feat: user pins for local model limits (context, max output, thinking)

A local endpoint that publishes no limits — LM Studio's /models says nothing — left every model on the 8192/4096 defaults with no way to say "this one has a 131k window". Each discovered model now gets three optional pins in the endpoint editor: context window, max output tokens, thinking (auto/on/off). A pin always wins: user pin → what the endpoint published → compiled default. Blank means no opinion, so discovery keeps ruling wherever you don't pin. Pins live in customProvider.overrides keyed by model id, stored apart from the discovery snapshot: a reconnect refreshes the models and prunes the pins of models that vanished while keeping the rest. Pins are written by PATCH /api/custom-provider, which never probes — the whole point is configuring a server that isn't answering /models correctly. Connect still probes and still fails loudly when the endpoint is down. A pin under 2048 tokens is rejected on the write side and clamped up on the read path, matching the compaction-floor discipline.
2026-10-27 16:25:03 -07:00
parent 286a95ba04
commit 695948271f
7 changed files with 455 additions and 28 deletions
@@ -1,24 +1,28 @@
 /**
 * The custom endpoint's connect surface (one provider in the Keys list).
- *   GET  → non-secret config { baseUrl, models } + keyPresent (masked).
- *   POST → connect: validate + normalize the base URL, probe /models (fail
- *          fast), store the key, persist { baseUrl, models }, re-sync the
- *          provider into piModels(). Discovery lives here because connecting
- *          and discovering are one user action.
+ *   GET    → non-secret config { baseUrl, models, overrides } + keyPresent (masked).
+ *   POST   → connect: validate + normalize the base URL, probe /models (fail
+ *            fast), store the key, persist { baseUrl, models }, re-sync the
+ *            provider into piModels(). Discovery lives here because connecting
+ *            and discovering are one user action. Stored user pins survive a
+ *            reconnect, pruned to the ids the endpoint still advertises.
+ *   PATCH  → store per-model pins only. No probe: pins must be editable with
+ *            the local server switched off (the case they exist for).
 *   DELETE → disconnect: drop the stored config + the stored key and unregister
- *          the provider (the mirror image of POST; idempotent).
+ *            the provider (the mirror image of POST; idempotent).
 */
 import {
 	deleteKey,
 	discoverModels,
 	effectiveSettings,
 	keyValue,
+	parseCustomModelOverridesBody,
 	readSettingsRecord,
 	saveKey,
 	syncCustomProvider,
 	writeSettingsRecord,
 } from "@/src/server";
-import { CUSTOM_PROVIDER, normalizeBaseUrl } from "@/src/core";
+import { CUSTOM_PROVIDER, normalizeBaseUrl, sanitizeCustomModelOverrides } from "@/src/core";

 export const dynamic = "force-dynamic";

@@ -27,6 +31,7 @@ export async function GET() {
 	return Response.json({
 		baseUrl: cp.baseUrl,
 		models: cp.models,
+		overrides: cp.overrides,
 		keyPresent: keyValue(CUSTOM_PROVIDER.envVar) !== undefined,
 	});
 }
@@ -54,11 +59,42 @@ export async function POST(request: Request) {
 	}

 	const record = readSettingsRecord() ?? { version: 1 };
-	const next = writeSettingsRecord({ ...record, customProvider: { baseUrl: base, models: found.models } });
+	// User pins are never rewritten by discovery — only pruned to what the
+	// endpoint still advertises (the single prune law, core/settings.ts).
+	const overrides = sanitizeCustomModelOverrides(
+		record.customProvider?.overrides,
+		found.models.map((m) => m.id),
+	);
+	const next = writeSettingsRecord({
+		...record,
+		customProvider: { baseUrl: base, models: found.models, overrides },
+	});
 	syncCustomProvider(next.customProvider);
 	return Response.json({ ok: true, models: next.customProvider?.models ?? [] });
 }

+/** Store the per-model pins (replace-whole). Pruned against the STORED model
+ *  list, never a probe — an offline endpoint keeps its pins editable. */
+export async function PATCH(request: Request) {
+	const parsed = parseCustomModelOverridesBody(await request.json().catch(() => null));
+	if (!parsed.ok) return Response.json({ error: parsed.error }, { status: 400 });
+
+	const cp = effectiveSettings().customProvider;
+	if (!cp.baseUrl) return Response.json({ error: "No custom endpoint configured" }, { status: 409 });
+
+	const overrides = sanitizeCustomModelOverrides(
+		parsed.settings,
+		cp.models.map((m) => m.id),
+	);
+	const record = readSettingsRecord() ?? { version: 1 };
+	const next = writeSettingsRecord({
+		...record,
+		customProvider: { baseUrl: cp.baseUrl, models: cp.models, overrides },
+	});
+	syncCustomProvider(next.customProvider);
+	return Response.json({ ok: true, overrides: next.customProvider?.overrides ?? {} });
+}
+
 /** Disconnect: forget the endpoint AND its key, then unregister the provider.
 *  Deleting the section (rather than blanking baseUrl) keeps settings.json the
 *  same shape as a never-connected install. Idempotent by construction. */
@@ -24,7 +24,15 @@
 */

