@gtrabanco/pi-nan-provider 0.8.0 → 0.10.2

package/scripts/generate-models.ts

@@ -8,16 +8,28 @@
  * package. Run before publishing (`bun run generate-models`, wired into
  * `prepublishOnly`).
  *
+ * Sources of truth (maintainer instruction, reaffirmed 2026-09-29):
+ * - https://nan.builders/docs/models decides WHICH models exist — which ids
+ *   belong in the chat catalog, which are premium/live-only, which were
+ *   removed by the provider, which are not chat models at all. Every
+ *   exclusion cites it.
+ * - models.dev (provider "nan") supplies the numeric limits for the ids it
+ *   documents; a divergence from the docs is recorded in MANUAL_OVERRIDES
+ *   with a source note.
+ *
  * Provenance rules (enforced, not decorative):
  * - Every emitted number must come from models.dev or an explicit
  *   MANUAL_OVERRIDES note recording where it was confirmed. Nothing invented.
- * - The founding models (qwen3.6, gemma4, deepseek-v4-flash, mimo-v2.5) MUST
- *   exist on models.dev with complete limits, or this script exits non-zero.
+ * - The official community chat models (qwen3.6, gemma4, deepseek-v4-flash,
+ *   mimo-v2.6-flash) MUST reach the catalog — from models.dev or from a
+ *   MANUAL_ONLY entry — or this script exits non-zero.
  * - Any other models.dev entry missing `limit.context`/`limit.output` is
- *   skipped and flagged "needs manual verification" — never guessed.
+ *   skipped and flagged "needs manual verification" — never guessed. Entries
+ *   that are not chat models are excluded via NON_CHAT_MODEL_IDS instead.
  * - models.dev modalities are intersected with pi's supported input set
- *   ("text" | "image"); e.g. mimo-v2.5's audio input is not representable in
- *   pi's Model type and is dropped from `input` (noted on the entry).
+ *   ("text" | "image"); e.g. mimo-v2.6-flash's audio input is not
+ *   representable in pi's Model type and is dropped from `input` (noted on
+ *   the entry).
  */
 
 import type { GeneratedModelEntry } from "../src/fetch-models.ts";
@@ -39,17 +51,23 @@ const REASONING_EFFORT_VALUES_FROM_DOCS: Record<string, string[]> = {
 	...REASONING_EFFORT_VALUES,
 	// deepseek-v4-flash: any value (no effect) — model decides per request
 	"deepseek-v4-flash": [],
-	// qwen3.8-flash, mimo-v2.5: accepted, depth not adjustable
+	// qwen3.8-flash, mimo-v2.6-flash: accepted, depth not adjustable
 	"qwen3.8-flash": [],
-	"mimo-v2.5": [],
+	"mimo-v2.6-flash": [],
 };
 
 const MODELS_DEV_API_URL = "https://models.dev/api.json";
 const SOURCE_PROVIDER_ID = "nan";
 const FETCH_TIMEOUT_MS = 15_000;
 
-/** Models the founding prompt requires in the catalog; absence is fatal. */
-const REQUIRED_MODEL_IDS = ["qwen3.6", "gemma4", "deepseek-v4-flash", "mimo-v2.5"] as const;
+/**
+ * Official community chat models per https://nan.builders/docs/models
+ * (checked 2026-09-29; glm5.3 is premium-tier and deliberately live-only).
+ * Each MUST reach the catalog — from models.dev or from a MANUAL_ONLY entry —
+ * or generation fails. The docs decide this set, not models.dev: models.dev
+ * dropped mimo-v2.5 together with the docs in 2026-09.
+ */
+const REQUIRED_MODEL_IDS = ["qwen3.6", "gemma4", "deepseek-v4-flash", "mimo-v2.6-flash"] as const;
 
 /** pi's Model.input only supports these values. */
 const PI_SUPPORTED_INPUT = new Set(["text", "image"]);
@@ -75,6 +93,20 @@ const MANUAL_NOTES: Record<string, string> = {
 const PROVIDER_REMOVED_MODEL_IDS: Record<string, string> = {
 	"glm5.2":
 		"removed by NaN (2026-09-05); absent from the official chat model list in https://nan.builders/openapi.json and https://nan.builders/docs/models (checked 2026-09-07) while models.dev provider nan still listed it — excluded so regeneration does not resurrect it",
+	"mimo-v2.5":
+		"removed by NaN (checked 2026-09-29): zero mentions in https://nan.builders/docs/models and in the model param description of https://nan.builders/openapi.json, and models.dev provider nan dropped it as well — superseded by mimo-v2.6-flash; excluded so regeneration does not resurrect it",
+};
+
+/**
+ * models.dev entries that are not chat models. They can never join a chat
+ * catalog (their output is an image, and models.dev reports limit.output 0),
+ * so they are excluded with the docs as the reason instead of being flagged
+ * "needs manual verification" forever — that flag would suggest limits worth
+ * confirming for a model this package must not serve as chat.
+ */
+const NON_CHAT_MODEL_IDS: Record<string, string> = {
+	"qwen-image-2.1":
+		"image-generation model (text→image; output modality image) documented in https://nan.builders/docs/models and https://nan.builders/openapi.json (checked 2026-09-29) — not a chat model, out of the static chat catalog (models.dev reports limit.output 0)",
 };
 
 /**
@@ -205,6 +237,7 @@ function buildManualOnlyModelEntry(
 			contextWindow: override?.contextWindow ?? 1_048_576,
 			maxTokens: override?.maxTokens ?? 131_072,
 			reasoningEffortValues,
+			...(override?.thinkingLevelMap ? { thinkingLevelMap: { ...override.thinkingLevelMap } } : {}),
 			compat: { ...NAN_COMPAT },
 			notes: [
 				NAN_COMPAT_NOTE,
@@ -221,6 +254,10 @@ function convertModel(modelId: string, m: ModelsDevModel): GeneratedModel | { sk
 	if (removedReason) {
 		return { skip: `provider-removed: "${modelId}" excluded from the catalog (${removedReason})` };
 	}
+	const nonChatReason = NON_CHAT_MODEL_IDS[modelId];
+	if (nonChatReason) {
+		return { skip: `non-chat: "${modelId}" excluded from the chat catalog (${nonChatReason})` };
+	}
 	const liveOnlyReason = LIVE_ONLY_MODEL_IDS[modelId];
 	if (liveOnlyReason) {
 		return { skip: `live-only: "${modelId}" kept out of the static catalog (${liveOnlyReason})` };
@@ -264,6 +301,7 @@ function convertModel(modelId: string, m: ModelsDevModel): GeneratedModel | { sk
 			contextWindow: override?.contextWindow ?? contextWindow,
 			maxTokens: override?.maxTokens ?? maxTokens,
 			reasoningEffortValues,
+			...(override?.thinkingLevelMap ? { thinkingLevelMap: { ...override.thinkingLevelMap } } : {}),
 			compat: { ...NAN_COMPAT },
 			notes: [
 				NAN_COMPAT_NOTE,
@@ -301,8 +339,20 @@ async function main(): Promise<void> {
 		}
 	}
 
