version refresh kept its bytes. No paid model requests were used.

Evidence: attached canonical 20261112T153934Z-codex-release-network-check.json (also committed under skills/multi-llm/tests/container/records/) and sanitized codex-public-verification.json. Promotion: successful workflow.