main.js · part 217

60 text occurrences from desktop/Cursor.app/Contents/Resources/app/extensions/cursor-agent-host/dist/main.js, part 217. Every entry preserves the shipped literal and its saved verdict or selection reason.

File contents and all parts · All files

Branch the subagent's own generated branch starts from. Defaults to the current branch. Uses the remote version, so uncommitted or unpushed work is not visible.

Exact environment build id (e.g. bld-YYYYMMDD-<uuid>) for the subagent's VM to boot from, instead of the environment's latest successful build. Must belong to the same team and environment; an invalid or inaccessible build fails the subagent.

Run on a dedicated cloud VM with its own clone and generated branch. After the subagent finishes, follow user instructions on whether to merge that branch, check it out, or neither.

Optional placement: where the subagent runs. Omit to run on this machine, which is almost always right. ONLY set this if the user explicitly asks for a cloud subagent or a specific machine.

Run the agent in the background (returns output_file path to check later).

Run the agent in the background. A background subagent cannot be polled or awaited; after spawning it, continue other work or end your turn, and its final result will be delivered to you automatically when it completes.

 Defaults true.

 If this is false, you will be blocked until the agent completes.

 If the user is currently in Multitask Mode, always set this parameter to True.

 When true, the background subagent will send a notification when it completes.

 When true, the background subagent will send a notification when it completes. That notification includes a user-visible summary portion; do not summarize or restate a completed background subagent's result unless the user asks, multiple background subagents need synthesis, or a background subagent reports a blocker requiring parent action outside of the user-visible high level summary.

This retired subagent type can only resume an existing child

subagent_type is required when creating a new subagent

cloud_base_branch may only be specified when environment equals cloud

cloud_requested_environment_build_id may only be specified when environment equals cloud

A short, user-friendly title for the subagent. This appears in the UI as the subagent's name. Make it concrete and distinct, consider recent titles to avoid reuse. For resumed subagents which you are prompting to work on a separate task, give an updated description based on the latest work the subagent is performing. (Do not rename if the subagent is continuing work on the same high-level task.)

  - If you want to read a specific file path, use the ${d}${p?` or ${p}`:""} tool instead of the ${l} tool, to find the match more quickly

 or ${p}

  - If you are searching for code within a specific file or set of 2-3 files, use the ${d} tool instead of the ${l} tool, to find the match more quickly

  - If you are searching for a specific class definition like "class Foo", use the ${p} tool instead, to find the match more quickly

- When the agent is done, it will return a single message back to you. Specify exactly what information the agent should return back in its final response to you.

- When the agent is done, it will return a single message back to you. Specify exactly what information the agent should return back in its final response to you. Background subagent completion messages already include a user-visible summary portion; do not summarize or restate a single background subagent's result by default. Respond only when the user asks, multiple background subagents need synthesis, or the background subagent reports a blocker requiring parent action outside of the user-visible high level summary.

- When the agent is done, it will return a single message back to you. Specify exactly what information the agent should return back in its final response to you. The result returned by the agent is not visible to the user. To show the user the result, you should send a text message back to the user with a concise summary of the result.

- Agents can be resumed using the `resume` parameter by passing the agent ID from a previous invocation. ${n||s?o?"This sends a follow-up message after the agent has completed, preserving existing context. If the agent is still running, the request fails unless `interrupt` is true. Set `interrupt` to true only when the user explicitly wants to interrupt the running agent.":"This sends a follow-up message after the agent has completed, preserving existing context. If the agent is still running, the request fails; wait for completion before resuming.":"When resumed, the agent continues with its full previous context preserved."}${c?` You can also set \`resume\` to "${bB}" to fork the current parent agent into a new child subagent.`:""} When NOT resuming, each invocation starts fresh and you should provide a detailed task description with all necessary context.

