Summary

Fixes #2279

Problem

All Ollama models return (no output) when used through OpenClaw. The TUI and messaging channels show no response even though the models produce valid output via direct API calls.

Root Cause

isReasoningTagProvider() in src/utils/provider-utils.ts returns true for Ollama, which causes:

1. enforceFinalTag: true — stripBlockTags() enforces strict <final> tag extraction. Since Ollama models don't emit <final> tags, all text content is discarded.

2. System prompt injection — <think>/<final> format instructions are injected, which most Ollama models don't follow reliably.

Why Ollama Doesn't Need Tag Enforcement

Ollama's OpenAI-compatible endpoint already handles reasoning natively via the reasoning field in streaming chunks:

{"choices":[{"delta":{"content":"","reasoning":"The user said hello"}}]}
{"choices":[{"delta":{"content":"Hello! How can I help?"}}]}

The pi-ai library correctly maps reasoning → thinking blocks and content → text blocks. Tag-based enforcement is unnecessary and actively harmful.

Fix

Remove "ollama" from the isReasoningTagProvider() check.

Tests

Added comprehensive test suite for isReasoningTagProvider() covering all providers.

Greptile Overview

Greptile Summary

Removes "ollama" from isReasoningTagProvider() in src/utils/provider-utils.ts to fix #2279, where all Ollama models returned (no output). The root cause was that isReasoningTagProvider("ollama") returned true, which triggered enforceFinalTag: true and injected <think>/<final> format instructions. Since Ollama models don't emit <final> tags, stripBlockTags() discarded all text content. Ollama's OpenAI-compatible endpoint already handles reasoning natively via the reasoning field in streaming chunks, making tag-based enforcement both unnecessary and harmful.

• Removed "ollama" from the provider list in isReasoningTagProvider() with a clear explanatory comment
• Added comprehensive test suite for isReasoningTagProvider() covering all provider branches (Ollama, Google variants, Minimax, null/undefined/empty, and standard providers)
• Impact is well-scoped: the function is consumed in 5 files (get-reply-run.ts, compact.ts, attempt.ts, agent-runner-utils.ts, and indirectly via the test), and removing Ollama from the check correctly prevents both enforceFinalTag and reasoning tag system prompt injection for Ollama sessions

Confidence Score: 5/5

• This PR is safe to merge — it's a minimal, well-targeted bug fix with comprehensive tests.
• The change is a single-line removal in a pure function with no side effects beyond its return value. The fix is well-justified: Ollama handles reasoning natively via its reasoning field, making tag enforcement both unnecessary and actively harmful (causing all output to be discarded). The new test file covers all code paths in the function. I traced all 5 downstream consumers and confirmed the change correctly prevents both enforceFinalTag: true and reasoningTagHint: true for Ollama sessions without affecting other providers.
• No files require special attention.

_{Last reviewed commit: 45de9e8}