Add tool_result pruning utility and split NL/T workflow into two passes #78
Conversation
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the You can disable this status message by setting the ✨ Finishing Touches🧪 Generate unit tests
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
There was a problem hiding this comment.
Greptile Summary
This pull request implements a comprehensive optimization strategy for the Unity MCP (Multi-Client Protocol) system focused on reducing payload sizes and improving conversation management. The primary innovation is the introduction of "lean tool responses" - a systematic approach where Unity MCP tools now return minimal, essential data by default instead of verbose outputs.
The changes include a new prune_tool_results.py utility that condenses tool_result JSON outputs into concise one-line summaries, making conversation logs more manageable and reducing token usage for AI interactions. The script processes common Unity MCP tool responses and generates targeted summaries while preserving essential information.
Multiple Unity MCP tools have been modified to emit trimmed responses: find_in_file now returns only match positions instead of text excerpts, validate_script provides diagnostic counts rather than full details, get_sha returns only hash and byte length, read_console strips timestamps and metadata, and read_resource returns metadata-only unless explicitly requested otherwise. These changes maintain API compatibility while dramatically reducing JSON payload sizes.
The CI workflow has been restructured to split NL/T testing into two separate passes using different Claude AI models (Sonnet for NL, Haiku for T), which keeps each session focused and lightweight. The workflow improvements include better Unity bridge port sharing through status directory mounting, more robust port detection using host-based JSON parsing, and refined tool permissions via .claude/settings.json.
Additional enhancements include structural validation improvements in the script editing system with unbalanced delimiter checking, updated documentation for the new lean response patterns, comprehensive test coverage for all trimmed output functionality, and clearer guidance for late-test editing operations in the automated testing prompts.
These changes integrate seamlessly with the existing Unity MCP bridge architecture while providing significant performance improvements for token-sensitive AI workflows and CI systems.
Important Files Changed
Changed Files
| Filename | Score | Overview |
|---|---|---|
README-DEV.md |
5/5 | Added comprehensive documentation for new tool result pruning utility and lean tool response features |
UnityMcpBridge/Editor/Tools/ManageScript.cs |
4/5 | Added structural balance checking for braces/parentheses before script validation to provide better error messages |
.claude/prompts/nl-unity-suite-full-additive.md |
5/5 | Added late-test editing rules for method body modifications during automated testing |
.github/workflows/claude-nl-suite.yml |
4/5 | Split workflow into two passes (NL/T) with different models and improved Unity bridge port sharing |
tests/test_read_console_trimmed.py |
4/5 | Added test coverage for trimmed console output functionality that removes timestamp fields |
tests/test_get_sha.py |
5/5 | Added assertion to validate pruned response format returning only essential hash and length data |
UnityMcpBridge/UnityMcpServer~/src/tools/resource_tools.py |
4/5 | Modified read_resource and find_in_file tools to return minimal output by default with opt-in for full content |
UnityMcpBridge/UnityMcpServer~/src/tools/manage_script.py |
4/5 | Added output trimming to validate_script and get_sha tools to return condensed counts and metadata only |
UnityMcpBridge/UnityMcpServer~/src/tools/read_console.py |
4/5 | Modified to return trimmed console output with only level and message fields, excluding timestamps |
tests/test_find_in_file_minimal.py |
4/5 | Added test coverage for find_in_file tool's new minimal position-only output format |
tests/test_validate_script_summary.py |
5/5 | Added test coverage for validate_script tool's condensed diagnostic count format |
tests/test_read_resource_minimal.py |
4/5 | Added test coverage for read_resource tool's metadata-only default behavior |
prune_tool_results.py |
4/5 | New utility script for condensing verbose JSON tool outputs into one-line summaries for conversation management |
Confidence score: 4/5
- This PR implements well-designed optimizations with comprehensive testing and documentation, making it relatively safe to merge
- Score reflects the complexity of workflow changes and potential for inter-pass communication issues in the new two-pass CI approach
- Pay close attention to