-	// Add manual-only models (not yet on models.dev).
+	// Add manual-only models (models.dev provider "nan" does not list them).
+	// When models.dev starts listing one, the manual entry is NOT emitted —
+	// otherwise the id would reach the catalog twice — and the reason plus the
+	// original manual provenance are attached to the models.dev-derived entry,
+	// so the manual data never disappears silently.
+	const entryById = new Map(entries.map((entry) => [entry.id, entry]));
 	for (const [modelId, detail] of Object.entries(MANUAL_ONLY_MODEL_IDS)) {
+		const existing = entryById.get(modelId);
+		if (existing) {
+			const note = `manual-only: "${modelId}" skipped: models.dev provider "${SOURCE_PROVIDER_ID}" now lists it, so the manual-only entry is not emitted and the models.dev values are used (${detail})`;
+			existing.notes = [...(existing.notes ?? []), note];
+			console.warn(`generate-models: skipped ${note}`);
+			continue;
+		}
 		const override = MANUAL_OVERRIDES[modelId];
 		if (override) {
 			console.log(`generate-models: manual override for "${modelId}" (${Object.keys(override).filter((k) => k !== "note").join(", ")})`);
@@ -318,8 +368,9 @@ async function main(): Promise<void> {
 	const missingRequired = REQUIRED_MODEL_IDS.filter((id) => !entryIds.has(id));
 	if (missingRequired.length > 0) {
 		fail(
-			`required models missing from models.dev provider "${SOURCE_PROVIDER_ID}": ${missingRequired.join(", ")} — ` +
-				"do NOT invent their data; confirm and add them via MANUAL_OVERRIDES with a source note.",
+			`required models missing from the catalog: ${missingRequired.join(", ")} (official community chat models per ` +
+				"https://nan.builders/docs/models) — do NOT invent their data; confirm them in the docs and add them " +
+				"via MANUAL_OVERRIDES (models.dev lists them) or MANUAL_ONLY_MODEL_IDS (models.dev does not), with a source note.",
 		);
 	}
 
@@ -333,6 +384,9 @@ async function main(): Promise<void> {
 			...Object.entries(PROVIDER_REMOVED_MODEL_IDS).map(
 				([id, reason]) => `provider-removed: "${id}" excluded from the catalog (${reason})`,
 			),
+			...Object.entries(NON_CHAT_MODEL_IDS).map(
+				([id, reason]) => `non-chat: "${id}" excluded from the chat catalog (${reason})`,
+			),
 			...Object.entries(LIVE_ONLY_MODEL_IDS).map(
 				([id, reason]) => `live-only: "${id}" kept out of the static catalog (${reason})`,
 			),

package/scripts/manual-overrides.ts

@@ -37,13 +37,20 @@ export interface ManualModelOverride {
 	maxTokens?: number;
 	/** Reasoning effort values as declared by NaN docs. An empty array means the parameter is accepted but depth is model-managed (not adjustable by the user). */
 	reasoningEffortValues?: string[];
+	/**
+	 * Mapping from pi-ai's user-facing thinking levels to model-specific effort
+	 * strings. Only the "off" key is needed to suppress runaway reasoning.
+	 * When `reasoning: "off"` maps to undefined in pi-ai, thinkingLevelMap.off
+	 * supplies the effort value that NaN understands (e.g. "none" or "minimal").
+	 */
+	thinkingLevelMap?: Partial<Record<string, string | null>>;
 	/** Required provenance note; emitted verbatim onto the generated entry. */
 	note: string;
 }
 
 /**
- * Reasoning effort values sourced from the NaN docs (https://nan.builders/docs/models
- * #controlling-reasoning, checked 2026-09-25).
+ * Manual reasoning effort values sourced from the NaN docs
+ * (https://nan.builders/docs/models #controlling-reasoning, checked 2026-09-29).
  *
  * models.dev has NO `reasoning_effort_values` field — only `reasoning_options`
  * (which is [{type:"toggle"}] or []), so this mapping must be hand-maintained
@@ -53,28 +60,48 @@ export interface ManualModelOverride {
  *
  * Per-model contract from the docs:
  * - glm5.3, glm5.3-flash: fully controllable (low/medium/high/max)
+ *   + minimal also accepted (issue #16, measured 2026-09-27): minimal
+ *     yields 38 reasoning tokens vs 2,288 for high — the docs simply
+ *     omitted it because it was not originally advertised, but it IS
+ *     accepted and effective by the gateway.
  * - qwen3.6, gemma4: none/minimal skip reasoning, others cap depth
- * - deepseek-v4-flash: any value accepted but model decides per-request
- * - qwen3.8-flash, mimo-v2.5, mimo-v2.6-flash: accepted, depth not adjustable
+ * - deepseek-v4-flash: any value accepted but only "none" deterministically
+ *   disables reasoning (issue #16, measured 2026-09-27); every other value
+ *   (minimal/low/medium/high/omitted) leaves 13-14K reasoning tokens with
+ *   runaway reasoning and finish_reason:"length" with ZERO answer. Since only
+ *   "none" has a measurable effect, the catalog declares ["none"].
+ * - qwen3.8-flash, mimo-v2.6-flash: accepted, depth not adjustable
+ *
+ * (mimo-v2.5, which used to share that last row, was removed by NaN — absent
+ * from the docs and from openapi.json, checked 2026-09-29.)
  *
  * A model with reasoning_effort_values=[] means the parameter is accepted but
  * the model manages its own reasoning depth — it is never an error.
  */
 export const REASONING_EFFORT_VALUES: Record<string, string[]> = {
 	"glm5.3": ["low", "medium", "high", "max"],
-	"glm5.3-flash": ["low", "medium", "high", "max"],
+	"glm5.3-flash": ["minimal", "low", "medium", "high", "max"],
 	"qwen3.6": ["none", "minimal", "low", "medium", "high", "max"],
 	"gemma4": ["none", "minimal", "low", "medium", "high", "max"],
 };
 
