How do I install the Expel Lesson Extractor skill?

Run npx skillmds@latest add dreamlab-ai/expel-lesson-extractor in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Expel Lesson Extractor safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Expel Lesson Extractor?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Expel Lesson Extractor free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Expel Lesson Extractor?

dreamlab-ai (@dreamlab-ai) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Expel Lesson Extractor by dreamlab-ai

Name: Expel Lesson Extractor
Author: dreamlab-ai

Expel Lesson Extractor

Fires as a post-task hook after a non-trivial completed task (3+ tool calls, observable terminal outcome) to distil 0-N generalisable IF/THEN lessons from the trajectory and store them in RuVector so they surface at the start of similar future tasks; also invocable manually to record a lesson after a complex or failed task. NOT for trivial (<3 tool calls) or interrupted tasks, and NOT a general memory-write tool — it writes only ex:DistilledLesson records to the code-harness-lessons namespace. Schema, gates, prompt, and contradiction/archive policy live in references/.

dreamlab-ai 63c0ce8 6 files · 15.9 KB Updated Sep 17, 2026

File contents

ExpeL Lesson Extractor

Status: Phase 1 — active (ExpeL, build cost S, no kernel dependency for lesson storage; kernel improves evidence quality but is not required).

See: ADR-019 §Mechanism 1, PRD-008 §3.4 / §7 Track B, DDD-005 §DistilledLesson. Multi-tier memory table: docs/developer/code-harness-multi-tier-memory.md.

When to Use

Invoke automatically via claude-flow hooks post-task when [features.expel_lesson_extraction].enabled = true. Do not invoke manually for trivial one-liner tasks. The minimum threshold is tasks with 3 or more tool calls — below this the trajectory contains too little signal for generalisation (PRD-008 §7 C4).

May also be invoked manually after a complex or failed task where the agent wishes to record a specific lesson explicitly, bypassing the automatic invocation threshold.

When NOT to Use

Tasks with fewer than 3 tool calls (trivial-task filter; no lesson stored).
Tasks that were interrupted mid-trajectory (ExpeL distillation only runs on tasks with an observable terminal outcome — success or explicit failure).
Contexts where [features.expel_lesson_extraction].enabled = false.
When the privacy filter (PrivacyFilterPort, ADR-008) is unreachable — the lesson is dropped, not written without redaction. Fail-closed is the contract. Emit LessonRedactionFailed event and return.

Mechanism at a glance

Post-task, the runtime calls the expel-distil binary with the trajectory. It privacy-filters the evidence, runs a templated extraction prompt to emit 0-N IF/THEN rules, validates them against a write gate, and writes surviving lessons to the code-harness-lessons RuVector namespace as ex:DistilledLesson records. At the start of similar future tasks the lessons are retrieved semantically and injected as a "Prior experience:" block. Contradictions decay confidence over time and demote stale lessons to an archive namespace.

All RuVector access is via mcp__claude-flow__memory_* only — never raw SQL, never the claude-flow memory * CLI (both bypass the bge-small-en-v1.5 (384-dim, via Xinference) embedding pipeline and are invisible to HNSW search, ADR-015).

Reference index (load on demand)

Topic	File
OWL2 classification, record schema, field definitions, write call pattern	references/record-schema.md
Extraction prompt template and inputs	references/extraction-prompt.md
Write-gate steps and task-start retrieval/injection	references/write-gate-and-retrieval.md
Contradiction detection, conflict ranking, archive policy	references/contradiction-and-archive.md
Manifest gates, validator rules, hook registration, implementation notes	references/manifest-and-hooks.md

Related Files

expel-distil (crate services/agentbox-ops) — post-task handler implementation.
skills/voyager-skill-library/SKILL.md — Phase 2 verified skill library.
skills/agentdb-memory-patterns/SKILL.md — memory_type discriminator details.
ontology/code-harness.ttl — OWL2 class declarations.
docs/developer/code-harness-multi-tier-memory.md — namespace / class table.
docs/archive/adr/ADR-019-experiential-skill-learning.md — canonical decision.
docs/archive/prd/PRD-008-code-as-harness-integration.md §3.4 / §7 Track B.
docs/archive/ddd/DDD-005-code-execution-domain.md §DistilledLesson aggregate.
tests/code-harness/lesson-retrieval-queries.json — C3 acceptance test fixture.

