feat(agents): add Grok Build (grok) agent support

[linhai0872]

Summary

I'd like to add Grok Build (grok, by xAI) as the 13th supported agent. Grok Build is a first-party terminal coding agent with a documented, open-source on-disk layout — similar in shape to Codex (TOML MCP) + Claude-like hooks, but with its own native enable/disable semantics.

Official product: https://x.ai/build
Source of truth for this proposal: https://github.com/xai-org/grok-build (checked at c2ad97f)

What Grok Build looks like on disk

Home is $GROK_HOME verbatim when set, otherwise ~/.grok (xai-grok-home).

$GROK_HOME/                    # default ~/.grok
  config.toml                  # settings + [mcp_servers.*] + [plugins] + [skills]
  pager.toml
  skills/<name>/SKILL.md
  hooks/*.json                 # Claude-like JSON
  disabled-hooks               # one Grok hook spec.name per line
  plugins/*/plugin.json
  installed-plugins/*/
  agents/*.md
  rules/*.md
  commands/*.md
  memory/
  AGENTS.md

<project>/.grok/
  config.toml                  # only [mcp_servers], [plugins], [permission]
  skills/
  hooks/
  plugins/
  agents/
  rules/
  commands/

Project marker: .grok/. Canonical skill write target: .grok/skills (not .agents/skills).

Proposed first contribution

A full native adapter (crates/hk-core/src/adapter/grok.rs) implementing AgentAdapter:

• Detect $GROK_HOME / ~/.grok without creating the directory.
• Skills — user ~/.grok/skills, project .grok/skills. Does not claim Claude/Cursor/.agents compat dirs (those stay with their own adapters). Disable via [skills].disabled is Grok-native; HK skill toggle can stay the existing rename model for files we own.
• MCP — TOML [mcp_servers.<name>] in user + project config.toml. Remote key is headers (not Codex http_headers); type = "sse" is supported. Native toggle via user-level disabled_mcp_servers + per-entry enabled, matching grok mcp enable/disable. Personal disable must not rewrite shared project files. New McpFormat::GrokToml / RemoteMcpSchema::GrokToml so Codex's writer is never reused.
• Hooks — scan $GROK_HOME/hooks/*.json and .grok/hooks/*.json (Claude-like). Install writes command hooks. Toggle uses $GROK_HOME/disabled-hooks with Grok's real spec name (global/<stem>:<event>[i].hooks[j]). Only $GROK_HOME/hooks is global/; a project under a custom $GROK_HOME stays project/. SubagentEnd uses upstream's subagent_stop display token and shares one group-index sequence. HTTP hooks are out of scope — HK's HookEntry is command-only; they are skipped as rows but still occupy Grok's hook index.
• Plugins — scan $GROK_HOME/plugins/, $GROK_HOME/installed-plugins/, and project .grok/plugins. Toggle via [plugins].disabled / [plugins].enabled using Grok's stable plugin id ({scope}/{hex8}/{name}). Personal toggle writes the user lists only.
• Rules / settings / subagents / memory / commands — Grok-native paths only. Do not surface auth.json or mcp_credentials.json.
• Frontend: AGENT_ORDER, display name Grok Build, colors, LobeHub Icons grok mascot, onboarding, EN/ZH README matrix 12 → 13.

Compat isolation

Grok also reads Claude Code and Cursor skills / MCP / hooks / plugins when [compat.*] cells are on. HarnessKit already owns those through the existing adapters. The Grok adapter will not claim .claude, .cursor, or .agents trees, matching the omp adapter's "don't double-claim" rule.

Known limitations (called out, not silent)

• HTTP hook handlers are not modeled.
• Inline [hooks] tables inside config.toml are not scanned in this first PR (JSON files under hooks/ are).
• Compat-sourced MCP (.mcp.json / Claude / Cursor) is not written through the Grok adapter.

Verification plan

• cargo test --workspace
• npm test + npm run build
• Isolated $GROK_HOME fixtures +, if available, grok inspect / grok mcp list smoke

Happy to do the work. Opening this for visibility; a PR will follow on feat/grok-build-agent-support.

[RealZST]

Thanks for the issue and pr! Shipped in v1.10.0.