[WenyuChiou]

Summary

• Preserve portable byte-identical Bash/PowerShell wrapper packaging, stable caller/repository-relative brief resolution, and platform-safe PowerShell paths
• Observe additional edits to already-dirty tracked or untracked files with NUL-delimited status and content fingerprints; include deletion, reversion, rename, and case-distinct paths
• Require HTTP/status context for numeric 429 quota classification, so an unrelated failure at source line 429 remains an error
• Add validated read-only sandbox forwarding while preserving the workspace-write default, legacy model, result schema, fallback contract, and inherited approval policies
• Clarify this repository's bounded compatibility role; no second broker or automatic fallback execution
• Keep proposed metadata 0.1.1 without a release or tag

Official integration comparison

The relevant official integration is openai/codex-plugin-cc v1.0.6, the Claude Code app-server/job integration. It is distinct from the OpenAI Developers documentation/API plugin and from generic native plugin packaging.

Prefer the official plugin as the main Claude Code integration when its runtime fits. Keep this adapter when a real consumer requires its synchronous Bash/PowerShell lane, on-disk brief convention, or legacy result/sentinel contract. Keep supervisor acceptance and agent-collab governance independent of transport.

No matched live quality or total-token-cost advantage is established. Existing delegation ratios used an estimated Claude-inline control, not a native/plugin control. Direct Codex already supplies native JSON events; the official plugin additionally supplies tracked jobs and resume.

Reproducible evidence

• Original exact head c40a0c1: 45 existing tests passed; 28 of 56 newly added adversarial regressions failed before these fixes
• Corrected candidate: 101 passed, 0 skipped locally, including actual Linux PowerShell Core and native command doubles
• Independent review reran the same 99 implementation cases and the original-wrapper control, then independently reviewed/reran the final 101-case suite with Windows fixture guards, verified case-distinct paths in all four mirrors, and replayed the Windows-default cp1252 fixture problem before/after
• Seven common deterministic engine scenarios across direct exec, this wrapper, and the official plugin were run before and after corrections. They establish transport behavior, not model intelligence or actual sandbox security
• Both integration paths accepted a deliberately out-of-scope fixture edit until independent supervisor review. Empty risks/tests fields and an exit-0 delegate process are not acceptance evidence
• A separately reviewed commandExecution/no-fileChange protocol fixture shows that this adapter's Git snapshot can observe an edit absent from the official plugin's touchedFiles. This is bounded event-coverage evidence; it does not claim that real Codex shell writes always omit events
• Official source was pinned to db52e28f, verified against all 63 Git blob identities, and its fixture/process suite passed 91/91

Validation at this head

• Published commit: 957ad60
• Reviewed and published tree: 1b21e821a0057bd1b909c2f1224a17be23e270d8 (production wrappers unchanged from accepted tree 911c7373)
• Incremental patch SHA256: a70a646c7262ddf91f8ea638490302b1ee6667567e454dc0f4a7e47fda1c7e21
• Windows fixture-only follow-up patch SHA256: ecacf0127514239c56084f08f1a71d7ed76bc2fae60ea41f67754fa1f95187b3
• Bash syntax, mirror byte parity, and whitespace checks passed
• Final exact-head Linux/Windows Python 3.10/3.11/3.12 CI passed 6/6 jobs. Each Linux job: 101 passed, 0 skipped. Each Windows job: 95 passed, 6 capability skips for physically invalid quote/angle-bracket filenames and case-distinct pairs on its case-insensitive filesystem
• Prior run 37024240167 failed on the unsupported Windows filename fixture before wrapper invocation; the one-file correction was independently reviewed before the successful new-head run

Limits and release gate

Real CLI and official task/setup calls failed before inference in the available cloud runtime with read-only app-server initialization, despite an existing reported ChatGPT login. Live model quality/cost, account-specific model availability, actual quota recovery, native Claude Code plugin loading, and native sandbox enforcement were not verified. No credentials were inspected, copied, or newly created.

Content/status deltas are best-effort observations, including concurrent writers; they do not prove authorship or semantic correctness. Ignored files, unreadable paths, and submodule contents need an independent candidate-tree audit.

Independent review and final exact-head CI are complete, and root cleared the conditional merge of this exact head. No release, tag, installation, credential/settings change, or deployment is included.