 import { useEffect, useRef, useState } from "react";
-import { AUTH_KEYS, CUSTOM_PROVIDER, type AuthKeyStatus, type CustomModel } from "@/src/core";
+import {
+	AUTH_KEYS,
+	CUSTOM_MODEL_DEFAULT_CONTEXT_WINDOW,
+	CUSTOM_MODEL_DEFAULT_MAX_TOKENS,
+	CUSTOM_PROVIDER,
+	type AuthKeyStatus,
+	type CustomModel,
+	type CustomModelOverride,
+} from "@/src/core";
 import { SettingsGroup } from "./controls.tsx";
 import { SubscriptionConnections } from "./subscription-connections.tsx";
 import { ProviderMenuList } from "./model-menu.tsx";
@@ -49,6 +57,8 @@ interface KeysResponse {
 interface CustomStatus {
 	baseUrl: string;
 	models: CustomModel[];
+	/** User pins by model id (context / max output / thinking). */
+	overrides: Record<string, CustomModelOverride>;
 	keyPresent: boolean;
 }

@@ -402,6 +412,7 @@ export function KeysSection({
 			setCustom({
 				baseUrl: cp.baseUrl ?? "",
 				models: cp.models ?? [],
+				overrides: cp.overrides ?? {},
 				keyPresent: Boolean(cp.keyPresent),
 			});
 		} catch {
@@ -500,11 +511,17 @@ export function KeysSection({
 					<CustomEndpointEditor
 						baseUrl0={custom?.baseUrl ?? ""}
 						keyPresent={custom?.keyPresent ?? false}
+						models={custom?.models ?? []}
+						overrides={custom?.overrides ?? {}}
 						onSaved={async () => {
 							setAddingCustom(false);
 							await reload();
 							notifySettingsSaved("keys");
 						}}
+						onPinsSaved={async () => {
+							await reload();
+							notifySettingsSaved("keys");
+						}}
 						onDelete={disconnectCustom}
 						onCancel={() => setAddingCustom(false)}
 					/>
@@ -663,18 +680,68 @@ function CustomEndpointRow({
 	);
 }

+/** A pin row's raw input state — "" means "no pin": the endpoint's own value,
+ *  then the compiled default, decides. `reasoning` is tri-state. */
+interface PinDraft {
+	context: string;
+	maxTokens: string;
+	reasoning: "" | "on" | "off";
+}
+
+const pinDraft = (pin: CustomModelOverride | undefined): PinDraft => ({
+	context: pin?.context === undefined ? "" : String(pin.context),
+	maxTokens: pin?.maxTokens === undefined ? "" : String(pin.maxTokens),
+	reasoning: pin?.reasoning === undefined ? "" : pin.reasoning ? "on" : "off",
+});
+
+/** One draft row per discovered model, seeded from the stored pins. */
+function draftsFor(
+	models: readonly CustomModel[],
+	overrides: Record<string, CustomModelOverride>,
+): Record<string, PinDraft> {
+	return Object.fromEntries(models.map((m) => [m.id, pinDraft(overrides[m.id])]));
+}
+
+/** Drafts → the PATCH payload (replace-whole): a blank field sends nothing and
+ *  a fully blank model is absent — which IS the un-pin. */
+function pinsPayload(drafts: Record<string, PinDraft>): Record<string, CustomModelOverride> {
+	const out: Record<string, CustomModelOverride> = {};
+	for (const [id, d] of Object.entries(drafts)) {
+		const pin: CustomModelOverride = {};
+		if (d.context.trim()) pin.context = Number(d.context);
+		if (d.maxTokens.trim()) pin.maxTokens = Number(d.maxTokens);
+		if (d.reasoning !== "") pin.reasoning = d.reasoning === "on";
+		if (Object.keys(pin).length) out[id] = pin;
+	}
+	return out;
+}
+
+/** testids must survive ids carrying slashes, colons, dots. */
+const pinTestid = (prefix: string, id: string) => `${prefix}-${id.replace(/[^a-z0-9-]/gi, "_")}`;
+
 /** The custom endpoint's connect editor: base URL + key → POST /api/custom-provider
- *  (validate + discover + register). Image capability is detected automatically. */
+ *  (validate + discover + register), and the per-model limits form → PATCH (pins,
+ *  no probe, so it works with the local server switched off). Image capability is
+ *  detected automatically. */
 function CustomEndpointEditor({
 	baseUrl0,
 	keyPresent,
+	models,
+	overrides,
 	onSaved,
+	onPinsSaved,
 	onCancel,
 	onDelete,
 }: {
 	baseUrl0: string;
 	keyPresent: boolean;
+	/** The discovered models — one limits row each. */
+	models: readonly CustomModel[];
+	/** Their stored user pins. */
+	overrides: Record<string, CustomModelOverride>;
 	onSaved: () => void | Promise<void>;
+	/** Pins saved — the editor stays open (a form, not a connect). */
+	onPinsSaved: () => void | Promise<void>;
 	onCancel: () => void;
 	onDelete?: () => void | Promise<void>;
 }) {
@@ -683,11 +750,46 @@ function CustomEndpointEditor({
 	const [keyTouched, setKeyTouched] = useState(false);
 	const [busy, setBusy] = useState(false);
 	const [error, setError] = useState<string | null>(null);
+	// The limits form: edited drafts over the stored pins, plus the last saved
+	// payload as the dirty baseline (the server's echo is what gets stored).
+	const [drafts, setDrafts] = useState<Record<string, PinDraft>>(() => draftsFor(models, overrides));
+	const [baseline, setBaseline] = useState(() => JSON.stringify(pinsPayload(draftsFor(models, overrides))));
+	const [pinsBusy, setPinsBusy] = useState(false);
+	const [pinsError, setPinsError] = useState<string | null>(null);
 	// A key is required to enable chatting; local servers accept any value. A
 	// stored key is kept when the field is left blank.
 	const needsKey = !keyPresent;
 	const canConnect = Boolean(baseUrl.trim()) && (!needsKey || Boolean(key.trim()));

