docs: add claude agent cheatsheet

CHEATSHEET.md

@@ -0,0 +1,87 @@
+# Claude Agent Cheatsheet — learn-claude-code
+
+## El Loop (nunca modificar)
+```python
+while True:
+    response = client.messages.create(model=MODEL, system=SYS, messages=msgs, tools=TOOLS, max_tokens=8000)
+    msgs.append({"role": "assistant", "content": response.content})
+    if response.stop_reason != "tool_use": return
+    results = []
+    for block in response.content:
+        if block.type == "tool_use":
+            output = HANDLERS[block.name](**block.input)
+            results.append({"type": "tool_result", "tool_use_id": block.id, "content": output})
+    msgs.append({"role": "user", "content": results})
+```
+
+## Sesiones y qué añaden
+
+| S | Mecanismo | Clave |
+|---|-----------|-------|
+| 01 | Loop base | exit: `stop_reason != "tool_use"` |
+| 02 | Tool dispatch + seguridad | dict handlers, safe_path, truncate 50k |
+| 03 | TodoWrite | 1 solo in_progress, nag cada 3 rondas |
+| 04 | Subagentes | contexto fresco, sin tool `task`, solo retorna texto |
+| 05 | Skills | SKILL.md, 2-layer injection |
+| 06 | Compression | micro + auto (50k) + explicit, transcripts en disco |
+| 07 | Task system | .tasks/*.json, DAG blockedBy/blocks |
+| 08 | Background | daemon threads, drain_notifications() pre-LLM |
+| 09 | Teams | JSONL inboxes, .team/config.json |
+| 10 | Protocols | request_id correlation, FSM pending→approved/rejected |
+| 11 | Autonomy | idle poll 5s/60s, task scan+claim, identity re-injection |
+| 12 | Worktrees | git worktree, EventBus JSONL, control↔execution planes |
+
+## Patrones de estado
+
+```
+messages[]          → volátil, solo en memoria
+.tasks/*.json       → durable, sobrevive compression
+.team/              → durable, teammates + inboxes
+.transcripts/       → durable, historial comprimido
+.worktrees/         → durable, aislamiento por tarea git
+```
+
+## Seguridad
+```python
+DANGEROUS = ["rm -rf /", "sudo", "shutdown", "reboot", "> /dev/"]
+safe_path: (WORKDIR / p).resolve() — raise si escapa
+output[:50000]  # truncar siempre
+timeout=120     # bash; 300 background
+```
+
+## Subagente vs Teammate
+```
+Subagente:  spawn → trabajo → muere  (s04)
+            contexto fresco, 30 iter max, retorna solo texto final
+
+Teammate:   spawn → WORK → IDLE → WORK → shutdown  (s09+)
+            thread daemon, JSONL inbox, identidad, 50 iter max
+            idle: poll 5s × 12 = 60s → shutdown si sin trabajo
+```
+
+## Compression
+```
+micro_compact:  cada turno, silencioso, retiene 3 últimos tool_results
+auto_compact:   >50k tokens, LLM resume, guarda transcript .jsonl
+explicit:       model llama tool compact(), misma lógica
+```
+
+## SKILL.md
+```markdown
+---
+name: skill-name
+description: one line
+tags: a, b
+---
+# Body
+```
+Layer 1: descripciones → system prompt
+Layer 2: contenido → tool_result on demand
+
+## Worktree workflow (s12)
+```
+task_create("feature") → .tasks/task_1.json
+worktree_create("feat-branch", task_id=1) → git worktree add -b wt/feat-branch
+worktree_run("feat-branch", "npm test")
+worktree_remove("feat-branch", complete_task=True) → task completo
+```