This sends a follow-up message after the agent has completed, preserving existing context. If the agent is still running, the request fails unless `interrupt` is true. Set `interrupt` to true only when the user explicitly wants to interrupt the running agent.

This sends a follow-up message after the agent has completed, preserving existing context. If the agent is still running, the request fails; wait for completion before resuming.

When resumed, the agent continues with its full previous context preserved.

 You can also set `resume` to "${bB}" to fork the current parent agent into a new child subagent.

Launch a new agent to handle complex, multi-step tasks autonomously.

The ${l} tool launches specialized subagents (subprocesses) that autonomously handle complex tasks. Each subagent_type has specific capabilities and tools available to it.

When using the ${l} tool, you must specify a subagent_type parameter to select which agent type to use.${t?`\n${h?`\nVERY IMPORTANT: When broadly exploring the codebase to gather context for a large task, it is recommended that you use the ${l} tool with subagent_type="${rg}" instead of running search commands directly.\n`:""}\nIf the query is a narrow or specific question, you should NOT use the ${l} and instead address the query directly using the other tools available to you.\n\nExamples:\n- user: "Where is the ClientError class defined?" assistant: [Uses Grep directly - this is a needle query for a specific class]\n- user: "Run this query using my database API" assistant: [Calls the MCP directly - this is not a broad exploration task]\n- user: "What is the codebase structure?" assistant: [Uses the ${l} tool with subagent_type="${rg}"]\n\nIf it is possible to explore different areas of the codebase in parallel, you should launch multiple agents concurrently.`:""}

When NOT to use the ${l} tool:
- Simple, single or few-step tasks that can be performed by a single agent (using parallel or sequential tool calls) -- just call the tools directly instead.
${f.length>0?`- For example:\n${f}`:""}

Usage notes:
- Always include a short description (3-5 words) summarizing what the agent will do
- Launch multiple agents concurrently whenever possible, to maximize performance; to do that, use a single message with multiple tool uses.
${g}${y}
${u?a?"- If you mention an agent or subagent in your response, link it with the `[Name](id)` Don't use generic label such as `[agent]`, `[worker]`, or `[subagent]`. For cloud subagents, when the agent has edited code, link to `[Review](bc-id#changes)`, or, if you know the exact added and deleted line counts, `[Review +A −D](bc-id#changes)`, replacing A and D with those counts. Never write A or D literally. Use `[Try Live](bc-id#desktop)` only when the agent used computer use.\n":"- In user-facing responses, you may link to agents and subagents with markdown chat links in the `[label](id)` format, using the agent ID as the link target. Do not print raw agent IDs separately.\n":""}- When using the ${l} tool, the subagent invocation does not have access to the user's message or prior assistant steps. Therefore, you should provide a highly detailed task description with all necessary context for the agent to perform its task autonomously.
- The subagent's outputs should generally be trusted
- Clearly tell the subagent which tasks you want it to perform, since it is not aware of the user's intent or your prior assistant steps (tool calls, thinking, or messages).
- If the subagent description mentions that it should be used proactively, then you should try your best to use it without the user having to ask for it first. Use your judgement.
- If the user specifies that they want you to run subagents "in parallel", you MUST send a single message with multiple ${l} tool use content blocks. For example, if you need to launch both a code-reviewer subagent and a test-runner subagent in parallel, send a single message with both tool calls.
- Avoid delegating the full query to the ${l} tool and returning the result. In these cases, you should address the query using the other tools available to you.

${h?`\nVERY IMPORTANT: When broadly exploring the codebase to gather context for a large task, it is recommended that you use the ${l} tool with subagent_type="${rg}" instead of running search commands directly.\n`:""}
If the query is a narrow or specific question, you should NOT use the ${l} and instead address the query directly using the other tools available to you.