+	const draftOf = (m: CustomModel): PinDraft => drafts[m.id] ?? pinDraft(overrides[m.id]);
+	const setDraft = (m: CustomModel, patch: Partial<PinDraft>) =>
+		setDrafts((prev) => ({ ...prev, [m.id]: { ...draftOf(m), ...patch } }));
+	const current = Object.fromEntries(models.map((m) => [m.id, draftOf(m)])) as Record<string, PinDraft>;
+	const dirty = JSON.stringify(pinsPayload(current)) !== baseline;
+
+	async function savePins() {
+		if (pinsBusy || !dirty) return;
+		setPinsBusy(true);
+		setPinsError(null);
+		const res = await fetch("/api/custom-provider", {
+			method: "PATCH",
+			headers: { "content-type": "application/json" },
+			body: JSON.stringify({ overrides: pinsPayload(current) }),
+		}).catch(() => null);
+		setPinsBusy(false);
+		if (!res || !res.ok) {
+			const b = (await res?.json().catch(() => ({}))) as { error?: string } | undefined;
+			setPinsError(res ? (b?.error ?? `Save failed (${res.status})`) : "Save failed. Check your connection.");
+			return;
+		}
+		// The server's stored map is the truth (it floors + prunes) — echo it back.
+		const body = (await res.json().catch(() => null)) as { overrides?: Record<string, CustomModelOverride> } | null;
+		const next = draftsFor(models, body?.overrides ?? pinsPayload(current));
+		setDrafts(next);
+		setBaseline(JSON.stringify(pinsPayload(next)));
+		await onPinsSaved();
+	}
+
 	async function save() {
 		if (busy || !baseUrl.trim()) return;
 		setBusy(true);
@@ -769,6 +871,75 @@ function CustomEndpointEditor({
 					</button>
 				)}
 			</div>
+			{models.length > 0 && (
+				<div className="flex flex-col gap-1" data-testid="custom-pins">
+					<span className="text-[11px] text-muted-foreground">
+						Model limits — blank uses the endpoint’s value, then the default
+					</span>
+					<div className="flex max-h-56 flex-col gap-1 overflow-y-auto pr-1">
+						{models.map((m) => {
+							const d = draftOf(m);
+							return (
+								<div key={m.id} className="flex items-center gap-2" data-testid={pinTestid("custom-pin-row", m.id)}>
+									<span className="min-w-0 flex-1 truncate font-mono text-[11px] text-foreground" title={m.id}>
+										{m.id}
+									</span>
+									<input
+										type="text"
+										inputMode="numeric"
+										value={d.context}
+										onChange={(e) => setDraft(m, { context: e.target.value })}
+										placeholder={String(m.context ?? CUSTOM_MODEL_DEFAULT_CONTEXT_WINDOW)}
+										aria-label={`${m.id} context window in tokens`}
+										data-testid={pinTestid("custom-pin-context", m.id)}
+										disabled={pinsBusy}
+										className="h-7 w-20 shrink-0 rounded-sm border border-border bg-background px-1.5 text-right font-mono text-[11px] text-foreground outline-none placeholder:text-muted-foreground/60 focus:border-muted-foreground/50"
+									/>
+									<input
+										type="text"
+										inputMode="numeric"
+										value={d.maxTokens}
+										onChange={(e) => setDraft(m, { maxTokens: e.target.value })}
+										placeholder={String(m.maxTokens ?? CUSTOM_MODEL_DEFAULT_MAX_TOKENS)}
+										aria-label={`${m.id} max output tokens`}
+										data-testid={pinTestid("custom-pin-maxtokens", m.id)}
+										disabled={pinsBusy}
+										className="h-7 w-20 shrink-0 rounded-sm border border-border bg-background px-1.5 text-right font-mono text-[11px] text-foreground outline-none placeholder:text-muted-foreground/60 focus:border-muted-foreground/50"
+									/>
+									<select
+										value={d.reasoning}
+										onChange={(e) => setDraft(m, { reasoning: e.target.value as PinDraft["reasoning"] })}
+										aria-label={`${m.id} thinking`}
+										data-testid={pinTestid("custom-pin-reasoning", m.id)}
+										disabled={pinsBusy}
+										className="h-7 shrink-0 rounded-sm border border-border bg-background px-1 font-ui text-[11px] text-foreground outline-none focus:border-muted-foreground/50"
+									>
+										<option value="">think: auto</option>
+										<option value="on">think: on</option>
+										<option value="off">think: off</option>
+									</select>
+								</div>
+							);
+						})}
+					</div>
+					<div className="flex items-center gap-2">
+						<button
+							type="button"
+							onClick={() => void savePins()}
+							disabled={pinsBusy || !dirty}
+							data-testid="custom-pins-save"
+							className="shrink-0 text-xs text-foreground underline-offset-2 hover:underline disabled:text-muted-foreground/40 disabled:no-underline"
+						>
+							{pinsBusy ? "Saving…" : "Save limits"}
+						</button>
+						{pinsError && (
+							<p role="alert" data-testid="custom-pins-error" className="text-[11px] text-red-500">
+								{pinsError}
+							</p>
+						)}
+					</div>
+				</div>
+			)}
 			{needsKey && (
 				<p className="text-[11px] text-muted-foreground">
 					A key is required to enable chatting — local servers (Ollama, LM Studio) accept any value.
@@ -38,6 +38,7 @@ import {
 	type DictationSettings,
 	type CompactionSettings,
 	type CustomModel,
+	type CustomModelOverride,
 	type GraphSettings,
 	type ModelSlot,
 	type ModelSlotName,
@@ -96,8 +97,9 @@ export interface EffectiveSettings {
 	/** Compaction (owner round 14): resolved concrete — absent record
 	 *  section resolves to the factory 150k. */
 	compaction: Required<CompactionSettings>;
-	/** Custom endpoint, resolved concrete: baseUrl defaults "", models default []. */
-	customProvider: { baseUrl: string; models: CustomModel[] };
+	/** Custom endpoint, resolved concrete: baseUrl defaults "", models and
+	 *  user pins default to empty. */
+	customProvider: { baseUrl: string; models: CustomModel[]; overrides: Record<string, CustomModelOverride> };
 }

