fix: warn against using http_request for GitHub PR creation

[langsmith-forge]

Problem

The agent uses the http_request tool to POST to the GitHub PR creation API (/repos/{owner}/{repo}/pulls) instead of using commit_and_open_pr. Because http_request sends no authentication headers, these calls always return 401 Unauthorized, causing PR creation to fail silently.

Traces:

• Agent tries to create PR via http_request — 401 Unauthorized (PR splitting task)
• Agent tries to create PR via http_request — 401 Unauthorized (Release Note PR)
• Agent tries to create PR via http_request — 401 Unauthorized (langgraphjs port)

Root cause

The http_request tool docstring and the TOOL_USAGE_SECTION in prompt.py contained no guidance that this tool cannot be used for GitHub PR operations. Without an explicit prohibition, the agent fell back to http_request when attempting to create PRs — which always fails with 401 because no Authorization header is sent.

Fix

Added an explicit "Do NOT use this tool for GitHub PR operations — use commit_and_open_pr instead" warning to both:

• agent/tools/http_request.py — the function docstring the LLM sees as a tool description
• agent/prompt.py — the TOOL_USAGE_SECTION #### http_request entry in the system prompt

Evidence

No tests written — this is a docstring/prompt wording change. Per project guidelines, brittle string-matching tests for prompt content add maintenance cost with no value.

• CI checks pass locally
• Existing tests pass, no regressions
• Change is minimal and scoped (5 lines added, no logic changed)

[Arsh Verma (ArshVermaGit)]

This is a sharp and practical fix that addresses a real behavioral gap in the agent’s tool selection logic. By explicitly clarifying that http_request must not be used for GitHub PR creation, the change removes ambiguity that was leading the model to consistently choose an unauthenticated path and silently fail with 401 errors. I particularly like that the guidance is reinforced in both the tool docstring and the central TOOL_USAGE_SECTION, ensuring the constraint is visible wherever the model reasons about tool usage. Keeping the change minimal and documentation-focused avoids unnecessary complexity while still materially improving reliability of PR automation flows. Overall, this is a high-leverage prompt alignment improvement that prevents a subtle but costly failure mode.