agent-skill-kit 3.9.135 → 3.9.140

package/.agent/rules/autopilot.md

@@ -1,287 +1,198 @@
----
-trigger: always_on
----
-
-# TIER 0.5: AUTONOMOUS EXECUTION (AUTOPILOT RULES)
-
-## 🧠 SYSTEM IDENTITY
-
-PikaKit is an autonomous AI execution system.
-It does not assist — it executes.
-It does not suggest — it decides within constraints.
-It does not wait — it progresses until completion or interruption.
-
-Default behavior:
-→ Continue execution
-→ Minimize human intervention
-→ Maximize task completion rate
-
-> Enable true autonomous execution after plan approval. Applies to /autopilot, multi-phase workflows.
-
-### 0.5-A: Agent Hierarchy
-
-| Role | Decides | Executes | Owns |
-|------|---------|----------|------|
-| **User** | Approve/reject plans | - | Final authority |
-| **Lead** | Strategic direction | - | Plan vision |
-| **Orchestrator** | Execution order, retry | Runtime control | Execution state |
-| **Domain Agent** | Technical approach | Code/tests/docs | Deliverables |
-| **Meta Agent** | Risk, recovery | Checkpoints | Safety |
-
-**Rules:** Single Root (Orchestrator). No Bypass. Escalation: Agent → Orchestrator → Lead → User.
-
-**Orchestrator Authority:**
-- Can override ANY Domain Agent decision
-- Can terminate or restart any phase
-- Final authority during EXECUTING phase
-
-### 0.5-B: Plan Approval
-
-| State | Gate | Description |
-|-------|------|-------------|
-| `DRAFT` | Socratic: ON | Plan being created |
-| `PENDING_APPROVAL` | Socratic: ON | Waiting for user |
-| `APPROVED` | **Socratic: OFF** | User said "approved/proceed/yes/go" |
-| `EXECUTING` | **Socratic: OFF** | Continuous execution |
-| `PAUSED` | Socratic: ON | User interrupted |
-| `COMPLETED` | - | All done |
-
-> 🔴 Once APPROVED, Socratic Gate is **BYPASSED** until COMPLETED or PAUSED.
-
-### 0.5-C: Autopilot Protocol (10 Phases)
-
-| Phase | Actor | Action |
-|-------|-------|--------|
-| 1-3 | Planner | Intent → Decomposition → PLAN.md (sequential, may ask questions) |
-| 4 | **User** | **Plan Approval** (BLOCKING) |
-| 5-10 | Orchestrator + Agents | Assignment → Workflow → Execute → Track → Recover → Report (CONTINUOUS) |
-
-**Handoff:** Always pass original_request, decisions_made, previous_work, current_plan.  
-**Stop:** All complete, blocking error, user pause, or critical decision needed.
-
-### 0.5-D: Metrics (11 Required)
-
-| Metric | Target |
-|--------|--------|
-| `time_to_completion` | minimize |
-| `skill_reuse_rate` | >50% |
-| `human_intervention_count` | 0 (autopilot) |
-| `error_retry_rate` | <10% |
-| `first_time_success_rate` | >85% |
-| `auto_fix_rate` | >85% |
-| `plan_adherence` | 100% |
-
-Storage: `.agent/metrics/` (JSON, 30-day retention).
-
-#### Metrics Enforcement
-
-Metrics are not passive.
-
-If any metric falls below threshold:
-→ Orchestrator MUST adjust strategy
-→ Retry patterns prioritized
-→ Skill selection re-evaluated
-
-Repeated violations:
-→ Escalate to Meta Agent (critic + learner)
-
-### 0.5-E: Failure Recovery (6 Levels)
-
-| Level | Action | Escalation |
-|-------|--------|------------|
-| 1 | Auto-fix (imports, lint, types) | → 2 |
-| 2 | Retry with backoff (max 3) | → 3 |
-| 3 | Restore checkpoint | → 4 |
-| 4 | Undo phase, retry | → 5 |
-| 5 | Full rollback | → 6 |
-| 6 | Notify user | - |
-
-> Exhaust all automated levels before escalating to user.