 /** The conversation pair's chain (BYOK, owner 2026-09-22): the chat's own
@@ -227,6 +229,7 @@ export function resolveSettings(record: UserSettings | null): EffectiveSettings
 		customProvider: {
 			baseUrl: record?.customProvider?.baseUrl ?? "",
 			models: record?.customProvider?.models ?? [],
+			overrides: record?.customProvider?.overrides ?? {},
 		},
 	};
 }
@@ -217,6 +217,12 @@ export interface ModelsSettings {
 export const CUSTOM_MODEL_DEFAULT_CONTEXT_WINDOW = 8192;
 export const CUSTOM_MODEL_DEFAULT_MAX_TOKENS = 4096;

+/** Floor for a user-pinned token count (owner, 2026-10-27): the read side
+ *  clamps up, the write side rejects below — the compaction-floor discipline.
+ *  Pins exist because endpoints that publish no limits (LM Studio) otherwise
+ *  sit on the 8192 default with no way to say better. */
+export const CUSTOM_MODEL_PIN_FLOOR = 2_048;
+
 /** One discovered custom model. `image`/`reasoning`/`context`/`maxTokens` are
 *  detected from the endpoint's /models metadata; anything it omits falls back
 *  to a default. */
@@ -233,12 +239,26 @@ export interface CustomModel {
 	maxTokens?: number;
 }

