Context
During a recent review of agent behavior, it was observed that agents were bypassing the pr-review skill (and thus the AGENTS.md guidelines, depth floor, and graph-ingestion tags) when conducting PR reviews. Because the manage_pr_review MCP tool is an atomic API call that did not mandate the skill as a pre-step, agents succumbed to "velocity-bias," preferring to execute the tool immediately rather than adhering to architectural discipline.
The Problem
Without a strict prompt-time guard in the tool description, the LLM naturally bypasses reading .agents/skills/pr-review/SKILL.md, relying entirely on pre-trained behavior and local context to review the PR. This causes a global failure in the review protocol, resulting in unstructured feedback that breaks Native Edge Graph ingestion and fails to meet the formal Depth Floor.
The Architectural Reality
The manage_pr_review endpoint in ai/mcp/server/github-workflow/openapi.yaml (line 637) defines the prompt block exposed to the agent. While tools like create_issue have an explicit **MANDATORY pre-step:** guard, manage_pr_review lacked this protection.
The Fix
Update the openapi.yaml description for manage_pr_review to explicitly mandate reading .agents/skills/pr-review/SKILL.md before execution.
Acceptance Criteria
Origin Session ID
fc8abc96-ce88-407e-9d92-7fcc494f8236
Context During a recent review of agent behavior, it was observed that agents were bypassing the
pr-reviewskill (and thus theAGENTS.mdguidelines, depth floor, and graph-ingestion tags) when conducting PR reviews. Because themanage_pr_reviewMCP tool is an atomic API call that did not mandate the skill as a pre-step, agents succumbed to "velocity-bias," preferring to execute the tool immediately rather than adhering to architectural discipline.The Problem Without a strict prompt-time guard in the tool description, the LLM naturally bypasses reading
.agents/skills/pr-review/SKILL.md, relying entirely on pre-trained behavior and local context to review the PR. This causes a global failure in the review protocol, resulting in unstructured feedback that breaks Native Edge Graph ingestion and fails to meet the formal Depth Floor.The Architectural Reality The
manage_pr_reviewendpoint inai/mcp/server/github-workflow/openapi.yaml(line 637) defines the prompt block exposed to the agent. While tools likecreate_issuehave an explicit**MANDATORY pre-step:**guard,manage_pr_reviewlacked this protection.The Fix Update the
openapi.yamldescription formanage_pr_reviewto explicitly mandate reading.agents/skills/pr-review/SKILL.mdbefore execution.Acceptance Criteria
openapi.yamldescription formanage_pr_reviewincludes a**MANDATORY pre-step:**block referencing thepr-reviewskill.Origin Session ID
fc8abc96-ce88-407e-9d92-7fcc494f8236