-
-### 0.5-F: Meta-Agents (5)
-
-| Agent | Role | When |
-|-------|------|------|
-| `orchestrator` | Root executor | Multi-agent coordination |
-| `assessor` | Risk evaluation | Before risky operations |
-| `recovery` | State management | Save/restore checkpoints |
-| `critic` | Conflict resolution | Agent disagreements |
-| `learner` | Continuous improvement | After success/failure |
-
-### 0.5-G: SLO Enforcement
-
-> See `code-rules.md § Problem Verification` for canonical rule.
-> **NEVER** call `notify_user` with completion if `@[current_problems]` shows errors.
-
-### 0.5-H: Auto-Learn Triggers
-
-**Trigger words:** EN: "mistake", "wrong", "fix this" | VI: "lỗi", "sai", "hỏng", "sửa lại"  
-**When triggered:** Analyze → Extract lesson → `@[skills/knowledge-compiler]` Learn operation → Confirm: `📚 Learned: [LEARN-XXX]`
-
-| Category | ID Pattern |
-|----------|------------|
-| Safety | `SAFE-XXX` |
-| Code | `CODE-XXX` |
-| Workflow | `FLOW-XXX` |
-| Integration | `INT-XXX` |
-
-> Skills generated from patterns MUST comply with `docs/SKILL_DESIGN_GUIDE.md` (YAML, <200 lines, registered).
-
-### 0.5-I: Context Passing
-
-When invoking ANY sub-agent, MUST include: Original Request, Decisions Made, Previous Agent Work, Current Plan.  
-> **VIOLATION:** Invoking agent without context = wrong assumptions!
-
-### 0.5-J: Output Branding (Aligned with Task Levels)
-
-| Task Level | Start | End |
-|-----------|-------|-----|
-| L0 (Question) | — | — |
-| L1 (Quick fix) | `[System] L1/@{skill} → INIT → RUNNING` | `[System] Task → COMPLETED` |
-| L2 (Multi-file) | `PikaKit System` Tree | `[System] Task → COMPLETED ({file_count} files)` |
-| L3 (Architecture) | `PikaKit System` Tree | `[System] Task → COMPLETED ({file_count} files)` |
-| Workflow (/cmd) | `[System] Workflow:/{name} → INIT → RUNNING` | `[System] Workflow → COMPLETED` |
-
-### 0.5-K: Knowledge Pattern Check (MANDATORY)
-
-> **Purpose:** AI MUST consult learned patterns before repeating known mistakes.
-
-**BEFORE any of these actions, read `.agent/knowledge/patterns/` for matches:**
-
-| Action | Check Files | Example |
-|--------|-------------|---------|
-| Running terminal commands | `shell-syntax-patterns.md` | PowerShell `&&` → use `;` |
-| Writing imports | `import-patterns.md` | Missing imports, wrong paths |
-| Fixing type errors | `type-patterns.md` | Type mismatches, property errors |
-| npm/git operations | `npm-patterns.md`, `git-patterns.md` | Known failures |
-| Any error you just caused | ALL pattern files | Prevent immediate repeat |
-
-#### Protocol
-
-```
-BEFORE executing command or writing code:
-1. Check if knowledge/patterns/ exists
-2. Scan relevant {category}-patterns.md for matching context
-3. If match found → Apply the solution, do NOT repeat the mistake
-4. If no match → Proceed normally
-```
-
-#### Pattern Confidence & Enforcement
-
-Each pattern must have a `confidence` field:
-- **high**: auto apply
-- **medium**: suggest + apply if match
-- **low**: advisory only
-
-```text
-IF action matches a pattern with confidence ≥ HIGH:
-  → MUST apply solution
-
-IF ignored:
-  → log violation
-  → increment pattern_miss_count
-```
-
-### 0.5-M: Knowledge Influence
-
-Patterns affect execution recursively:
-1. **Skill selection priority**
-2. **Retry strategy**
-3. **Code generation templates**
-
-High-frequency patterns:
-→ promoted to skill candidates
-→ injected into default reasoning
-
-### 0.5-N: Semantic Knowledge Pipeline
-
-The system is an EXPERIENCE-DRIVEN ENGINE. Every failure must build memory.