+/** A user pin for one custom model: it beats what the endpoint publishes
+ *  (owner ruling — explicit user intent wins; a local server's number is
+ *  often wrong or absent). Absent field = no pin. `reasoning` is tri-state:
+ *  true forces thinking on, false forces it off (a server that advertises
+ *  reasoning but rejects reasoning_effort), absent = what discovery found. */
+export interface CustomModelOverride {
+	context?: number;
+	maxTokens?: number;
+	reasoning?: boolean;
+}
+
 /** The single custom endpoint's stored config (delta section). */
 export interface CustomProviderSettings {
 	/** Normalized base URL (see normalizeBaseUrl). Empty/absent = not configured. */
 	baseUrl?: string;
 	/** Models discovered from the endpoint (refreshed on connect). */
 	models?: CustomModel[];
+	/** User pins keyed by model id — pure user intent, never rewritten by
+	 *  discovery, pruned to the endpoint's live ids on connect. */
+	overrides?: Record<string, CustomModelOverride>;
 }

 /** Canonical base-URL form (single source): trims, requires absolute http(s),
@@ -562,6 +582,38 @@ function positiveInt(value: unknown): number | undefined {
 	return typeof value === "number" && Number.isInteger(value) && value > 0 ? value : undefined;
 }

+/** One pin, clamped up to the floor (clamp-never-reject on read). */
+function sanitizeCustomModelOverride(raw: unknown): CustomModelOverride | undefined {
+	if (raw === null || typeof raw !== "object" || Array.isArray(raw)) return undefined;
+	const rec = raw as Record<string, unknown>;
+	const out: CustomModelOverride = {};
+	const context = positiveInt(rec.context);
+	if (context) out.context = Math.max(CUSTOM_MODEL_PIN_FLOOR, context);
+	const maxTokens = positiveInt(rec.maxTokens);
+	if (maxTokens) out.maxTokens = Math.max(CUSTOM_MODEL_PIN_FLOOR, maxTokens);
+	if (typeof rec.reasoning === "boolean") out.reasoning = rec.reasoning;
+	return Object.keys(out).length ? out : undefined;
+}
+
+/** THE pins map + THE prune law: well-formed entries only, ids trimmed, and
+ *  when `liveIds` is given every entry for a model the endpoint no longer
+ *  advertises drops. POST passes the discovered ids, PATCH the stored list. */
+export function sanitizeCustomModelOverrides(
+	raw: unknown,
+	liveIds?: Iterable<string>,
+): Record<string, CustomModelOverride> {
+	const out: Record<string, CustomModelOverride> = {};
+	if (raw === null || typeof raw !== "object" || Array.isArray(raw)) return out;
+	const live = liveIds ? new Set(liveIds) : undefined;
+	for (const [key, value] of Object.entries(raw)) {
+		const id = key.trim();
+		if (!id || (live && !live.has(id))) continue;
+		const pin = sanitizeCustomModelOverride(value);
+		if (pin) out[id] = pin;
+	}
+	return out;
+}
+
 /** The customProvider section: normalized base URL (invalid → dropped) and
 *  discovered models (ids trimmed, dupes/empties dropped; image kept only when
 *  true; token counts only as positive ints). Clamp-never-reject. */
@@ -590,6 +642,8 @@ function sanitizeCustomProvider(raw: Record<string, unknown>): CustomProviderSet
 		}
 		out.models = [...byId.values()];
 	}
+	const overrides = sanitizeCustomModelOverrides(raw.overrides);
+	if (Object.keys(overrides).length) out.overrides = overrides;
 	return out;
 }

@@ -15,6 +15,7 @@ import {
 	customProviderUsable,
 	normalizeBaseUrl,
 	type CustomModel,
+	type CustomModelOverride,
 	type CustomProviderSettings,
 } from "../core/settings.ts";
 import { keyValue } from "./keys.ts";
@@ -36,23 +37,31 @@ const customEndpointAuth: ApiKeyAuth = {
 };

