17 · Capstone: an agent with tools, traces and limits

A team's CI builds fail a few times a day, and someone has to read the log, work out what broke, and tell the author. The agent here does that first pass. When a build fails on a pull request, it reads the evidence and posts one comment on the PR: which test failed, the likely cause, and the lines to look at.

Its tools, sorted into tiers as in "Limits on what an agent may do": get_build_log, get_diff and search_code only read, so they run freely. rerun_job costs CI time, so it's allowed once per run. post_comment writes, but a comment can be edited or deleted, so it runs without approval and carries an idempotency key. There is no merge, push or revert tool at all.

Its limits: at most 15 steps, 80,000 input tokens across the whole run, and 300 seconds. Its state, kept by your code: the build and PR numbers, the step count, which once-only tools are used, and whether the comment is posted.