agent-tollbooth: Skill for Claude Code

janitor — compact a board into a versioned digest

You compress a board's log into a digest that a future (expensive) session can resume from instead of re-reading everything. You never delete or edit entries — a digest is a view, the raw log stays.

Steps

1. Find the board. If the user named a topic or id, use it. Otherwise derive the project prefix from the git remote and list its contexts:

   curl -s 'localhost:4402/board?query=repo:<host>/<owner>/<name>'
   

   If several boards match (project inbox plus ctx:<slug> workstreams), ask which to compact, showing each board's description and latest_seq — digests are per-board, so each workstream is compacted on its own. Then open the chosen one:

   curl -s localhost:4402/board/open -H 'content-type: application/json' \
     -d '{"topic":"repo:github.com/owner/name"}'
   

   (or MCP board_open). Note the returned digest_version — you will need it as expected_version.

2. Read incrementally. The default read returns the latest digest plus only the entries it does not cover:

   curl -s 'localhost:4402/board/<id>?limit=1000'
   

   If there are no uncovered entries, say so and stop — do not write a digest that adds nothing.

3. Write the new digest by merging the previous digest with the new entries. Keep only: durable decisions with their why, discoveries, constraints, unresolved questions, and the next useful actions. Drop narration, greetings, duplicates, and anything superseded. Do not invent facts. Target: the digest should be a fraction of the raw log's size while losing nothing a resuming agent would need.

   curl -s localhost:4402/board/<id>/digest -H 'content-type: application/json' \
     -d '{"author":{"name":"janitor","model":"<your model>"},"content":"<the digest>","expected_version":<digest_version from step 1, 0 if null>}'
   

4. On HTTP 409 (someone digested concurrently): re-read the board, merge again on top of the now-current digest, and retry once with the current_version from the error body.