 /** Map a discovered model to a pi-ai entry: image→input, reasoning→thinking
- *  levels, context/maxTokens→limits (else CUSTOM_MODEL_DEFAULT_*). baseUrl is
- *  stamped per-model. */
-function toPiModel(model: CustomModel, baseUrl: string): Model<Api> {
+ *  levels, context/maxTokens→limits. Resolution ladder (owner, 2026-10-27):
+ *  the user's pin wins, then the endpoint's own metadata, then the compiled
+ *  default — a local server that says nothing (or says 8k) can be corrected.
+ *  baseUrl is stamped per-model. */
+function toPiModel(
+	model: CustomModel,
+	baseUrl: string,
+	overrides: Record<string, CustomModelOverride>,
+): Model<Api> {
+	const pin = overrides[model.id];
 	return {
 		id: model.id,
 		name: model.id,
 		api: "openai-completions",
 		provider: CUSTOM_PROVIDER.id,
 		baseUrl,
-		// Detected from the endpoint's metadata (see detectReasoning). true makes
-		// pi-ai offer the standard levels and send reasoning_effort on the wire;
-		// no thinkingLevelMap → the default off/minimal/low/medium/high set.
-		reasoning: Boolean(model.reasoning),
+		// Pin wins; else detected from the endpoint's metadata (see
+		// detectReasoning). true makes pi-ai offer the standard levels and send
+		// reasoning_effort on the wire; no thinkingLevelMap → the default
+		// off/minimal/low/medium/high set.
+		reasoning: pin?.reasoning ?? Boolean(model.reasoning),
 		input: model.image ? ["text", "image"] : ["text"],
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-		contextWindow: model.context ?? CUSTOM_MODEL_DEFAULT_CONTEXT_WINDOW,
-		maxTokens: model.maxTokens ?? CUSTOM_MODEL_DEFAULT_MAX_TOKENS,
+		contextWindow: pin?.context ?? model.context ?? CUSTOM_MODEL_DEFAULT_CONTEXT_WINDOW,
+		maxTokens: pin?.maxTokens ?? model.maxTokens ?? CUSTOM_MODEL_DEFAULT_MAX_TOKENS,
 		type: "chat",
 	};
 }
@@ -61,13 +70,16 @@ function toPiModel(model: CustomModel, baseUrl: string): Model<Api> {
 export function buildOpenAiCompatibleProvider(input: {
 	baseUrl: string;
 	models: readonly CustomModel[];
+	/** User pins by model id (see CustomModelOverride) — absent = none. */
+	overrides?: Record<string, CustomModelOverride>;
 }): Provider<Api> {
+	const overrides = input.overrides ?? {};
 	return createProvider<Api>({
 		id: CUSTOM_PROVIDER.id,
 		name: CUSTOM_PROVIDER.label,
 		baseUrl: input.baseUrl,
 		auth: { apiKey: customEndpointAuth },
-		models: input.models.map((m) => toPiModel(m, input.baseUrl)),
+		models: input.models.map((m) => toPiModel(m, input.baseUrl, overrides)),
 		api: openAICompletionsApi(),
 	});
 }
@@ -81,7 +93,11 @@ export function syncCustomProvider(cp: CustomProviderSettings | undefined): void
 		return;
 	}
 	models.setProvider(
-		buildOpenAiCompatibleProvider({ baseUrl: normalizeBaseUrl(cp.baseUrl), models: cp.models ?? [] }),
+		buildOpenAiCompatibleProvider({
+			baseUrl: normalizeBaseUrl(cp.baseUrl),
+			models: cp.models ?? [],
+			overrides: cp.overrides ?? {},
+		}),
 	);
 }

@@ -18,13 +18,14 @@ import type {
 	AppearanceSettings,
 	BackgroundModelRole,
 	CompactionSettings,
+	CustomModelOverride,
 	GeneralSettings,
 	DictationSettings,
 	GraphSettings,
 	ModelsSettings,
 	RoleModelOverride,
 } from "../core/settings.ts";
-import { BACKGROUND_MODEL_ROLES, COMPACTION_THRESHOLD_FLOOR, EXTRACTION_CHUNK_TOKENS_FLOOR, MODEL_SLOTS } from "../core/settings.ts";
+import { BACKGROUND_MODEL_ROLES, COMPACTION_THRESHOLD_FLOOR, CUSTOM_MODEL_PIN_FLOOR, EXTRACTION_CHUNK_TOKENS_FLOOR, MODEL_SLOTS } from "../core/settings.ts";
 import { WRITABLE_TOOL_KEYS } from "../core/pending.ts";
 import { isDictationLanguage } from "../core/dictation.ts";

@@ -386,6 +387,45 @@ export function parseDictationSettingsBody(body: unknown): SettingsBodyResult<Di
 	return { ok: true, settings: { language: raw.language } };
 }

