Adopt ACP session/resume and the 16-subagent cap in Grok CLI
Grok Build ed6d543 adds ACP session/resume and session/close, caps live workflow subagents at 16 per run, allows model switch during plan approval, fixes sandbox start on large deny-glob workspaces, extracts MCP images before truncation, and streams fork copies for bounded memory.
Why this loop
This release changes multi-agent and long-session workflow limits. Live workflow subagents are hard-capped at 16 per run, so parallel plans must be batched. ACP session/resume and session/close support durable control-plane sessions instead of full restarts. Model switch during plan approval avoids reject-and-restart loops. Sandbox no longer refuses to start when deny globs walk large workspaces, so deny patterns can stay strict. MCP images are extracted before output truncation, and large session forks stream to bound memory—both matter for vision and long-context runs.
Proposed actions
- Upgrade to Grok CLI ed6d543, then on disconnect call ACP session/resume for the same session id instead of opening a new session; call session/close when the task finishes.
- Split any live workflow that spawns more than 16 concurrent subagents into sequential batches of at most 16 to match the new per-run cap.
- At plan approval, switch the model inline when the plan needs a different model rather than rejecting the plan and restarting the run.
- Re-enable the sandbox on large workspaces that use deny globs; confirm startup no longer fails when those globs walk the tree.
- For MCP tools that return images, send image-bearing results through the normal path and verify images are preserved before truncation; confirm file-mention input chunks replay as user attachment links after fork.
Agent prompt
Paste into your agent or query via MCP (`get_agent_prompt`) — free, no extra AI costPaste into Claude Code / CLAUDE.md task
# DevAgentRadar → Claude Code You are helping me adopt a real coding-assistant change. Work only from the facts below. Do not invent features. ## Context Assistant: xAI Grok CLI (Grok Build) Proposal: Adopt ACP session/resume and the 16-subagent cap in Grok CLI Summary: Grok Build ed6d543 adds ACP session/resume and session/close, caps live workflow subagents at 16 per run, allows model switch during plan approval, fixes sandbox start on large deny-glob workspaces, extracts MCP images before truncation, and streams fork copies for bounded memory. Primary source: https://github.com/xai-org/grok-build/commit/ed6d543643628663873c5de28298e022ed634238 ## Why it matters This release changes multi-agent and long-session workflow limits. Live workflow subagents are hard-capped at 16 per run, so parallel plans must be batched. ACP session/resume and session/close support durable control-plane sessions instead of full restarts. Model switch during plan approval avoids reject-and-restart loops. Sandbox no longer refuses to start when deny globs walk large workspaces, so deny patterns can stay strict. MCP images are extracted before output truncation, and large session forks stream to bound memory—both matter for vision and long-context runs. ## Suggested actions 1. Upgrade to Grok CLI ed6d543, then on disconnect call ACP session/resume for the same session id instead of opening a new session; call session/close when the task finishes. 2. Split any live workflow that spawns more than 16 concurrent subagents into sequential batches of at most 16 to match the new per-run cap. 3. At plan approval, switch the model inline when the plan needs a different model rather than rejecting the plan and restarting the run. 4. Re-enable the sandbox on large workspaces that use deny globs; confirm startup no longer fails when those globs walk the tree. 5. For MCP tools that return images, send image-bearing results through the normal path and verify images are preserved before truncation; confirm file-mention input chunks replay as user attachment links after fork. ## Your job 1. Restate the change in one sentence. 2. Propose a minimal plan for my repo (or a throwaway pilot). 3. Implement only what I approve; prefer small diffs and tests. 4. Call out risks (permissions, breaking APIs, cost). Start by confirming you understood the proposal.
Your loop
This browser · no sign-in · not shared as “you”Community signal · 0 approved · 0 deferred · 0 ignored
Anonymous public tally (one vote per visitor; changing choice updates your vote). Not moderated product truth.