Examples
June 5, 2026 · View on GitHub
This directory contains sample traces, skill outputs, and worked end-to-end examples. All data is synthetic — no real usernames, tokens, or PII.
Traces
Expected candidate count is what LessonDetector.detect() produces for the
trace. Detection is conservative, so "boring" traces produce zero candidates.
| Trace | Scenario | Expected candidates |
|---|---|---|
traces/github_pr_review_failure.json | Human correction shows diff inspection was skipped. | 1 |
traces/external_chatbot_policy_failure.json | Failed evaluation shows stale policy-version usage. | 1 |
traces/specialist_agent_governance_miss.json | Governance miss corrected by a human reviewer. | 1 |
traces/voice_slot_correction.json | Voice slot correction for time and location. | 1 |
traces/tool_api_fallback.json | API call is rate-limited (429); agent falls back to a cached result. | 1 |
traces/repo_check_finding.json | A repository-check finding plus a human correction. | 1 |
traces/workflow_validation_order.json | Workflow validation order with no candidate signal (the boring case). | 0 |
traces/workflow_validation_failure.json | A workflow step promotes to production before validation, then fails. | 1 |
Detection corpus
detection_corpus/— a labeled corpus of should-detect / should-not-detect traces and a precision/recall/F1 scorecard. Run it withlessonweaver eval-detection examples/detection_corpus/corpus.json.
Skills
skills/github_pr_review_discipline.md— example exported skill markdown.skills/chatbot_policy_version_check.md— example policy-version skill markdown.skills/voice_slot_repair.md— example voice slot-repair skill markdown.
Worked examples
closed_loop_contextweaver/— the closed-loop keystone: a coding-agent failure → reviewed skill card → loaded back into an agent's context for the next run. The flagship "sum > parts" demo.coding_agent_pr_review/— the main end-to-end example: multiple PR-review traces, an approved skill, a validation suite, and the exported instruction fragment.usefulness_report/— count how often one reviewed skill would have been retrieved across a set of traces.
Framework integration examples
Each is dependency-optional and runs the lessonweaver portion even when the framework is not installed.
llamaindex_runtime_loader/— load skills into a LlamaIndex system prompt.openai_agents_runtime_loader/— load skills into OpenAI Agents SDK instructions.
See also docs/integrations/pipecat.md for the
voice-agent pattern using traces/voice_slot_correction.json and
skills/voice_slot_repair.md.