feat: support custom model info overrides

docs/superpowers/specs/2026-08-04-custom-model-settings-design.md

@@ -0,0 +1,301 @@
+# Custom Model Settings Override
+
+**Date:** 2026-08-04
+**Status:** Approved, ready for implementation plan
+
+## Problem
+
+A user selects a model ID that is not present in a router provider's fetched
+model list — for example typing `anthropic/claude-sonnet-4-6` into the
+OpenRouter model picker via the "use custom model" affordance
+(`webview-ui/src/components/settings/ModelPicker.tsx:277`). Three distinct
+defects follow.
+
+### Defect 1 — context window collapses to `1`, percentage renders as 7000%
+
+`webview-ui/src/components/chat/TaskHeader.tsx:72`:
+
+```ts
+const contextWindow = model?.contextWindow || 1
+```
+
+When `useSelectedModel` cannot resolve the model, `model` is `undefined` and
+`contextWindow` becomes `1`. That value flows into the percentage at
+`TaskHeader.tsx:253-258`:
+
+```ts
+const availableInputSpace = contextWindow - reservedForOutput
+const percentage =
+    availableInputSpace > 0
+        ? Math.round(((contextTokens || 0) / availableInputSpace) * 100)
+        : 0
+```
+
+With `contextWindow === 1` and `reservedForOutput === 0`, `availableInputSpace`
+is `1`, so the percentage equals `contextTokens * 100`. 70 context tokens
+render as **7000%**. There is no upper clamp on this path.
+
+**Precise trigger.** For OpenRouter an unknown ID is normally rewritten to the
+default model by `getValidatedModelId`
+(`webview-ui/src/components/ui/hooks/useSelectedModel.ts:56`), which yields a
+valid `info`. The `undefined` case therefore arises when the router model list
+is empty rather than merely missing the ID: no API key configured, a failed or
+in-flight fetch, or offline. In that state the default-model lookup also misses,
+`info` is `undefined`, and the `|| 1` fallback produces both the "token limit
+shows 1" symptom and the 7000% reading. They are two faces of one fault.
+
+### Defect 2 — webview and extension host disagree on the model
+
+`getValidatedModelId` silently substitutes the provider default when the
+configured ID is absent from the list, while `openRouterModelId` continues to
+hold the user's typed value. The extension host does not perform the same
+substitution — `src/api/providers/openrouter.ts:554`:
+
+```ts
+let info = this.models[id] ?? openRouterDefaultModelInfo
+```
+
+The host sends the user's real ID with a 200K-context default profile; the
+webview displays a different model entirely. Requests may succeed while the UI
+describes something else.
+
+### Defect 3 — no override UI outside the OpenAI-compatible provider
+
+`openAiCustomModelInfo` (`packages/types/src/provider-settings.ts:245`) is the
+only user-facing way to supply `contextWindow` / `maxTokens`, and it is wired
+solely to the `openai` provider's settings panel
+(`webview-ui/src/components/settings/providers/OpenAICompatible.tsx:286-347`).
+OpenRouter, Requesty, Unbound, Vercel AI Gateway and Zoo Gateway offer no
+equivalent, so a custom model on those providers can never be given correct
+token limits.
+
+## Goals
+
+1. Let the user override context window and max output tokens for any model on
+   the router providers, and have that override govern both the UI and the real
+   request/truncation path.
+2. Ensure the UI never displays a nonsensical figure when no override is set.
+
+## Non-goals
+
+- Editing per-token pricing. Overridden prices would corrupt cost reporting;
+  that is separate work.
+- Migrating or removing `openAiCustomModelInfo`. It keeps working unchanged.
+- Reworking `ModelPicker`'s custom-model entry flow.
+
+## Architecture
+
+Two layers resolve model info independently and must not diverge:
+
+| Layer | Resolver |
+|---|---|
+| Webview | `getSelectedModel()` in `useSelectedModel.ts:132` |
+| Extension host | each provider's `getModel()` (30 implementations) |
+
+An override applied to only one layer would fix the display while leaving
+context truncation wrong. The design therefore applies one shared helper at both
+layers, each through a single chokepoint.
