Frugal Subagents

Frugal Subagents

Why

Delegation is where a session's budget leaks. Subagents inherit the parent's model unless told otherwise, so a top-tier session that fans out three research agents is running three more copies of its most expensive model — and if one of them fans out again, the cost multiplies silently. On subscription plans that can exhaust the limit in minutes, with most results lost to rate limits. Written rules ("use sonnet for scans") don't survive context compaction; this plugin turns them into a hook, which does.

What the hook enforces

Every Agent call passes through hooks/guard.js before it runs. Nothing is silent: when the hook fills in a model it says so in the tool result, and a blocked call returns the reason as tool feedback.

Variables go into the user's or project's settings.json under "env", or the shell environment. The user changes them, not the session: a denied call is policy, and splitting or rephrasing a call to get around it is not an option.

How to delegate under the guard

Pick the tier by what the work costs, not by how important it feels. Search/fetch volume, page scanning, listing comparison, extraction, renames, tests-to-a-spec → sonnet or haiku (leaving model out gives the default, sonnet). The session's own top-tier model is for judgment that cannot be delegated — decomposition, contested decisions, final synthesis — and normally stays in the main session; name it explicitly on a subagent only when that judgment genuinely has to run in a separate context.

Use the bundled workers before writing a custom brief:

• web-scout (sonnet, web + files, no Bash, no Agent) — research passes. Give it the question, the target list or search terms, and the file to write into.
• extractor (haiku, files only) — turning collected material into tables, lists, JSON, deduplicated notes.

Both write their full output to files and return a short digest, so the expensive context receives conclusions, not dumps.

Brief so the worker can't drift: goal, scope (files, sites, count), constraints, definition of done, the output path, and the user's request quoted verbatim where it matters. Workers don't see the conversation.

One agent, then reuse it. A follow-up in the same area goes to the agent already running, not to a new spawn — that keeps its context and stays within the session budget. Don't spawn to double-check yourself, and don't spawn what a handful of your own tool calls would finish.

Results go to files first. Anything worth keeping is written to disk by the worker as it goes; a worker cut off by a rate limit then loses minutes, not the pass.

When the guard denies a call, do what the reason says: finish inline, reuse a running agent, or tell the user which limit is in the way and let them decide. Never work around a hook.

Workflows are outside the guard. Agents started by the Workflow tool (agent() calls in a workflow script) don't pass through the Agent hook. On a top-tier session, before running any workflow, put model: "sonnet" (or "haiku") in the options of every agent() call in the script unless that step genuinely needs top-tier judgment, and keep the agent count of the script in view — a research workflow can spawn a hundred agents in minutes.

Talking to a non-technical user

Many users of this plugin are not developers — they delegate web research (housing, travel, purchases, suppliers). With them, say "helpers" rather than "subagents", name the model plainly ("the cheaper model, Sonnet"), and never quote hook JSON. When a limit kicks in, one sentence: what stopped, why, what you'll do instead, and — only if it matters — which setting changes it. If the user wants a setting changed, offer to edit settings.json → "env" for them rather than explaining where the file is.

If the guard can't run (Node.js missing)

The hook is a Node.js script. If Node.js is not installed, Claude Code shows a hook error when a subagent is spawned (typically node not found / not recognized) and the spawn proceeds unguarded. When you see that:

1. Tell the user in one plain sentence: the plugin's guard needs Node.js, it isn't installed, so the limits aren't enforced yet — the guidance still applies and you'll follow it manually.
2. Offer to install Node.js LTS and wait for a yes. Use the platform's package manager: Windows winget install OpenJS.NodeJS.LTS; macOS brew install node (or the installer from nodejs.org if Homebrew is absent); Debian/Ubuntu via the NodeSource setup script or sudo apt install nodejs npm; otherwise point to nodejs.org. Don't install without consent, and don't use sudo without saying so.
3. After installation, the user starts a new Claude Code session; the guard is active from then on. Until then, apply this skill's rules yourself: cheap tier, no nesting, results in files.

Companion settings (first-party, optional)

Claude Code has its own levers that combine well with the guard; recommend them when the user asks how to lock things down harder, and put them in settings.json → "env":

• CLAUDE_CODE_SUBAGENT_MODEL=sonnet — Claude Code's own default for subagents whose definition doesn't set a model; the guard's default does the same job at the call level and additionally tells the session which model was used. CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 (Claude Code 2.1.257+) makes that model mandatory for every subagent, overriding both agent definitions and per-call choices — the strictest setting, at the cost of any explicit override.
• CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 — nesting depth (default 3); with 1, subagents don't even receive the Agent tool.
• CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS (default 20), CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION (default 200).

Verify the hook

After installing or editing the plugin, confirm the guard answers as expected (Node.js required — it is what runs the hook):

node -e "
const {spawnSync}=require('child_process');
const run=(ev)=>spawnSync('node',[process.argv[1]],{input:JSON.stringify(ev),encoding:'utf8'}).stdout;
const base={hook_event_name:'PreToolUse',tool_name:'Agent',session_id:'selftest-'+Date.now()};
const d=(s)=>{try{return JSON.parse(s).hookSpecificOutput}catch{return null}};
const a=d(run({...base,tool_input:{prompt:'x'}}));
console.log('no model  →', a&&!a.permissionDecision&&a.updatedInput&&a.updatedInput.model==='sonnet' ? 'defaulted to sonnet' : 'UNEXPECTED');
console.log('opus      →', run({...base,tool_input:{prompt:'x',model:'opus'}})==='' ? 'passed through' : 'UNEXPECTED');
const n=d(run({...base,agent_id:'a1',tool_input:{prompt:'x',model:'haiku'}}));
console.log('nested    →', n&&n.permissionDecision==='deny' ? 'denied' : 'UNEXPECTED');
" <path-to-plugin>/hooks/guard.js

Expected: defaulted to sonnet, passed through, denied. Live check: with the plugin loaded, ask Claude to spawn a subagent without naming a model — it runs on sonnet, Claude sees a note saying so, and the user gets a one-time notice about the defaults for the session.