autocontext-consumer

Autocontext: Using Existing Knowledge

Overview

Autocontext writes what it learns to a knowledge directory. This skill covers reading and moving that knowledge. To produce it, use autocontext-creator.

Nothing here assumes a particular agent host, and most of it is plain file reading - the layout is documented below precisely so an agent can go straight to the file it wants.

When to Use

• You want to know whether Autocontext has learned anything about a task.
• You want the current playbook or lessons for a scenario.
• You want to move knowledge from one checkout or machine to another.

Do not use this skill to run scenarios or judge output. That is autocontext-creator.

Where Knowledge Lives

The root defaults to ./knowledge and moves with AUTOCONTEXT_KNOWLEDGE_ROOT. Inside it, each scenario owns a directory:

<knowledge_root>/
  <scenario>/
    playbook.md              the current approach, rewritten as the loop learns
    lessons.json             accumulated lessons, newest last
    hints.md                 hints carried into the next attempt
    mutation_log.jsonl       one line per change, append-only
    package_metadata.json    present once the scenario has been exported
    reports/<run_id>.md      per-run written reports
  analytics/                 cross-scenario analytics
  _hub/                      shared research hub state
  _evaluator_epochs/         evaluator versioning

Directories starting with _ are shared across scenarios rather than owned by one. playbook.md is the file to read first: it is the current answer, where lessons.json is the history of how it got there.

Reading Knowledge

The playbook and lessons are plain files. Read them directly:

cat "${AUTOCONTEXT_KNOWLEDGE_ROOT:-knowledge}/grid_ctf/playbook.md"

An absent file means nothing has been learned for that scenario yet. That is a normal state, not an error.

Finding Runs

autoctx list --json
autoctx status "$RUN_ID" --json
autoctx show "$RUN_ID"