Commit remaining changes, push the branch, open a pull request, and run Codex review
Full development cycle โ plan a feature, implement it, and open a PR
Process unhandled PR review comments on the current branch's open PR โ answer questions, modify code, or push back
Like /develop, but skips test-writer invocation and post-impl verification (for repos without a test suite โ see ADR-014 ยง12)
Run deep-learning experiments as disciplined hypothesis tests โ frame the question, set up a self-contained per-experiment folder, launch training/eval runs (confirming before heavy GPU jobs), track metrics, analyze results against a baseline, and write a human-readable report with tables and plots. Use this whenever the user is doing experimental ML work: launching or preparing a training/finetuning/sampling/eval run, an ablation, or a hyperparameter sweep; saying things like "let's try X and see if it helps," "does this change improve FID/accuracy/loss," "compare these two runs/checkpoints," "track this experiment," "analyze the results," or "write up what we found." Trigger even when the user doesn't say the word "experiment" but is clearly testing whether a change moves a metric, or wants results organized, compared, or reported reproducibly.
Post-merge cleanup โ update main, prune the remote, and print one paste-ready block of follow-up commands
Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
Run a Codex code review (regular or adversarial) by calling the codex-companion runtime directly, with a Claude Code subagent fallback when Codex is unavailable