# ADR-008: Intrinsics Consolidation
## Status
Accepted
## Date
2024-12-20
## Context
Delia had multiple overlapping systems for pre-execution checks and orchestration decisions:
| System | Purpose | Location |
|--------|---------|----------|
| **Intent Detection** | Classify task type and orchestration mode | `orchestration/intent.py` |
| **Meta-learning** | Decide when to use ToT | `orchestration/meta_learning.py` |
| **Frustration Tracker** | Detect user frustration, escalate | `frustration.py` |
| **Intrinsics Engine** | Answerability, groundedness checks | `orchestration/intrinsics.py` |
| **Profiles** | Specialist model routing | `profiles.py` |
This created:
- Redundant code paths (frustration + intrinsics both trigger escalation)
- Multiple singletons to initialize
- Unclear execution order
- Maintenance burden
## Decision
Consolidate into a simpler pipeline:
```
Before: Intent → Meta-learning → Frustration → Routing → Profiles → Intrinsics → Execute
After: Intent (includes ToT) → Intrinsics (includes user state) → Execute
↓
Profiles (specialist routing)
```
### Changes Made
1. **Frustration → Intrinsics**
- Added `FrustrationLevel` enum to `intrinsics.py`
- Added `UserStateResult` dataclass
- Added `check_user_state()` method (no LLM needed, pure regex)
- Added `ESCALATE_VOTING` and `ESCALATE_DEEP` actions
2. **Meta-learning already in Intent**
- `should_use_tot()` was already called from `intent.py:245`
- No changes needed
3. **Service.py Simplified**
- Removed `FrustrationTracker` dependency
- Added inline `_check_repeat()` for session tracking
- STEP 2-3 now use single `intrinsics.check_user_state()` call
4. **frustration.py Deprecated**
- Added deprecation notice
- No active imports remain
### New IntrinsicsEngine API
```python
from delia.orchestration.intrinsics import (
get_intrinsics_engine,
IntrinsicAction,
FrustrationLevel,
)
intrinsics = get_intrinsics_engine()
# Pre-execution checks (all in one place)
answerability = await intrinsics.check_answerability(task, context)
user_state = intrinsics.check_user_state(message, is_repeat=True, repeat_count=2)
# Post-execution checks
groundedness = await intrinsics.check_groundedness(response, sources)
# Action-based routing
if user_state.action == IntrinsicAction.ESCALATE_DEEP:
intent.orchestration_mode = OrchestrationMode.DEEP_THINKING
elif user_state.action == IntrinsicAction.ESCALATE_VOTING:
intent.orchestration_mode = OrchestrationMode.VOTING
```
### Frustration Detection Logic
```python
# Scoring (from check_user_state):
score = 0.0
if repeat_count > 0: score += repeat_count * 1.5
if has_angry_keywords: score += 3.5
if has_negative_feedback: score += 2.5
# Thresholds:
# score >= 5.0 → HIGH → ESCALATE_DEEP
# score >= 3.0 → MEDIUM → ESCALATE_VOTING
# score >= 1.5 → LOW → ESCALATE_VOTING
# else → NONE → PROCEED
```
## Consequences
### Positive
- Single location for all pre-execution checks
- Clearer mental model: "intrinsics" = sanity checks
- Fewer singletons (removed FrustrationTracker)
- Easier to add new checks (just add method to IntrinsicsEngine)
### Negative
- Session-level repeat tracking now in OrchestrationService (simple dict)
- `frustration.py` still exists (deprecated, for backwards compat)
### Neutral
- Meta-learning remains separate (already integrated into intent)
- Profiles remain separate (orthogonal concern: which model, not whether to escalate)
## Files Changed
| File | Change |
|------|--------|
| `orchestration/intrinsics.py` | Added FrustrationLevel, UserStateResult, check_user_state() |
| `orchestration/service.py` | Removed FrustrationTracker, use intrinsics |
| `frustration.py` | Marked DEPRECATED |
