submit-experiment-results

name	submit-experiment-results
description	Post material interim assignment updates, then commit and submit a terminal experiment result through the typed Senpai GitHub tools.
argument-hint	<pr-number> <problem-dir>
model	claude-sonnet-4-6
effort	high

Submit experiment results

Interim assignment comments

Use post_assignment_comment when the advisor asked for a response or the experiment is blocked or materially delayed. Use a fresh comment_id for each new message and reuse that ID only for an exact retry. Keep the PR concise rather than mirroring routine internal narration.

First make the repository worktree clean by committing only the assigned change. Collect the current local commit SHA and the current remote assignment-branch SHA. Build the strict ExperimentResult required by submit_experiment_result:

• the assignment repository, PR, assignment ID, revision ID, student, and current expected head SHA;
• terminal status, hypothesis, and bounded summary;
• every W&B run ID, URL, and terminal state;
• the primary metric comparison; and
• the same local commit SHA.

Call submit_experiment_result with exactly this operation-specific payload. Do not add PR-body text, per-run metrics, hyperparameters, or aliases such as head_sha, previous_head_sha, success, or min; put those details in the bounded summary instead.

{
  "branch": "student/experiment",
  "remote_branch_sha_before_push": "REMOTE_SHA_BEFORE_PUSH",
  "result": {
    "assignment": {
      "repo": "owner/repo",
      "pr_number": 123,
      "assignment_id": "assignment-id",
      "revision_id": "LATEST_REVISION_ID",
      "expected_head_sha": "LOCAL_COMMIT_SHA",
      "student": "student-name"
    },
    "status": "succeeded",
    "hypothesis": "The falsifiable hypothesis tested.",
    "summary": "The conclusion, evidence, caveats, and important per-run metrics (maximum 4,000 characters).",
    "runs": 
      
         
         
         
      
    
     
       
       
       
       
       
    
     
  

Use one of succeeded, failed, inconclusive, or cancelled for result status; minimize or maximize for metric direction; and finished, failed, crashed, or killed for each run. primary_metric may be omitted when no finite comparison exists. Refresh the PR first and use its latest assignment revision_id; a mid-turn advisor revision supersedes an earlier one.

That single tool call derives the PR and local result head from result, lease-pushes the clean assignment branch, verifies the new PR head, upserts the authenticated structured result, marks the PR ready, and reconciles status:review. That label is the durable advisor notification. Exact replay of the same result is safe. Once result evidence is published for that revision and commit, changed evidence requires a new commit or assignment revision; the tool will not rewrite the reviewed result in place. Do not run git push, edit labels, write result markers, or call gh pr ready yourself.

If any run is still active or could change the conclusion, keep the assignment in progress and register it with monitor_training; do not submit a terminal result.