| name | knowledge-compiler |
| description | Install a per-repo, self-building knowledge base into the current repository. Captures Claude Code sessions into daily logs and compiles them into a queryable knowledge/ wiki (concepts + connections + index), injected back at session start. Trigger when the user says "knowledge compiler", "install knowledge compiler", "set up a knowledge base", "knowledge compiler einrichten", "distil my sessions into knowledge", or wants per-repo memory that builds itself and can seed from an existing repo. |
Knowledge Compiler — Install Skill
Installs the knowledge-compiler engine into the current repo: SessionEnd/PreCompact
hooks capture each session into <kdir>/daily/, a compiler distils those into
<kdir>/knowledge/ (concepts, connections, index), and SessionStart injects the
index back. Compilation runs on a manual command or a 6-hour SessionStart gate.
Knowledge always lives inside the repo, never under .claude/.
Authentication
The compiler uses the Claude Agent SDK, which needs ANTHROPIC_API_KEY (or
CLAUDE_CODE_OAUTH_TOKEN) in the environment. Subscription credentials are NOT
sanctioned for third-party plugins — public/customer installs must set an API key.
Capture (hooks/scaffold) works without it; only compile/query/seed make API calls.
Naming note
Invoke this as neurawork-cc-harness:knowledge-compiler (the fully-qualified name)
so it always resolves to this plugin regardless of what else is enabled.
Phase A — Recon (read-only)
Run the recon script and read its RECON_JSON blob:
python3 "${CLAUDE_PLUGIN_ROOT}/engines/knowledge-compiler/recon.py"
status: NOT_A_GIT_REPO → stop; tell the user to run inside a git repo (offer
git init).existing_kdir set → this is an ADOPT (refresh) install; reuse that name.- Note
seed_recommended, timezone, and clean.
Phase B — Ask
Use AskUserQuestion to confirm:
- Knowledge dir name — default
knowledge-base (or the detected existing_kdir).
- Timezone — confirm the detected one (used only for display; no hardcoding).
- Seed now? — offer only when
seed_recommended is true AND the tree is clean.
Seeding builds initial articles from the repo's README/docs and makes API calls.
Phase C — Execute
python3 "${CLAUDE_PLUGIN_ROOT}/engines/knowledge-compiler/install.py" \
--knowledge-dir <NAME> [--seed]
uv sync --directory <NAME>
Then tell the user to commit <NAME>/ and .claude/settings.json. After install:
- Sessions now capture automatically.
- Manual compile:
/neurawork-cc-harness:kc-compile.
- Query:
uv run --directory <NAME> python scripts/query.py "...".