+
+Two facts from the codebase make the host-side chokepoint viable: there is
+exactly one factory, `buildApiHandler` (`src/api/index.ts:153`), and no
+`instanceof <X>Handler` check exists anywhere in `src/`. A wrapper around the
+returned handler is therefore safe.
+
+### Data model
+
+Add one field to `baseProviderSettingsSchema`
+(`packages/types/src/provider-settings.ts:176`):
+
+```ts
+customModelInfo: modelInfoSchema.partial().nullish(),
+```
+
+`partial()` is deliberate. The field is an **overlay**, not a replacement: a user
+who sets only `contextWindow` keeps the fetched values for price, image support
+and reasoning. Placing it on the base schema means every provider inherits it,
+avoiding the five near-identical fields that a per-provider approach would need.
+
+`openAiCustomModelInfo` remains as-is. Where both are present, `customModelInfo`
+is applied second and wins on the fields it defines.
+
+### Shared helper
+
+In `packages/types` (importable by both webview and host):
+
+```ts
+applyCustomModelInfo(
+    info: ModelInfo | undefined,
+    settings: { customModelInfo?: Partial<ModelInfo> | null } | undefined,
+): ModelInfo | undefined
+```
+
+Behaviour:
+
+- `info` present → return `info` with the override's **defined and valid** keys
+  merged over it.
+- `info` absent but the override supplies a positive `contextWindow` → synthesise
+  a `ModelInfo` from a synthesis base plus the override. This is what makes a
+  genuinely unknown model usable.
+- Neither → return `undefined`, preserving today's "invalid selection" signal.
+
+The synthesis base is defined locally rather than reusing
+`openAiModelInfoSaneDefaults`, whose `maxTokens: -1` sentinel
+(`packages/types/src/providers/openai.ts:692-693`) would propagate a negative
+value into arithmetic:
+
+```ts
+const CUSTOM_MODEL_SYNTHESIS_BASE = {
+    maxTokens: undefined,
+    supportsImages: false,
+    supportsPromptCache: false,
+} satisfies Partial<ModelInfo>
+```
+
+`contextWindow` is deliberately absent from the base: synthesis only runs when
+the override supplies a positive one, so the merged result always has a real
+value and never a fabricated default.
+
+Leaving `maxTokens` undefined is safe rather than lossy. `getModelMaxOutputTokens`
+(`src/shared/api.ts:131-133`) supplies `ANTHROPIC_DEFAULT_MAX_TOKENS` whenever the
+model ID contains `claude` and `maxTokens` is absent — which covers the reported
+`anthropic/claude-sonnet-4-6` case. For non-Anthropic IDs it returns `undefined`
+(line 158-160), which `TaskHeader` already handles by reserving nothing.
+
+A key is treated as "valid" when it is not `undefined`/`null`, and — for the
+numeric fields `contextWindow` and `maxTokens` — is a finite number greater than
+zero. Invalid entries are dropped, never coerced to `0`, because
+`contextWindow: 0` would reproduce the original division fault.
+
+### Integration points
+
+**Webview** — apply the helper to the `{ id, info }` produced by the ternary at
+`useSelectedModel.ts:98-113`, not to `getSelectedModel()`'s return. That ternary
+has three branches: the resolved call, a `kimi-code` fallback, and a
+not-ready/invalid-provider fallback that yields `info: undefined`. The override
+must cover all three — the third is precisely the still-loading state that
+produces the reported symptom, and `getSelectedModel()` is not called there at
+all. Applying it after the ternary covers every branch and leaves the 30 `switch`
+cases untouched.
+
+**Host** — in `buildApiHandler`, wrap the constructed handler in a `Proxy` that
+decorates `getModel()` and forwards everything else. Forwarding uses
+`Reflect.get(target, prop, target)` — passing `target` rather than the proxy as
+receiver, so private class fields continue to resolve. All twelve
+`this.api.getModel().info` consumers in `Task.ts` inherit the corrected value,
+including the context-window-exceeded and condense paths.
+
+### Display hardening
+
+Independent of any override, so the UI is correct when the user sets nothing:
+
+- `TaskHeader.tsx:72` — drop `|| 1`. When no context window is known, skip
+  rendering the percentage entirely rather than printing a fabricated number.
