Grok Build runtime spawns bare `grok` (TUI) instead of `grok agent --always-approve stdio` — agents fail with ENXIO

[madhu-korada]

Summary

Selecting Grok Build as the managed-agent runtime starts Fizz/Honey/Bumble with:

agent_cmd=/Users/.../grok

and empty agent_args. That launches the interactive TUI, not the ACP server. Under Buzz (no controlling TTY) every pool worker exits immediately:

Error: Device not configured (os error 6)
ERROR buzz_acp: agent initialize failed: Agent process exited unexpectedly
Error: all 10 agents failed to start — cannot continue

Claude / other ACP-native adapters work on the same machine. The Grok CLI is installed and authenticated; this is a harness arg wiring bug, not auth.

Environment

• Buzz: 0.5.0 (macOS, desktop)
• Grok CLI: 0.2.114 (~/.local/bin/grok → ~/.grok/bin/grok)
• OS: macOS arm64
• Config: preferred_runtime: "grok" in global-agent-config.json
• Managed agents end up with:

"agent_command": "claude-agent-acp",
"agent_command_override": "grok",
"agent_args": []

Steps to reproduce

1. Install Grok CLI (grok on PATH / ~/.grok/bin) and grok login (or set XAI_API_KEY).
2. Install latest Buzz desktop; create/start Welcome Team agents (Fizz etc.).
3. Set agent runtime / preferred runtime to Grok Build.
4. Start agents and open the harness log.

Expected

Spawn should match the preset harness definition on main:

// desktop/src-tauri/src/managed_agents/discovery.rs — PRESET_HARNESSES
PresetHarness {
    id: "grok",
    label: "Grok Build",
    command: "grok",
    args: &["agent", "--always-approve", "stdio"],
    ...
}

i.e. effectively:

grok agent --always-approve stdio

ACP initialize should succeed (same as a manual pipe of JSON-RPC over stdio).

Actual

• Log line: agent_cmd=.../grok with no agent stdio args.
• Child prints Error: Device not configured (os error 6) (ENXIO) and exits.
• All parallel workers fail; harness exits status 1: all 10 agents failed to start.

Manual repro of the bad spawn:

# bare TUI, no TTY — same ENXIO Buzz sees
/Users/$USER/.local/bin/grok </dev/null
# → Error: Device not configured (os error 6)

# correct ACP entrypoint — works
printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":1,"clientCapabilities":{},"clientInfo":{"name":"test","version":"0"}}}' \
  | grok agent stdio
# → initialize result with protocolVersion 1, authMethods, models, etc.

Root cause (from current main)

Two related gaps:

1. Preset args are not applied when only the command override is set

UI / preferred_runtime = "grok" pins agent_command_override: "grok" but leaves agent_args: [] on the managed agent record. Spawn uses the override binary + the stored (empty) args, so the preset's ["agent", "--always-approve", "stdio"] never reach the process.

2. default_agent_args("grok") does not fill in the preset defaults

In both:

• desktop/src-tauri/src/managed_agents/discovery.rs → default_agent_args
• crates/buzz-acp/src/config.rs → default_agent_args

only goose / Claude / Codex / buzz-agent are special-cased. grok falls through to None, so normalize_agent_args("grok", []) returns [] instead of the preset ACP args.

Same class of risk for other Tier-2 presets that need non-empty args when the agent record has empty agent_args (omp → acp, opencode → acp, kimi → acp, etc.) if spawn only sees the command + empty stored args.

Suggested fix

1. Include Grok (and other preset harnesses) in default_agent_args, mirroring PRESET_HARNESSES, e.g.:

   "grok" => Some(vec![
       "agent".into(),
       "--always-approve".into(),
       "stdio".into(),
   ]),
   "omp" | "opencode" | "kimi" | "cursor" | "cursor-agent" => Some(vec!["acp".into()]),
   // ...

   Prefer a single source of truth: derive defaults from PRESET_HARNESSES rather than a second match table.

2. When applying preferred_runtime / command override from a preset, also persist (or resolve at spawn) that preset's args, not only the command.

3. Regression test: spawn path / normalize_agent_args("grok", vec![]) must yield ["agent", "--always-approve", "stdio"]; integration smoke that BUZZ_ACP_AGENT_COMMAND=grok with empty args can complete ACP initialize without a TTY.

Workaround (works today)

Edit managed agents so args are set explicitly, then restart agents:

"agent_command_override": "/Users/<you>/.local/bin/grok",
"agent_args": ["agent", "stdio"]

or point the override at a tiny wrapper:

#!/bin/bash
exec "$HOME/.local/bin/grok" agent stdio "$@"

(--always-approve is recommended for managed agents so Grok-side tool prompts do not deadlock headless ACP; Buzz already documents this in the Grok preset PRs.)

Related

• Closed feature: feat(agents): add Grok Build as a first-class ACP runtime #2347 (register Grok Build as first-class ACP runtime)
• Closed PRs that defined the intended args: feat(agents): add Grok Build as a first-class ACP runtime #2812, feat(desktop): register Grok Build as native ACP runtime #2918, feat(desktop): register Grok Build as native ACP runtime + onboarding harness #2546 (agent --always-approve stdio)
• Related enhancement (not this bug): Add Grok CLI option #3279 “Add Grok CLI option”
• Note: some earlier Grok PR discussion mentioned /dev/tty / spawn-context issues on the CLI side; current Grok agent stdio completes initialize fine over pipes — the Buzz-side empty-args spawn is sufficient to explain this failure.

Acceptance criteria

• Selecting Grok Build as preferred/agent runtime starts agents without ENXIO
• Harness log shows ACP init for Grok (agentInfo / capabilities), not TUI ENXIO
• Empty agent_args + command grok normalizes to agent --always-approve stdio (or equivalent from preset)
• Unit coverage for normalize_agent_args("grok", [])