Skip to main content
OpenClaw is the only currently supported agent framework. Integration is zero-config: Critiqor ships a bundled plugin that auto-loads into OpenClaw, captures runtime events from startup to teardown, and writes structured evidence locally — all without any changes to your agent code.

​
How It Works

Critiqor ships a bundled OpenClaw plugin at critiqor/clawhub/critiqor-openclaw/. When you run critiqor monitor openclaw, Critiqor sets the OPENCLAW_BUNDLED_PLUGINS_DIR environment variable, which tells OpenClaw to automatically discover and load the plugin before launching. The plugin:
  • Registers itself as "critiqor" (id: critiqor, name: Critiqor Evidence Collector)
  • Activates immediately when OpenClaw loads it — no manual activation required
  • Observes all runtime events passively; it does not score runs, generate diagnoses, or alter agent behaviour
  • Writes a structured session.json evidence file to your local runs/ directory
You do not need to change your agent code, configure environment variables, or install anything beyond Critiqor itself.

​
Prerequisites

  • Critiqor: pip install critiqor
  • OpenClaw: installed and available on your PATH, compatible with plugin API >=2026.6.1
  • Python: 3.10 or newer

​
Setup

The complete workflow from install to results: 
# 1. Install Critiqor
pip install critiqor

# 2. Verify OpenClaw is available
openclaw --version

# 3. Start a monitored session
#    Critiqor initialises the session, enables the plugin, and launches OpenClaw
critiqor monitor openclaw

# 4. Use OpenClaw as normal — Critiqor observes silently
#    (interact with the OpenClaw TUI as you normally would)

# 5. Exit the OpenClaw TUI, then finalize the session
critiqor finalize
critiqor monitor openclaw creates the run session and launches the OpenClaw TUI in the same terminal. critiqor finalize stops the observer, generates diagnosis.json, and opens the local dashboard automatically.

​
What the Plugin Collects

The plugin collects runtime evidence across two layers:
LayerEvents
Extension API (api.on(...))agent_start, agent_end, turn_start, turn_end, session_start, session_end, before_provider_request, after_provider_response, message_received, message_sent, message_start, message_update, message_end, input, user_bash
Tool Hookstool_call, tool_result, tool_execution_start, tool_execution_update, tool_execution_end
Memory tool calls (memory_search, memory_get) are additionally re-emitted as a memory_event type entry so that memory integrity analysis can run as a dedicated diagnostic pass. Each collected event is normalised and appended to runs/<run_id>/session.json. For tool hook events, the plugin records tool_name, tool_call_id, error status, and wall-clock duration_ms. Older OpenClaw builds that do not expose a specific event type will gracefully skip that subscription — the collector is forward- and backward-compatible.

​
Plugin Configuration

The plugin has no user-configurable settings. All runtime configuration is supplied via environment variables set automatically by critiqor monitor openclaw:
VariablePurpose
CRITIQOR_RUN_IDThe identifier for the current run (e.g. run_001)
CRITIQOR_RUNS_DIRThe directory where session.json is written
You should not need to set these manually.

​
Advanced: Custom OpenClaw Command

If your OpenClaw binary is not named openclaw, or you need to pass additional arguments when launching, use the --openclaw-command flag: 
critiqor monitor openclaw --openclaw-command "openclaw chat --model gpt-4"
The CRITIQOR_RUN_ID environment variable is injected into the child process environment regardless of the command used, so the plugin will always write to the correct session.

​
Compatibility

  • Plugin API: >=2026.6.1 (as specified in the plugin’s package.json compat field)
  • Build target: OpenClaw 2026.6.1
  • Older OpenClaw builds that do not expose specific event types will skip those subscriptions silently — no errors, no data loss for supported events