Skip to content

pr.md

Latest commit

 

History

History
90 lines (54 loc) · 4.27 KB

pr.md

File metadata and controls

90 lines (54 loc) · 4.27 KB
Raw

Pull Request Summary

Replaces the allowlist-based fetch_docs tool with Tavily MCP, giving the review agent unrestricted web search and URL extraction via OpenAI's server-side MCP dispatch. Switches the agent loop from Chat Completions to the Responses API (required for MCP tool support) and bumps openai to 1.66.0.

Changes Made

  • app/services/review_service.py (rewritten): switched from client.chat.completions.create to client.responses.create; replaced fetch_docs function tool with a Tavily MCP tool entry (type: "mcp"); updated _dispatch_tool() to handle only local function calls; updated input_items accumulation to match Responses API conversation format; updated system prompt to reference tavily-search and tavily-extract
  • app/services/docs_service.py (deleted): functionality superseded by Tavily MCP
  • app/config.py: added required TAVILY_API_KEY: str field
  • .env.example: added TAVILY_API_KEY=tvly-... entry
  • requirements.txt: bumped openai from 1.51.0 to 1.66.0
  • app/main.py: bumped FastAPI version param to 0.2.1

Payload/Response Changes

GET /

Response Changes:

  • Modified: version — "0.2.0" to "0.2.1" (once the hardcoded string in root() is corrected — see breaking changes)

No changes to webhook or review API payloads. The Finding model schema is unchanged; post_review() in github_service.py requires no updates.

Breaking Changes

1. TAVILY_API_KEY is now a required startup field

TAVILY_API_KEY: str has no default in config.py. Any existing deployment that restarts without adding this variable to .env will fail at import time with a pydantic ValidationError — the app will not start. Migration: add TAVILY_API_KEY=tvly-... to .env before restarting.

2. openai 1.51.0 → 1.66.0

The Responses API (client.responses.create) does not exist in 1.51.0. Rolling back the package without also reverting review_service.py will cause an AttributeError at review time. The two changes must be deployed together.

Known Bugs in This PR

Bug 1 — Version string mismatch in app/main.py (line 23)

FastAPI(version="0.2.1") (used by /docs Swagger UI) but the root() handler returns "version": "0.2.0". These should match. The root() return dict needs updating to "0.2.1".

Bug 2 — Responses API assistant turn format (review_service.py line 174)

The conversation continuation after a local function call appends:

input_items.append({"role": "assistant", "content": response.output})

The Responses API expects prior output items to be passed back as top-level input entries (each output item already carries type, id, etc.), not nested inside a role/content wrapper. The correct pattern is:

input_items.extend(response.output)

This bug is silent on non-Python PRs (where run_linter is never called) but will cause a malformed request error on the follow-up API call whenever the linter is dispatched.

New Dependencies

requirements.txt:

  • openai==1.66.0 — upgraded from 1.51.0; required for Responses API and MCP tool type support

.env variables:

  • TAVILY_API_KEY — Tavily API key, embedded in the MCP server URL at startup. Required; no default. Obtain from app.tavily.com.

Removed:

  • docs_service.py deleted; httpx remains in requirements.txt but is now unused — safe to remove in a follow-up cleanup.

API Latency Impact

NA — Tavily searches are resolved server-side by OpenAI before the response returns, so latency characteristics are similar to the previous agent loop. Background task timing is unchanged from the caller's perspective.

Testing Instructions

  1. Open a PR that modifies a Python file with a known ruff violation (e.g. an unused import). Confirm server logs show [linter] ruff returned N finding(s) and the GitHub review comment includes the linter finding.

  2. Open a PR that introduces a new third-party import (e.g. import requests). Confirm logs show a Tavily search occurring and the final review comment reflects documentation-backed analysis.

  3. Start the service without TAVILY_API_KEY in .env. Confirm it fails immediately at startup with a pydantic ValidationError (expected, intentional) rather than failing silently at review time.