docs: add Ollama streaming NDJSON issue guide and workaround

[rooftop-Owl]

Summary

Resolves #1124 - Ollama integration fails with streaming NDJSON responses

This PR provides documentation and utility code for the Ollama streaming issue.

Problem

When using Ollama with oh-my-opencode agents that use tools, the error occurs:

JSON Parse error: Unexpected EOF

Root Cause: Ollama returns NDJSON when stream: true, but SDK expects single JSON.

Solution

1. Troubleshooting Guide - Explains issue and provides immediate workaround
2. Configuration Documentation - Documents stream: false setting
3. NDJSON Parser Utility - For future SDK integration

Verification

• ✅ Tested with real Ollama API (ministral-3:14b, qwen3-coder)
• ✅ NDJSON parser tested with real streaming data (9 lines)
• ✅ Tool calls verified working with stream=false
• ✅ No JSON parse errors

Files Changed

• docs/troubleshooting/ollama-streaming-issue.md (NEW - 3.4 KB)
• docs/configurations.md (MODIFIED - +60 lines)
• src/shared/ollama-ndjson-parser.ts (NEW - 4.9 KB)

Checklist

• Code follows project conventions
• bun run typecheck passes
• bun run build succeeds
• Tested locally with OpenCode
• Updated documentation
• No version changes in package.json
• Clear commit messages (present tense)
• Issue [Bug]: Ollama integration fails - streaming NDJSON response not parsed correctly #1124 referenced

---

Summary by cubic

Documents and works around Ollama’s streaming NDJSON responses that caused “JSON Parse error: Unexpected EOF” during tool calls. Adds an NDJSON parser utility and documents setting stream: false to restore stability (fixes #1124).

• New Features

  • NDJSON parser utility for Ollama streaming (parseOllamaStreamResponse, isNDJSONResponse).
  • Merges tool_calls and content, skips malformed lines, and collects basic stats.

• Migration

  • For Ollama agents, set "stream": false in config.

^{Written for commit c177f56. Summary will update on new commits.}

[github-actions]

All contributors have signed the CLA. Thank you! ✅
_{Posted by the CLA Assistant Lite bot.}

[cubic-dev-ai]

No issues found across 3 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

[rooftop-Owl]

I have read the CLA Document and I hereby sign the CLA

[code-yeongyu]

LGTM! This is a well-documented fix for the Ollama streaming NDJSON issue.

Review summary:

• Troubleshooting guide is clear and actionable
• NDJSON parser utility is well-tested with real Ollama data
• Configuration documentation properly explains the stream: false workaround
• cubic review passed with no issues

The immediate workaround (stream: false) helps users now, and the parser utility sets up future SDK-level integration.

Approving for merge.

[rooftop-Owl]

Thank you so much for the quick review and merge! 🎉

I'm incredibly grateful for your work on oh-my-opencode. As a graduate student in Korea, this project has been transformative in helping me orchestrate AI models while working within limited subscription budgets.

The Ollama streaming issue (#1124) was blocking my hybrid local/cloud workflow, where I route high-volume tasks to local GPU (Ollama) to preserve cloud tokens for quality-critical work. This fix enables 60-80% cost savings for budget-conscious users like myself.

I hope this contribution helps other researchers and students who are exploring cost-effective AI orchestration with local models.

Thank you for building and maintaining such an incredible tool that democratizes access to advanced AI orchestration! 🙏

P.S. I owe you more than a cup of coffee if we happen to meet in the real world (I also live in Korea 🇰🇷)

[rooftop-Owl]

FYI: Filed a follow-up issue #1508 regarding category delegation model resolution for Ollama.

Your PR #1197 successfully fixed the NDJSON parsing issue - direct agent routing now works perfectly. However, category-based delegation (delegate_task(category='quick')) still fails due to a separate model resolution issue in executor.ts.

Issue: #1508