+/** The custom endpoint's per-model pins (PATCH /api/custom-provider): the
+ *  whole desired map, replace-whole like every section. Strict write side — a
+ *  token count must be a whole number at or above the floor (the read side
+ *  clamps, this rejects — the compaction discipline). `reasoning` is the
+ *  tri-state pin: true forces thinking on, false forces it off. */
+export function parseCustomModelOverridesBody(
+	body: unknown,
+): SettingsBodyResult<Record<string, CustomModelOverride>> {
+	const raw = asRecord(body);
+	if (!raw) return { ok: false, error: "Invalid JSON body" };
+	const map = asRecord(raw.overrides);
+	if (!map) return { ok: false, error: "overrides must be a { modelId: { context?, maxTokens?, reasoning? } } map" };
+	const out: Record<string, CustomModelOverride> = {};
+	for (const [key, value] of Object.entries(map)) {
+		const id = key.trim();
+		if (!id) return { ok: false, error: "override keys must be non-empty model ids" };
+		const entry = asRecord(value);
+		if (!entry) return { ok: false, error: `override for ${id} must be an object` };
+		const pin: CustomModelOverride = {};
+		for (const field of ["context", "maxTokens"] as const) {
+			const v = entry[field];
+			if (v === undefined) continue;
+			if (typeof v !== "number" || !Number.isInteger(v) || v < CUSTOM_MODEL_PIN_FLOOR) {
+				return {
+					ok: false,
+					error: `${field} for ${id} must be a whole number of tokens at least ${CUSTOM_MODEL_PIN_FLOOR}`,
+				};
+			}
+			pin[field] = v;
+		}
+		if (entry.reasoning !== undefined) {
+			if (typeof entry.reasoning !== "boolean") return { ok: false, error: `reasoning for ${id} must be true or false` };
+			pin.reasoning = entry.reasoning;
+		}
+		out[id] = pin;
+	}
+	return { ok: true, settings: out };
+}
+
 /** The section→parser map for PUT /api/settings/[section]. The custom endpoint
 *  is NOT a section — it connects via POST /api/custom-provider. */
 export const SETTINGS_SECTIONS = {
@@ -2,8 +2,9 @@
 * Tests for the custom OpenAI-compatible endpoint (custom-endpoint wave).
 * Covers the pure core (normalizeBaseUrl, customProviderUsable, the sanitizer,
 * resolution), the deep server module (provider build + live sync + model
- * discovery incl. image-modality and reasoning detection), and the catalog
- * federation row. No network: the discovery probe runs against a stubbed fetch.
+ * discovery incl. image-modality, reasoning detection and the per-model user
+ * pins), the pins' strict write-side parser, and the catalog federation row.
+ * No network: the discovery probe runs against a stubbed fetch.
 */

 import assert from "node:assert/strict";
@@ -11,11 +12,14 @@ import { after, describe, it } from "node:test";
 import {
 	CUSTOM_MODEL_DEFAULT_CONTEXT_WINDOW,
 	CUSTOM_MODEL_DEFAULT_MAX_TOKENS,
+	CUSTOM_MODEL_PIN_FLOOR,
 	customProviderUsable,
 	normalizeBaseUrl,
+	sanitizeCustomModelOverrides,
 	sanitizeUserSettings,
 } from "../src/core/settings.ts";
 import { resolveSettings } from "../src/core/settings-resolution.ts";
+import { parseCustomModelOverridesBody } from "../src/server/settings-body.ts";
 import { CUSTOM_PROVIDER } from "../src/core/auth-keys.ts";
 import {
 	buildOpenAiCompatibleProvider,
@@ -84,10 +88,86 @@ describe("sanitizeUserSettings — customProvider", () => {
 	});
 });

