Reading Buildsworth logs
Buildsworth builds produce pi session artifacts: .jsonl files written by the pi agent that record everything it did. This skill covers how to get them and how to read them without wrecking your context.
Step 1: Download the artifact
Taks a sub-agent with downloading the log. Never download the log in your own context.
Use the using-buildkite skill for the full CLI reference. The short version:
# Find the artifact ID
bk artifacts list <build-number> -p buildkite/buildsworth -o json
# Create ./tmp if it doesn't exist, then download into it. Do not pass -p to bk
mkdir -p ./tmp && cd ./tmp && bk artifacts download <artifact-id>
The downloaded file will be a .jsonl at a path like:
./tmp/artifact-<id>/2026-03-05T00-42-16-011Z_<uuid>.jsonl
The artifact is on the "Review pull request" job. If there are multiple artifacts, look for the one with a path matching root/.pi/agent/sessions/.
Step 2: Always delegate reading to a sub-agent
Never load the .jsonl into your own context. It can be hundreds of kilobytes. Spin up a sub-agent with the file path and a specific question; let it pay the context cost and return only the answer.
Example prompt to a sub-agent:
Read the pi session log at ./tmp/artifact-019cbb75.../2026-03-05T...jsonl. Did the agent find any security issues, and if so what were they?
Keep the question narrow. "What did the agent do?" produces a wall of text. "Were there any failing checks?" produces a sentence.
Step 3: Ask one question per sub-agent
The log is a transcript. One focused question per sub-agent call is more reliable than one open-ended one. Spin up multiple sub-agents in parallel if you have multiple questions.
Good questions:
- What was the agent's final verdict?
- Did it find any security issues?
- Were there any tool call errors?
- What files did it look at?
- Did it leave any comments on the PR?
Avoid "summarise everything" unless the user genuinely asked for a full summary. Even then, a focused set of questions produces a better result than one open-ended one.