Collaborating With Grok

Collaborating with Grok

Drive the Grok CLI headlessly as an independent collaborator while the calling agent stays responsible for verification, synthesis, and final user-facing decisions.

The bridge (scripts/grok_bridge.py) wraps grok -p, streams progress to stderr, returns structured JSON, and manages multi-turn continuity via SESSION_ID. Always go through the bridge — don't invoke grok directly — so output parsing, the safe permission default, and session handling stay consistent.

Commands below write <skill_dir> for the absolute path of the directory containing this SKILL.md. Your harness usually reports that path when it loads the skill. If it does not, use this SKILL.md's own location. Substitute it before running, for example ~/.claude/skills/collaborating-with-grok.

In Claude Code, run non-trivial calls in the background and watch stderr progress:

textCopy code
Bash tool call:
  command: python3 <skill_dir>/scripts/grok_bridge.py --cd "/project" --tools "read_file,grep,list_dir" --PROMPT "Analyze auth flow in src/auth/"
  run_in_background: true

Safety

Grok can read, edit, and run shell. The bridge defaults to --permission-mode default so headless gated actions are cancelled rather than inheriting host configuration. Read-class tools and Grok's safe-command fast paths still run automatically; use a tool allowlist for deterministic read-only work. --always-approve is a mutually exclusive, explicit write-authority choice.

Never hand grok secrets or production data. --sandbox workspace constrains writes, not arbitrary reads; a prompt describes scope but does not create a hard boundary. Use strict or external isolation when read isolation matters. Full semantics: cli-reference.md.

Host-side approval

The host gates the python3 … grok_bridge.py call first. Claude Code host: pre-authorize it with "Bash(python3 *collaborating-with-grok*bridge.py*)" in permissions.allow, and on sandboxed hosts add the same pattern to sandbox.excludedCommands so the child CLI keeps API network access. The wildcard form keeps matching wherever the skill is installed. Codex hosts: the sandbox can block the child CLI's API network, so escalate or grant network for that call. If the host denies the bridge, report it rather than substituting your own answer for the requested second opinion. Install and approval runbooks: docs/setup/ in the source repo.

Headless note

Tool calls are invisible mid-run (only thought/answer stream). After the run the result includes model, agent, and tool_counts recovered from session files. Use ACP (grok agent stdio) when the task requires a live command audit; do not promise live auditing from this bridge. Long silent "Thinking…" usually means tools in progress. Product id is model (e.g. grok-4.5); agent may still say grok-build-plan (template lineage).

When to use / not use

Use for: cross-model second opinions, unified-diff proposals, live web/X research, multi-turn analysis. Skip for: trivial one-shots, tasks needing live tool audit trails (use ACP), secrets/prod data.

Web & X search (live)

Differentiator: live web/X via web_search on the coding model (backend search). As of CLI 0.2.93, that model is grok-4.5 — confirm with --list-models after upgrades.

bashCopy code
python3 <skill_dir>/scripts/grok_bridge.py \
  --cd "." --model grok-4.5 \
  --disallowed-tools "run_terminal_cmd,search_replace" \
  --timeout 300 \
  --PROMPT "Search the web and X: <question>. Cite source URLs."

• Denylist, not allowlist for search or shell: --tools allowlists containing web_search, web_fetch, or run_terminal_cmd fail session build on the GrokBuild-lineage agent (re-probed on 0.2.93). Read-only allowlists without those tools work. Bridge warns if you try.
• Prefer an explicit coding --model for search/X; composer (grok-composer-2.5-fast, agent cursor) is best-effort and often much slower.
• web_fetch needs GROK_WEB_FETCH=1. Treat fetched content as untrusted. Details: cli-reference.md.