 /**
- * Models that NaN serves but models.dev provider nan hasn't listed yet.
- * Mirrors the generator's MANUAL_ONLY_MODEL_IDS — kept in sync so the
- * generator and the tests share the same source.
+ * Provenance for the manual-only mimo-v2.6-flash entry, shared by
+ * MANUAL_ONLY_MODEL_IDS (generator) and MANUAL_ONLY_MODELS (tests) so both
+ * sides assert the exact same string.
+ */
+const MANUAL_ONLY_NOTE =
+	"omnimodal model (text, image, audio input) served by NaN; limits and modalities from https://nan.builders/docs/models#mimo-v2-6-flash (checked 2026-09-25, quotas re-checked 2026-09-29): 1,048,576 context / 131,072 max output / 1.0B monthly quota per member. Reasoning: accepted, depth not adjustable (docs table, checked 2026-09-29). Tool calling: yes. Streaming: yes. Input modalities (pi-representable): text, image (audio not representable in pi's Model type). Fallback entry: emitted only while models.dev provider nan does not list the model — models.dev started listing it on 2026-09-29, so the generator then keeps the models.dev values and attaches this note to that entry instead of emitting a second one.";
+
+/**
+ * Models NaN serves that models.dev provider nan may not list. Mirrors the
+ * generator's MANUAL_ONLY_MODEL_IDS — kept in sync so the generator and the
+ * tests share the same source. An entry stops being emitted as soon as
+ * models.dev lists the model; the note above then rides on the models.dev
+ * entry, so the manual data never vanishes silently.
  */
 export const MANUAL_ONLY_MODEL_IDS: Record<string, string> = {
-	"mimo-v2.6-flash":
-		"omnimodal model (text, image, audio input) served by NaN, not yet listed on models.dev provider nan (checked 2026-09-25); included so the live /models refresh does not hand it UNKNOWN_MODEL_LIMITS. Same limits as mimo-v2.5: 1,048,576 / 131,072 / 1.0B monthly quota per member. Reasoning: accepted, depth not adjustable. Tool calling: yes. Streaming: yes. Input modalities (pi-representable): text, image (audio not representable in pi's Model type).",
+	"mimo-v2.6-flash": MANUAL_ONLY_NOTE,
 };
 
 /**
@@ -89,15 +116,52 @@ export const MANUAL_ONLY_MODELS: Record<string, ManualModelOverride> = {
 		reasoningEffortValues: [],
 		contextWindow: 1_048_576,
 		maxTokens: 131_072,
-		note: "omnimodal model (text, image, audio input) served by NaN, not yet listed on models.dev provider nan (checked 2026-09-25); included so the live /models refresh does not hand it UNKNOWN_MODEL_LIMITS. Same limits as mimo-v2.5: 1,048,576 / 131,072 / 1.0B monthly quota per member. Reasoning: accepted, depth not adjustable. Tool calling: yes. Streaming: yes. Input modalities (pi-representable): text, image (audio not representable in pi's Model type).",
+		note: MANUAL_ONLY_NOTE,
 	},
 };
 
+/**
+ * Manual overrides keyed by model id. Fields override the models.dev-derived
+ * entry one-for-one. Every override carries a provenance note.
+ *
+ * Issue #16 (measured 2026-09-27 against live gateway, repro included):
+ * - deepseek-v4-flash: reasoning_effort_values=["none"] (only "none" suppresses);
+ *   thinkingLevelMap={off:"none"} (supplies reasoning_effort when pi maps
+ *   reasoning:"off" → undefined).
+ * - glm5.3-flash: reasoningEffortValues includes "minimal" (not documented but
+ *   effective); thinkingLevelMap={off:"minimal"} (supplies reasoning_effort
+ *   when pi maps reasoning:"off" → undefined).
+ * - qwen3.6, gemma4: the NaN docs already document `none` as "skip the
+ *   reasoning phase entirely" and state that with no parameter the model
+ *   defaults to reasoning ON (16,384-token budget), so off must send
+ *   `none` explicitly — thinkingLevelMap={off:"none"} (NaN docs
+ *   #controlling-reasoning, checked 2026-10-01). No measurement needed: the
+ *   docs are explicit about both the default and the `none` behavior.
+ */
 export const MANUAL_OVERRIDES: Record<string, ManualModelOverride> = {
 	"deepseek-v4-flash": {
 		input: ["text", "image"],
-		// Any value accepted but model decides per-request (no effect).
-		reasoningEffortValues: [],
-		note: "input includes image: NaN serves the Vision-Exp variant ('takes images as input', https://nan.builders/docs/models; the image_url content-parts in https://nan.builders/openapi.json list deepseek-v4-flash among the vision models). models.dev provider nan also lists text+image now (DeepSeek V4.1 Flash entry, checked 2026-09-13; its 2026-09-07 snapshot listed text only), so this override is kept as a pin for the vision capability rather than as a divergence. reasoning_effort_values=[]: the model decides per-request how much to reason (https://nan.builders/docs/models, checked 2026-09-25).",
+		reasoningEffortValues: ["none"],
+		thinkingLevelMap: { off: "none" },
+		note:
+			"input includes image: NaN serves the Vision-Exp variant ('takes images as input', https://nan.builders/docs/models; the image_url content-parts in https://nan.builders/openapi.json list deepseek-v4-flash among the vision models). models.dev provider nan also lists text+image now (DeepSeek V4.1 Flash entry, checked 2026-09-13; its 2026-09-07 snapshot listed text only), so this override is kept as a pin for the vision capability rather than as a divergence. " +
+			"reasoning_effort_values=[\"none\"] (issue #16, measured 2026-09-27 against live gateway, repro included in the issue): only \"none\" deterministically disables reasoning (0 reasoning tokens, answer produced). Every other value leaves 13-14K reasoning tokens and causes runaway reasoning (finish_reason:\"length\" with ZERO answer text). The catalog declares only [\"none\"] because the remaining values are useless noise. " +
+			"thinkingLevelMap { off: \"none\" } (issue #16, measured 2026-09-27): when pi maps reasoning:\"off\" to reasoningEffort:undefined, the thinkingLevelMap.off=\"none\" supplies the effort so the gateway receives reasoning_effort:\"none\" and does NOT enter runaway reasoning mode.",
+	},
+	"glm5.3-flash": {
+		thinkingLevelMap: { off: "minimal" },
+		note:
+			"thinkingLevelMap { off: \"minimal\" } (issue #16, measured 2026-09-27 against live gateway, repro included in the issue): when pi maps reasoning:\"off\" to reasoningEffort:undefined, the thinkingLevelMap.off=\"minimal\" supplies the effort so the gateway receives reasoning_effort:\"minimal\" (38 reasoning tokens + answer). Note: \"none\" does NOT suppress reasoning on glm5.3-flash (13,382 tokens) — only \"minimal\" was measured to be effective. " +
+			"reasoning_effort_values includes \"minimal\" (not documented by NaN docs but accepted and effective by the gateway, issue #16).",
+	},
+	"qwen3.6": {
+		thinkingLevelMap: { off: "none" },
+		note:
+			"thinkingLevelMap { off: \"none\" } (issue #16): NaN docs state that with no reasoning_effort parameter every model uses its own default and that default is reasoning ON for qwen3.6 (16,384-token budget), while `none` and `minimal` skip the reasoning phase entirely (https://nan.builders/docs/models#controlling-reasoning, checked 2026-10-01). pi maps reasoning:\"off\" to reasoningEffort:undefined, so without this map `thinking: off` sent no parameter and the model reasoned anyway; the map sends reasoning_effort:\"none\" and actually disables reasoning.",
+	},
+	"gemma4": {
+		thinkingLevelMap: { off: "none" },
+		note:
+			"thinkingLevelMap { off: \"none\" } (issue #16): NaN docs state that with no reasoning_effort parameter gemma4 defaults to reasoning ON (16,384-token budget), while `none`/`minimal` skip reasoning entirely (https://nan.builders/docs/models#controlling-reasoning, checked 2026-10-01). pi maps reasoning:\"off\" to reasoningEffort:undefined, so without this map `thinking: off` sent no parameter and the model reasoned anyway; the map sends reasoning_effort:\"none\" and actually disables reasoning.",
 	},
 };

