Files
learnharness/components/model-picker.tsx
T
amosblomqvist 3e9a85d60f per-session models: every session owns its model + thinking level from birth
The reported bug (2026-09-28): every chat ran glm-5.3-flash regardless of
any pick, while composer and settings UIs agreed with each other and lied
about the run. Root cause: a fossil default-settings record (settings.json
chat section, frozen pre-slots-wave) was stamped verbatim into every new
chat at birth by the new-session flow, and resolveConversationSpec prefers
the per-chat record over the strong slot forever.

Owner redesign — models are per session, always; the strong slot is the
new-session default:

- POST /api/chats stamps the newborn verbatim from the strong slot
  (model + thinking level); no client settings PUT on create anymore
- the composer control is dual-bound: in-session it reads/writes the
  session's own record (GET /api/chats/:id, PUT .../settings); on the
  draft surfaces it binds the strong slot (picking there = changing the
  default)
- thinking levels follow the identical mechanics (confirmed by owner)
- the zombie default-settings system dies whole: useDefaultSettings,
  /api/settings/defaults, src/store/default-settings.ts, the chat
  section of UserSettings (sanitizer strips fossil copies), the legacy
  default-settings.json fold, and DEFAULT_SETTINGS_FILE
- catalogDefaultsFor collapses to the strong slot alone
- thinking display for a legacy unstamped session resolves exactly like
  the engine (session ?? strong ?? model ladder)

Tests: new chat-birth-stamp suite pins the stamp, the unconfigured boot,
the unset level, and the copy-not-live-view regression; defaults pins
rewritten strong-only; legacy-migration suite deleted.
2026-09-28 23:30:11 +03:00

427 lines
17 KiB
TypeScript

"use client";
import { useCallback, useEffect, useRef, useState } from "react";
import type { ChatSettings, ModelCatalogEntry, ModelCatalogResponse, ModelsSettings, UserSettings } from "@/src/core";
import type { ThinkingLevel } from "@earendil-works/pi-agent-core";
import { prettyModel } from "@/components/usage-format";
import { ModelMenuList, ThinkingMenuList } from "@/components/settings/model-menu";
import { composeSlotModel } from "@/components/settings/agent-section";
import { MenuShell } from "@/components/menu-shell";
import { Hint } from "@/components/ui/hint";
import { notifySettingsSaved, useSettingsSaved } from "@/components/settings-signal";
import { defaultOrderLevel, populateLevel, THINKING_LABEL } from "@/components/thinking-default";
/** GET /api/models is cached per page load for instant menu opens, but
* the federation is LIVE per call — keys saved mid-session widen the
* catalog immediately server-side. ModelControl therefore re-fetches on
* every model-menu open (stale-while-revalidate: the cached list shows
* instantly, the fresh one swaps in while the menu is open). */
let catalogCache: ModelCatalogResponse | undefined;
let catalogInflight: Promise<ModelCatalogResponse> | undefined;
function fetchCatalog(force = false): Promise<ModelCatalogResponse> {
if (force) {
catalogCache = undefined;
catalogInflight = undefined;
}
catalogInflight ??= fetch("/api/models")
.then((r) => {
if (!r.ok) throw new Error(`catalog request failed: ${r.status}`);
return r.json() as Promise<ModelCatalogResponse>;
})
.then((c) => {
catalogCache = c;
return c;
})
.catch((err) => {
catalogInflight = undefined; // retry on next mount/open
throw err;
});
return catalogInflight;
}
export function useModelCatalog(): { catalog: ModelCatalogResponse | undefined; refresh: () => void } {
const [catalog, setCatalog] = useState(catalogCache);
useEffect(() => {
if (catalogCache) return;
let alive = true;
fetchCatalog()
.then((c) => alive && setCatalog(c))
.catch(() => {
/* the picker stays closed on failure; the buttons still
display the strong slot from the record fetch */
});
return () => {
alive = false;
};
}, []);
const refresh = useCallback(() => {
fetchCatalog(true)
.then(setCatalog)
.catch(() => {
/* refresh is best-effort: the cached list stays on failure */
});
}, []);
return { catalog, refresh };
}
/** provider:id → contextWindow over the cached catalog — the client-side
* twin of the server's catalogContextLimits join (context-meter
* amendment, owner 2026-09-23: pi's live catalog is the ONLY limit
* source — no hardcoded fallback exists anywhere, so a model missing
* from the index reads limit-null and the meter renders usage-only,
* the honest-render precedent). Consumed by the stats bar: the same
* cached /api/models fetch the picker rides feeds the denominator. */
export function catalogLimits(catalog: Pick<ModelCatalogResponse, "models">): Map<string, number> {
return new Map(catalog.models.map((m) => [`${m.provider}:${m.id}`, m.contextWindow]));
}
/**
* THE CONVERSATION MODEL CONTROL — dual-bound (owner, per-session models,
* 2026-09-28). WITH a chatId (the in-session stats bar): the pills bind
* THE SESSION'S OWN record — reads GET /api/chats/:id (its settings field;
* a legacy unstamped record falls back to the strong slot's display,
* exactly what the engine resolves), writes PUT /api/chats/:id/settings
* (validated + level-clamped server-side). The session owns its model +
* thinking level from birth; the strong slot never reaches it again.
