/model-switch

name: model-switch
description: "Change the core's model without the CLI: send /model to the live core's pane through the shared sender, wait for the CLI to accept it (answering its confirm dialog only under --confirm), record only once accepted; persistence is the CLI's own."
user-invocable: true

model-switch

`scripts/switch-model.sh <model> [--dry-run] [--confirm] [--accept-timeout S]` — `<model>` is one of the CLI's aliases (`default|opus|sonnet|haiku|fable`) or a `claude-*` id, either with an optional `[1m]` tag (`default` takes none); anything else is refused before any write (exit 2).

Order, under one lock per brain (`<workspace>/state/.model-switch.lock`, held across preflight, send, observation and record so overlapping switches linearize): (1) preflight through the shared sender (`scripts/tmux-send-line.sh --dry-run --refuse-if-pending`) — refuse on a Codex runtime (exit 4) or when the core's input box carries text (exit 5), nothing written; exit 3 when no live pane exists; (2) count the `Set model to` lines already on screen (`scripts/pane-observe.sh --count`), so a stale acceptance from an earlier switch cannot pass as this one; (3) `/model <model>` + Enter through the shared sender; (4) **wait for the CLI to accept**: a NEW `Set model to` line is the switch. A warm core (cached conversation) first shows *"Yes, switch / No, go back"*; with `--confirm` — pass it on an owner instruction, the flag is the authorization — the helper answers Enter once and waits again; without it the dialog is cancelled (Escape) and the script exits 6 with nothing recorded. No acceptance within `--accept-timeout` (default 20 s) is exit 8, nothing recorded. (5) Only then record `<workspace>/state/model-switch.json` `{model, previous, previous_source, accepted, confirmed, ts, by, settings_read}` — `previous` from the runtime's `settings.json`, else the last record (the CLI does not persist every `/model` pick, so the file can lag the live session), else null. **The script never writes `settings.json`: Claude Code's own `/model` persists the choice there** (measured 2026-09-04: the key changed hours into a session with no repo writer). Measured the same day on a live core: the send alone exited 0 while the CLI sat at the confirm dialog and the model had not changed — which is why the send is treated as initiation and the record follows acceptance.

**`--offline` (core down).** With no core process for the session there is no CLI to take `/model`, so `--offline` writes `model` into the runtime's `settings.json` (other keys kept) and records `model-switch.json` with `accepted: false`; the core starts on it at its next launch. It is refused (exit 9) while a `claude` process named for the session runs (the launcher's own test, from `src/agent/claude/cli/session-launch.sh`; the tmux server, whose argv also names the session, is not a core), and never overwrites a settings file it cannot read (exit 1). The launcher's provider-limit hold uses it for the owner's "switch to <model>" reply.

**Held for a provider limit (no flag needed).** When no core process runs and `src/core_limit_hold.py held` reports a hold in force, a plain switch (the desktop's Model picker) takes the `--offline` write, releases the switch lock, and starts the core through `src/agent/start-cli.sh` (`--launcher` overrides it), whose limit gate lets a launch on another model family through. Exit 0: the core is starting on the new model; exit 10: the model is set but the launcher did not start the core (its reason and log path on stderr). Without a hold, a down core is never started from here.

**A picker left on screen.** Right before the send the script writes an attribution record through `src/self_opened_gate.py` (`<workspace>/state/self-opened-gate.<session>.json`) and clears it once the CLI accepted or the script cancelled its own dialog. On exit 8 the record stays, its claim window closed a few seconds after the exit (a `/model` the owner opens after that is never attributed to the script): if the send left a picker up (e.g. *"Enter to set as default · s to use this session only · Esc to cancel"*), the core supervisor (`src/core-input-watch.py`) presses Escape once after it has sat unchanged for `MODEL_SWITCH_PICKER_DISMISS_AFTER_S` (manifest default 300; env, then `--picker-dismiss-after`, override; 0 = never). A picker first seen outside the record's claim window, one someone is moving through, a picker with no record (a human opened it) and any usage/spend gate are never touched; until then the owner's chat card can still answer it.

Defaults come from the runtime descriptor (`sutando-config.sh runtime`: `brain`, `socket`, `session`); `--brain/--socket/--session/--descriptor-file` and `SUTANDO_TMUX_*` are explicit overrides.

Python comes from `sutando-config.sh python-bin`, never bare `python3`. The core-side entry `scripts/switch-model.sh` only execs this script. An owner message like "switch model to opus" is handled by running it; the dashboard's Quota tile shows the model the proxy then sees.