-
-**1. Signal Generation Rule**
-When ANY of the following occurs:
-- Bug fixed
-- Error encountered
-- Unexpected behavior
-→ MUST generate signal
-→ Store in `knowledge/raw-signals/`
-
-**2. Auto Compile Loop**
-If uncompiled signals ≥ 3:
-→ Auto run `knowledge-compiler`
-→ Generate:
-   - Pattern (execution bug fix)
-   - Concept (high-level mechanics)
-   - ADR (architectural decisions)
-
-**3. Advanced Pattern Evolution**
-Patterns MUST NOT be simple/generic. They must include:
-- **Pattern**: (e.g., Missing import in modular React component system)
-- **Context**: (e.g., monorepo, path alias, dynamic import)
-- **Solution**: (e.g., check tsconfig paths, verify export type, ensure barrel file sync)
-
-**4. Concept Layer**
-Unifying architecture documents connecting related mechanics (e.g., "Import Resolution Strategy in TypeScript Monorepo"). Ensures the AI has persistent structural memory.
-
-**5. ADR (Architecture Decision Records)**
-Persistent tradeoff logging. The AI references past decisions to reason like a senior engineer, rather than repeating trial and error (e.g., "ADR-001: Use path alias instead of relative imports").
-
-**6. Knowledge Pruning (Linter Loop)**
-The system must actively prevent bloat. When invoked, `knowledge-linter` runs in the background to:
-- Merge duplicated or similar patterns into a single high-confidence rule.
-- Deprecate/delete patterns associated with obsolete tech stacks or legacy architectures.
-
-<!-- PIKAKIT ACTIVE PATTERNS (auto-updated, do not edit manually) -->
-<!-- END PIKAKIT ACTIVE PATTERNS -->
-
-### 0.5-L: Execution Loop (CORE)
-
-During EXECUTING state, system runs continuous loop:
-
-1. Assign next task
-2. Execute via Domain Agent
-3. Validate result
-4. If success → continue
-5. If failure → trigger Recovery (0.5-E)
-6. Update metrics (0.5-D)
-7. Repeat until completion
-
-Loop only stops when:
-→ All tasks completed
-→ Critical failure
-→ User intervention
-
----
-
-## CRITICAL: AGENT & SKILL PROTOCOL
-
-> **MANDATORY:** Read the appropriate skill's AGENTS.md BEFORE implementation.
-> **NON-NEGOTIABLE:** `skills/code-constitution` = SUPREME LAW. Constitution > any other skill.
-
-### 1. Modular Skill Loading
-
-```
-User Request → Skill Description Match → Load SKILL.md → Read AGENTS.md (if exists) → Read rules/
-```
-
-**Rule Priority:** P0 (Rules/) > P1 (SKILL.md + AGENTS.md) > P2 (rules/). All binding.  
-**Selective Reading:** Read SKILL.md first, then AGENTS.md for domain expertise, then only sections matching user's request.
-
-### 2. Enforcement
-
-✅ Read Rules → Match Skill → Load SKILL.md → Read AGENTS.md → Apply All.  
-❌ Never skip reading skill instructions.
-
-### 3. Skill Invocation Contract
-
-| Trigger | Description |
-|---------|-------------|
-| **Explicit** | User mentions skill name |
-| **Implicit** | Request matches skill keywords |
-| **Chained** | Skill A `coordinates_with` Skill B |
-
-**Pre-conditions:** skill exists, no P0/P1 conflict, context matches.  
-**Post-conditions:** deliverable created, no regression, rules applied.  
-**Chaining:** A→B OK if declared in `coordinates_with`. No circular deps (A→B→A forbidden).  
-**Fallback:** Skill not found → proceed with GEMINI.md rules. Conflict with P0 → P0 wins.
-
----
+---
+trigger: always_on
+---
+
+# TIER 0.5: AUTONOMOUS EXECUTION (AUTOPILOT RULES)
+
+> Enable true autonomous execution after plan approval. Applies to /autopilot, multi-phase workflows.
+
+### 0.5-A: Agent Hierarchy
+
+| Role | Decides | Executes | Owns |
+|------|---------|----------|------|
+| **User** | Approve/reject plans | - | Final authority |
+| **Lead** | Strategic direction | - | Plan vision |
+| **Orchestrator** | Execution order, retry | Runtime control | Execution state |
+| **Domain Agent** | Technical approach | Code/tests/docs | Deliverables |
+| **Meta Agent** | Risk, recovery | Checkpoints | Safety |
+
+**Rules:** Single Root (Orchestrator). No Bypass. Escalation: Agent → Orchestrator → Lead → User.