*
* WITHOUT a chatId (the draft surfaces — home + /chat/new): the pills
* bind the STRONG SLOT (reads the raw models section, GET /api/settings;
* writes PUT /api/settings/models via composeSlotModel — the same seam
* the settings Agent section and the SetupDialog write): picking there IS
* changing the new-session default. Unpicked reads "—" (the
* no-factory-fallback language: the UI never presents a default model).
*/
export function ModelControl({ chatId }: { chatId?: string }) {
const { catalog, refresh } = useModelCatalog();
const [models, setModels] = useState<ModelsSettings | undefined>();
const [session, setSession] = useState<ChatSettings | undefined>();
const [open, setOpen] = useState<"model" | "thinking" | null>(null);
const [error, setError] = useState<string | null>(null);
const containerRef = useRef<HTMLDivElement>(null);
// The record read — BOTH mount and the cross-surface signal land
// here (fix wave: another surface's models write re-reads into the
// pill live; own writes also notify, and the refetch is idempotent).
// Dual-bound: chat mode reads BOTH the session record (the source of
// truth) and the models section (the display fallback for a legacy
// unstamped session — exactly what the engine resolves); draft mode
// reads the models section alone.
const readRecord = useCallback(() => {
if (chatId) {
fetch(`/api/chats/${chatId}`)
.then((r) => r.json() as Promise<{ settings?: ChatSettings }>)
.then((body) => setSession(body.settings))
.catch(() => {
/* same read-only posture as a models-section failure */
});
}
fetch("/api/settings")
.then((r) => r.json() as Promise<{ settings: { models?: ModelsSettings } | null }>)
.then((body) => setModels(body.settings?.models ?? undefined))
.catch(() => {
/* the record fetch failing leaves the control read-only
("—" triggers, closed menus) — same posture as a
catalog failure */
});
}, [chatId]);
useEffect(() => {
readRecord();
}, [readRecord]);
useSettingsSaved((section) => {
if (section === "models") readRecord();
});
// Dual-bound truth: chat mode displays the session's own record,
// falling back to the strong slot for legacy unstamped sessions —
// EXACTLY what resolveConversationSpec runs. Draft mode is the strong
// slot itself.
const strong = models?.strong;
const override: ChatSettings | undefined = chatId
? {
model: session?.model ?? strong?.model,
thinkingLevel: session?.thinkingLevel ?? strong?.thinkingLevel,
}
: strong;
const effectiveId = override?.model?.id;
const effectiveProvider = override?.model?.provider;
const storedLevel = override?.thinkingLevel;
const entry = catalog?.models.find(
(m) => m.id === effectiveId && m.provider === effectiveProvider,
);
const levels = entry?.supportedLevels ?? [];
// Legacy display (chunk 2 item 4, owner): a record with a model but no
// stored level shows the RESOLVED level — the ordering's first supported
// — because that is exactly what the engine build would run. Display-
// only: the record stays unset until a fresh pick populates it.
const effectiveLevel = storedLevel ?? (entry ? defaultOrderLevel(levels) : undefined);
// The intuitive dropdown contract (owner, stats-control round 1): any
// click outside the open menu — or an Esc keypress — closes it. The
// menu's own surface keeps it open: the toggle buttons live inside the
// container, and the menu BODY lives in a portal on document.body
// (MenuShell) — never a container descendant, so its rows are carved
// out via [data-lh-menu] (the settings sections' discipline; without
// it a row's pointerdown counts as outside and eats the pick —
// live-caught in the fix wave, broken since the MenuShell migration).
useEffect(() => {
if (open === null) return;
const onPointerDown = (e: PointerEvent) => {
const t = e.target as Element;
if (t instanceof Element && t.closest("[data-lh-menu]")) return;
if (containerRef.current && !containerRef.current.contains(t)) {
setOpen(null);
}
};
const onKeyDown = (e: KeyboardEvent) => {
if (e.key === "Escape") {
// Claim the Esc (keyboard wave A): the window-level chain must
// not ALSO stop dictation / abort the turn under this menu.
