[BUG FIX] [MER-5820] Harden DOT GenAI failure handling

[nicocirio]

Jira

https://eliterate.atlassian.net/browse/MER-5820

Summary

Fixes DOT failure paths that could leave students on an infinite loading spinner.

• Reject invalid or undecryptable registered-model credentials before invoking a provider; OpenAI no longer falls back to the global API key.
• Monitor and cancel dialogue tasks, and translate task crashes, stream failures, invalid tool JSON, and tool failures into the existing generic DOT error.
• Add a request-scoped watchdog based on configured model timeouts.
• Correlate dialogue events with an engagement ID so late events from a timed-out request cannot affect the next response.
• Emit safe, actionable failure logs with model/service context and normalized categories (for example, 429 insufficient_quota), without provider bodies, prompts, tool arguments, or credentials.

Failure mode and fix

A bad encrypted credential previously reached provider execution (or silently used the global OpenAI key), while a stalled task could outlive the UI watchdog. The UI could then re-enable submission and receive stale stream events from the old task. This change fails invalid credentials explicitly, bounds the task lifecycle, and ignores stale events.

Video

Video: DOT tested with an OpenAI API key that has no remaining credits. The request fails gracefully, the chat displays the generic error and correctly re-enables input for another message. At the same time, the server logs the failure safely and actionably as 429 insufficient_quota, without exposing credentials or sensitive content.

Screen.Recording.2026-08-11.at.9.52.36.AM.mov

[warning] GenAI request failed service_config_id=1 registered_model_id=4 provider=open_ai request_type=stream http_status=429 error_category=insufficient_quota

Validation

• mix format
• mix compile
• yarn --cwd assets run format
• yarn --cwd assets run check-types
• yarn --cwd assets run deploy
• mix test test/oli/gen_ai/dialogue/server_test.exs test/oli/gen_ai/execution_test.exs test/oli/gen_ai/completions_test.exs test/oli/gen_ai/completions/open_ai_compliant_provider_test.exs test/oli/encrypted/binary_test.exs test/oli_web/live/dialogue/window_live_test.exs (66 tests, 0 failures)

[github-actions]

	Warnings
⚠️	PR is large (892 LOC changed). Consider splitting.

Risk score: 11 → risk/high

Generated by 🚫 dangerJS against a19c2a4

[github-actions]

AI Review — ui

No issues found

[github-actions]

AI Review — performance

No issues found

[github-actions]

AI Review — elixir

No issues found

[github-actions]

AI Review — security

No issues found