Add local skills, durable tasks, and bounded subagents

Stage 6 adds the Option B-min slice from the cc product roadmap: local SKILL.md loading, a store-backed durable task graph, and a synchronous stateless run_subagent tool with an exact child-tool allowlist. The implementation keeps richer Claude Code agent runtime concepts deferred while wiring the new surfaces through the existing LangChain create_agent tool system.

Constraint: LangChain-first implementation; no custom query loop, background agents, mailbox, worktrees, remote/team runtime, sidechain resume, or forked skill execution.

Constraint: TodoWrite remains session-local and separate from durable task records.

Rejected: Full AgentTool parity now | would require background/resume/mailbox/worktree behavior outside the approved Option B-min slice.

Rejected: Plugin/MCP skill loading now | Stage 6 is local skills only and keeps extension platform work for a later stage.

Confidence: high

Scope-risk: moderate

Reversibility: clean

Directive: Future multi-agent work must explicitly expand the subagent contract before adding mailbox/background/worktree/resume semantics.

Tested: cd coding-deepgent && python -m pytest -q -> 87 passed

Tested: cd coding-deepgent && ruff check . -> all checks passed

Tested: cd coding-deepgent && ruff format --check . -> 107 files already formatted

Tested: cd coding-deepgent && python -m mypy src/coding_deepgent tests -> success

Tested: forbidden runtime-creep grep guards -> passed

Tested: Architect verification -> APPROVE

Not-tested: real background/remote subagent execution; intentionally deferred.

Author: Kun

coding-deepgent/PROJECT_PROGRESS.md

@@ -2,8 +2,8 @@
 
 ## Current product stage
 
-- `current_product_stage`: `stage-5-memory-context-compact-foundation`
-- `compatibility_anchor`: `memory-context-compact-foundation`
+- `current_product_stage`: `stage-6-skills-subagents-task-graph`
+- `compatibility_anchor`: `skills-subagents-task-graph`
 - Status: control-plane foundation implemented as one cumulative LangChain cc product surface
 - Last updated: 2026-04-12
 
@@ -34,3 +34,7 @@ Stage 4 adds deterministic permission/safety decisions, local lifecycle hooks, a
 ## Stage 5 memory/context/compact foundation
 
 Stage 5 adds a store-backed long-term memory foundation seam, the model-visible `save_memory` tool, bounded memory context injection, and deterministic tool-result budget helpers. Message-history projection/pruning, LLM autocompact, session-memory side-agent writing, subagents, durable tasks, and MCP/plugin memory sync remain future work.
+
+## Stage 6 skills/subagents/task graph
+
+Stage 6 adds local skill loading, a store-backed durable task graph, and a minimal synchronous/stateless `run_subagent` tool. Background agents, SendMessage/mailbox, worktrees, remote/team runtime, sidechain resume, forked skill execution, MCP/plugin marketplace, and custom query loops remain future work.

coding-deepgent/README.md

@@ -4,8 +4,8 @@ Independent cumulative LangChain cc product surface.
 
 ## Current product stage
 
-- `current_product_stage`: `stage-5-memory-context-compact-foundation`
-- `compatibility_anchor`: `memory-context-compact-foundation`
+- `current_product_stage`: `stage-6-skills-subagents-task-graph`
+- `compatibility_anchor`: `skills-subagents-task-graph`
 - Upgrade policy: advance by explicit product-stage plan approval, not tutorial chapter completion.
 