package/scripts/models.generated.ts

@@ -1,7 +1,7 @@
 // This file is auto-generated by scripts/generate-models.ts
 // Do not edit manually — run `bun run generate-models` to update.
 //
-// Source: https://models.dev/api.json (provider "nan"), fetched 2026-09-27T21:25:24.946Z
+// Source: https://models.dev/api.json (provider "nan"), fetched 2026-10-01T10:18:48.654Z
 // Provenance: every contextWindow/maxTokens/input/cost value traces to
 // models.dev or to the per-entry notes below. Nothing is invented; entries
 // models.dev documents incompletely are omitted and flagged instead.
@@ -28,7 +28,12 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 		},
 		"contextWindow": 1000000,
 		"maxTokens": 384000,
-		"reasoningEffortValues": [],
+		"reasoningEffortValues": [
+			"none"
+		],
+		"thinkingLevelMap": {
+			"off": "none"
+		},
 		"compat": {
 			"supportsDeveloperRole": false,
 			"supportsReasoningEffort": true,
@@ -38,7 +43,7 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 		},
 		"notes": [
 			"compat matches the maintainer's working ~/.pi/agent/models.json LiteLLM config for api.nan.builders (2026-09-04): supportsDeveloperRole false, supportsReasoningEffort true, maxTokensField max_tokens. NaN's docs example sets only supportsDeveloperRole: true and is not battle-tested. supportsFinishReason true (2026-09-13, issue #2): the LiteLLM gateway intermittently closes SSE streams before emitting finish_reason; with true pi-ai raises 'Stream ended without finish_reason', which matches pi-ai's retryable-provider pattern ('ended without') and is retried automatically, whereas false silently synthesized stop/toolUse and stalled the turn mid-answer. supportsUsageInStreaming true (2026-09-16, issue #7): NaN's published schema is silent about stream_options, but the live gateway honors it — two identical streaming calls per model, differing only in stream_options: { include_usage: true }, returned 0 usage chunks without it and exactly 1 with it (prompt/completion/reasoning/cached token counts) on deepseek-v4-flash, glm5.3-flash, qwen3.6, mimo-v2.5 and gemma4, and a real pi session then recorded token counts where it recorded zeros. pi-ai only sends stream_options when this is not false, and the sanitizer forwards it when the model declares true, so chat models opt in by default and usage is reported (issue #4). A model that does not report streaming usage can still opt out per model with a models.json compat override (supportsUsageInStreaming: false); the sanitizer then strips stream_options and the payload stays strict.",
-			"input includes image: NaN serves the Vision-Exp variant ('takes images as input', https://nan.builders/docs/models; the image_url content-parts in https://nan.builders/openapi.json list deepseek-v4-flash among the vision models). models.dev provider nan also lists text+image now (DeepSeek V4.1 Flash entry, checked 2026-09-13; its 2026-09-07 snapshot listed text only), so this override is kept as a pin for the vision capability rather than as a divergence. reasoning_effort_values=[]: the model decides per-request how much to reason (https://nan.builders/docs/models, checked 2026-09-25)."
+			"input includes image: NaN serves the Vision-Exp variant ('takes images as input', https://nan.builders/docs/models; the image_url content-parts in https://nan.builders/openapi.json list deepseek-v4-flash among the vision models). models.dev provider nan also lists text+image now (DeepSeek V4.1 Flash entry, checked 2026-09-13; its 2026-09-07 snapshot listed text only), so this override is kept as a pin for the vision capability rather than as a divergence. reasoning_effort_values=[\"none\"] (issue #16, measured 2026-09-27 against live gateway, repro included in the issue): only \"none\" deterministically disables reasoning (0 reasoning tokens, answer produced). Every other value leaves 13-14K reasoning tokens and causes runaway reasoning (finish_reason:\"length\" with ZERO answer text). The catalog declares only [\"none\"] because the remaining values are useless noise. thinkingLevelMap { off: \"none\" } (issue #16, measured 2026-09-27): when pi maps reasoning:\"off\" to reasoningEffort:undefined, the thinkingLevelMap.off=\"none\" supplies the effort so the gateway receives reasoning_effort:\"none\" and does NOT enter runaway reasoning mode."
 		],
 		"extras": {
 			"id": "deepseek-v4-flash",
@@ -71,7 +76,8 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"cost": {
 				"input": 0,
 				"output": 0
-			}
+			},
+			"canonical_model_id": "deepseek/deepseek-v4.1-flash"
 		}
 	},
 	{
@@ -98,6 +104,9 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"high",
 			"max"
 		],
+		"thinkingLevelMap": {
+			"off": "none"
+		},
 		"compat": {
 			"supportsDeveloperRole": false,
 			"supportsReasoningEffort": true,
@@ -106,7 +115,8 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"maxTokensField": "max_tokens"
 		},
 		"notes": [
-			"compat matches the maintainer's working ~/.pi/agent/models.json LiteLLM config for api.nan.builders (2026-09-04): supportsDeveloperRole false, supportsReasoningEffort true, maxTokensField max_tokens. NaN's docs example sets only supportsDeveloperRole: true and is not battle-tested. supportsFinishReason true (2026-09-13, issue #2): the LiteLLM gateway intermittently closes SSE streams before emitting finish_reason; with true pi-ai raises 'Stream ended without finish_reason', which matches pi-ai's retryable-provider pattern ('ended without') and is retried automatically, whereas false silently synthesized stop/toolUse and stalled the turn mid-answer. supportsUsageInStreaming true (2026-09-16, issue #7): NaN's published schema is silent about stream_options, but the live gateway honors it — two identical streaming calls per model, differing only in stream_options: { include_usage: true }, returned 0 usage chunks without it and exactly 1 with it (prompt/completion/reasoning/cached token counts) on deepseek-v4-flash, glm5.3-flash, qwen3.6, mimo-v2.5 and gemma4, and a real pi session then recorded token counts where it recorded zeros. pi-ai only sends stream_options when this is not false, and the sanitizer forwards it when the model declares true, so chat models opt in by default and usage is reported (issue #4). A model that does not report streaming usage can still opt out per model with a models.json compat override (supportsUsageInStreaming: false); the sanitizer then strips stream_options and the payload stays strict."
