feat(models): clamp-on-write for pinned thinking levels + DEFAULT_THINKING_ORDER in the core barrel
Fix 1 — clamp, never reject (owner ruling, interlude 2026-09-22): a model swap with a pinned thinking level must JUST WORK. The owner hit the write-side 400 live (Claude Fable 5.1/xhigh → GLM 5.3 flash: "z-ai/glm-5.3-flash does not support thinking level xhigh"). The supported-level verdicts in validateChatSettings and validateModelsSettings are DEAD — the WRITE clamps the pin to the new model's nearest supported level (pi's clampThinkingLevel, the same translation piEngineForSpec applies at runtime) and stores the clamped result; the server is the truth — the record shows the clamped level after write. New pure helpers in model-catalog.ts, applied by ALL THREE write routes (section, defaults, per-chat — one verdict seam, so every surface that used the old 400 gets the clamp uniformly): - clampModelsLevels: every slot's level clamps against the SLOT'S OWN model (the parser already requires a model on each slot write). - clampChatLevel: the body's own model, else the record's resolved strong slot (the chain's fallback); no model anywhere → rides unchecked (runtime clamps at engine build; the turn gate refuses the model-less chat anyway). Fix 2 — DEFAULT_THINKING_ORDER moves from config.ts to core/settings (core barrel, AUTH_KEYS precedent — one source of truth, no drift). The UI imports it to auto-populate a thinking level on FRESH model picks; the ordering has no other client role. Legacy unset records keep the runtime defaultThinkingLevel fallback at engine build untouched. Import sites updated: agent/models.ts, tests. Verified live on 3004: the owner's exact case stores max (not 400); a bare chat-level write clamps to low against the strong slot's glm; unknown provider/model still 400s with the provider list. Tests: the level-verdict suites became clamp suites (the owner's case pinned verbatim: xhigh→max, minimal→low, off→low, supported pass-through, level-less slots ride); the machine-aware expectBareLevelVerdict chain is retired — clampChatLevel takes the strong slot as an explicit param, so the tests are pure again. Gates: typecheck, lint, 839/839, build.
This commit is contained in:
amosblomqvist
+62
-51
@@ -17,7 +17,7 @@
|
||||
* set; nothing here changes.
|
||||
*/
|
||||
|
||||
import { getSupportedThinkingLevels, type Api, type Model } from "@earendil-works/pi-ai";
|
||||
import { clampThinkingLevel, getSupportedThinkingLevels, type Api, type Model } from "@earendil-works/pi-ai";
|
||||
import type { ThinkingLevel } from "@earendil-works/pi-agent-core";
|
||||
import { piModels } from "../agent/models.ts";
|
||||
import { settingsStore } from "../store/settings.ts";
|
||||
@@ -34,7 +34,7 @@ import type {
|
||||
} from "../core/contracts/models.ts";
|
||||
import type { EffectiveModelSpec, EffectiveSlot } from "../core/settings-resolution.ts";
|
||||
import { AUTH_PROVIDER_KEYS } from "../core/auth-keys.ts";
|
||||
import { BACKGROUND_MODEL_ROLES, type BackgroundModelRole, type ModelSlotName, type ModelsSettings } from "../core/settings.ts";
|
||||
import { BACKGROUND_MODEL_ROLES, MODEL_SLOTS, type BackgroundModelRole, type ModelSlotName, type ModelsSettings } from "../core/settings.ts";
|
||||
|
||||
/** The picker's client-safe model entry (contract: core/contracts/models.ts). */
|
||||
export type ModelDto = ModelCatalogEntry;
|
||||
@@ -196,38 +196,17 @@ export function validateChatSettings(settings: ChatSettings): SettingsValidation
|
||||
return { ok: false, error: `No model "${id}" in the ${provider} catalog` };
|
||||
}
|
||||
}
|
||||
if (settings.thinkingLevel !== undefined) {
|
||||
// The conversation chain's fallback is the strong slot — the model
|
||||
// a settings-less chat would run. BYOK (owner, 2026-09-22): with no
|
||||
// model to check against (no chat.model, no strong slot), the level
|
||||
// rides unchecked — it clamps against whatever the chain later holds.
|
||||
const strong = effectiveSettings().models.strong;
|
||||
const provider = settings.model?.provider ?? strong?.provider;
|
||||
const id = settings.model?.id ?? strong?.id;
|
||||
if (!provider || !id) return { ok: true };
|
||||
const model = piModels().getModel(provider, id);
|
||||
if (!model) return { ok: false, error: `No model "${id}" in the ${provider} catalog` };
|
||||
const supported = getSupportedThinkingLevels(model);
|
||||
if (!supported.includes(settings.thinkingLevel)) {
|
||||
return {
|
||||
ok: false,
|
||||
error: `${id} does not support thinking level "${settings.thinkingLevel}"`,
|
||||
supported,
|
||||
};
|
||||
}
|
||||
}
|
||||
// BYOK interlude (owner, 2026-09-22): thinking levels NEVER reject a
|
||||
// write — a model swap with a pinned level JUST WORKS. The WRITE clamps
|
||||
// the pin to the new model's nearest supported level (clampChatLevel
|
||||
// below, applied by the routes) and stores the clamped result; a bare
|
||||
// level with no model anywhere rides unchecked and clamps against
|
||||
// whatever the chain later holds at engine build.
