Summary
On Windows, a Desktop-managed agent (buzz-acp → Hermes Agent ACP harness) hangs on its first terminal tool invocation for up to the full tool timeout (420 s), capturing only ~55 chars of output, then the tool times out. The exact same harness + config executed from a normal console completes the same terminal task in ~25 s. Turns end up taking 15-25 minutes because the model keeps retrying terminal commands.
Environment
- Buzz Desktop v0.5.20 (Windows, Tauri), hosted relay (buzz.xyz communities, relay 0.2.1)
- Agent harness: Hermes Agent v0.20.6 spawned by
buzz-acp.exe (agent pool), Windows 11 10.0.22631
- Model: zai/glm-5.3-flash
Reproduction / observations
- Start a managed agent from Desktop → Agents → Start.
- DM the agent with an @mention. The turn dispatches fine (ACP prompt arrives, model responds).
- As soon as the agent uses the
terminal tool, the agent log shows:
API call #1 ... out=222
tools.terminal_tool: Creating new local environment for task session:...
... 420 seconds later ...
agent.tool_executor: sequential tool terminal timed out after 420.0s
agent.tool_executor: tool terminal completed (420.01s, 55 chars)
The deterministic "55 chars" partial output repeats across attempts, suggesting the spawned local shell never produces its ready/handshake output in this detached (no console, stdio wired to ACP pipes) context.
- Control test on the same machine, same config, Hermes one-shot CLI (
hermes chat -q "run: echo ...") from a normal console: terminal tool completes, whole turn ~25 s.
Impact
First turns of Desktop-managed agents on Windows take 15-25 minutes when any tool uses the terminal; the model often retries the terminal several times, multiplying the 420 s timeouts.
Additional note
Agents also try to use the buzz CLI from the terminal to reply, but BUZZ_PRIVATE_KEY / BUZZ_RELAY_URL are not present in the terminal subprocess environment (fast auth failure instead of the hang). Passing those through the harness's terminal env would let CLI-based flows work; replying via plain assistant text already works and is preferable.
Question
Is the buzz-acp agent spawn on Windows expected to give child tools a console/DETACHED_PROCESS environment that local-shell tools can use? If there is a recommended env/flag for the harness (e.g. console allocation, persistent-shell off), we'd love to configure it.
Happy to provide full agent logs if useful.
Summary
On Windows, a Desktop-managed agent (buzz-acp → Hermes Agent ACP harness) hangs on its first
terminaltool invocation for up to the full tool timeout (420 s), capturing only ~55 chars of output, then the tool times out. The exact same harness + config executed from a normal console completes the same terminal task in ~25 s. Turns end up taking 15-25 minutes because the model keeps retrying terminal commands.Environment
buzz-acp.exe(agent pool), Windows 11 10.0.22631Reproduction / observations
terminaltool, the agent log shows:hermes chat -q "run: echo ...") from a normal console: terminal tool completes, whole turn ~25 s.Impact
First turns of Desktop-managed agents on Windows take 15-25 minutes when any tool uses the terminal; the model often retries the terminal several times, multiplying the 420 s timeouts.
Additional note
Agents also try to use the
buzzCLI from the terminal to reply, butBUZZ_PRIVATE_KEY/BUZZ_RELAY_URLare not present in the terminal subprocess environment (fast auth failure instead of the hang). Passing those through the harness's terminal env would let CLI-based flows work; replying via plain assistant text already works and is preferable.Question
Is the buzz-acp agent spawn on Windows expected to give child tools a console/DETACHED_PROCESS environment that local-shell tools can use? If there is a recommended env/flag for the harness (e.g. console allocation, persistent-shell off), we'd love to configure it.
Happy to provide full agent logs if useful.