Claude Code. Any Model. The most powerful AI coding agent now speaks every language.
/**
* should-open-picker.ts — the model picker's launch gate, as one pure predicate.
*
* THE DEFECT THIS CLOSES. Until now nothing asked whether a terminal was there.
* `index.ts` opened the picker on `cliConfig.interactive` alone, and `cli.ts`
* sets that flag true whenever there is no positional prompt, no `--stdin` and
* no `-p`:
*
* if (!config._hasPositionalPrompt && !config.stdin && !config._hasPrintFlag)
* config.interactive = true;
*
* So `claudish < /dev/null`, `claudish` on a CI runner, and `claudish` detached
* all reached a prompt that drew into a stream nobody reads and then blocked on a
* keypress that could never arrive. `-p` was already excluded — `_hasPrintFlag`
* exists precisely to stop that default flipping on — and
* `ai-docs/architecture/headless-vs-interactive.md` is why that matters: `-p
* --input-format stream-json` is the mode where a wrong flag yields silence and
* exit 0, so the picker must never appear in it. This predicate makes the piped
* and detached cases fail the same way, by asking the one question that was
* missing.
*
* WHY NOT A FALLBACK TO THE LINE-ORIENTED PROMPT. Inquirer in a non-TTY is the
* bug, not the safety net: it draws into an unread stream and waits for a key
* that cannot come. "Falling back" would be falling back to a hang. So
* `requiresExplicitModel` routes the same cases to the existing non-interactive
* error — four actionable lines naming `--model`, `CLAUDISH_MODEL`, `--profile`
* and `claudish --models` — and exit 1.
*
* THE TTY ANSWER IS A PARAMETER, NOT AN IMPORT. Both functions here are total
* over their inputs and read nothing from the process, which is what makes the
* whole 2^5 grid table-testable without running `runCli`. `index.ts` supplies the
* answer from `canDrawTui()`, the single isTTY oracle next door. For the same
* reason as that file: THIS MODULE IMPORTS NOTHING, and keeps `index.ts`'s cold
* static graph clean.
*/
/**
* Everything the gate is allowed to look at, derived at the call site from
* `ClaudishConfig`.
*
* `hasProfileTiers` is an input rather than four fields so the predicate's domain
* stays small enough to enumerate: five booleans, which is the 32-cell grid its
* test walks, plus `advisorNativeSession`, which the test pins as behaving exactly
* like `monitor` rather than doubling the grid to 64 rows.
*/
export interface ModelPickerGate {
/**
* `cli.ts`'s interactive default: no positional prompt, no `--stdin`, no `-p`.
* False means some other channel already carries the prompt, so there is no
* point in the launch sequence at which a human is waiting to choose.
*/
interactive?: boolean;
/** `--monitor` proxies to the real Anthropic API and runs no model of ours. */
monitor?: boolean;
/**
* `--advisor` with no main model: Claude Code runs natively and picks its own
* model, so there is nothing to select and nothing to demand. A term of the
* gate rather than a condition beside it — see `noModelConfigured`.
*/
advisorNativeSession?: boolean;
/** `--model` / `CLAUDISH_MODEL` / a profile's single spec — already chosen. */
model?: string;
/**
* True when `--model-opus/-sonnet/-haiku/-subagent` (or a profile supplying
* them) is present: Claude Code picks per tier internally, so claudish needs no
* single model and must not ask for one.
*/
hasProfileTiers?: boolean;
}
/**
* True when no model configuration of any kind exists for this run, so SOMETHING
* has to supply one before the proxy can start.
*
* Not exported: it is the shared left-hand side of the two gates below, and both
* of them are what call sites should ask. Exporting a third name would invite a
* third gate that drifts from these two.
*
* `advisorNativeSession` belongs HERE, beside `monitor`, and not AND-ed onto one
* call site: it is a reason no model is needed, so it must suppress the picker
* and the error together. A term carried by only one of the two gates is exactly
* the drift this file exists to prevent — it would let a run skip the picker
* silently and still be told to name a model, or the reverse.
*/
function noModelConfigured(gate: ModelPickerGate): boolean {
return !gate.monitor && !gate.advisorNativeSession && !gate.model && !gate.hasProfileTiers;
}
/**
* May the interactive model picker be opened for this run?
*
* @param gate - The run's model configuration and interactivity.
* @param tty - `canDrawTui()` — both of our own streams are terminals.
* @returns True only when a model is missing, a human is waiting, and there is a
* screen to paint on. `{interactive: true, tty: false}` is the case this whole
* file exists for, and it is false.
*/
export function shouldOpenPicker(gate: ModelPickerGate, tty: boolean): boolean {
return noModelConfigured(gate) && Boolean(gate.interactive) && tty;
}
/**
* Must this run be told to name a model on the command line instead?
*
* The exact complement of `shouldOpenPicker` over the runs that need a model:
* when `noModelConfigured(gate)` holds, exactly one of the two is true, and when
* it does not, both are false. That law is what makes the second call site safe
* to widen — there is no input for which the picker is skipped silently AND the
* error stays quiet.
*
* @param gate - The run's model configuration and interactivity.
* @param tty - `canDrawTui()` — both of our own streams are terminals.
* @returns True when a model is still needed and no picker can supply it, either
* because the run is non-interactive or because there is no terminal.
*/
export function requiresExplicitModel(gate: ModelPickerGate, tty: boolean): boolean {
return noModelConfigured(gate) && !shouldOpenPicker(gate, tty);
}