+- `TaskHeader.tsx:253-258` — clamp the upper bound with `Math.min(100, …)` and
+  render at/over 100% in a warning colour. Keep the existing
+  `availableInputSpace > 0` guard: it is the lower bound, and an over-large
+  `maxTokens` override can still drive `availableInputSpace` to zero or below.
+- `useSelectedModel.ts:51-57` — stop substituting the provider default for a
+  configured-but-unlisted ID on the router providers. The condition is *the
+  configured ID is absent from the list*, which covers both an empty list and a
+  populated list that lacks the user's custom ID; the current guard conflates
+  them. The litellm case (lines 178-189) is the in-repo precedent for returning
+  the configured ID untouched.
+
+  This aligns the webview with the host, which never substitutes — closing
+  Defect 2's divergence. It does not make the two produce identical `info`: the
+  host still falls back to `openRouterDefaultModelInfo` (200K) at
+  `openrouter.ts:554` while the webview yields `undefined`. Full convergence is
+  what the shared helper delivers once an override exists, and is why the helper
+  must be bound at both layers rather than the webview alone.
+
+  Callers that assume a non-empty, listed ID must be checked. `ModelPicker`
+  already tolerates it: `modelIds` explicitly retains `selectedModelId`
+  (lines 122-127) and the initialization effect at 187-194 only fires when
+  `selectedModelId` is falsy, so a preserved custom ID is displayed rather than
+  overwritten.
+
+### UI
+
+New shared component `CustomModelInfoSettings.tsx`, following the field pattern
+already established in `OpenAICompatible.tsx` (text field, green/red border
+validation, label plus description). Rendered beneath `ModelPicker` for the
+router providers: OpenRouter, Requesty, Unbound, Vercel AI Gateway, Zoo Gateway.
+
+Collapsible, collapsed by default. It auto-expands, with an explanatory note,
+when the selected model has no resolved info — the exact situation this feature
+addresses.
+
+Fields: **context window**, **max output tokens**, **supportsImages**,
+**supportsPromptCache**. A "reset to detected values" control clears the
+override.
+
+New i18n keys under `settings:providers.customModelInfo.*` in
+`webview-ui/src/i18n/locales/en/settings.json`. Only English is authored; other
+locales fall back until translated.
+
+## Error handling
+
+| Input | Result |
+|---|---|
+| Empty string | Key omitted from overlay |
+| `NaN` / non-numeric | Key omitted, red border |
+| `<= 0` | Key omitted, red border |
+| Valid positive integer | Applied, green border |
+
+`maxTokens` exceeding `contextWindow` is accepted but flagged with an inline
+warning. The 20% context-window clamp in `getModelMaxOutputTokens`
+(`src/shared/api.ts:154`) is **not** a reliable backstop here — three earlier
+branches return before reaching it: reasoning-budget models (line 117), Anthropic
+contexts with `supportsReasoningBudget` or absent `maxTokens` (lines 126-133),
+and `supportsMaxTokens` models honouring an explicit `modelMaxTokens` (line 138).
+The first two are exactly the `anthropic/claude-*` path in this bug report.
+
+Since the clamp cannot be relied on, the inline warning is the actual guard, and
+`TaskHeader`'s `availableInputSpace` must tolerate `reservedForOutput >=
+contextWindow`. Its existing `> 0` guard already returns `0%` rather than a
+negative percentage; the display-hardening change must preserve that guard rather
+than replace it with the new `Math.min(100, …)` clamp.
+
+## Testing
+
+- `applyCustomModelInfo` unit tests: overlay onto existing info; synthesis from
+  absent info; empty/invalid/zero/negative input dropped rather than coerced;
+  `undefined` returned when nothing is available.
+- `TaskHeader` regression test: with `info === undefined`, assert no `7000%`-class
+  output — the percentage element is absent. This is the lock on the reported bug.
+- `TaskHeader` clamp test: `contextTokens` exceeding the window renders `100%`,
+  not more.
+- `useSelectedModel` tests: an empty router model list preserves the configured
+  custom ID rather than substituting the default; a *populated* list that lacks
+  the configured ID also preserves it. The second case is the one the current
+  guard gets wrong.