+			"compat matches the maintainer's working ~/.pi/agent/models.json LiteLLM config for api.nan.builders (2026-09-04): supportsDeveloperRole false, supportsReasoningEffort true, maxTokensField max_tokens. NaN's docs example sets only supportsDeveloperRole: true and is not battle-tested. supportsFinishReason true (2026-09-13, issue #2): the LiteLLM gateway intermittently closes SSE streams before emitting finish_reason; with true pi-ai raises 'Stream ended without finish_reason', which matches pi-ai's retryable-provider pattern ('ended without') and is retried automatically, whereas false silently synthesized stop/toolUse and stalled the turn mid-answer. supportsUsageInStreaming true (2026-09-16, issue #7): NaN's published schema is silent about stream_options, but the live gateway honors it — two identical streaming calls per model, differing only in stream_options: { include_usage: true }, returned 0 usage chunks without it and exactly 1 with it (prompt/completion/reasoning/cached token counts) on deepseek-v4-flash, glm5.3-flash, qwen3.6, mimo-v2.5 and gemma4, and a real pi session then recorded token counts where it recorded zeros. pi-ai only sends stream_options when this is not false, and the sanitizer forwards it when the model declares true, so chat models opt in by default and usage is reported (issue #4). A model that does not report streaming usage can still opt out per model with a models.json compat override (supportsUsageInStreaming: false); the sanitizer then strips stream_options and the payload stays strict.",
+			"thinkingLevelMap { off: \"none\" } (issue #16): NaN docs state that with no reasoning_effort parameter gemma4 defaults to reasoning ON (16,384-token budget), while `none`/`minimal` skip reasoning entirely (https://nan.builders/docs/models#controlling-reasoning, checked 2026-10-01). pi maps reasoning:\"off\" to reasoningEffort:undefined, so without this map `thinking: off` sent no parameter and the model reasoned anyway; the map sends reasoning_effort:\"none\" and actually disables reasoning."
 		],
 		"extras": {
 			"id": "gemma4",
@@ -117,7 +127,14 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"reasoning": true,
 			"reasoning_options": [
 				{
-					"type": "toggle"
+					"type": "effort",
+					"values": [
+						"none",
+						"low",
+						"medium",
+						"high",
+						"max"
+					]
 				}
 			],
 			"tool_call": true,
@@ -142,7 +159,8 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"cost": {
 				"input": 0,
 				"output": 0
-			}
+			},
+			"canonical_model_id": "google/gemma-4-26b-a4b-it"
 		}
 	},
 	{
@@ -162,11 +180,15 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 		"contextWindow": 1000000,
 		"maxTokens": 131072,
 		"reasoningEffortValues": [
+			"minimal",
 			"low",
 			"medium",
 			"high",
 			"max"
 		],
+		"thinkingLevelMap": {
+			"off": "minimal"
+		},
 		"compat": {
 			"supportsDeveloperRole": false,
 			"supportsReasoningEffort": true,
@@ -175,7 +197,8 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"maxTokensField": "max_tokens"
 		},
 		"notes": [
-			"compat matches the maintainer's working ~/.pi/agent/models.json LiteLLM config for api.nan.builders (2026-09-04): supportsDeveloperRole false, supportsReasoningEffort true, maxTokensField max_tokens. NaN's docs example sets only supportsDeveloperRole: true and is not battle-tested. supportsFinishReason true (2026-09-13, issue #2): the LiteLLM gateway intermittently closes SSE streams before emitting finish_reason; with true pi-ai raises 'Stream ended without finish_reason', which matches pi-ai's retryable-provider pattern ('ended without') and is retried automatically, whereas false silently synthesized stop/toolUse and stalled the turn mid-answer. supportsUsageInStreaming true (2026-09-16, issue #7): NaN's published schema is silent about stream_options, but the live gateway honors it — two identical streaming calls per model, differing only in stream_options: { include_usage: true }, returned 0 usage chunks without it and exactly 1 with it (prompt/completion/reasoning/cached token counts) on deepseek-v4-flash, glm5.3-flash, qwen3.6, mimo-v2.5 and gemma4, and a real pi session then recorded token counts where it recorded zeros. pi-ai only sends stream_options when this is not false, and the sanitizer forwards it when the model declares true, so chat models opt in by default and usage is reported (issue #4). A model that does not report streaming usage can still opt out per model with a models.json compat override (supportsUsageInStreaming: false); the sanitizer then strips stream_options and the payload stays strict."
+			"compat matches the maintainer's working ~/.pi/agent/models.json LiteLLM config for api.nan.builders (2026-09-04): supportsDeveloperRole false, supportsReasoningEffort true, maxTokensField max_tokens. NaN's docs example sets only supportsDeveloperRole: true and is not battle-tested. supportsFinishReason true (2026-09-13, issue #2): the LiteLLM gateway intermittently closes SSE streams before emitting finish_reason; with true pi-ai raises 'Stream ended without finish_reason', which matches pi-ai's retryable-provider pattern ('ended without') and is retried automatically, whereas false silently synthesized stop/toolUse and stalled the turn mid-answer. supportsUsageInStreaming true (2026-09-16, issue #7): NaN's published schema is silent about stream_options, but the live gateway honors it — two identical streaming calls per model, differing only in stream_options: { include_usage: true }, returned 0 usage chunks without it and exactly 1 with it (prompt/completion/reasoning/cached token counts) on deepseek-v4-flash, glm5.3-flash, qwen3.6, mimo-v2.5 and gemma4, and a real pi session then recorded token counts where it recorded zeros. pi-ai only sends stream_options when this is not false, and the sanitizer forwards it when the model declares true, so chat models opt in by default and usage is reported (issue #4). A model that does not report streaming usage can still opt out per model with a models.json compat override (supportsUsageInStreaming: false); the sanitizer then strips stream_options and the payload stays strict.",
+			"thinkingLevelMap { off: \"minimal\" } (issue #16, measured 2026-09-27 against live gateway, repro included in the issue): when pi maps reasoning:\"off\" to reasoningEffort:undefined, the thinkingLevelMap.off=\"minimal\" supplies the effort so the gateway receives reasoning_effort:\"minimal\" (38 reasoning tokens + answer). Note: \"none\" does NOT suppress reasoning on glm5.3-flash (13,382 tokens) — only \"minimal\" was measured to be effective. reasoning_effort_values includes \"minimal\" (not documented by NaN docs but accepted and effective by the gateway, issue #16)."
 		],
 		"extras": {
 			"id": "glm5.3-flash",
@@ -184,7 +207,17 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"family": "glm-flash",
 			"attachment": true,
 			"reasoning": true,
-			"reasoning_options": [],
+			"reasoning_options": [
+				{
+					"type": "effort",
+					"values": [
+						"low",
+						"medium",
+						"high",
+						"max"
+					]
+				}
+			],
 			"tool_call": true,
 			"structured_output": true,
 			"temperature": true,
@@ -207,12 +240,13 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"cost": {
 				"input": 0,
 				"output": 0
-			}
+			},
+			"canonical_model_id": "zhipuai/glm-5.3-flash"
 		}
 	},
 	{
-		"id": "mimo-v2.5",
-		"name": "MiMo-V2.5",
+		"id": "mimo-v2.6-flash",
+		"name": "MiMo-V2.6-Flash",
 		"reasoning": true,
 		"input": [
 			"text",
