subagent-models

A Claude Code mod that decides which models and agent types subagents may run as, and the effort they run at.

What it does

• A spawn that names a turned-off model runs on the default model instead. With offAction set to deny, the spawn is refused with a message that names the model to use.
• A full model id counts by its family. claude-fable-5-1 is fable.
• A turned-off agent type leaves the model's agent listing, and a spawn that names it is refused.
• A turned-off skill still appears in the model's skill listing. When the model or you run it, the skill's text is replaced with a note that says it is off, so the model skips it. A short name such as tdd also matches pstack:tdd.
• The default model is always allowed, so a turned-off model always has somewhere to move.
• Every request a subagent makes runs at the effort you set. The main conversation keeps its own model and effort.
• With applyToRunning on, an effort change reaches running subagents on their next request. With it off, each subagent keeps the effort it started with until apply, or the bar's apply now button, moves them all.
• A subagent whose model comes from its agent definition is caught on its first request and moved the same way.
• A fork inherits its parent and passes untouched. So does a spawn that names no model.
• The footer pill shows the default model and effort, as opus/medium. A * marks a session that differs from the defaults. Clicking it opens or closes the bar.

Settings

/config holds the defaults every new session starts with:

| Setting | Options | Default | |---|---|---| | Opus subagents | on, off | on | | Fable subagents | on, off | off | | Sonnet subagents | on, off | off | | Haiku subagents | on, off | on | | Default subagent model | opus, sonnet, haiku, fable | opus | | Subagent effort | inherit, low, medium, high, xhigh, max | medium | | When a spawn names a turned-off model | move, deny | move | | Effort changes reach running subagents | on, off | on | | Turned-off agent types | comma-separated agent types, such as Explore, Plan | none | | Turned-off skills | comma-separated skills, such as pstack:tdd, simplify | none | | Days before a skill counts as unused | a whole number of days, the starting cutoff of the unused after menu | 14 |

The command refuses to turn off the default model, and refuses a default model that is turned off.

This session only

/subagent-models changes the current session from its next spawn, and leaves other sessions alone.

• /subagent-models prints the values in force and marks the ones this session changed.
• /subagent-models fable on sets one value for this session. Any setting name works, with one of its options.
• /subagent-models agents lists the agent types offered so far in this session, each with its switch.
• /subagent-models agent Explore off turns one agent type off for this session. on turns it back on.
• /subagent-models skills lists the skills the session has, each with its switch. /subagent-models skill pstack:tdd off turns one skill off for this session.
• A framed bar above the prompt holds every setting in a grid of four equal columns. The header row has the title, a summary such as opus · medium · 2 of 4 models on, and a minimize button. Under it a strip of four colored segments shows which models are on, and a row of four model entries below it has a dot, the name and an on or off button each. The next row has dropdowns for the default model, effort, the turned-off action, and running subagents. The last row has the agents and skills buttons, apply to running, and a more menu that saves or resets. A pill in the footer, ◈ opus/medium ▴, opens and closes the bar with one click.
• The agents and skills buttons open a panel with enable all, disable all, and one row per agent type or skill: a dot, the name, and an on or off button at the right edge. Every row sits in a cell as wide as the longest name, so the rows line up in columns. A name longer than 44 characters keeps its tail and loses its middle. The rows sit under a small-caps heading per namespace, such as PSTACK or BRAND-VOICE, with a count. Press a heading to fold or open its group. A group of more than 16 starts folded, and opens 40 rows at a time with a show more button. A list of more than 16 opens in a side pane beside the transcript, laid out in as many equal columns as the pane is wide. The button shows ▸ for a side pane and ▴ for an inline panel. A panel stays open across picks until you press its button again or close the pane. A pick applies at once and a toast says what changed.
• The mod records the time of each skill run in its own store, so the skills listing and panel can show when a skill was last used. The skills panel has sort: recent, which groups the chips as used in the last N days, unused N+ days, and never seen, newest first, with the age on each chip. The unused after menu picks the cutoff from 14, 28 or 56 days, starting at the staleDays setting, and turn off unused Nd+ (count) turns off the skills with a recorded date older than that cutoff. A skill with no recorded date is never turned off by that button.
• /subagent-models bar minimizes the bar or expands it.
• /subagent-models seed <file> loads last-used dates from a backfill file into the mod's own store. The store lives in the Claude config folder of the account that runs the command, so run it once in each account.
• /subagent-models apply moves every running subagent to the current effort now.
• /subagent-models save writes this session's values to the /config defaults.
• /subagent-models reset drops this session's values, so the defaults apply again.

Backfill last-used dates

The mod only records runs from the day it loads. scripts/backfill-last-used.mjs reads the saved transcripts of every Claude account on the machine and writes the newest date of each skill run, for skills the model ran and slash commands you typed:

node packages/subagent-models/scripts/backfill-last-used.mjs --out dates.json

It scans every ~/.claude* folder and every ~/.claude-profiles/* folder that holds a projects folder, skips a folder that is a link to one it already scanned, and prints the transcript and skill-run count for each. Add --root <config dir> to scan only the folders you name, or --exclude <text> to skip a folder whose path contains the text. Then run /subagent-models seed dates.json in each account. Run the script's tests with node --test packages/subagent-models/scripts/backfill-last-used.test.mjs.

Install

A full npx claude-dev-env install adds the claude-dev-env marketplace and installs the mod at user scope, so every Claude Code session under that config folder loads it. --no-subagent-models or CDE_INSTALL_SUBAGENT_MODELS=0 skips that step.

The installer removes and reinstalls the plugin on each run. Claude Code drops a plugin's saved /config values when it uninstalls the plugin, so after an installer run the defaults in the Settings table apply again.

To install it by hand:

claude plugin marketplace add jl-cmd/claude-dev-env
claude plugin install subagent-models@claude-dev-env

Or load it from a clone for one session:

claude --plugin-dir packages/subagent-models

Test

claude plugin validate .
claude plugin test .

Built and tested on Claude Code 2.1.288. The mod API is in early access and can change between releases.