feat: Add comprehensive Ollama integration with JSON repair

[t-chamb]

Pull Request: Add comprehensive Ollama integration with JSON repair

Branch Information

• Branch name: feat/ollama-json-repair
• Base branch: main
• Commits: 4 new commits

PR Title

feat: Add comprehensive Ollama integration with JSON repair

PR Description

Summary

This PR adds comprehensive Ollama support to Codex with robust handling of malformed JSON responses. Ollama models can now execute shell commands, apply patches, and use other Codex tools reliably.

Key Features

1. Dynamic JSON Repair System

• Automatically repairs common JSON issues from Ollama:
  • Invalid escape sequences (e.g., \( → \\()
  • Trailing commas
  • Missing quotes around property names
  • Single quotes instead of double quotes
  • Incomplete JSON (missing closing braces)

2. Multi-line JSON Parsing

• Implements brace-counting parser to handle JSON split across multiple lines
• Consistent parsing between streaming response handler and UI layer
• Handles complex nested structures

3. Ollama-Specific Enhancements

• Automatic detection based on baseURL (port 11434)
• Custom system prompts optimized for Ollama's function calling
• Special streaming logic to buffer and process complete responses
• UI filtering to ensure JSON is never visible to users

4. Error Recovery

• ENOTDIR fix for invalid workdir parameters
• Three-level fallback system for JSON parsing
• Graceful degradation when repair fails

Changes

• New file: repair-json.ts - JSON repair utilities
• Updated: responses.ts - Ollama detection and multi-line JSON parsing
• Updated: parsers.ts - JSON repair for tool arguments/output
• Updated: extract-applied-patches.ts - JSON repair for patch extraction
• Updated: agent-loop.ts - Ollama-specific system prompts
• Updated: handle-exec-command.ts - ENOTDIR error handling
• Updated: exec.ts - Shell built-in detection
• Updated: terminal-chat-response-item.tsx - Enhanced UI filtering
• Documentation: Comprehensive OLLAMA_INTEGRATION.md

Testing

Tested with multiple Ollama models including:

• qwen2.5-coder:32b
• deepseek-coder-v2
• llama3.1

All models successfully:

• Execute shell commands with special characters
• Apply patches correctly
• Handle multi-line JSON responses
• Work with malformed JSON output

Usage

# With Ollama provider
codex --provider ollama --model qwen2.5-coder:32b

# With OpenAI-compatible endpoint
codex --provider openai --base-url http://localhost:11434/v1 --model qwen2.5-coder:32b

Test Plan

• Basic function calling works with Ollama
• Malformed JSON is repaired correctly
• Multi-line JSON is parsed properly
• No JSON visible in UI during execution
• ENOTDIR errors are handled gracefully
• Shell built-ins execute correctly
• Patches can be extracted from malformed responses

Files Changed (12 files, +763 -65)

• OLLAMA_INTEGRATION.md (new file, +121)
• codex-cli/src/components/chat/terminal-chat-response-item.tsx (+88 -6)
• codex-cli/src/components/model-overlay.tsx (+22 -0)
• codex-cli/src/utils/agent/agent-loop.ts (+28)
• codex-cli/src/utils/agent/exec.ts (+12)
• codex-cli/src/utils/agent/handle-exec-command.ts (+10 -0)
• codex-cli/src/utils/extract-applied-patches.ts (+11 -0)
• codex-cli/src/utils/parsers.ts (+23 -0)
• codex-cli/src/utils/repair-json.ts (new file, +116)
• codex-cli/src/utils/responses.ts (+351 modifications)
• codex-rs/core/src/chat_completions.rs (+36 modifications)
• run-codex.sh (+10)

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

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

---

Todd Chamberlain seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}

[t-chamb]

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