@@ -235,21 +269,21 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"maxTokensField": "max_tokens"
 		},
 		"notes": [
-			"compat matches the maintainer's working ~/.pi/agent/models.json LiteLLM config for api.nan.builders (2026-09-04): supportsDeveloperRole false, supportsReasoningEffort true, maxTokensField max_tokens. NaN's docs example sets only supportsDeveloperRole: true and is not battle-tested. supportsFinishReason true (2026-09-13, issue #2): the LiteLLM gateway intermittently closes SSE streams before emitting finish_reason; with true pi-ai raises 'Stream ended without finish_reason', which matches pi-ai's retryable-provider pattern ('ended without') and is retried automatically, whereas false silently synthesized stop/toolUse and stalled the turn mid-answer. supportsUsageInStreaming true (2026-09-16, issue #7): NaN's published schema is silent about stream_options, but the live gateway honors it — two identical streaming calls per model, differing only in stream_options: { include_usage: true }, returned 0 usage chunks without it and exactly 1 with it (prompt/completion/reasoning/cached token counts) on deepseek-v4-flash, glm5.3-flash, qwen3.6, mimo-v2.5 and gemma4, and a real pi session then recorded token counts where it recorded zeros. pi-ai only sends stream_options when this is not false, and the sanitizer forwards it when the model declares true, so chat models opt in by default and usage is reported (issue #4). A model that does not report streaming usage can still opt out per model with a models.json compat override (supportsUsageInStreaming: false); the sanitizer then strips stream_options and the payload stays strict."
+			"compat matches the maintainer's working ~/.pi/agent/models.json LiteLLM config for api.nan.builders (2026-09-04): supportsDeveloperRole false, supportsReasoningEffort true, maxTokensField max_tokens. NaN's docs example sets only supportsDeveloperRole: true and is not battle-tested. supportsFinishReason true (2026-09-13, issue #2): the LiteLLM gateway intermittently closes SSE streams before emitting finish_reason; with true pi-ai raises 'Stream ended without finish_reason', which matches pi-ai's retryable-provider pattern ('ended without') and is retried automatically, whereas false silently synthesized stop/toolUse and stalled the turn mid-answer. supportsUsageInStreaming true (2026-09-16, issue #7): NaN's published schema is silent about stream_options, but the live gateway honors it — two identical streaming calls per model, differing only in stream_options: { include_usage: true }, returned 0 usage chunks without it and exactly 1 with it (prompt/completion/reasoning/cached token counts) on deepseek-v4-flash, glm5.3-flash, qwen3.6, mimo-v2.5 and gemma4, and a real pi session then recorded token counts where it recorded zeros. pi-ai only sends stream_options when this is not false, and the sanitizer forwards it when the model declares true, so chat models opt in by default and usage is reported (issue #4). A model that does not report streaming usage can still opt out per model with a models.json compat override (supportsUsageInStreaming: false); the sanitizer then strips stream_options and the payload stays strict.",
+			"manual-only: \"mimo-v2.6-flash\" skipped: models.dev provider \"nan\" now lists it, so the manual-only entry is not emitted and the models.dev values are used (omnimodal model (text, image, audio input) served by NaN; limits and modalities from https://nan.builders/docs/models#mimo-v2-6-flash (checked 2026-09-25, quotas re-checked 2026-09-29): 1,048,576 context / 131,072 max output / 1.0B monthly quota per member. Reasoning: accepted, depth not adjustable (docs table, checked 2026-09-29). Tool calling: yes. Streaming: yes. Input modalities (pi-representable): text, image (audio not representable in pi's Model type). Fallback entry: emitted only while models.dev provider nan does not list the model — models.dev started listing it on 2026-09-29, so the generator then keeps the models.dev values and attaches this note to that entry instead of emitting a second one.)"
 		],
 		"extras": {
-			"id": "mimo-v2.5",
-			"name": "MiMo-V2.5",
-			"description": "Open MiMo model for multimodal coding agents and long-context automation",
+			"id": "mimo-v2.6-flash",
+			"name": "MiMo-V2.6-Flash",
+			"description": "MiMo Flash model for multimodal coding agents and long-context automation",
 			"family": "mimo",
 			"attachment": true,
 			"reasoning": true,
 			"reasoning_options": [],
 			"tool_call": true,
 			"temperature": true,
-			"knowledge": "2024-12",
-			"release_date": "2026-04-22",
-			"last_updated": "2026-04-22",
+			"release_date": "2026-09-22",
+			"last_updated": "2026-09-22",
 			"modalities": {
 				"input": [
 					"text",
@@ -268,38 +302,10 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"cost": {
 				"input": 0,
 				"output": 0
-			}
+			},
+			"canonical_model_id": "xiaomi/mimo-v2.6-flash"
 		}
 	},
-	{
-		"id": "mimo-v2.6-flash",
-		"name": "mimo-v2.6-flash",
-		"reasoning": true,
-		"input": [
-			"text",
-			"image"
-		],
-		"cost": {
-			"input": 0,
-			"output": 0,
-			"cacheRead": 0,
-			"cacheWrite": 0
-		},
-		"contextWindow": 1048576,
-		"maxTokens": 131072,
-		"compat": {
-			"supportsDeveloperRole": false,
-			"supportsReasoningEffort": true,
-			"supportsUsageInStreaming": true,
-			"supportsFinishReason": true,
-			"maxTokensField": "max_tokens"
-		},
-		"notes": [
-			"compat matches the maintainer's working ~/.pi/agent/models.json LiteLLM config for api.nan.builders (2026-09-04): supportsDeveloperRole false, supportsReasoningEffort true, maxTokensField max_tokens. NaN's docs example sets only supportsDeveloperRole: true and is not battle-tested. supportsFinishReason true (2026-09-13, issue #2): the LiteLLM gateway intermittently closes SSE streams before emitting finish_reason; with true pi-ai raises 'Stream ended without finish_reason', which matches pi-ai's retryable-provider pattern ('ended without') and is retried automatically, whereas false silently synthesized stop/toolUse and stalled the turn mid-answer. supportsUsageInStreaming true (2026-09-16, issue #7): NaN's published schema is silent about stream_options, but the live gateway honors it — two identical streaming calls per model, differing only in stream_options: { include_usage: true }, returned 0 usage chunks without it and exactly 1 with it (prompt/completion/reasoning/cached token counts) on deepseek-v4-flash, glm5.3-flash, qwen3.6, mimo-v2.5 and gemma4, and a real pi session then recorded token counts where it recorded zeros. pi-ai only sends stream_options when this is not false, and the sanitizer forwards it when the model declares true, so chat models opt in by default and usage is reported (issue #4). A model that does not report streaming usage can still opt out per model with a models.json compat override (supportsUsageInStreaming: false); the sanitizer then strips stream_options and the payload stays strict.",