|
||||
return { ok: true };
|
||||
}
|
||||
/** Validate a models-section write (PUT /api/settings/models) against the
|
||||
* catalog — the same gate the chat section gets. A slot must NAME a real
|
||||
* provider+model (owner, 2026-09-22: no factory floor exists, so a level
|
||||
* alone can never run — the parser rejects it, this gate re-checks the
|
||||
* model); a thinkingLevel must be supported by the slot's own model.
|
||||
* keyPresent is NOT gated here (a stored pin for a provider whose key
|
||||
* lapsed stays valid — it fails honestly at stream time, and the picker
|
||||
* stops offering it). */
|
||||
|
||||
export function validateModelsSettings(settings: ModelsSettings): SettingsValidation {
|
||||
for (const name of ["strong", "cheap", "vision"] as const) {
|
||||
for (const name of MODEL_SLOTS) {
|
||||
const slot = settings[name];
|
||||
if (!slot) continue;
|
||||
const verdict = validateSlotSpec(name, slot);
|
||||
@@ -247,25 +226,57 @@ function validateSlotSpec(name: ModelSlotName, slot: { model?: { provider: strin
|
||||
return { ok: false, error: `No model "${id}" in the ${provider} catalog` };
|
||||
}
|
||||
}
|
||||
if (slot.thinkingLevel !== undefined) {
|
||||
// The write-side parser requires slot.model (owner, 2026-09-22 — a
|
||||
// level-only slot can never run), so this fallback is tolerance for
|
||||
// future callers; with no model to check against, accept (the
|
||||
// engine build clamps whatever the chain later holds).
|
||||
const resolved = effectiveSettings().models[name];
|
||||
const provider = slot.model?.provider ?? resolved?.provider;
|
||||
const id = slot.model?.id ?? resolved?.id;
|
||||
if (!provider || !id) return { ok: true };
|
||||
const model = piModels().getModel(provider, id);
|
||||
if (!model) return { ok: false, error: `No model "${id}" in the ${provider} catalog` };
|
||||
const supported = getSupportedThinkingLevels(model);
|
||||
if (!supported.includes(slot.thinkingLevel)) {
|
||||
return {
|
||||
ok: false,
|
||||
error: `${id} does not support thinking level "${slot.thinkingLevel}"`,
|
||||
supported,
|
||||
};
|
||||
}
|
||||
}
|
||||
// BYOK interlude (owner, 2026-09-22): the supported-level verdict is
|
||||
// DEAD — a slot write with a pinned level never rejects; the WRITE
|
||||
// clamps the level to the model's nearest supported (clampModelsLevels
|
||||
// below, applied by the route) and stores the clamped result.
|
||||
return { ok: true };
|
||||
}
|
||||
|
||||
// ─── clamp-on-write (owner ruling, interlude 2026-09-22) ─────────────────────
|
||||
// A model swap with a pinned thinking level must JUST WORK: the pin
|
||||
// translates to the new model's nearest supported level — pi's
|
||||
// clampThinkingLevel, the SAME translation piEngineForSpec applies at
|
||||
// runtime — and the clamped value is what lands in the record. The
|
||||
// server is the truth: the record shows the clamped level after write.
|
||||
|
||||
/** Clamp one level against one model (identity when the model is unknown
|
||||
* or the level already supported — the existence gate rejects unknown
|
||||
* models separately, so this only ever sees real ones in practice). */
|
||||
function clampLevelFor(provider: string, id: string, level: ThinkingLevel): ThinkingLevel {
|
||||
const model = piModels().getModel(provider, id);
|
||||
return model ? clampThinkingLevel(model, level) : level;
|
||||
}
|
||||
|
||||
/** The models-section write's level translation: every slot's pinned
|
||||
* thinkingLevel clamps against THE SLOT'S OWN model (the parser requires
|
||||
* a model on every slot write — owner, 2026-09-22). Pure: catalog in,
|
||||
* clamped section out; the route applies it after validation. */
|
||||
export function clampModelsLevels(settings: ModelsSettings): ModelsSettings {
|
||||
const out: ModelsSettings = {};
|
||||
for (const name of MODEL_SLOTS) {
|
||||
const slot = settings[name];
|
||||
if (!slot?.model) continue;
|
||||
out[name] =
|
||||
slot.thinkingLevel !== undefined
|
||||
? {
|
||||
model: slot.model,
|
||||
thinkingLevel: clampLevelFor(slot.model.provider, slot.model.id, slot.thinkingLevel),
|
||||
}
|
||||
: slot;
|
||||
}
|
||||
return out;
|
||||
}
|
||||
|
||||
/** The chat-section write's level translation: a pinned thinkingLevel
|
||||
* clamps against the body's own model, else the record's resolved strong
|
||||
* slot (the conversation chain's fallback — what a settings-less chat
|
||||
* would run). No model anywhere → rides unchecked (runtime clamps at
|
||||
* engine build; the turn gate refuses the model-less chat anyway). */
|
||||
export function clampChatLevel(settings: ChatSettings, strong: EffectiveSlot | null): ChatSettings {
|
||||
if (settings.thinkingLevel === undefined) return settings;
|
||||
const provider = settings.model?.provider ?? strong?.provider;
|
||||
const id = settings.model?.id ?? strong?.id;
|
||||
if (!provider || !id) return settings;
|
||||
return { ...settings, thinkingLevel: clampLevelFor(provider, id, settings.thinkingLevel) };
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user