Bound OpenAI Codex auth JSON responses

[ooiuuii]

Summary

• add a streamed bounded JSON reader for OpenAI Codex device-code auth responses
• route CLI device-code creation, auth polling, and token exchange through the bounded reader
• route the dashboard Codex full-login worker through the same bounded reader
• preserve existing 429/non-200 status handling while capping successful JSON bodies at 1 MiB
• add regression tests for oversized JSON at all three CLI stages plus the dashboard entry point

Fixes #55253.

Validation

• PYTHONPATH=$PWD C:\Users\Administrator\Documents\Codex\2026-06-29\hermes-main-latest-scan\.venv\Scripts\python.exe -m pytest tests\hermes_cli\test_auth_codex_provider.py tests\hermes_cli\test_web_oauth_dispatch.py -q --basetemp .pytest-tmp-codex-auth-json-web -> 61 passed
• PYTHONPATH=$PWD C:\Users\Administrator\Documents\Codex\2026-06-29\hermes-main-latest-scan\.venv\Scripts\python.exe -m ruff check hermes_cli\auth.py hermes_cli\web_server.py tests\hermes_cli\test_auth_codex_provider.py tests\hermes_cli\test_web_oauth_dispatch.py -> passed
• git diff --check -> passed

Notes

This is separate from #54750, which bounds hermes_cli.copilot_auth. This PR covers the openai-codex device-code flow in both hermes_cli.auth and the dashboard login worker in hermes_cli.web_server.

[tonydwb]

Code Review Summary

Verdict: LGTM

Good security hardening — bounds OpenAI Codex auth JSON response reads to prevent memory exhaustion. The _read_codex_auth_json_response helper with iter_bytes and max_bytes check follows the established pattern. Clean dataclass extraction for the response type.

Security: Limits auth JSON responses to prevent oversized upstream responses.
Code Quality: Well-structured with proper error codes and AuthError wrapping.

---

Reviewed by Hermes Agent

[tonydwb]

Code Review Summary

Verdict: LGTM

Adds streamed bounded JSON reader for OpenAI Codex device-code auth responses. Routes CLI device-code creation, auth polling, and token exchange through the bounded reader.

• Clean streaming pattern with size enforcement
• Well-scoped: 4 files, 256 additions

Reviewed 4 files, 256 additions. Approved.

[teknium1]

Thanks for the focused hardening work. The device-code and dashboard premise still holds on current main: those flows eagerly use post() followed by .json() in hermes_cli/auth.py:7284-7439 and hermes_cli/web_server.py:9277-9345.

Problems

• refresh_codex_oauth_pure() is another Codex OAuth token-endpoint path, but remains unbounded at hermes_cli/auth.py:3450; it parses non-200 and successful JSON at :3489 and :3531.
• The dashboard context has changed since the PR base. Current hermes_cli/web_server.py:9283 passes start failures to _codex_device_code_start_error(), which reads resp.json()/resp.text at :9203-9226. The submitted wrapper retains neither a non-200 payload nor those response methods, so salvage must preserve the current dashboard-specific authorization guidance.

Suggested changes

• Cover the refresh-token response path with the same bounded-reader approach while retaining its 429 and structured-error behavior.
• Integrate the dashboard change with the current bounded error-detail path and retain the existing device-authorization guidance test.

Automated hermes-sweeper review.

[teknium1]

When salvaging this onto current main, preserve the non-200 response detail consumed by _codex_device_code_start_error() (hermes_cli/web_server.py:9203-9248). Current main invokes that helper for this response, while _CodexAuthJsonResponse retains only status, headers, and a 200-only payload.