ollama-streaming

name	ollama-streaming
description	Guidelines for modifying Ollama streaming and reasoning extraction in src/ollamaStreamParser.ts. Covers NDJSON parsing, <think> tag extraction, and integration with the Copilot Chat participant.

ollama-streaming

OllamaStreamParser parses Ollama /api/chat NDJSON streams and extracts both regular content and reasoning blocks from <think>...</think> tags.

Scope

• src/ollamaStreamParser.ts — the parser.
• src/copilotChatParticipant.ts — participant usage.

Rules

1. NDJSON format. Ollama streams newline-delimited JSON objects, not SSE. Each line is a separate JSON object. The parser splits on \n, not \n\n.
2. Stateful parsing. The parser maintains internal state (buffer, isThinking, thinkBuffer) so it can handle partial lines across multiple parse() calls.
3. tag extraction. Reasoning is extracted from <think>...</think> tags inside message.content. The parser tracks whether it is inside a think block and emits reasoning separately from content.
4. Partial think blocks. If </think> has not arrived yet, accumulate text in thinkBuffer and wait for the next chunk. Do not emit partial reasoning until the block closes.
5. done flag. When json.done === true or line is [DONE], flush any remaining think buffer and emit { done: true }.
6. error handling. If json.error is present, emit { error: string } immediately.
7. No assumptions about field names. Ollama's response shape may vary. Always access json.message?.content as the primary text source.

Interface

export interface OllamaStreamChunk {
  reasoning?: string;
  content?: string;
  done?: boolean;
  error?: string;
}

Common mistakes

• Treating Ollama stream as SSE (splitting on \n\n instead of \n).
• Emitting partial <think> content as content before the closing tag arrives.
• Not handling the case where <think> and </think> appear in the same chunk.
• Forgetting to flush thinkBuffer on stream end, losing the final reasoning block.

Testing

Test the parser by feeding it known NDJSON strings:

const parser = new OllamaStreamParser();
const chunks = parser.parse('{"message":{"content":"<think>Let"}}\n{"message":{"content":" me think</think>Hello"}}\n');
// Expect: [{reasoning:"Let me think"}, {content:"Hello"}]

Add unit tests in a new src/ollamaStreamParser.test.ts if modifying parsing logic.