MAMA Context - Auto-Injection Skill

Overview

This skill provides automatic background context injection using MAMA's hook system. It runs silently and surfaces relevant past decisions via the UserPromptSubmit hook (~150ms latency with HTTP embedding server).

Philosophy: Gentle hints, not intrusive walls of text. Claude sees topic + time, decides if relevant.

---

How It Works

The skill uses a multi-hook system for comprehensive context injection:

UserPromptSubmit Hook (active, ~150ms — no plugin-side script; handled by Claude Code host)

• Triggers: On every user message submission
• Purpose: Inject relevant decisions as context before Claude responds
• Latency: ~150ms (HTTP embedding server keeps model in memory)
• Token budget: 40 tokens (teaser format)

SessionStart Hook (scripts/sessionstart-hook.js)

• Triggers: Once per session
• Purpose: Initialize MAMA, pre-warm embedding model
• Timeout: 15s

PreToolUse Hook (scripts/pretooluse-hook.js) - disabled (script retained)

• Previously: Injected contracts before Edit/Write operations
• Status: Disabled for performance. Script retained for future use.

PostToolUse Hook (scripts/posttooluse-hook.js) - disabled (script retained)

• Previously: Tracked code changes, suggested decision saves
• Status: Disabled for performance. Script retained for future use.

PreCompact Hook (scripts/precompact-hook.js)

• Triggers: Before context compaction
• Purpose: Preserve unsaved decisions in compacted context
• Timeout: 10s

---

Teaser Format (40 tokens)

💡 MAMA: 2 related
   • authentication_strategy (85%, 3 days ago)
   • mesh_detail (78%, 1 week ago)
   /mama:search <topic> for details

Why teaser?

• Hooks fire on user messages (UserPromptSubmit) → Must be lightweight
• Claude infers relevance from topic + similarity + time
• Full details available via /mama:search if needed
• Avoids token bloat (250 tokens → 40 tokens)

---

Status Transparency

Every injection shows current tier status:

Tier 1 (Full Features):

🔍 System Status: ✅ Full Features Active (Tier 1)
   - Vector Search: ✅ ON (Transformers.js, 3ms latency)
   - Search Quality: HIGH (80% accuracy)

Tier 2 (Degraded):

🔍 System Status: ⚠️ DEGRADED MODE (Tier 2)
   - Vector Search: ❌ OFF (embedding model failed)
   - Search Quality: BASIC (40% accuracy, exact match only)

⚠️ Fix: Check embedding model installation

---

Configuration

Disable Skill:

# Environment variable
export MAMA_DISABLE_HOOKS=true

# Or in config file (~/.mama/config.json)
{
  "disable_hooks": true
}

Adjust Thresholds:

{
  "similarity_threshold": 0.7,
  "token_budget": 40,
  "rate_limit_ms": 1000
}

Note: similarity_threshold: 0.7 applies to explicit searches (/mama:search). The disabled pretooluse-hook.js uses a looser SIMILARITY_THRESHOLD = 0.6 (kept from its original contract-injection design). The active UserPromptSubmit hook delegates search to the MCP server, which uses its own default threshold. Hook thresholds are intentionally separate from the config value shown above.

---

When Claude Should Use This

✅ Automatic (no action needed):

• Context appears when relevant decisions exist
• Claude notices hints and can request details
• User sees transparent status (Tier 1/2)

❌ Don't explicitly invoke this skill:

• It's always-on (background process)
• Hooks handle triggering automatically
• Use /mama:search for explicit lookups

---

Technical Details

Hook Integration:

• UserPromptSubmit: Active (~150ms, no plugin-side script; handled by Claude Code host via MCP)
• SessionStart: scripts/sessionstart-hook.js (initialization)
• PreToolUse: scripts/pretooluse-hook.js (disabled, script retained)
• PostToolUse: scripts/posttooluse-hook.js (disabled, script retained)
• PreCompact: scripts/precompact-hook.js (decision preservation)

Performance:

• Hook latency: ~150ms (HTTP embedding server, model stays in memory)
• Cold start: ~1500ms (embedding model initialization, first session only)
• Warm: ~50ms (HTTP embedding request)
• Timeout: 1800ms (graceful degradation if exceeded)

Search Algorithm:

• Vector search: Transformers.js (3ms embedding)
• Hybrid scoring: 20% recency + 50% importance + 30% semantic
• Graph expansion: Follows supersedes edges
• Recency boost: Gaussian decay (30-day half-life)

---

Acceptance Criteria Mapping

• ✅ AC1: Declared in plugin.json, references hook outputs
• ✅ AC2: Similarity thresholds (70%) + token budgets (40 teaser / 250 full)
• ✅ AC3: Disable via config (MAMA_DISABLE_HOOKS)
• ✅ AC4: Status indicator (Tier 1/2, accuracy, fix instructions)
• ✅ AC5: Smoke test - fires during normal coding session

---

Example Output

User edits a file related to authentication:

Skill injects (via UserPromptSubmit hook):

💡 MAMA: 1 related
   • auth_strategy (90%, 2 days ago)
   /mama:search auth_strategy for full decision

🔍 System Status: ✅ Full Features Active (Tier 1)

Claude sees the hint and can:

1. Ignore (if not relevant)
2. Suggest /mama:search auth_strategy to user
3. Continue with general advice

---

For Developers

Testing:

# Test SessionStart hook (the only testable standalone hook)
node packages/claude-code-plugin/scripts/sessionstart-hook.js

# UserPromptSubmit hook is triggered automatically by Claude Code
# on every user message — no manual invocation needed.

Architecture:

User submits prompt
    ↓
UserPromptSubmit Hook (~150ms, 1800ms timeout)
    ↓
Decision search (generate embedding, search, score)
    ↓
Context injection to Claude (40-token teaser)
    ↓
Claude sees context

---

Key Principles

1. Lightweight: 40 tokens teaser format
2. Transparent: Always show tier status and latency
3. Non-intrusive: Hints, not walls of text
4. Opt-out: User control via config (MAMA_DISABLE_HOOKS)
5. Graceful degradation: Tier 2 fallback if embeddings unavailable
6. Multi-hook system: UserPromptSubmit (active) + SessionStart + PreCompact + PreToolUse/PostToolUse (disabled)

---

Related

• Story M3.2 (this skill)
• Story M2.2 (PreToolUse hook — disabled, scripts retained)
• Story M2.4 (Transparency banner)
• Architecture: docs/MAMA-ARCHITECTURE.md (Decision 4 - Hook Implementation)