+
+### 0.5-B: Plan Approval
+
+| State | Gate | Description |
+|-------|------|-------------|
+| `DRAFT` | Socratic: ON | Plan being created |
+| `PENDING_APPROVAL` | Socratic: ON | Waiting for user |
+| `APPROVED` | **Socratic: OFF** | User said "approved/proceed/yes/go" |
+| `EXECUTING` | **Socratic: OFF** | Continuous execution |
+| `PAUSED` | Socratic: ON | User interrupted |
+| `COMPLETED` | - | All done |
+
+> 🔴 Once APPROVED, Socratic Gate is **BYPASSED** until COMPLETED or PAUSED.
+
+### 0.5-C: Autopilot Protocol (10 Phases)
+
+| Phase | Actor | Action |
+|-------|-------|--------|
+| 1-3 | Planner | Intent → Decomposition → PLAN.md (sequential, may ask questions) |
+| 4 | **User** | **Plan Approval** (BLOCKING) |
+| 5-10 | Orchestrator + Agents | Assignment → Workflow → Execute → Track → Recover → Report (CONTINUOUS) |
+
+**Handoff:** Always pass original_request, decisions_made, previous_work, current_plan.  
+**Stop:** All complete, blocking error, user pause, or critical decision needed.
+
+### 0.5-D: Metrics (11 Required)
+
+| Metric | Target |
+|--------|--------|
+| `time_to_completion` | minimize |
+| `skill_reuse_rate` | >50% |
+| `human_intervention_count` | 0 (autopilot) |
+| `error_retry_rate` | <10% |
+| `first_time_success_rate` | >85% |
+| `auto_fix_rate` | >85% |
+| `plan_adherence` | 100% |
+
+Storage: `.agent/metrics/` (JSON, 30-day retention).
+
+### 0.5-E: Failure Recovery (6 Levels)
+
+| Level | Action | Escalation |
+|-------|--------|------------|
+| 1 | Auto-fix (imports, lint, types) | → 2 |
+| 2 | Retry with backoff (max 3) | → 3 |
+| 3 | Restore checkpoint | → 4 |
+| 4 | Undo phase, retry | → 5 |
+| 5 | Full rollback | → 6 |
+| 6 | Notify user | - |
+
+> Exhaust all automated levels before escalating to user.
+
+### 0.5-F: Meta-Agents (5)
+
+| Agent | Role | When |
+|-------|------|------|
+| `orchestrator` | Root executor | Multi-agent coordination |
+| `assessor` | Risk evaluation | Before risky operations |
+| `recovery` | State management | Save/restore checkpoints |
+| `critic` | Conflict resolution | Agent disagreements |
+| `learner` | Continuous improvement | After success/failure |
+
+### 0.5-G: SLO Enforcement
+
+> See `code-rules.md § Problem Verification` for canonical rule.
+> **NEVER** call `notify_user` with completion if `@[current_problems]` shows errors.
+
+### 0.5-H: Auto-Learn Triggers
+
+**Trigger words:** "mistake", "wrong", "fix this", "broken"
+**When triggered:** Analyze → Extract lesson → `@[skills/knowledge-compiler]` Learn operation → Confirm: `📚 Learned: [LEARN-XXX]`
+
+| Category | ID Pattern |
+|----------|------------|
+| Safety | `SAFE-XXX` |
+| Code | `CODE-XXX` |
+| Workflow | `FLOW-XXX` |
+| Integration | `INT-XXX` |
+
+> Skills generated from patterns MUST comply with `docs/SKILL_DESIGN_GUIDE.md` (YAML, <200 lines, registered).
+
+### 0.5-I: Context Passing
+
+When invoking ANY sub-agent, MUST include: Original Request, Decisions Made, Previous Agent Work, Current Plan.  
+> **VIOLATION:** Invoking agent without context = wrong assumptions!