 ## Current architecture
@@ -35,3 +35,7 @@ Stage 4 adds deterministic permission/safety decisions, local lifecycle hooks, a
 ## Stage 5 memory/context/compact foundation
 
 Stage 5 adds a store-backed long-term memory foundation seam, the model-visible `save_memory` tool, bounded memory context injection, and deterministic tool-result budget helpers. Message-history projection/pruning, LLM autocompact, session-memory side-agent writing, subagents, durable tasks, and MCP/plugin memory sync remain future work.
+
+## Stage 6 skills/subagents/task graph
+
+Stage 6 adds local skill loading, a store-backed durable task graph, and a minimal synchronous/stateless `run_subagent` tool. Background agents, SendMessage/mailbox, worktrees, remote/team runtime, sidechain resume, forked skill execution, MCP/plugin marketplace, and custom query loops remain future work.

coding-deepgent/docs/cc-alignment-roadmap.md

@@ -32,6 +32,13 @@ Aligning Stage 4 should improve safety, maintainability, context quality, and te
 - `/root/claude-code-haha/src/utils/queryContext.ts:fetchSystemPromptParts` -> prompt builder accepts rendered memory context as a distinct prompt section -> align -> use existing `create_agent` prompt path.
 - `/root/claude-code-haha/src/query.ts:applyToolResultBudget` -> deterministic `apply_tool_result_budget()` helper for oversized tool-result strings -> partial -> no message-history projection/pruning in Stage 5.
 
+## Stage 6 skills/tasks/subagents rows
+
+- `/root/claude-code-haha/src/tools/SkillTool/SkillTool.ts` -> local `load_skill` tool loads one `SKILL.md` by explicit name -> partial -> no plugin/MCP/remote skills.
+- `/root/claude-code-haha/src/utils/tasks.ts` and `/root/claude-code-haha/src/tools/Task*Tool/*` -> store-backed `task_create/get/list/update` with strict transitions -> partial -> no coordinator/team runtime yet.
+- `/root/claude-code-haha/src/tools/AgentTool/AgentTool.tsx` -> minimal synchronous/stateless `run_subagent` tool with exact child-tool allowlist -> partial -> no background/worktree/mailbox/resume.
+- `/root/claude-code-haha/src/tools/SendMessageTool/*` -> mailbox/send-message semantics -> defer -> requires later multi-agent runtime.
+
 ## Next candidates
 
 1. Stage 5: memory + context budget + compact seam.

coding-deepgent/project_status.json

@@ -1,7 +1,7 @@
 {
   "project": "coding-deepgent",
-  "current_product_stage": "stage-5-memory-context-compact-foundation",
-  "compatibility_anchor": "memory-context-compact-foundation",
+  "current_product_stage": "stage-6-skills-subagents-task-graph",
+  "compatibility_anchor": "skills-subagents-task-graph",
   "shape": "staged_langchain_cc_product",
   "public_shape": "single cumulative app",
   "upgrade_policy": "Advance by explicit product-stage plan approval, not tutorial chapter completion.",

coding-deepgent/src/coding_deepgent/containers/tool_system.py

@@ -7,6 +7,9 @@
 
 from coding_deepgent.memory import save_memory
 from coding_deepgent.permissions import PermissionManager
+from coding_deepgent.skills import load_skill
+from coding_deepgent.subagents import run_subagent
+from coding_deepgent.tasks import task_create, task_get, task_list, task_update
 from coding_deepgent.tool_system import (
     CapabilityRegistry,
     ToolCapability,
@@ -34,6 +37,12 @@ def _tool_domain(name: str) -> str:
         return "todo"
     if name == "save_memory":
         return "memory"
+    if name == "load_skill":
+        return "skills"
+    if name.startswith("task_"):
+        return "tasks"
+    if name == "run_subagent":
+        return "subagents"
     if name in READ_ONLY_TOOL_NAMES | DESTRUCTIVE_TOOL_NAMES:
         return "filesystem"
     return "unknown"
@@ -63,11 +72,22 @@ class ToolSystemContainer(containers.DeclarativeContainer):
     filesystem_tools: Any = providers.Dependency(default=providers.Object([]))
     todo_tools: Any = providers.Dependency(default=providers.Object([]))
     memory_tools: Any = providers.Dependency(default=providers.Object([save_memory]))
+    skill_tools: Any = providers.Dependency(default=providers.Object([load_skill]))
+    task_tools: Any = providers.Dependency(
+        default=providers.Object([task_create, task_get, task_list, task_update])
+    )
+    subagent_tools: Any = providers.Dependency(default=providers.Object([run_subagent]))
     permission_mode: Any = providers.Dependency(default=providers.Object("default"))
     event_sink: Any = providers.Dependency(default=providers.Object(None))
 
