Skip to content

openclaw-core-validation-checklist.md

Latest commit

 

History

History
50 lines (37 loc) · 2.33 KB

openclaw-core-validation-checklist.md

File metadata and controls

50 lines (37 loc) · 2.33 KB

OpenClaw Core Validation Checklist

ExperienceEngine v3 keeps multiple supported hosts, but OpenClaw is the core-learning validation baseline. Use this checklist whenever candidate capture, async distillation, or governance behavior changes.

Preconditions

  • openclaw CLI is installed and the local gateway can load experienceengine
  • ee doctor openclaw reports the plugin wiring as healthy
  • Distillation environment variables are configured if remote LLM distillation should be exercised
  • if OpenClaw loads ExperienceEngine from ~/.openclaw/extensions/experienceengine, confirm the copied plugin tree is not world-writable
  • if ee doctor openclaw reports install_drift = true, run ee repair openclaw before trusting any live-host result

Core Learning Baseline

  1. Run a supported OpenClaw task that produces a successful candidate.
  2. Confirm experience_input_records gets a new record immediately after finalize.
  3. Confirm experience_candidates gets a new row with lifecycle_state = pending.
  4. Confirm distillation_jobs gets a pending job for the same candidate.
  5. Drain the distillation queue or wait for auto-drain.
  6. Confirm the candidate becomes distilled.
  7. Confirm a formal experience_nodes row is created or updated.
  8. Confirm before_prompt_build can retrieve the distilled node on the next similar task.

Failure / Retry Baseline

  1. Force the distiller to fail.
  2. Confirm retry_count increments on both candidate and job.
  3. Confirm the candidate moves to failed before retry exhaustion.
  4. Confirm the candidate moves to discarded once retry budget is exhausted.
  5. Confirm discarded candidates do not appear in normal inspect or intervention surfaces.

Governance Baseline

  1. Run a successful injected follow-up task.
  2. Confirm the injected node increments usage_count and helped_count.
  3. Run a failed injected follow-up task with a relevant failure signature.
  4. Confirm the injected node increments harmed_count.
  5. Confirm node state transitions still behave as expected (active -> cooling -> retired).

Host Reuse Regression

After OpenClaw passes the core baseline:

  • Run ee doctor claude-code
  • Run claude mcp get experienceengine
  • Run ee doctor codex
  • Run codex mcp get experienceengine

These follow-up checks are regressions only. They do not replace the OpenClaw baseline.