sam-grok-worker

name	sam-grok-worker
description	Delegate a bounded implementation task to the fixed Grok worker (grok-4.6) under workspace sandbox and headless execution. Use when a host agent or user needs Grok to implement, fix, or verify a scoped coding task; default effort high unless explicitly supplied.

Sam Grok Worker

Delegate one bounded task to Grok. Keep the calling agent responsible for scope, final decision, publication, and proof.

Non-Negotiable Contract

• Use model grok-4.6 exactly. Do not substitute another model.
• Use effort high unless the user explicitly supplies low, medium, high, xhigh, or max; use the supplied value exactly.
• Do not infer a lower effort from urgency, simplicity, cost, or latency.
• Invoke headless Grok with workspace sandbox and always-approve so the worker can edit inside the working tree without interactive prompts.
• Disable memory, subagents, and auto-update for the worker session.
• Transport the task only through --prompt-file. Headless Grok does not read piped stdin as the prompt.
• Ask for one bounded task. Do not hand over the entire multi-phase workflow or authorization to publish.
• Pass only the minimum context required and exclude secrets or credentials.
• Do not silently fall back when the CLI, model, effort, sandbox, or authentication is unavailable.
• Treat worker output as candidate work, not proof. Verify material claims and diffs before accepting or publishing.

Token Saver inheritance

If RC_TOKEN_SAVER_EXECUTION_RECEIPT_V1 is present, preserve it and the host-provided RC_TOKEN_SAVER_* capability/lane variables in every child process you start. Do not derive a new admission decision, enable a disabled lane, prepend an unmanaged wrapper path, or copy content into the receipt. Malformed/denied/cross-user/provider-mismatched receipts remain raw. Skills, exact-output commands, and provider responses must stay complete; Token Saver is an optimization seam, not permission to truncate output.

When this worker is a controlled nested task, bracket the real worker lifetime with the provider-neutral telemetry bridge, preserving the returned run id:

telemetry_command="${REMOTE_CODE_SUBAGENT_TELEMETRY_COMMAND:-distill}"
child_run="$("$telemetry_command" subagent begin --node '<stable-worker-node-id>')"
# run the bounded worker
"$telemetry_command" subagent end --run-id "$child_run" --status completed

Use failed or cancelled on the corresponding terminal path. If the bridge is unavailable, continue the task raw and report the missing Subagents proof; never fabricate a receipt or mark an unobserved child done.

1. Freeze the Worker Request

Record:

• Exact task goal and acceptance criteria.
• In-scope paths, commands, and surfaces.
• Out-of-scope and no-go surfaces (publish, push, remote comments, secrets).
• Relevant facts, files, and evidence already known.
• Desired return: summary of changes, verification run, residual risks.

If the request contains an explicit effort, preserve it. Otherwise select high. Do not treat words such as “deep”, “quick”, or “careful” as effort values.

2. Materialize the Prompt File

Write the full worker prompt to a temporary absolute path owned by the caller. Do not put the prompt in process argv via -p.

Require the prompt to say:

• Act only as a worker for this bounded task.
• Stay inside the frozen scope and no-go list.
• Do not publish, push, open remote proposals, send messages, or perform other external writes unless the frozen request explicitly authorizes that action.
• Prefer the smallest safe change that satisfies acceptance criteria.
• Return: what changed, how it was verified, residual risks, and blockers.
• Separate completed work from unverified claims.

3. Resolve the Invocation

Run the deterministic resolver before invoking the worker:

SAM_GROK_WORKER_DIR="<absolute directory containing this SKILL.md>"
python3 "$SAM_GROK_WORKER_DIR/scripts/resolve_worker.py" --prompt-file /abs/path/prompt.txt
python3 "$SAM_GROK_WORKER_DIR/scripts/resolve_worker.py" --prompt-file /abs/path/prompt.txt --effort high

Use the first form when effort is absent. Use the second form with the user's exact effort when supplied. The resolver returns an argv array fixed to Grok, grok-4.6, the selected effort, workspace sandbox, always-approve, no-memory, no-subagents, JSON output, and the absolute prompt file.

4. Invoke Safely

Execute the returned argv directly without a shell. Do not rewrite model, effort, sandbox, or approval flags. Do not swap --prompt-file for -p or stdin.

Wait for completion and parse the JSON text field as the worker report. If invocation fails, return the exact blocker and do not retry with another model or effort. Delete the temporary prompt file after the run when it is safe to do so.

5. Reconcile

Inspect the working tree and any verification the worker claims. Reject unsupported claims, scope expansion, invented evidence, and unauthorized external writes. Resolve disagreement with direct proof or present the tradeoff to the user; do not defer blindly.

Output

Return:

• Worker: Grok grok-4.6.
• Effort: selected effort and whether it was defaulted or user-specified.
• Summary: concise report of work performed.
• Changes: files or behaviors touched, if any.
• Verification: proofs run and their results.
• Risks: residual risks, hard assumptions, and blockers.
• Caller decision: accepted, rejected, or unresolved, with reason.

Do not claim the worker ran unless the invocation completed successfully.