chati-dev 4.3.1 → 4.4.1

package/framework/intelligence/context-engine.md

@@ -33,7 +33,7 @@ The orchestrator injects context through 6 hierarchical layers:
 
 | Layer | Name | Source | When Active |
 |-------|------|--------|-------------|
-| **L0** | Constitution | `chati.dev/constitution.md` (Articles I-XXIII) | ALWAYS (non-negotiable) |
+| **L0** | Constitution | `chati.dev/constitution.md` (Articles I-XXV) | ALWAYS (non-negotiable) |
 | **L1** | Mode + Global | `config.yaml` + mode governance (planning/build/deploy) | ALWAYS |
 | **L2** | Agent Scope | `chati.dev/agents/{agent}/` — mission, inputs, outputs, criteria | When agent is active |
 | **L3** | Pipeline State | `.chati/session.yaml` — pipeline position, scores, backlog | When session is active |
@@ -42,7 +42,7 @@ The orchestrator injects context through 6 hierarchical layers:
 
 ### Layer Activation by Bracket
 
-| Bracket | Active Layers | Budget Ratio | Claude (200K) | Gemini (1M) | Codex (128K) |
+| Bracket | Active Layers | Budget Ratio | Claude (1M) | Gemini (1M) | Codex (128K) |
 |---------|--------------|-------------|---------------|-------------|--------------|
 | FRESH | L0, L1 | 1.5% | 3,000 | 15,000 | 1,920 |
 | MODERATE | L0, L1, L2, L3 | 2.5% | 5,000 | 25,000 | 3,200 |
@@ -60,7 +60,7 @@ The orchestrator produces a structured XML block injected into agent prompts:
 ```xml
 <chati-context bracket="MODERATE">
   <constitution>
-    Articles I-XXIII governing agent behavior.
+    Articles I-XXV governing agent behavior.
     Key: Self-validation required. Loop until quality threshold.
     Guided options (1,2,3). Persistent session state.
     Two-layer handoff. Language protocol. Deviation protocol.
@@ -116,7 +116,7 @@ estimatedTokens = promptText.length / 4
 remainingPercent = (1 - estimatedTokens / providerLimit) * 100
 ```
 
-This replaces the legacy heuristic (`turnCount / maxTurns`) which was inaccurate for sessions with large tool results or long agent outputs. The token-based approach scales correctly across providers (Claude 200K, Gemini 1M, Codex 128K).
+This replaces the legacy heuristic (`turnCount / maxTurns`) which was inaccurate for sessions with large tool results or long agent outputs. The token-based approach scales correctly across providers (Claude 1M, Gemini 1M, Codex 128K).
 
 Fallback: If prompt text is unavailable, the turn-count heuristic is used.
 
@@ -150,11 +150,15 @@ If Smart Continuation is insufficient, the orchestrator spawns a new session:
 
 | Condition | Threshold | Rationale |
 |-----------|-----------|-----------|
+| BUILD phase + DEPLETED/CRITICAL | Immediate (default to Level 2) | During code generation a depleted window produces regressions faster than a reset costs; a clean slate beats in-place compaction |
 | Multiple compactions | 3+ in single session | Context is churning too fast |
 | Post-compact quality | Agent score drops >15% | Compaction lost critical context |
 | Critical bracket persists | CRITICAL for 3+ consecutive interactions | Recovery not working |
+| Echo loop detected | 2+ echo entries in the current phase (session.yaml decision_trail) | The model is looping; a fresh session breaks the cycle |
 | Agent handoff pending | Current agent done, next agent needs fresh context | Clean start is better |
 
+**BUILD-phase reset preference:** During the BUILD phase, when the active agent reaches DEPLETED or CRITICAL, the orchestrator defaults to Level 2 (Autonomous Session Spawn, a clean reset) rather than Level 1 (Smart Continuation, in-place compaction). Compaction preserves continuity but not a clean slate, so the model can keep wrapping up work prematurely while code is being written; a reset carrying a full handoff is safer there. Level 1 stays the default for the MODERATE bracket and for the DISCOVER and PLAN phases.
+
 ### IDE-Specific Behavior
 
 | IDE | Level 1 | Level 2 |
@@ -185,7 +189,7 @@ Memory injection follows the same progressive reinforcement principle: as contex
 
 ---
 
-*Context Engine v4.3.0 — Chati.dev Intelligence Layer (Progressive Reinforcement Model)*
+*Context Engine v4.4.0 — Chati.dev Intelligence Layer (Progressive Reinforcement Model)*
 
 ---
 

package/framework/intelligence/decision-engine.md

@@ -96,7 +96,7 @@ npx chati-dev health
 Checks:
 1. All registered entities exist on disk
 2. Schema files are valid JSON
-3. Constitution has >= 23 articles
+3. Constitution has >= 25 articles
 4. Agents contain required protocol references
 5. Entity count matches actual file count
 6. Checksums match (if calculated)