Examples:
- user: "Where is the ClientError class defined?" assistant: [Uses Grep directly - this is a needle query for a specific class]
- user: "Run this query using my database API" assistant: [Calls the MCP directly - this is not a broad exploration task]
- user: "What is the codebase structure?" assistant: [Uses the ${l} tool with subagent_type="${rg}"]

If it is possible to explore different areas of the codebase in parallel, you should launch multiple agents concurrently.

VERY IMPORTANT: When broadly exploring the codebase to gather context for a large task, it is recommended that you use the ${l} tool with subagent_type="${rg}" instead of running search commands directly.

- For example:
${f}

- If you mention an agent or subagent in your response, link it with the `[Name](id)` Don't use generic label such as `[agent]`, `[worker]`, or `[subagent]`. For cloud subagents, when the agent has edited code, link to `[Review](bc-id#changes)`, or, if you know the exact added and deleted line counts, `[Review +A −D](bc-id#changes)`, replacing A and D with those counts. Never write A or D literally. Use `[Try Live](bc-id#desktop)` only when the agent used computer use.

- In user-facing responses, you may link to agents and subagents with markdown chat links in the `[label](id)` format, using the agent ID as the link target. Do not print raw agent IDs separately.

 (Auto-resumes most recent agent of this type; `resume` arg is ignored)

 (Auto-resumes most recent agent; `resume` arg is ignored)

- ${r}: ${s}

Available subagent_types and a quick description of what they do:
${e}

Available subagent_types and descriptions are listed in <available_subagent_types> in the initial user-info message at the start of this conversation${A}.

No alternative models are available. Subagents will inherit the parent model.

If the user explicitly asks for the model of a subagent/task, you may ONLY use model slugs from this list:
${t?"- inherit (default; required unless the user explicitly requested another model)\n":""}${[...e.modelsBySlug.keys()].sort().map(e=>`- ${e}`).join("\n")}

If the user isn't asking for a specific version, prefer the latest version of the model family. As an example, if the user just says "gpt" or "claude", use the latest available version of GPT or Claude.

IMPORTANT: If the user requests a model that is NOT in the list above, do NOT substitute a different model or guess. Instead, skip launching the subagent with that model and tell the user which model was unavailable and which models are available.

When speaking to the USER about which model you selected for a subagent, do NOT use the kebab-case model names unless the user requested the model using that format. Ue the same naming scheme the user used to discuss the model when they requested it.

- inherit (default; required unless the user explicitly requested another model)

Available model slugs for subagents are listed in <available_subagent_models> in the initial user-info message at the start of this conversation${A}.

When an agent runs in the background, you will be automatically notified when it completes after you end your own turn - do NOT ${D?`${e} or poll it`:`${e}, poll, or proactively check on its progress`}. Continue with other work or end your turn instead. Don't mention this to the user.

${e} or poll it

${e}, poll, or proactively check on its progress

When an agent runs in the background, you will be automatically notified when it completes after you end your own turn - do NOT ${D?"poll it":"poll or proactively check on its progress"}. Continue with other work or end your turn instead. Don't mention this to the user.

poll it

poll or proactively check on its progress

IMPORTANT: Do NOT use this tool unless the user has explicitly asked you to use subagents, delegate to agents, or use the ${ie} tool. You should perform tasks directly using your own tools instead of delegating to subagents. Only use this tool when the user specifically requests it.

${e.path.length>0?e.path.join("."):"argument"}: ${e.message}

Invalid preToolUse updated_input:
${e.join("\n")}

Getting Task tool config

Failed to get Task tool config

Error executing subagentStart hook

Invalid arguments:
Self-hosted machine placement requires server-side first-class subagent execution and is not supported by a fresh client-executed Task.

Invalid arguments:
Unsupported machine placement for a fresh client-executed Task: ${JSON.stringify(t)}.

Using client-side subagent execution

You've made {idle_count} responses without tool calls. If you're done, respond with exactly: {escape_token}
If not, continue working.

The following background agents have completed:

{summaries}

Review their results. If more work is needed, spawn additional workers. Otherwise, wrap up.