Environment: macOS (Apple Silicon), Traycer Desktop, epic with multiple GUI child agents. Actual: Creating an agent via the agent-creation API (harnessId=hermes, model=opencode-free:muse-spark-1.3-contributor-free) produces an agent whose first turn — and every subsequent turn on that same agent — fails with: API call failed after 3 retries: HTTP 500: Internal server error Zero tool calls run, no partial output. Reproduced 3x in a row on one agent, and identically on a second, separately-created agent before it. Reconfiguring the agent via the API with identical harness/model/profile/permission settings (a no-op API reconfigure) does NOT fix it. Manually opening the same agent in the GUI and reselecting the identical model in the model-picker dropdown DOES fix it — the agent responds normally afterward (confirmed with a test message and a full multi-tool-call task). Impact: Any workflow that creates or configures these agents programmatically (rather than through the GUI picker) gets a silently broken agent that looks created successfully but can never complete a turn, with no error surfaced until the first message fails. Suggestion: Whatever session/auth initialization the GUI model-picker triggers on selection should also run when the same model is set via the create/configure API — right now there's an initialization step that's GUI-only. Repro: 1. Create an agent via the agent-creation API with harnessId=hermes, model=opencode-free:muse-spark-1.3-contributor-free. 2. Send it any first message. 3. Turn fails: "API call failed after 3 retries: HTTP 500: Internal server error", zero tools run. 4. Retry with a fresh message — same failure, repeatable. 5. Open the agent in the GUI, reselect the same model in the dropdown. 6. Agent now responds normally.