learner

name	learner
description	Extract trigger patterns from conversation history and update SKILL.md files. Analyzes recent sessions to find new keywords that should activate existing skills.
triggers	{"keywords":["apprendre les triggers","learner","extract patterns","analyse les conversations"],"exact":["learn"],"patterns":["(learn|extract|analyz).{0,10}(pattern|trigger|keyword)"],"priority":"low"}
metadata	{"author":"Roo Extensions Team","version":"1.0.0","compatibility":{"surfaces":["claude-code"],"restrictions":"Read-only analysis. Changes require manual approval."},"issue":"#1854"}

Skill: Learner — Pattern Extraction from Conversations

Version: 1.0.0 Created: 2026-05-01 Issue: #1854 Phase 3 ADR: docs/harness/adr/006-skill-auto-injection-triggers.md

---

Objective

Analyze recent conversation history to extract trigger patterns — user prompts that should have activated a skill but didn't because the trigger keywords are missing. Suggest additions to existing SKILL.md trigger lists.

Principle: The learner NEVER modifies files directly. It proposes changes for human review.

---

Workflow

Phase 1: Inventory Current Triggers

Read all SKILL.md files and catalog existing triggers:

# For each skill in .claude/skills/*/SKILL.md
# Extract: name, triggers.keywords, triggers.exact, triggers.patterns
# Build a map of skill → known triggers

Output: Table of all skills with their current trigger counts.

Phase 2: Analyze Recent Conversations

Use conversation_browser to scan recent sessions for skill-invocation patterns:

conversation_browser(action: "list", limit: 20, source: "all")

For each session:

1. Check firstUserMessage for natural language that matches a skill intent
2. Identify sessions where a skill was eventually invoked (via /skill-name)
3. Note the original user prompt that preceded the skill invocation

What to look for:

Pattern	Example	Should trigger
French commands	"synchronise le repo"	git-sync
Informal requests	"fais un tour"	sync-tour
Indirect mentions	"le build est cassé"	validate
Partial matches	"status des PRs"	github-status
Missing synonyms	"nettoyer la mémoire"	redistribute-memory

Phase 3: Gap Analysis

Compare user prompts against existing triggers:

1. For each prompt that led to a skill invocation, check if any existing trigger would have matched
2. If no trigger matched → gap found
3. Group gaps by skill
4. Score gaps by frequency

Scoring:

Factor	Weight	Description
Frequency	3.0	How often this prompt pattern appears
Specificity	2.0	How uniquely it maps to one skill
Brevity	1.0	Shorter triggers are better (easier to match)

Phase 4: Generate Proposals

For each gap with score >= 2.0, propose a new trigger:

# Proposed additions for skill "validate"
triggers:
  keywords:
    - "build cassé"      # Found 3 times in sessions #X, #Y, #Z
    - "marche plus"      # Found 2 times in sessions #A, #B

Format per proposal:

### [SKILL-NAME] — N new triggers proposed

| Trigger | Type | Score | Evidence (sessions) |
|---------|------|-------|---------------------|
| "build cassé" | keyword | 5.0 | #session1, #session2, #session3 |
| "marche plus" | keyword | 3.0 | #session3, #session5 |

**Recommendation:** Add as keywords (substring match). No false positives expected.

Phase 5: Present for Approval

Output all proposals as a structured report. The user must approve each addition before any SKILL.md is modified.

## Learner Report — YYYY-MM-DD

**Sessions analyzed:** N
**Gaps found:** M
**Proposals:** K

[Proposals table]

To apply: `/learner --apply` or manually edit SKILL.md files.

---

Rules

Read-Only

• NEVER modify SKILL.md files directly
• NEVER modify any project file
• Output is a report only

Evidence-Based

• Every proposed trigger MUST cite at least 1 conversation session as evidence
• Triggers without evidence are REJECTED

No Speculation

• Only analyze REAL conversations from the last 30 days
• Do not invent hypothetical triggers
• Do not add triggers for skills that already have 10+ keywords (diminishing returns)

Deduplication

• Before proposing, check if the trigger already exists (case-insensitive)
• Before proposing, check if an existing trigger already covers the pattern (substring check)

---

Scope Limits

Metric	Limit
Sessions analyzed	20 max
Proposals per skill	3 max
Total proposals	10 max
Minimum evidence sessions	1
Maximum trigger length	30 chars

---

Integration with Trigger Detector

Proposed triggers follow the same format as scripts/claude/skill-trigger-detector.ps1 expects. Once approved and added to SKILL.md, they are automatically picked up by the detector on the next prompt.

---

Invocation

# Analyze recent conversations
/learner

# Apply approved proposals (requires confirmation)
# (Future: not implemented in v1.0)

---

Last updated: 2026-05-01