---name: expel-lesson-extractordescription: Fires as a post-task hook after a non-trivial completed task (3+ tool calls, observable terminal outcome) to distil 0-N generalisable IF/THEN lessons from the trajectory and store them in RuVector so they surface at the start of similar future tasks; also invocable manually to record a lesson after a complex or failed task. NOT for trivial (<3 tool calls) or interrupted tasks, and NOT a general memory-write tool — it writes only ex:DistilledLesson records to the code-harness-lessons namespace. Schema, gates, prompt, and contradiction/archive policy live in references/.---# ExpeL Lesson Extractor**Status: Phase 1 — active (ExpeL, build cost S, no kernel dependency for lessonstorage; kernel improves evidence quality but is not required).**See: ADR-019 §Mechanism 1, PRD-008 §3.4 / §7 Track B, DDD-005 §DistilledLesson.Multi-tier memory table: `docs/developer/code-harness-multi-tier-memory.md`.---## When to UseInvoke automatically via `claude-flow hooks post-task` when`[features.expel_lesson_extraction].enabled = true`. Do not invoke manuallyfor trivial one-liner tasks. The minimum threshold is tasks with **3 or moretool calls** — below this the trajectory contains too little signal forgeneralisation (PRD-008 §7 C4).May also be invoked manually after a complex or failed task where the agentwishes to record a specific lesson explicitly, bypassing the automaticinvocation threshold.## When NOT to Use- Tasks with fewer than 3 tool calls (trivial-task filter; no lesson stored).- Tasks that were interrupted mid-trajectory (ExpeL distillation only runs on tasks with an observable terminal outcome — success or explicit failure).- Contexts where `[features.expel_lesson_extraction].enabled = false`.- When the privacy filter (PrivacyFilterPort, ADR-008) is unreachable — the lesson is **dropped**, not written without redaction. Fail-closed is the contract. Emit `LessonRedactionFailed` event and return.---## Mechanism at a glancePost-task, the runtime calls the `expel-distil` binary with the trajectory. Itprivacy-filters the evidence, runs a templated extraction prompt to emit 0-NIF/THEN rules, validates them against a write gate, and writes survivinglessons to the `code-harness-lessons` RuVector namespace as`ex:DistilledLesson` records. At the start of similar future tasks the lessonsare retrieved semantically and injected as a "Prior experience:" block.Contradictions decay confidence over time and demote stale lessons to anarchive namespace.All RuVector access is via `mcp__claude-flow__memory_*` only — never raw SQL,never the `claude-flow memory *` CLI (both bypass the bge-small-en-v1.5(384-dim, via Xinference) embedding pipeline and are invisible to HNSW search,ADR-015).---## Reference index (load on demand)| Topic | File ||---|---|| OWL2 classification, record schema, field definitions, write call pattern | [references/record-schema.md](references/record-schema.md) || Extraction prompt template and inputs | [references/extraction-prompt.md](references/extraction-prompt.md) || Write-gate steps and task-start retrieval/injection | [references/write-gate-and-retrieval.md](references/write-gate-and-retrieval.md) || Contradiction detection, conflict ranking, archive policy | [references/contradiction-and-archive.md](references/contradiction-and-archive.md) || Manifest gates, validator rules, hook registration, implementation notes | [references/manifest-and-hooks.md](references/manifest-and-hooks.md) |---## Related Files- `expel-distil` (crate `services/agentbox-ops`) — post-task handler implementation.- `skills/voyager-skill-library/SKILL.md` — Phase 2 verified skill library.- `skills/agentdb-memory-patterns/SKILL.md` — memory_type discriminator details.- `ontology/code-harness.ttl` — OWL2 class declarations.- `docs/developer/code-harness-multi-tier-memory.md` — namespace / class table.- `docs/archive/adr/ADR-019-experiential-skill-learning.md` — canonical decision.- `docs/archive/prd/PRD-008-code-as-harness-integration.md` §3.4 / §7 Track B.- `docs/archive/ddd/DDD-005-code-execution-domain.md` §DistilledLesson aggregate.- `tests/code-harness/lesson-retrieval-queries.json` — C3 acceptance test fixture.

Frequently asked questions