feat: add Intent system for semi-automated attack loop

docs/skills/pentest-with-weaponized/SKILL.md

@@ -63,6 +63,17 @@ them, read their output, send commands, and even create pre-configured terminals
 | `send_to_terminal` | Send a command to a terminal | Running commands, interacting with shells |
 | `create_terminal` | Create a new terminal (with optional profile) | Starting tools — profiles: `netcat`, `msfconsole`, `meterpreter`, `web-delivery`, `shell` |
 
+### Goal & Intent（半自动渗透闭环）
+
+| Tool | Purpose | When to use |
+|------|---------|-------------|
+| `set_goal` | Set engagement goal and phase constraints | Start of engagement or phase switch |
+| `get_goal` | Read current goal | Before every Reason step |
+| `create_intent` | Write reasoned action intent to queue (pending) | After each Reason round — create multiple |
+| `list_intents` | List intent queue, filterable by status | Confirm which intents are approved |
+| `update_intent_status` | Update intent status (dismiss/elevate) | When hypothesis invalidated or confirmed |
+| `execute_intent` | Execute approved intent, auto-capture output | When intent has status=approved |
+
 ### Available Resources (read-only)
 
 | URI | Content |
@@ -121,6 +132,32 @@ them, read their output, send commands, and even create pre-configured terminals
 7. get_graph → check if new attack paths opened
 ```
 
+### Pattern 5: AI-Driven Intent Loop（半自动渗透闭环）
+
+**Full loop:**
+
+1. `get_goal`
+   → Confirm current goal and phase constraints
+
+2. `get_engagement_summary(include_intents=true)`
+   → One call for full context: Findings + Hosts + Users + Graph + existing Intents
+
+3. `create_intent × N` (batch — never one at a time)
+   → All action intents from this Reason round
+   → `reasoning` MUST reference specific Finding IDs or confirmed state
+
+4. [Wait for human approval in TreeView]
+   → `list_intents(status="approved")` to confirm
+
+5. `execute_intent(id)`
+   → Auto-executes + captures initial output (no manual send+read needed)
+
+6. Analyze output:
+   - Hypothesis confirmed → `create_finding(...)` + `update_intent_status(id, "elevated", finding_id=...)`
+   - Hypothesis invalidated → `update_intent_status(id, "dismissed", dismissed_reason="...")`
+
+7. Return to step 2 (loop until Goal achieved or user stops)
+
 ---
 
 ## Best Practices
@@ -227,3 +264,19 @@ and the tester has explicit authorization. TLS certificate verification is
 intentionally disabled for lab/range environments. Treat the workspace as a live
 engagement — document everything, maintain operational security, and follow
 responsible disclosure practices.
+
+## Intent Usage Rules
+
+**Batch creation, never single**
+Create all intents from a Reason round in one batch. The human sees the full attack plan.
+
+**reasoning MUST cite specific evidence**
+✗ "This target may have vulnerabilities"
+✓ "Finding F-003 confirms SMB signing is disabled; credential user@corp.local valid; NTLM relay conditions met"
+
+**Confirm approved before execute_intent**
+Call `list_intents(status="approved")` before `execute_intent` to confirm the human approved.
+
+**Findings and Intents are decoupled**
+Do NOT pass `intent_id` to `create_finding` — Findings are pure fact records.
+After elevating, call `update_intent_status(id, "elevated", finding_id=...)` to link them.