e.stopPropagation();
setOpen(null);
}
};
document.addEventListener("pointerdown", onPointerDown);
document.addEventListener("keydown", onKeyDown);
return () => {
document.removeEventListener("pointerdown", onPointerDown);
document.removeEventListener("keydown", onKeyDown);
};
}, [open]);
// One write seam: composeSlotModel preserves the cheap/vision siblings
// (replace-whole-section discipline) and the per-field independence of
// model + thinkingLevel within the strong slot.
const save = useCallback(
async (next: ModelsSettings) => {
setError(null);
try {
const res = await fetch("/api/settings/models", {
method: "PUT",
headers: { "Content-Type": "application/json" },
body: JSON.stringify(next),
});
if (!res.ok) {
const body = (await res.json().catch(() => ({}))) as { error?: string };
setError(body.error ?? `settings rejected (${res.status})`);
return;
}
// Server truth, never the request body (fix wave): a pick
// that rides an unsupported level lands CLAMPED in the record
// — the pill must show the stored value in this interaction.
const stored = (
(await res.json().catch(() => null)) as { settings?: UserSettings } | null
)?.settings?.models;
setModels(stored ?? next);
setOpen(null);
notifySettingsSaved("models");
} catch {
setError("Could not reach the server — try again.");
}
},
[],
);
// The session write seam (chat mode): PUT the per-chat settings route
// — validated + level-clamped server-side; the response is the stored
// truth (a pick riding an unsupported level lands CLAMPED — the pill
// shows the stored value). No notifySettingsSaved: the write touches
// one session's record, not the slots — no other surface reads it.
const saveChat = useCallback(
async (next: ChatSettings) => {
setError(null);
try {
const res = await fetch(`/api/chats/${chatId}/settings`, {
method: "PUT",
headers: { "Content-Type": "application/json" },
body: JSON.stringify(next),
});
if (!res.ok) {
const body = (await res.json().catch(() => ({}))) as { error?: string };
setError(body.error ?? `settings rejected (${res.status})`);
return;
}
const stored = (await res.json().catch(() => null)) as { settings?: ChatSettings } | null;
setSession(stored?.settings ?? next);
setOpen(null);
} catch {
setError("Could not reach the server — try again.");
}
},
[chatId],
);
// One apply seam over both write paths: a pick names provider + id +
// level; chat mode writes the session's own record (any pick pins it),
// draft mode composes the strong slot (siblings preserved,
// replace-whole-section discipline).
const apply = useCallback(
(provider: string, id: string, level: ThinkingLevel | undefined) => {
if (chatId) {
void saveChat({ model: { provider, id }, thinkingLevel: level });
} else {
// composeSlotModel's empty-string convention: no level = unset.
void save(composeSlotModel(models, "strong", provider, id, level ?? ""));
}
},
[chatId, models, save, saveChat],
);
// Shortcut seams (keyboard wave A): ⌘K opens the model menu (the
// universal command convention); Shift+Tab cycles the thinking level
// up the ladder (wrapping) — the owner's replacement for browser
// tab-focus. Both work from any focus, including mid-typing.
useEffect(() => {
const onOpenPicker = () => {
setOpen("model");
setError(null);
refresh(); // mid-session key changes widen the menu (cache fix)
};
window.addEventListener("lh:kbd-open-model-picker", onOpenPicker);
return () => window.removeEventListener("lh:kbd-open-model-picker", onOpenPicker);
}, [refresh]);
useEffect(() => {
const onCycleThinking = () => {
if (levels.length === 0) return;
const idx = effectiveLevel === undefined ? -1 : levels.indexOf(effectiveLevel);
const next = levels[(idx + 1) % levels.length];
if (next === undefined) return;
apply(effectiveProvider ?? "", effectiveId ?? "", next);
};
window.addEventListener("lh:kbd-cycle-thinking", onCycleThinking);
return () => window.removeEventListener("lh:kbd-cycle-thinking", onCycleThinking);
}, [levels, effectiveLevel, effectiveProvider, effectiveId, apply]);
// Selecting a model (click and Enter share one path): keep the
// thinking level only when the new model supports it — otherwise
// drop it (a carry-over would 400 at the server's catalog gate).
const pickModel = (m: ModelCatalogEntry) => {
// Auto-populate (chunk 2, owner: "populate everywhere"): a fresh pick
// ALWAYS writes a level — the STORED one rides as-is (the write clamps
// server-side when the new model doesn't support it; the signal
// re-reads and the clamped truth displays), and a record with no
// stored level populates the ordering's level. The unset state can
// never re-enter through this door. (The DERIVED display level never
// rides — it belongs to the OLD model; the fresh pick's level comes
// from the fresh model's own ladder.)
