CLI Reference¶

Complete reference for all openclaw learning commands.

Overview¶

openclaw learning <command> [options]

Commands¶

`learning status`¶

Display learning layer summary with posteriors and baseline comparison.

openclaw learning status [--host <host>] [--port <port>]

When the gateway is reachable, data is fetched from the gateway API (live data). If the gateway is unreachable, falls back to the local SQLite database.

Options:

Option	Default	Description
`--host`	`127.0.0.1`	Gateway host (or set `OPENCLAW_GATEWAY_HOST`)
`--port`	`18789`	Gateway port

Output:

Phase badge (PASSIVE or ACTIVE)
Config summary (budget, baseline rate, min pulls)
Trace count, arm count, total tokens, date range
Run distribution (baseline vs. selected, with percentages)
Token savings (percentage reduction vs. baseline)
Top 5 arms (highest posterior mean)
Bottom 5 arms (candidates for exclusion)

Example output:

Learning Layer Status  [PASSIVE]
  Budget: 8,000  |  Baseline: 10%  |  Min pulls: 5

  Traces: 247    Arms: 18    Tokens: 1,284,000    Range: 1/15/2025 – 2/5/2025

Run Distribution
  Baseline: 24 (9.7%)    Selected: 223 (90.3%)
  Token Savings: +12.3% (baseline avg: 5200, selected avg: 4560)

Top Arms (highest posterior mean)
  Arm               Mean     Pulls    Last Updated
  tool:fs:Read      0.923       187    2/5/2025
  tool:exec:Bash    0.891       165    2/5/2025
  tool:fs:Edit      0.845       142    2/4/2025
  tool:fs:Grep      0.812       128    2/4/2025
  tool:fs:Write     0.798       119    2/3/2025

Bottom Arms (candidates for exclusion)
  Arm                          Mean     Pulls    Last Updated
  file:workspace:old-notes.md  0.124       31    2/1/2025
  memory:project:legacy-api    0.187       22    1/30/2025
  skill:debug:verbose          0.234       18    1/29/2025
  file:workspace:scratch.ts    0.267       15    1/28/2025
  memory:project:draft-spec    0.312       12    1/27/2025

`learning reset`¶

Reset all arm posteriors back to uninformative priors Beta(1,1).

openclaw learning reset [--host <host>] [--port <port>] [--confirm]

This is a destructive operation -- all learned data is lost. The bandit starts fresh as if no observations had been recorded. Use this when:

You have significantly changed your tool inventory or skills
Poisoned data has skewed posteriors (e.g., a bug caused incorrect reward signals)
You want to re-run the learning phase from scratch after a major config change

Options:

Option	Default	Description
`--host`	`127.0.0.1`	Gateway host (or set `OPENCLAW_GATEWAY_HOST`)
`--port`	`18789`	Gateway port
`--confirm`	—	Skip the confirmation prompt

Without --confirm, you will be prompted to confirm before the reset proceeds.

Example:

$ openclaw learning reset
? Reset all arm posteriors to Beta(1,1)? (Y/n) Y
Reset 18 arm(s) for learner "openclaw".

# Non-interactive (CI, scripts)
openclaw learning reset --confirm

`learning export`¶

Export traces and posteriors to stdout.

openclaw learning export [options]

Options:

Option	Default	Description
`--format`	`json`	Export format: `json` or `csv`
`--traces`	`true`	Include run traces in export
`--posteriors`	`true`	Include arm posteriors in export

Examples:

# Full JSON export
openclaw learning export --format json

# CSV posteriors only
openclaw learning export --format csv --no-traces

# JSON traces only
openclaw learning export --format json --no-posteriors

# Pipe to file
openclaw learning export --format json > learning-data.json

`learning dashboard`¶

Print the URL of the Learning dashboard.

openclaw learning dashboard [--host <host>] [--port <port>]

The gateway serves the dashboard HTML on-the-fly at /__openclaw__/api/learning/dashboard. No files are written to disk.

Options:

Option	Default	Description
`--host`	`127.0.0.1`	Gateway host (or set `OPENCLAW_GATEWAY_HOST`)
`--port`	`18789`	Gateway port

Example:

$ openclaw learning dashboard
Dashboard: http://localhost:18789/__openclaw__/api/learning/dashboard

See the Dashboard Guide for details on dashboard sections, themes, and troubleshooting.

Global Options¶

Option	Description
`--help`	Show help for command

Exit Codes¶

Code	Description
`0`	Success
`1`	General error
`2`	Invalid arguments

Examples¶

Daily Workflow¶

# Check learning layer status
openclaw learning status

# Open the dashboard for visual analysis
openclaw learning dashboard

# Export data for offline analysis
openclaw learning export --format json > daily-snapshot.json

Monitoring Convergence¶

# Watch posterior means stabilize over time
openclaw learning status

# Check if bottom arms have enough pulls for confident exclusion
# Look for "high" confidence in the posteriors table on the dashboard
openclaw learning dashboard

# When ready, switch to active mode in openclaw.json
# Then monitor token savings
openclaw learning status

Resetting After Config Changes¶

# You've added several new tools and removed old ones.
# Existing posteriors no longer reflect the current inventory.

# Reset all posteriors to Beta(1,1)
openclaw learning reset --confirm

# Verify the reset
openclaw learning status
# Should show 0 traces and all arms at mean 0.500

Remote Gateway¶

# Point at a remote gateway
openclaw learning status --host 10.0.0.5 --port 9999

# Or set the environment variable
export OPENCLAW_GATEWAY_HOST=10.0.0.5
openclaw learning status
openclaw learning dashboard

CLI Reference¶

Overview¶

Commands¶

learning status¶

learning reset¶

learning export¶

learning dashboard¶

Global Options¶

Exit Codes¶

Examples¶

Daily Workflow¶

Monitoring Convergence¶

Resetting After Config Changes¶

Remote Gateway¶

`learning status`¶

`learning reset`¶

`learning export`¶

`learning dashboard`¶