-			"manual-only: \"mimo-v2.6-flash\" not yet on models.dev provider nan (omnimodal model (text, image, audio input) served by NaN, not yet listed on models.dev provider nan (checked 2026-09-25); included so the live /models refresh does not hand it UNKNOWN_MODEL_LIMITS. Same limits as mimo-v2.5: 1,048,576 / 131,072 / 1.0B monthly quota per member. Reasoning: accepted, depth not adjustable. Tool calling: yes. Streaming: yes. Input modalities (pi-representable): text, image (audio not representable in pi's Model type).)."
-		],
-		"extras": {}
-	},
 	{
 		"id": "qwen3.6",
 		"name": "Qwen3.6 35B-A3B",
@@ -324,6 +330,9 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"high",
 			"max"
 		],
+		"thinkingLevelMap": {
+			"off": "none"
+		},
 		"compat": {
 			"supportsDeveloperRole": false,
 			"supportsReasoningEffort": true,
@@ -332,7 +341,8 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"maxTokensField": "max_tokens"
 		},
 		"notes": [
-			"compat matches the maintainer's working ~/.pi/agent/models.json LiteLLM config for api.nan.builders (2026-09-04): supportsDeveloperRole false, supportsReasoningEffort true, maxTokensField max_tokens. NaN's docs example sets only supportsDeveloperRole: true and is not battle-tested. supportsFinishReason true (2026-09-13, issue #2): the LiteLLM gateway intermittently closes SSE streams before emitting finish_reason; with true pi-ai raises 'Stream ended without finish_reason', which matches pi-ai's retryable-provider pattern ('ended without') and is retried automatically, whereas false silently synthesized stop/toolUse and stalled the turn mid-answer. supportsUsageInStreaming true (2026-09-16, issue #7): NaN's published schema is silent about stream_options, but the live gateway honors it — two identical streaming calls per model, differing only in stream_options: { include_usage: true }, returned 0 usage chunks without it and exactly 1 with it (prompt/completion/reasoning/cached token counts) on deepseek-v4-flash, glm5.3-flash, qwen3.6, mimo-v2.5 and gemma4, and a real pi session then recorded token counts where it recorded zeros. pi-ai only sends stream_options when this is not false, and the sanitizer forwards it when the model declares true, so chat models opt in by default and usage is reported (issue #4). A model that does not report streaming usage can still opt out per model with a models.json compat override (supportsUsageInStreaming: false); the sanitizer then strips stream_options and the payload stays strict."
+			"compat matches the maintainer's working ~/.pi/agent/models.json LiteLLM config for api.nan.builders (2026-09-04): supportsDeveloperRole false, supportsReasoningEffort true, maxTokensField max_tokens. NaN's docs example sets only supportsDeveloperRole: true and is not battle-tested. supportsFinishReason true (2026-09-13, issue #2): the LiteLLM gateway intermittently closes SSE streams before emitting finish_reason; with true pi-ai raises 'Stream ended without finish_reason', which matches pi-ai's retryable-provider pattern ('ended without') and is retried automatically, whereas false silently synthesized stop/toolUse and stalled the turn mid-answer. supportsUsageInStreaming true (2026-09-16, issue #7): NaN's published schema is silent about stream_options, but the live gateway honors it — two identical streaming calls per model, differing only in stream_options: { include_usage: true }, returned 0 usage chunks without it and exactly 1 with it (prompt/completion/reasoning/cached token counts) on deepseek-v4-flash, glm5.3-flash, qwen3.6, mimo-v2.5 and gemma4, and a real pi session then recorded token counts where it recorded zeros. pi-ai only sends stream_options when this is not false, and the sanitizer forwards it when the model declares true, so chat models opt in by default and usage is reported (issue #4). A model that does not report streaming usage can still opt out per model with a models.json compat override (supportsUsageInStreaming: false); the sanitizer then strips stream_options and the payload stays strict.",
+			"thinkingLevelMap { off: \"none\" } (issue #16): NaN docs state that with no reasoning_effort parameter every model uses its own default and that default is reasoning ON for qwen3.6 (16,384-token budget), while `none` and `minimal` skip the reasoning phase entirely (https://nan.builders/docs/models#controlling-reasoning, checked 2026-10-01). pi maps reasoning:\"off\" to reasoningEffort:undefined, so without this map `thinking: off` sent no parameter and the model reasoned anyway; the map sends reasoning_effort:\"none\" and actually disables reasoning."
 		],
 		"extras": {
 			"id": "qwen3.6",
@@ -343,7 +353,14 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"reasoning": true,
 			"reasoning_options": [
 				{
-					"type": "toggle"
+					"type": "effort",
+					"values": [
+						"none",
+						"low",
+						"medium",
+						"high",
+						"max"
+					]
 				}
 			],
 			"tool_call": true,
@@ -368,7 +385,8 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"cost": {
 				"input": 0,
 				"output": 0
-			}
+			},
+			"canonical_model_id": "alibaba/qwen3.6-35b-a3b"
 		}
 	},
 	{
@@ -428,7 +446,8 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			"cost": {
 				"input": 0,
 				"output": 0
-			}
+			},
+			"canonical_model_id": "alibaba/qwen3.8-flash"
 		}
 	}
 ];
@@ -436,15 +455,20 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 export const GENERATED_CATALOG_META = {
 	source: "https://models.dev/api.json",
 	modelsDevProvider: "nan",
-	fetchedAt: "2026-09-27T21:25:24.946Z",
-	modelCount: 7,
-	models: ["deepseek-v4-flash","gemma4","glm5.3-flash","mimo-v2.5","mimo-v2.6-flash","qwen3.6","qwen3.8-flash"],
+	fetchedAt: "2026-10-01T10:18:48.654Z",
+	modelCount: 6,
+	models: ["deepseek-v4-flash","gemma4","glm5.3-flash","mimo-v2.6-flash","qwen3.6","qwen3.8-flash"],
 	notes: [
+	"non-chat: \"qwen-image-2.1\" excluded from the chat catalog (image-generation model (text→image; output modality image) documented in https://nan.builders/docs/models and https://nan.builders/openapi.json (checked 2026-09-29) — not a chat model, out of the static chat catalog (models.dev reports limit.output 0))",
 	"live-only: \"glm5.3\" kept out of the static catalog (premium-tier model (models.dev now documents it with 1M context / 131,072 max output; NaN docs https://nan.builders/docs/models + https://nan.builders/openapi.json, checked 2026-09-13) kept live-only so a non-premium key never sees a model it cannot call when the live /models fetch is unavailable; premium keys still get it via the /models refresh with conservative placeholder limits)",
 	"provider-removed: \"glm5.2\" excluded from the catalog (removed by NaN (2026-09-05); absent from the official chat model list in https://nan.builders/openapi.json and https://nan.builders/docs/models (checked 2026-09-07) while models.dev provider nan still listed it — excluded so regeneration does not resurrect it)",