.github/workflows/claude-nl-suite.ymlfor workflow isolation andUnityMcpBridge/UnityMcpServer~/src/tools/resource_tools.pyfor API compatibility
Sequence Diagram
sequenceDiagram
participant User
participant GitHub as "GitHub Actions"
participant Docker as "Unity Container"
participant Unity as "Unity Editor"
participant Claude as "Claude Agent"
participant MCP as "MCP Server"
participant Tools as "Unity MCP Tools"
User->>GitHub: Trigger NL/T workflow
GitHub->>GitHub: Setup Python/UV environment
GitHub->>GitHub: Install MCP server dependencies
GitHub->>GitHub: Configure Unity licensing
GitHub->>Docker: Start Unity container with MCP bridge
Docker->>Unity: Launch headless Unity Editor
Unity->>Unity: Start MCP bridge on TCP port
Unity->>GitHub: Write status JSON with port info
GitHub->>GitHub: Wait for Unity bridge ready
GitHub->>GitHub: Write MCP config and permissions
Note over GitHub: NL Pass
GitHub->>Claude: Run NL pass with Sonnet model
Claude->>MCP: Connect via stdio transport
MCP->>Unity: Establish TCP connection to bridge
loop NL Tests (NL-0 to NL-4)
Claude->>Tools: Call manage_script/apply_text_edits
Tools->>Unity: Execute script operations
Unity->>Tools: Return results
Tools->>Claude: Return lean response
Claude->>Tools: Call validate_script
Tools->>Unity: Validate syntax
Unity->>Tools: Return diagnostics summary
Tools->>Claude: Return warnings/errors counts
Claude->>Tools: Call read_console
Tools->>Unity: Read console messages
Unity->>Tools: Return trimmed entries
Tools->>Claude: Return level/message only
Claude->>GitHub: Write test result XML fragment
end
Note over GitHub: T Pass
GitHub->>Claude: Run T pass with Haiku model
Claude->>MCP: Connect via stdio transport
MCP->>Unity: Reuse TCP connection to bridge
loop T Tests (T-A to T-J)
Claude->>Tools: Call script_apply_edits (structured)
Tools->>Unity: Execute method operations
Unity->>Tools: Return results with SHA
Tools->>Claude: Return minimal payload
Claude->>Tools: Call get_sha
Tools->>Unity: Get file hash
Unity->>Tools: Return hash/length only
Tools->>Claude: Return SHA without content
Claude->>GitHub: Write test result XML fragment
end
GitHub->>GitHub: Merge XML fragments into JUnit
GitHub->>GitHub: Generate markdown summary
GitHub->>GitHub: Upload artifacts
GitHub->>Docker: Stop Unity container
GitHub->>User: Publish test results
13 files reviewed, 3 comments
| ignore_case: bool | None = True, | ||
| project_root: str | None = None, | ||
| max_results: int | None = 200, | ||
| max_results: int | None = 1, |
There was a problem hiding this comment.
style: Default max_results=1 is very restrictive and may limit usefulness for legitimate search scenarios where multiple matches are expected
| data = obj.get("data", {}) or {} | ||
| msg = obj.get("message") or obj.get("status") or "" | ||
| # Common tool shapes | ||
| if "sha256" in str(data): |
There was a problem hiding this comment.
style: Using str(data) to check for 'sha256' presence is inefficient and could match unrelated strings containing 'sha256'
| if "sha256" in str(data): | |
| if "sha256" in data: |
| elif isinstance(convo, list): | ||
| convo=[prune_message(m) for m in convo] | ||
| json.dump(convo, sys.stdout, ensure_ascii=False) | ||
| main() |
There was a problem hiding this comment.
style: Script calls main() at module level without if __name__ == '__main__': guard, preventing safe imports
| main() | |
| if __name__ == '__main__': | |
| main() |
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
This PR is being reviewed by Cursor Bugbot
Details
You are on the Bugbot Free tier. On this plan, Bugbot will review limited PRs each billing cycle.
To receive Bugbot reviews on all of your PRs, please upgrade to Bugbot Pro by visiting the Cursor dashboard. Your first 14 days will be free!
| if m: | ||
| first = m[0] | ||
| return f"find_in_file: {len(m)} match(es) first@{first.get('line',0)}:{first.get('col',0)}" | ||
| return "find_in_file: 0 matches" |
There was a problem hiding this comment.
Bug: Field Name Mismatch Causes Incorrect Summarization
The find_in_file summarization in prune_tool_results.py looks for line and col fields, but the tool now returns startLine and startCol. This causes the summary to always show first@0:0 instead of the actual match position.
Summary
prune_tool_results.pyto condense tool_result JSON outputs into one-line summariesfind_in_filereturns first match positions only,validate_script&get_shaprovide condensed counts and hashes,read_consoletrims entries, andread_resourcereturns only hash and byte length unless text is requestedallowed_toolsso each pass can access the required Unity edit and console tools.claude/settings.jsonand explicit allow/deny lists to grant only Unity MCP tools and edits underreports/Testing
pytesthttps://chatgpt.com/codex/tasks/task_e_68bcbbf232cc832791102f8596bd5683