const nextLevel = populateLevel(storedLevel, m.supportedLevels);
apply(m.provider, m.id, nextLevel);
};
// Stale-while-revalidate on manual open too (⌘K covers the keyboard
// path above): the cached list shows instantly, the fresh federation
// swaps in when it lands.
const openMenu = (kind: "model" | "thinking") => {
setOpen((o) => (o === kind ? null : kind));
setError(null);
if (kind === "model") refresh();
};
return (
<div ref={containerRef} className="relative flex items-center gap-1.5" data-testid="model-control">
<Hint
// Trimmed (owner, subtlety round 2026-09-25): "Model selector"
// — the long sentence was way too much. ⌘K opens this same
// menu (registry, override-aware).
label="Model selector"
action="open-model-picker"
side="bottom"
>
<button
type="button"
data-testid="stats-model"
onClick={() => openMenu("model")}
aria-expanded={open === "model"}
// Chrome look (owner, fix wave round 3): matches the graph
// view's floating top-right buttons (import notes, gear,
// coverage) — blurred, fuller opacity, shadowed — so the
// graph no longer bleeds through these pills.
className="flex items-center gap-1.5 rounded-full border bg-card/85 px-2.5 py-1 shadow-lg backdrop-blur transition-colors hover:border-border hover:bg-muted"
>
<span className="text-card-foreground">{effectiveId ? prettyModel(effectiveId) : "—"}</span>
<span aria-hidden className="text-xs text-muted-foreground">▾</span>
</button>
</Hint>
<Hint
// "Thinking" + its real shortcut via the registry (owner interlude
// 2026-09-25): ⇧⇥ cycles the level, so "Thinking (⇧⇥)" —
// rebind-live like every Hint suffix.
label="Thinking"
action="cycle-thinking"
side="bottom"
>
<button
type="button"
data-testid="stats-thinking"
onClick={() => {
// No model picked → nothing to set a level FOR: the click
// is a no-op, never an empty menu.
if (entry === undefined) return;
openMenu("thinking");
}}
aria-expanded={open === "thinking"}
// Chrome look (owner, fix wave round 3): matches the graph
// view's floating top-right buttons (import notes, gear,
// coverage) — blurred, fuller opacity, shadowed — so the
// graph no longer bleeds through these pills.
className="flex items-center gap-1.5 rounded-full border bg-card/85 px-2.5 py-1 shadow-lg backdrop-blur transition-colors hover:border-border hover:bg-muted"
>
{/* Just the level — the button reads like the model button reads
the model name (owner, stats-control round 1). */}
<span className="text-card-foreground">
{effectiveLevel === undefined ? "—" : (THINKING_LABEL[effectiveLevel] ?? effectiveLevel)}
</span>
<span aria-hidden className="text-xs text-muted-foreground">▾</span>
</button>
</Hint>
{/* The shared dropdown body (extracted 2026-09-20): the agent
settings' slot selects and the setup dialog use the exact same
one. */}
{open === "model" && (
<MenuShell
anchorRef={containerRef}
testid="model-menu"
preferredSide="up"
menuClass="flex w-[22rem] flex-col rounded-xl border bg-card p-2 text-xs shadow-xl"
>
{error && <div className="mb-1.5 text-destructive">{error}</div>}
<ModelMenuList
models={catalog?.models}
effectiveId={effectiveId}
effectiveProvider={effectiveProvider}
onPick={pickModel}
footer={
catalog && catalog.models.length === 0
? "No models — add a provider key in Keys"
: chatId
? "This session's model"
: "Strong tier — the default for new sessions"
}
/>
</MenuShell>
)}
{open === "thinking" && (
<MenuShell
anchorRef={containerRef}
testid="thinking-menu"
preferredSide="up"
menuClass="flex w-44 flex-col gap-0.5 rounded-xl border bg-card p-2 text-xs shadow-xl"
>
<ThinkingMenuList
levels={catalog ? levels : []}
effectiveLevel={effectiveLevel}
onPick={(level) => apply(effectiveProvider ?? "", effectiveId ?? "", level as ThinkingLevel)}
empty={catalog ? (effectiveId ? `${prettyModel(effectiveId)} has no thinking levels.` : "Pick a model first.") : "Loading…"}
footer="Levels this model actually supports"
/>
{error && <div className="mt-1 text-destructive">{error}</div>}
</MenuShell>
)}
</div>
);
}