Composer 2.5 parallel read-only subagents

Writes stay on the parent. Reads use explore-standard (Composer 2.5, non-Fast).

Ensure agents exist (do this first)

Cursor loads custom subagents from ~/.cursor/agents/ or the project's .cursor/agents/, not from the skill folder.

Before any Task call, check both locations for composer-standard.md and explore-standard.md. If a file exists in either place, leave it.

If composer-standard is missing, write ~/.cursor/agents/composer-standard.md using the exact contents of composer-standard.md in this skill (create ~/.cursor/agents/ if needed). Do not overwrite an existing file. Do not write it only into the current repo unless the user asked for a project agent.

Required reminder: if you just created composer-standard (file did not exist before this turn), stop after writing it and tell the user to restart Cursor so the new subagent is loaded. Do not assume the current chat can use it. Example: “已写入 ~/.cursor/agents/composer-standard.md。请重启 Cursor（关掉再打开），新开一轮对话后再继续查询。”

If explore-standard is missing, write ~/.cursor/agents/explore-standard.md from explore-standard.md the same way. If you created it in this turn, also tell the user to restart Cursor.

After creating files, later chats use explore-standard for reads. Never set model on Task. Never use Fast.

This skill still does not launch composer-standard to implement. Auto-add only registers the agent so the machine has Composer 2.5 non-Fast available.

Pin

When launching Task:

• Reads/queries: subagent_type: explore-standard only (not built-in explore, not generalPurpose).
• Do not set model on Task (parent model overrides the pin and often becomes Fast).
• Never composer-2.5-fast or any Fast variant.

Subagents never edit

Subagents must not write files, change config, or commit. Investigation is read-only.

Do not launch write/coding subagents to implement (generalPurpose, composer-standard, or similar). After explore returns paths/symbols to change, the parent Reads those files and edits them.

Query tasks — launch immediately

Treat as a query task: where/how/who/why, find usages, map a flow, compare modules, explain a failure, answer from the repo.

Default: launch explore-standard immediately. Do not first Grep/Glob/Read in the parent “to see if it is simple.”

Skip a subagent only if the answer is already in the current message, an open file, or a path the user just named — not because the query looks small.

Question plus fix: parallel explore-standard to locate; parent Reads the files to change and implements. If the user already named the exact change, parent Reads those files and edits — no write subagent.

Multiple loci → multiple agents

Split independent axes in one message (several parallel explore-standard calls), not one mega-explorer then serial parent reads:

• definition vs callers vs config vs tests
• frontend vs backend, module A vs module B
• different directories / packages / services

Keep dumps (source, search hits, trees) in subagent context. Parent keeps: paths, how it works, key symbols, next steps.

After subagents return:

• Explain/cite: at most one targeted parent Read/Grep to confirm.
• Files that will be edited: parent must Read them (this is the allowed parent Read, not a “is it simple?” scan).

Subagent prompt

Each read agent must get: scope (dirs/symbols/question); return file list + mechanism + line citations + candidate paths if a change is needed; no whole-file dumps; do not modify the repo. Launch extra axes in the same turn; synthesize when they finish.