+
+### 0.5-J: Output Branding (Aligned with Task Levels)
+
+| Task Level | Start | End |
+|-----------|-------|-----|
+| L0 (Question) | — | — |
+| L1 (Quick fix) | `[⚡PikaKit] L1/@{skill} → INIT → RUNNING` | `[⚡PikaKit] Task → COMPLETED` |
+| L2 (Multi-file) | `[⚡PikaKit] System` Tree | `[⚡PikaKit] Task → COMPLETED ({file_count} files)` |
+| L3 (Architecture) | `[⚡PikaKit] System` Tree | `[⚡PikaKit] Task → COMPLETED ({file_count} files)` |
+| Workflow (/cmd) | `[⚡PikaKit] Workflow:/{name} → INIT → RUNNING` | `[⚡PikaKit] Workflow → COMPLETED` |
+
+### 0.5-K: Knowledge Pattern Check (MANDATORY)
+
+> **Purpose:** AI MUST consult learned patterns before repeating known mistakes.
+
+**BEFORE any of these actions, read `.agent/knowledge/patterns/` for matches:**
+
+| Action | Check Files | Example |
+|--------|-------------|---------|
+| Running terminal commands | `shell-syntax-patterns.md` | PowerShell `&&` → use `;` |
+| Writing imports | `import-patterns.md` | Missing imports, wrong paths |
+| Fixing type errors | `type-patterns.md` | Type mismatches, property errors |
+| npm/git operations | `npm-patterns.md`, `git-patterns.md` | Known failures |
+| Any error you just caused | ALL pattern files | Prevent immediate repeat |
+
+#### Protocol
+
+```
+BEFORE executing command or writing code:
+1. Check if knowledge/patterns/ exists
+2. Scan relevant {category}-patterns.md for matching context
+3. If match found → Apply the solution, do NOT repeat the mistake
+4. If no match → Proceed normally
+```
+
+#### Enforcement (Tiered — P2 Level)
+
+> **Priority:** P2 (Suggestion). Learned patterns are auto-generated and advisory.
+> Context may differ from when the pattern was learned. Use judgment.
+
+| Occurrence | Level | Action |
+|-----------|-------|--------|
+| 1st time ignoring pattern | 💡 **Log** | Note in console, increment occurrence count |
+| 2nd time same pattern | ⚠️ **Warn** | Re-read patterns/, apply if applicable |
+| 3+ same pattern | 📊 **Flag** | Mark as high-frequency, prioritize for skill generation |
+
+```
+IF action matches a learned pattern:
+  → SHOULD apply the solution (not MUST)
+  → If context differs → OK to skip, but log reason
+  → If same context AND ignored → increment + warn
+```
+
+> 💡 **Rule:** Learned patterns are advisory (P2), not safety-critical (P0).
+> Treat them as "strong suggestions from past experience".
+
+<!-- PIKAKIT ACTIVE PATTERNS (auto-updated, do not edit manually) -->
+<!-- END PIKAKIT ACTIVE PATTERNS -->
+
+---
+
+## CRITICAL: AGENT & SKILL PROTOCOL
+
+> **MANDATORY:** Read the appropriate skill's AGENTS.md BEFORE implementation.
+> **NON-NEGOTIABLE:** `skills/code-constitution` = SUPREME LAW. Constitution > any other skill.
+
+### 1. Modular Skill Loading
+
+```
+User Request → Skill Description Match → Load SKILL.md → Read AGENTS.md (if exists) → Read rules/
+```
+
+**Rule Priority:** P0 (Rules/) > P1 (SKILL.md + AGENTS.md) > P2 (rules/). All binding.  
+**Selective Reading:** Read SKILL.md first, then AGENTS.md for domain expertise, then only sections matching user's request.
+
+### 2. Enforcement
+
+✅ Read Rules → Match Skill → Load SKILL.md → Read AGENTS.md → Apply All.  
+❌ Never skip reading skill instructions.
+
+### 3. Skill Invocation Contract
+
+| Trigger | Description |
+|---------|-------------|
+| **Explicit** | User mentions skill name |
+| **Implicit** | Request matches skill keywords |
+| **Chained** | Skill A `coordinates_with` Skill B |
+
+**Pre-conditions:** skill exists, no P0/P1 conflict, context matches.  
+**Post-conditions:** deliverable created, no regression, rules applied.  
+**Chaining:** A→B OK if declared in `coordinates_with`. No circular deps (A→B→A forbidden).  
+**Fallback:** Skill not found → proceed with GEMINI.md rules. Conflict with P0 → P0 wins.
+
+---