[claude] Consume hai-agents SDK local mode; drop owned binary lifecycle

README.md

@@ -85,23 +85,25 @@ The trigger and the lever are decoupled through a single one-line file, `~/.holo
 - **Reading it (the lever):** every turn records its own `started_at` and, while running, polls the file ~4×/s. It acts **only if the file's timestamp is newer than its `started_at`** — then it pauses, then cancels at the next action boundary.
 - **Clearing it:** the file is *never deleted*. It's cleared by time — the next turn starts later, so a leftover request is automatically stale and can't kill it. This is also why a `holo stop` fired *before* a run begins is ignored: nothing was running to stop.
 
-`holo stop --force` is the exception to the step-bounded model: it reads the runtime's pid file (`~/.holo/agent-pid-<port>`) and SIGKILLs the process directly, so it doesn't wait for an action boundary.
+`holo stop --force` is the exception to the step-bounded model: it discovers the runtime through the hai-agents SDK (`LocalRuntime.attach(...).force_kill()`) and SIGKILLs the process directly, so it doesn't wait for an action boundary. For one release it also honours pre-SDK pid files (`~/.holo/agent-pid-<port>`) so an upgrade mid-session can still be stopped.
 
 ## Use from Python
 
-`holo_desktop.agent_client` is the same client every CLI surface is built on: it spawns (or attaches to) the `hai-agent-runtime` binary on loopback and drives sessions over the agent API.
+`holo_desktop.agent_client` is the same client every CLI surface is built on: it spawns (or attaches to) the `hai-agent-runtime` binary on loopback via the hai-agents local runtime and drives sessions over the agent API.
 
 ```python
 import asyncio
 
-from holo_desktop.agent_client import AgentApiClient, SpawnConfig, ensure_running
+from holo_desktop.agent_client import AgentApiClient, SpawnConfig, ensure_local_runtime
 from holo_desktop.agent_client.requests import build_session_request
+from holo_desktop.settings import load_holo_settings
 
 
 async def main() -> None:
-    daemon = await ensure_running(SpawnConfig(port=18795))
+    settings = load_holo_settings()
+    runtime = await asyncio.to_thread(ensure_local_runtime, SpawnConfig(port=18795), settings=settings)
     try:
-        async with AgentApiClient(daemon.base_url, daemon.token) as client:
+        async with AgentApiClient(runtime) as client:
             request = build_session_request(
                 task="Tell me how many unread emails I have", max_steps=None, max_time_s=None
             )
@@ -110,7 +112,8 @@ async def main() -> None:
                 print(event.type)
             print(stream.answer)
     finally:
-        await daemon.aclose()
+        if runtime.owned:
+            await asyncio.to_thread(runtime.shutdown)
 
 
 asyncio.run(main())

pyproject.toml

@@ -27,6 +27,8 @@ dependencies = [
     # Agent-API wire contract (Pydantic specs + agp_types) for talking to the closed
     # hai-agent-runtime binary; the wheel vendors agent_interface + agp_types as the only source of the spec types.
     "hai-agent-api>=0.1.15",
+    # SDK local mode (plan 002): owns runtime download/spawn/auth/discovery; Holo delegates lifecycle to it.
+    "hai-agents[local]>=0.1",
     # CLI / serve (A2A) / acp / mcp adapter surface.
     "a2a-sdk[http-server]>=1.0.3",
     "agent-client-protocol>=0.10",
@@ -78,6 +80,12 @@ dev = [
 [tool.uv.workspace]
 members = ["examples/expense_report"]
 
+[tool.uv.sources]
+# SHIP-GATE: replace with the published hai-agents[local] version before merge.
+# Dev-time override onto the plan-002 implementation worktree; the [local] extra
+# is empty today, so the editable path install satisfies it as-is.
+hai-agents = { path = "/Users/charlie.masters/.config/superpowers/worktrees/hai-agents-python/claude-sdk-local-mode", editable = true }
+
 [tool.hatch.build.targets.wheel]
 packages = ["src/holo_desktop"]
 

src/holo_desktop/agent_client/__init__.py

@@ -1,20 +1,16 @@
-"""Thin client to the hai-agent-runtime binary."""
+"""Thin client to the hai-agent-runtime binary (lifecycle delegated to hai-agents local mode)."""
 
 from holo_desktop.agent_client.client import AgentApiClient, SessionStream
-from holo_desktop.agent_client.launcher import (
-    AgentDaemon,
+from holo_desktop.agent_client.sdk_runtime import (
     SpawnConfig,
-    ensure_running,
-    runtime_log_path,
-    runtime_log_tail,
+    ensure_local_runtime,
+    ensure_local_runtime_from_env,
 )
 
 __all__ = [
     "AgentApiClient",
-    "AgentDaemon",
     "SessionStream",
     "SpawnConfig",
-    "ensure_running",
-    "runtime_log_path",
-    "runtime_log_tail",
+    "ensure_local_runtime",
+    "ensure_local_runtime_from_env",
 ]