+- `useSelectedModel` test: the override applies in the not-ready branch (router
+  models still loading), where `getSelectedModel()` is never called.
+- `buildApiHandler` proxy test: `getModel()` reflects the override while other
+  methods and private field access remain intact.
+- `CustomModelInfoSettings` component tests: validation borders, persistence,
+  reset, auto-expansion when info is unresolved.
+
+## Files affected
+
+| File | Change |
+|---|---|
+| `packages/types/src/provider-settings.ts` | Add `customModelInfo` to base schema |
+| `packages/types/src/model.ts` (or sibling) | Add `applyCustomModelInfo` + synthesis base |
+| `webview-ui/src/components/ui/hooks/useSelectedModel.ts` | Apply helper after the ternary (98-113); stop substituting the default for an unlisted ID |
+| `src/api/index.ts` | Proxy-wrap handler in `buildApiHandler` |
+| `webview-ui/src/components/chat/TaskHeader.tsx` | Remove `\|\| 1`; clamp; conditional render |
+| `webview-ui/src/components/settings/CustomModelInfoSettings.tsx` | New component |
+| `webview-ui/src/components/settings/providers/{OpenRouter,Requesty,Unbound,VercelAiGateway,ZooGateway}.tsx` | Mount component |
+| `webview-ui/src/i18n/locales/en/settings.json` | New keys |
+
+## Risks
+
+- **Proxy overhead** — `getModel()` is called frequently (twelve sites in
+  `Task.ts` alone). The decoration is a shallow object spread over a plain
+  object; negligible, but the overlay should not be recomputed per call beyond
+  that.
+- **Stale override after model switch** — an override set for one custom model
+  persists when the user picks a different one. Accepted: the reset control and
+  the collapsed-by-default panel keep this visible. Auto-clearing on model change
+  risks discarding deliberate configuration.

packages/types/src/__tests__/custom-model-info.test.ts

@@ -0,0 +1,68 @@
+import { applyCustomModelInfo, customModelInfoSchema, type ModelInfo } from "../model.js"
+import { providerIdentifiers, providerSettingsSchemaDiscriminated } from "../index.js"
+
+describe("custom model info", () => {
+	it("overlays only supported metadata and preserves provider-owned fields", () => {
+		const model: ModelInfo = {
+			maxTokens: 4096,
+			contextWindow: 8192,
+			supportsImages: false,
+			supportsPromptCache: false,
+			inputPrice: 0.1,
+			outputPrice: 0.2,
+			description: "Provider metadata",
+		}
+
+		expect(
+			applyCustomModelInfo(model, {
+				customModelInfo: {
+					contextWindow: 128_000,
+					maxTokens: 16_384,
+					supportsImages: true,
+					supportsPromptCache: true,
+				},
+			}),
+		).toEqual({
+			...model,
+			contextWindow: 128_000,
+			maxTokens: 16_384,
+			supportsImages: true,
+			supportsPromptCache: true,
+		})
+	})
+
+	it("does not synthesize model info without a valid context window", () => {
+		expect(
+			applyCustomModelInfo(undefined, {
+				customModelInfo: {
+					contextWindow: 0,
+					maxTokens: -1,
+					supportsImages: true,
+				},
+			}),
+		).toBeUndefined()
+	})
+
+	it("synthesizes safe defaults when only a context window is supplied", () => {
+		expect(
+			applyCustomModelInfo(undefined, {
+				customModelInfo: { contextWindow: 64_000, supportsImages: true },
+			}),
+		).toEqual({
+			maxTokens: undefined,
+			contextWindow: 64_000,
+			supportsImages: true,
+			supportsPromptCache: false,
+		})
+	})
+
+	it("rejects unsupported pricing fields in the persisted override schema", () => {
+		expect(customModelInfoSchema.safeParse({ contextWindow: 64_000, inputPrice: 1 }).success).toBe(false)
+		expect(
+			providerSettingsSchemaDiscriminated.safeParse({
+				apiProvider: providerIdentifiers.openrouter,
+				customModelInfo: { contextWindow: 64_000, outputPrice: 1 },
+			}).success,
+		).toBe(false)
+	})
+})