Qjudge Ta Protocol

QJudge TA Execution Protocol

Sources of truth

• Read the current problem, contest, answer, and test-case state through QJudge MCP before reasoning about it.
• Use qjudge_code_runner for platform execution. Local reasoning or scratch execution is not platform verification.
• Write platform changes only through the corresponding MCP create/update/grade action.
• Use qjudge-mcp-tool-operator for exact tool routing and payload fields.

Files and artifacts

The DeepAgent backend has two different storage boundaries:

• write_file and edit_file may write temporary scratch content to the run's StateBackend.
• The mounted /app/.deepagents/** tree contains deployed AGENTS.md and skills. Runtime permissions deny writes to it.
• User-visible or resumable deliverables belong in Artifact Service through artifact_write and the specialized artifact tools.

Scratch files are not QJudge platform state and are not durable grading evidence.

Problem and test-case flow

1. Use the matching MCP get action and treat its response as current.
2. Align samples and constraints; use coding-problem-ta-skill for problem-design quality.
3. Draft a reference solution and test cases in scratch or an artifact.
4. Run qjudge_code_runner with the current problem_id, language, code, and cases.
5. Only after successful runner evidence, request/obtain required write approval and call the matching update action.
6. Report exactly what was validated and what was written.

Failure handling

• Correct a recoverable payload/tool error once. Stop after the second failure.
• If routing is unclear, call qjudge_browse(action="get_help", tool_name="<tool_name>").
• If a write is rejected or interrupted at the approval boundary, do not resend it automatically.
• Never claim success from an intended tool call; require its successful result.