+	"provider-removed: \"mimo-v2.5\" excluded from the catalog (removed by NaN (checked 2026-09-29): zero mentions in https://nan.builders/docs/models and in the model param description of https://nan.builders/openapi.json, and models.dev provider nan dropped it as well — superseded by mimo-v2.6-flash; excluded so regeneration does not resurrect it)",
 	"compat matches the maintainer's working ~/.pi/agent/models.json LiteLLM config for api.nan.builders (2026-09-04): supportsDeveloperRole false, supportsReasoningEffort true, maxTokensField max_tokens. NaN's docs example sets only supportsDeveloperRole: true and is not battle-tested. supportsFinishReason true (2026-09-13, issue #2): the LiteLLM gateway intermittently closes SSE streams before emitting finish_reason; with true pi-ai raises 'Stream ended without finish_reason', which matches pi-ai's retryable-provider pattern ('ended without') and is retried automatically, whereas false silently synthesized stop/toolUse and stalled the turn mid-answer. supportsUsageInStreaming true (2026-09-16, issue #7): NaN's published schema is silent about stream_options, but the live gateway honors it — two identical streaming calls per model, differing only in stream_options: { include_usage: true }, returned 0 usage chunks without it and exactly 1 with it (prompt/completion/reasoning/cached token counts) on deepseek-v4-flash, glm5.3-flash, qwen3.6, mimo-v2.5 and gemma4, and a real pi session then recorded token counts where it recorded zeros. pi-ai only sends stream_options when this is not false, and the sanitizer forwards it when the model declares true, so chat models opt in by default and usage is reported (issue #4). A model that does not report streaming usage can still opt out per model with a models.json compat override (supportsUsageInStreaming: false); the sanitizer then strips stream_options and the payload stays strict.",
-	"input includes image: NaN serves the Vision-Exp variant ('takes images as input', https://nan.builders/docs/models; the image_url content-parts in https://nan.builders/openapi.json list deepseek-v4-flash among the vision models). models.dev provider nan also lists text+image now (DeepSeek V4.1 Flash entry, checked 2026-09-13; its 2026-09-07 snapshot listed text only), so this override is kept as a pin for the vision capability rather than as a divergence. reasoning_effort_values=[]: the model decides per-request how much to reason (https://nan.builders/docs/models, checked 2026-09-25).",
-	"manual-only: \"mimo-v2.6-flash\" not yet on models.dev provider nan (omnimodal model (text, image, audio input) served by NaN, not yet listed on models.dev provider nan (checked 2026-09-25); included so the live /models refresh does not hand it UNKNOWN_MODEL_LIMITS. Same limits as mimo-v2.5: 1,048,576 / 131,072 / 1.0B monthly quota per member. Reasoning: accepted, depth not adjustable. Tool calling: yes. Streaming: yes. Input modalities (pi-representable): text, image (audio not representable in pi's Model type).).",
+	"input includes image: NaN serves the Vision-Exp variant ('takes images as input', https://nan.builders/docs/models; the image_url content-parts in https://nan.builders/openapi.json list deepseek-v4-flash among the vision models). models.dev provider nan also lists text+image now (DeepSeek V4.1 Flash entry, checked 2026-09-13; its 2026-09-07 snapshot listed text only), so this override is kept as a pin for the vision capability rather than as a divergence. reasoning_effort_values=[\"none\"] (issue #16, measured 2026-09-27 against live gateway, repro included in the issue): only \"none\" deterministically disables reasoning (0 reasoning tokens, answer produced). Every other value leaves 13-14K reasoning tokens and causes runaway reasoning (finish_reason:\"length\" with ZERO answer text). The catalog declares only [\"none\"] because the remaining values are useless noise. thinkingLevelMap { off: \"none\" } (issue #16, measured 2026-09-27): when pi maps reasoning:\"off\" to reasoningEffort:undefined, the thinkingLevelMap.off=\"none\" supplies the effort so the gateway receives reasoning_effort:\"none\" and does NOT enter runaway reasoning mode.",
+	"thinkingLevelMap { off: \"none\" } (issue #16): NaN docs state that with no reasoning_effort parameter gemma4 defaults to reasoning ON (16,384-token budget), while `none`/`minimal` skip reasoning entirely (https://nan.builders/docs/models#controlling-reasoning, checked 2026-10-01). pi maps reasoning:\"off\" to reasoningEffort:undefined, so without this map `thinking: off` sent no parameter and the model reasoned anyway; the map sends reasoning_effort:\"none\" and actually disables reasoning.",
+	"thinkingLevelMap { off: \"minimal\" } (issue #16, measured 2026-09-27 against live gateway, repro included in the issue): when pi maps reasoning:\"off\" to reasoningEffort:undefined, the thinkingLevelMap.off=\"minimal\" supplies the effort so the gateway receives reasoning_effort:\"minimal\" (38 reasoning tokens + answer). Note: \"none\" does NOT suppress reasoning on glm5.3-flash (13,382 tokens) — only \"minimal\" was measured to be effective. reasoning_effort_values includes \"minimal\" (not documented by NaN docs but accepted and effective by the gateway, issue #16).",
+	"manual-only: \"mimo-v2.6-flash\" skipped: models.dev provider \"nan\" now lists it, so the manual-only entry is not emitted and the models.dev values are used (omnimodal model (text, image, audio input) served by NaN; limits and modalities from https://nan.builders/docs/models#mimo-v2-6-flash (checked 2026-09-25, quotas re-checked 2026-09-29): 1,048,576 context / 131,072 max output / 1.0B monthly quota per member. Reasoning: accepted, depth not adjustable (docs table, checked 2026-09-29). Tool calling: yes. Streaming: yes. Input modalities (pi-representable): text, image (audio not representable in pi's Model type). Fallback entry: emitted only while models.dev provider nan does not list the model — models.dev started listing it on 2026-09-29, so the generator then keeps the models.dev values and attaches this note to that entry instead of emitting a second one.)",
+	"thinkingLevelMap { off: \"none\" } (issue #16): NaN docs state that with no reasoning_effort parameter every model uses its own default and that default is reasoning ON for qwen3.6 (16,384-token budget), while `none` and `minimal` skip the reasoning phase entirely (https://nan.builders/docs/models#controlling-reasoning, checked 2026-10-01). pi maps reasoning:\"off\" to reasoningEffort:undefined, so without this map `thinking: off` sent no parameter and the model reasoned anyway; the map sends reasoning_effort:\"none\" and actually disables reasoning.",
 	"contextWindow 262,144: the earlier 1,000,000 override (maintainer-confirmed 2026-09-05) was withdrawn 2026-09-07 — the updated https://nan.builders/docs/models still states '262K token context, the model's native window' and models.dev agrees at 262,144; NaN docs are treated as the most reliable source (maintainer instruction, 2026-09-07)."
 ],
 } as const;