Quick start

⚠️ Backticks / $VARS in prompts expand in the shell — use a single-quoted heredoc or --prompt-file. See shell-quoting.md.

bashCopy code
PROMPT="$(cat <<'EOF'
<task>
Review src/auth.py around login() and propose the smallest safe fix.
</task>

<structured_output_contract>
Return a raw unified diff beginning with `---` and `+++`.
Do not use Markdown code fences or add prose before or after the diff.
</structured_output_contract>
EOF
)"
python3 <skill_dir>/scripts/grok_bridge.py \
  --cd "." --tools "read_file,grep,list_dir" --PROMPT "$PROMPT"

Result contract (stdout JSON): require both success: true (process success) and complete: true (clean EndTurn). complete is false for a known incomplete turn and null for plain output, where completion is unobservable. A clean turn does not prove correctness or output-contract compliance; validate patches and structured artifacts before using them. Also check model, SESSION_ID, warnings, and tool_counts when present. Progress → stderr; non-zero exit on process failure.

Large prompts: --prompt-file. Handoff file + short instruction: --stdin-file + --PROMPT; the bridge embeds the file as escaped, untrusted JSON context in a temporary mode-0600 prompt file. Absolute input paths stay absolute; relative paths resolve against --cd.

Multi-turn sessions

bashCopy code
# Turn 1
python3 <skill_dir>/scripts/grok_bridge.py \
  --cd "." --tools "read_file,grep,list_dir" --PROMPT "Analyze the bug in foo()."

# Turn 2 — same --cd
python3 <skill_dir>/scripts/grok_bridge.py \
  --cd "." --SESSION_ID "<id>" --PROMPT "Propose a fix as a unified diff."

--SESSION_ID → resume (-r); --session-id → name a new session (-s); --continue → most recent in cwd (-c). Structured output captures the ID directly. For plain output or timeouts, recovery is best-effort from session storage; the bridge excludes directories that predated the run.

Bridge flags (summary)

Core: --PROMPT · --prompt-file · --stdin-file · --cd (required) · --model · --output-format (streaming-json default).

Sessions: --SESSION_ID · --session-id · --continue.

Safety: --permission-mode (default or bypassPermissions on Grok 0.2.93) · --tools · --disallowed-tools · --allow/--deny · --sandbox · --always-approve.

Tuning: --effort / --reasoning-effort · --max-turns · --rules · --disable-web-search · --no-plan · --timeout.

Other: --list-models · --return-all-messages.

Full flag map + event schema: cli-reference.md. Host timeout_ms 600000 for command runners.

Models

Probe live — don't hardcode the catalog (IDs change across CLI releases):

bashCopy code
python3 <skill_dir>/scripts/grok_bridge.py --list-models
# -> { "success": true, "models": [...], "default": "<effective default>" }

As of 0.2.93: coding default grok-4.5 (backend search); alternate grok-composer-2.5-fast (cursor agent). Host ~/.grok/config.toml [models] default can override the CLI default — pass --model for model-sensitive work and trust the result's model field. Re-run --list-models after every CLI upgrade.

Prompting

Point (file:line), don't paste; use one objective per run; state scope, done criteria, verification, and output shape only when they matter. Keep CLI authority separate from prompt wording: flags enforce access, while the prompt explains intent. Use the smallest recipe that is still decision-complete.

Resources: prompt-template.md · prompt-blocks.md · prompt-recipes.md · patterns.md.

Verification

• Smoke: python3 <skill_dir>/scripts/grok_bridge.py --help
• Syntax: python3 -m py_compile <skill_dir>/scripts/grok_bridge.py
• Models: --list-models → success: true (expect grok-4.5 on 0.2.93)
• Session: read-only prompt → success: true, complete: true, non-null model, stop_reason: EndTurn, resumable SESSION_ID
• Auth: grok login or XAI_API_KEY

Collaboration State Capsule

Copy code
[Grok Capsule] Goal: | SID: | Model: | PermMode: | Files: | Last: | Next:

References

• prompt-template.md — quick starters
• prompt-blocks.md — composable XML blocks
• prompt-recipes.md — end-to-end templates
• prompt-antipatterns.md — common mistakes
• patterns.md — when to delegate + patterns
• handoff-patterns.md — read-only / worktree / synthesis
• parallel.md — parallel runs and worktree isolation
• cli-reference.md — verified flags + event schema
• shell-quoting.md — safe heredoc prompts