Onlooker Ecosystem
Composable observability, memory, and quality-gate plugins for Claude Code — all built on the Onlooker event substrate.
The ecosystem is a Claude Code plugin marketplace. Every plugin writes to a shared, schema-validated event log, derives a stable project key from your git remote, and stores artifacts under ~/.onlooker/ — so plugins compose without stepping on each other, and every event is queryable in one place.
Core concepts
- Shared event substrate. The
ecosystem plugin is always-on infrastructure. Every other plugin emits canonical Onlooker events through scripts/lib/onlooker-event.mjs; nothing writes the log directly. Event types follow <plugin>.<noun>.<verb> (e.g. tribunal.gate.blocked, inspector.run.completed).
- Plugins compose, they don't couple. Plugins coordinate only by reading and writing the JSONL event bus — no plugin calls another directly. Every plugin depends on
ecosystem; none depends on another plugin. So bursar can roll up governor.session.complete events without importing governor, and degrade gracefully when it's absent.
- Stable project keys. Artifacts are partitioned by a project key — the first 12 hex chars of
SHA256(remote:<origin-url>), falling back to the repo root for remote-less checkouts — so history stays attached to a project across clones and worktrees.
- One storage root. Runtime artifacts live under
$ONLOOKER_DIR (default $HOME/.onlooker/), namespaced per plugin and project. Plugins fail soft when it's absent — a plugin never blocks a session it wasn't invited to.
- Opt-in by default. Most plugins ship disabled and are enabled per-project or globally via
settings.json. The substrate, and a couple of low-cost reporters, are the exceptions (see the table).
For how these fit together, see docs/architecture.md and the ecosystem-level ADRs.
Plugins
Seventeen plugins, grouped by what they do. Each links to its own README and config.
Substrate
| Plugin | Description | Default |
|---|
ecosystem | Observability substrate: $ONLOOKER_DIR storage, canonical schema-validated event emission, session/tool tracking hooks, and prompt rules. Required by every other plugin. | Always on |
Memory & context
| Plugin | Description | Default |
|---|
archivist | Structured session memory across context truncation. Extracts decisions, dead ends, and open questions on PreCompact; reinjects the most relevant items at the next SessionStart. | Opt-in |
librarian | Consolidation layer between archivist's per-session artifacts and your durable typed memory store. Detects which decisions deserve to persist, classifies them, and queues proposals for explicit confirmation. | Opt-in |
curator | Maintenance layer for the typed auto-memory store. Runs cheap heuristic checks (date-decayed, broken paths, orphaned entries) within a wall-clock budget and points you at /curator review. Never edits memory directly. | Opt-in |
historian | Episodic memory. Chunks and sanitizes the transcript at SessionEnd, then embeds each prompt on UserPromptSubmit to retrieve relevant past context. | Opt-in |
scribe | Intent documentation from agent activity. Captures why changes were made — problem context, decisions, tradeoffs — and distills them into readable artifacts at session end. | Enabled |
cartographer | Proactive auditor of the instruction layer (CLAUDE.md, AGENTS.md, .claude/rules/). Maps relationships and surfaces contradictions, shadowing, gaps, and drift before they cause misbehavior. | Opt-in |
Quality & verification
