Summary
A single local run exposed a chain of runtime failures that make the project difficult to complete end-to-end when using an OpenAI-compatible third-party provider and Zep Cloud:
- OpenAI-compatible provider requests can be rejected by provider-side request filtering in both direct
OpenAI(...) clients and CAMEL ModelFactory.create(...) simulation clients.
- Zep graph search rejects long generated queries with
query cannot be longer than 400 characters (max).
- Completed simulations can be marked as
stopped during backend reload/shutdown cleanup, which breaks later interaction state.
- Interview endpoints can wait for IPC timeout when
env_status.json is stale but the child simulation process is already gone.
- Initial seed posts can be duplicated into the next simulation round because the action cursor starts before the initial
env.step(...) rows.
- Multiple initial posts assigned to the same agent can collapse to one action in Twitter paths.
- Some malformed JSON requests surface as 500s instead of validation-style 400s.
- BOM-prefixed persisted JSON state can fail to load with
Unexpected UTF-8 BOM.
Related issues and PRs
This overlaps with existing reports but does not appear to be fully covered by one issue/PR:
Reproduction notes
Observed locally on Windows with:
LLM_BASE_URL pointing at an OpenAI-compatible third-party provider- Zep Cloud enabled
- dual-platform simulation enabled
- simulation followed by post-run interview/deep interaction
Representative symptoms:
- Step 1/5 fails with frontend
Request failed with status code 500.
- Backend logs include provider-side OpenAI SDK blocking errors.
- Zep logs include
query cannot be longer than 400 characters (max).
- Interview/deep interaction waits until timeout if the backend has reloaded and only stale
env_status.json remains.
- Initial seed posts appear again in the next logged round.
- Simulation action logs show only the initial rounds when later LLM action calls fail.
Expected behavior
- All configured OpenAI-compatible LLM clients should consistently use the configured key/base URL and provider-safe headers.
- Zep search calls should respect the provider query length limit.
- Completed simulation state should remain completed after backend cleanup/reload.
- Interview APIs should fail fast when the actual simulation process is not alive.
- Initial seed posts should be logged once and not reprocessed as next-round actions.
- Same-agent initial posts should all be submitted.
- Bad JSON requests should return 400 instead of producing 500 tracebacks.
- Persisted state should be readable even if written with a UTF-8 BOM.
Proposed fix
A corresponding PR will harden the runtime paths above without changing the public workflow or UI.
Summary
A single local run exposed a chain of runtime failures that make the project difficult to complete end-to-end when using an OpenAI-compatible third-party provider and Zep Cloud:
OpenAI(...)clients and CAMELModelFactory.create(...)simulation clients.query cannot be longer than 400 characters (max).stoppedduring backend reload/shutdown cleanup, which breaks later interaction state.env_status.jsonis stale but the child simulation process is already gone.env.step(...)rows.Unexpected UTF-8 BOM.Related issues and PRs
This overlaps with existing reports but does not appear to be fully covered by one issue/PR:
Reproduction notes
Observed locally on Windows with:
LLM_BASE_URLpointing at an OpenAI-compatible third-party providerRepresentative symptoms:
Request failed with status code 500.query cannot be longer than 400 characters (max).env_status.jsonremains.Expected behavior
Proposed fix
A corresponding PR will harden the runtime paths above without changing the public workflow or UI.