Fixed
- OpenAI-compatible dispatch with a configured
base_url/api_key_envno longer rejects itself (#659)._validate_openai_compatible_agent_command()now accepts the--base-urland--api-key-envflagsdispatch.shemits for env-configured providers, with strict format validation (a non-empty HTTP(S) host, a safe environment-variable name) so the two functions stay in sync. - Concurrent same-second spawns no longer share a task_id and interleave provider output (#661).
spawn_agent()andspawn_agent_capture_pid()derived an unsuppliedtask_idfromdate +%salone, so two spawns starting in the same second collided and wrote to the same temp files, silently attributing one provider's answer to another's result. Default task_id generation now uses an OS-guaranteed-uniquemktempreservation. - Gemini dispatch now preserves CLI dotenv fallback and custom CA certificates (#660). The isolated provider environment omits empty API-key variables so gemini-cli can load
~/.gemini/.env, while forwarding non-emptyGOOGLE_GEMINI_BASE_URLandNODE_EXTRA_CA_CERTSvalues needed by relays and custom trust chains.