fix: llm:choose reliability + surface silently swallowed errors - #34
Conversation
… model support - Replace bidirectional contains matching with case-insensitive exact match - Throw clear error on no match instead of silent random fallback - Use chatWithFullResponse() with thinking field fallback for thinking models - Store only original prompt in history, not internal constraint text - Don't mutate history until API call succeeds (ChatReporter, ChatAsyncReporter, ChooseReporter) - Centralize timeout lookup via getTimeoutSeconds helper - Fix hardcoded 30s timeout in ChatReporter and ChatWithTemplateReporter - Add 8 deterministic tests and 2 Ollama integration tests for choose behavior - Remove unused Random import
Replace silent error swallowing with warnings (System.err.println) and cause-chained exceptions across 6 files (18 locations). Users can now debug response parsing failures, connection issues, and invalid config values that were previously hidden. - GeminiProvider: guard empty candidates with promptFeedback context, warn on unparseable parts/finishReason, chain exception causes - ClaudeProvider: warn on malformed thinking/text blocks, chain causes - OllamaProvider: warn on thinking field parse failures, log connection and model list errors, chain causes - OpenAIProvider: chain exception cause in parse failures - LLMExtension: warn on invalid timeout_seconds, log Ollama reachability failures, require template field in YAML templates - ReasoningModelDetector: warn on invalid thinking_budget_tokens
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f95fbf7c71
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| // Add response to history when completed | ||
|
|
||
| // Send with user message included, but don't mutate history yet | ||
| val responseFuture = provider.chat(history.toSeq :+ userMessage).map { responseMessage => |
There was a problem hiding this comment.
llm:chat-async now builds each request from history.toSeq :+ userMessage but defers writing userMessage into history until the future succeeds, so if an agent launches multiple async chats before awaiting results, later requests are sent without earlier pending turns in context. This can change model behavior and produce inconsistent conversation state for concurrent async workflows that rely on per-agent history continuity.
Useful? React with 👍 / 👎.
1 participant
Summary
Closes #21, closes #24
llm:chooseprimitiveSystem.err.println) and cause-chained exceptions across 6 files (18 locations) — users can now debug response parsing failures, connection issues, and invalid config values that were previously hiddenError handling fixes by file:
promptFeedbackcontext, warn on unparseable parts/finishReason, chain exception causestimeout_seconds, log Ollama reachability failures, requiretemplatefield in YAML templatesthinking_budget_tokensTest plan
sbt compilepassessbt test)OLLAMA_TESTS=true sbt "testOnly org.nlogo.extensions.llm.OllamaIntegrationTests")case _: Exception =>without warning remaining in modified codeRuntimeExceptionconstructors chain the causeegetOrElse("")without accompanying warning