     tools: Any = providers.Callable(
-        _combine_tools, filesystem_tools, todo_tools, memory_tools
+        _combine_tools,
+        filesystem_tools,
+        todo_tools,
+        memory_tools,
+        skill_tools,
+        task_tools,
+        subagent_tools,
     )
     capability_registry: Any = providers.Callable(_capability_registry, tools)
     permission_manager: Any = providers.Factory(

coding-deepgent/src/coding_deepgent/memory/schemas.py

@@ -1,6 +1,8 @@
 from __future__ import annotations
 
 from typing import Literal
+
+from langchain.tools import ToolRuntime
 from pydantic import BaseModel, ConfigDict, Field, field_validator
 
 MemoryNamespace = Literal["project", "user", "local"]
@@ -25,7 +27,7 @@ def _text_must_not_be_blank(cls, value: str) -> str:
 
 
 class SaveMemoryInput(BaseModel):
-    model_config = ConfigDict(extra="forbid")
+    model_config = ConfigDict(extra="forbid", arbitrary_types_allowed=True)
 
     content: str = Field(
         ...,
@@ -39,6 +41,7 @@ class SaveMemoryInput(BaseModel):
     source: str = Field(
         default="agent", description="Source label for this memory entry."
     )
+    runtime: ToolRuntime
 
     @field_validator("content", "source")
     @classmethod

coding-deepgent/src/coding_deepgent/memory/tools.py

@@ -14,6 +14,7 @@
 
 @tool(
     "save_memory",
+    args_schema=SaveMemoryInput,
     description=(
         "Save durable reusable knowledge or preferences as long-term memory. "
         "Do not save transient todos, current plans, task status, or one-off observations."
@@ -28,7 +29,10 @@ def save_memory(
     """Save reusable long-term memory through the LangGraph store seam."""
 
     validated = SaveMemoryInput(
-        content=content, namespace=cast(MemoryNamespace, namespace), source=source
+        content=content,
+        namespace=cast(MemoryNamespace, namespace),
+        source=source,
+        runtime=runtime,
     )
     store = runtime.store
     if store is None:

coding-deepgent/src/coding_deepgent/runtime/context.py

@@ -12,4 +12,5 @@ class RuntimeContext:
     workdir: Path
     entrypoint: str
     agent_name: str
+    skill_dir: Path
     event_sink: RuntimeEventSink

coding-deepgent/src/coding_deepgent/runtime/invocation.py

@@ -35,6 +35,7 @@ def build_runtime_context(
         workdir=settings.workdir,
         entrypoint=entrypoint or settings.entrypoint,
         agent_name=agent_name or settings.agent_name,
+        skill_dir=settings.skill_dir,
         event_sink=event_sink,
     )
 

coding-deepgent/src/coding_deepgent/settings.py

@@ -48,6 +48,7 @@ class Settings(BaseSettings):
 
     workdir: Path = Field(default_factory=resolve_workdir)
     session_dir: Path = Field(default=Path(".coding-deepgent/sessions"))
+    skill_dir: Path = Field(default=Path("skills"))
     model_name: str = Field(default_factory=deepgent_model_name)
     openai_api_key: SecretStr | None = Field(default=None, alias="OPENAI_API_KEY")
     openai_base_url: str | None = Field(default=None, alias="OPENAI_BASE_URL")
@@ -83,6 +84,11 @@ def _normalize_paths(self) -> "Settings":
             self.session_dir = (self.workdir / self.session_dir).resolve()
         else:
             self.session_dir = self.session_dir.expanduser().resolve()
+
+        if not self.skill_dir.is_absolute():
+            self.skill_dir = (self.workdir / self.skill_dir).resolve()
+        else:
+            self.skill_dir = self.skill_dir.expanduser().resolve()
         return self