-// ─── resolution ──────────────────────────────────────────────────────────────
+// ─── per-model pins: the read side + THE prune law ───────────────────────────
+describe("sanitizeUserSettings / sanitizeCustomModelOverrides — pins", () => {
+	it("keeps well-formed pins, clamps token counts up to the floor, keeps reasoning:false", () => {
+		const out = sanitizeUserSettings({
+			customProvider: {
+				baseUrl: "https://h/v1",
+				overrides: {
+					llama: { context: 131072, maxTokens: 8192, reasoning: true },
+					qwen: { context: 10 }, // under the floor → clamped up (clamp-never-reject on read)
+					quiet: { reasoning: false }, // a real pin, not an absent one
+					junk: { context: "big", maxTokens: -1, reasoning: "yes" }, // every field junk
+					" ": { context: 4096 }, // no usable id
+					nope: "not-an-object",
+				},
+			},
+		});
+		assert.deepEqual(out?.customProvider?.overrides, {
+			llama: { context: 131072, maxTokens: 8192, reasoning: true },
+			qwen: { context: CUSTOM_MODEL_PIN_FLOOR },
+			quiet: { reasoning: false },
+		});
+	});
+	it("prunes pins for models the endpoint no longer advertises — the single prune law", () => {
+		const pins = { a: { context: 32768 }, b: { context: 32768 } };
+		assert.deepEqual(Object.keys(sanitizeCustomModelOverrides(pins, ["a"])), ["a"]);
+		assert.deepEqual(
+			Object.keys(sanitizeCustomModelOverrides(pins)),
+			["a", "b"],
+			"no liveIds → no prune (that is the routes' call, not the record sanitizer's)",
+		);
+		assert.deepEqual(sanitizeCustomModelOverrides("nope"), {}, "a non-object map resolves empty");
+	});
+	it("drops an invalid overrides section from the record entirely", () => {
+		const out = sanitizeUserSettings({ customProvider: { baseUrl: "https://h/v1", overrides: "nope" } });
+		assert.equal(out?.customProvider?.overrides, undefined);
+	});
+});
+
+// ─── pins: the strict write side (PATCH) ──────────────────────────────────
+describe("parseCustomModelOverridesBody — the strict write side", () => {
+	it("accepts whole counts at or above the floor and the tri-state reasoning pin", () => {
+		const r = parseCustomModelOverridesBody({
+			overrides: {
+				a: { context: CUSTOM_MODEL_PIN_FLOOR, reasoning: false },
+				b: { maxTokens: 8192 },
+				c: {}, // an empty pin is legal — the sanitizer drops it on write
+			},
+		});
+		assert.ok(r.ok);
+		assert.deepEqual(r.settings, {
+			a: { context: CUSTOM_MODEL_PIN_FLOOR, reasoning: false },
+			b: { maxTokens: 8192 },
+			c: {},
+		});
+	});
+	it("rejects a count under the floor, naming the floor", () => {
+		const r = parseCustomModelOverridesBody({ overrides: { a: { context: CUSTOM_MODEL_PIN_FLOOR - 1 } } });
+		assert.equal(r.ok, false);
+		if (!r.ok) assert.ok(r.error.includes(String(CUSTOM_MODEL_PIN_FLOOR)), r.error);
+	});
+	it("rejects a fractional count, a non-boolean reasoning, and a missing overrides map", () => {
+		assert.equal(parseCustomModelOverridesBody({ overrides: { a: { context: 4096.5 } } }).ok, false);
+		assert.equal(parseCustomModelOverridesBody({ overrides: { a: { reasoning: "on" } } }).ok, false);
+		assert.equal(parseCustomModelOverridesBody({ overrides: { a: "nope" } }).ok, false);
+		assert.equal(parseCustomModelOverridesBody({}).ok, false);
+		assert.equal(parseCustomModelOverridesBody(null).ok, false);
+	});
+});
+
+// ─── resolution ──────────────────────────────────────────────────────────
 describe("resolveSettings — customProvider", () => {
 	it("resolves an absent section to concrete defaults", () => {
-		assert.deepEqual(resolveSettings(null).customProvider, { baseUrl: "", models: [] });
+		assert.deepEqual(resolveSettings(null).customProvider, { baseUrl: "", models: [], overrides: {} });
+	});
+	it("carries stored pins through to the concrete view", () => {
+		const resolved = resolveSettings({
+			version: 1,
+			customProvider: { baseUrl: "http://h/v1", models: [{ id: "a" }], overrides: { a: { context: 32768 } } },
+		});
+		assert.deepEqual(resolved.customProvider.overrides, { a: { context: 32768 } });
 	});
 });

@@ -124,6 +204,33 @@ describe("buildOpenAiCompatibleProvider / syncCustomProvider", () => {
 		assert.equal(big!.maxTokens, 393216);
 	});

+	it("a user pin beats the endpoint, which beats the compiled default", () => {
+		const provider = buildOpenAiCompatibleProvider({
+			baseUrl: "http://h/v1",
+			models: [{ id: "pinned", context: 32768, maxTokens: 4096 }, { id: "detected", context: 65536 }, { id: "bare" }],
+			overrides: { pinned: { context: 131072, maxTokens: 16384 } },
+		});
+		const [pinned, detected, bare] = provider.getModels();
+		assert.equal(pinned!.contextWindow, 131072, "the pin wins over the endpoint's own number");
+		assert.equal(pinned!.maxTokens, 16384);
+		assert.equal(detected!.contextWindow, 65536, "no pin → the endpoint's value");
+		assert.equal(detected!.maxTokens, CUSTOM_MODEL_DEFAULT_MAX_TOKENS, "endpoint silent → the compiled default");
+		assert.equal(bare!.contextWindow, CUSTOM_MODEL_DEFAULT_CONTEXT_WINDOW);
+	});
+
+	it("the thinking pin is tri-state: false beats a detected-true endpoint, true beats silence", () => {
+		const provider = buildOpenAiCompatibleProvider({
+			baseUrl: "http://h/v1",
+			models: [{ id: "liar", reasoning: true }, { id: "silent" }],
+			overrides: { liar: { reasoning: false }, silent: { reasoning: true } },
+		});
+		const [liar, silent] = provider.getModels();
+		assert.equal(liar!.reasoning, false, "a server that advertises reasoning but rejects it can be pinned off");
+		assert.deepEqual(getSupportedThinkingLevels(liar!), ["off"]);
+		assert.equal(silent!.reasoning, true, "LM Studio publishes nothing; the user says otherwise");
+		assert.deepEqual(getSupportedThinkingLevels(silent!), ["off", "minimal", "low", "medium", "high"]);
+	});
+
 	it("gives a reasoning-capable model the standard thinking levels (others stay \"off\")", () => {
 		const provider = buildOpenAiCompatibleProvider({
 			baseUrl: "http://h/v1",