@gtrabanco/pi-nan-provider 0.6.10 → 0.7.0

package/AGENTS.md

@@ -93,10 +93,13 @@ Every PR that changes code MUST bump `package.json` version in the same PR; CI p
   `~/.pi/agent/npm/node_modules`; its `estimateMessageTokens` lacks the `system`
   branch and crashes pi 0.87's string-content `system` transcript with
   `block.name.length`. `src/pi-ai-loader.ts` therefore binds the streaming
-  factory to the same package instance as the bare-root import: use the root
-  export when present, else derive a FILE URL from
-  `import.meta.resolve("@earendil-works/pi-ai")`
-  (`api/openai-completions.lazy.js`, then `compat.js`). No bare pi-ai subpath
+  factory to the same package instance the host loaded: use the root export
+  when present, else resolve the bare root **from the host process entrypoint**
+  (`process.argv[1]`) via
+  `import.meta.resolve("@earendil-works/pi-ai", hostAnchor)` — an
+  extension-relative resolve returns the extension tree's stale copy (the
+  v0.6.10 regression that left #8 open) — then derive a FILE URL for
+  `api/openai-completions.lazy.js` (then `compat.js`). No bare pi-ai subpath
   specifier is imported anywhere in `src/` (static or dynamic); failure is loud
   (`PiAiStreamingApiResolutionError`). Guarded by `test/extension-load.test.ts`
   and `test/issue-8-pi-ai-instance.test.ts`.

package/README.md

@@ -236,9 +236,25 @@ Baseline catalog (verified against [NaN docs](https://nan.builders/docs/models)
 | `gemma4` | 262,144 | 32,768 | text, image | ✅ |
 | `deepseek-v4-flash` | 1,000,000 | 384,000 | text, image | ✅ |
 | `mimo-v2.5` | 1,048,576 | 131,072 | text, image | ✅ |
+| `mimo-v2.6-flash` | 1,048,576 | 131,072 | text, image | ✅ |
 | `glm5.3-flash` | 1,000,000 | 131,072 | text, image | ✅ |
 | `qwen3.8-flash` | 262,144 | 131,072 | text, image | ✅ |
 
+> [!NOTE]  
+> `mimo-v2.6-flash` is served by NaN but not yet listed on models.dev provider `nan`; it enters the catalog through a manual-only entry with the same limits as `mimo-v2.5`. The model will be auto-detected from models.dev once added there.
+
+---
+
+## 🧠 Reasoning controls
+
+NaN's `reasoning_effort` parameter controls how much the model thinks before answering — but the degree of control varies by model:
+
+| Model | Reasoning effort | How it works |
+| :--- | :--- | :--- |
+| `glm5.3`, `glm5.3-flash` | `low` · `medium` · `high` · `max` | Fully controllable — higher values let the model reason longer |
+| `qwen3.6`, `gemma4` | `none` · `minimal` · `low` · `medium` · `high` · `max` | `none`/`minimal` skip reasoning entirely; others cap at 2K / 8K / 16K / 32K tokens |
+| `deepseek-v4-flash`, `qwen3.8-flash`, `mimo-v2.5`, `mimo-v2.6-flash` | *(accepted but not adjustable)* | The parameter is accepted and never rejected, but the model manages its own reasoning depth — it is never an error to send a value these models don't adjust |
+
 ---
 
 ## 🚀 Development

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@gtrabanco/pi-nan-provider",
-	"version": "0.6.10",
+	"version": "0.7.0",
 	"description": "NaN Builders (api.nan.builders) model provider for pi - OpenAI-compatible registration with a models.dev-generated fallback, tier-aware live catalog, and MCP bridges (official web search + optional community media server)",
 	"keywords": [
 		"pi",
@@ -38,7 +38,8 @@
 		"check-nan-mcp-server": "bun run scripts/check-nan-mcp-server.ts",
 		"prepublishOnly": "bun run generate-models && bun test && bun run typecheck",
 		"test": "bun test",
-		"typecheck": "bunx tsc --noEmit"
+		"typecheck": "bunx tsc --noEmit",
+		"check-pi-sdk-versions": "node scripts/check-pi-sdk-versions.mjs --report"
 	},
 	"peerDependencies": {
 		"@earendil-works/pi-ai": ">=0.83.0 <1",

package/scripts/generate-models.ts

@@ -21,7 +21,28 @@
  */
 
 import type { GeneratedModelEntry } from "../src/fetch-models.ts";
-import { MANUAL_OVERRIDES } from "./manual-overrides.ts";
+import type { ManualModelOverride } from "./manual-overrides.ts";
+import {
+	MANUAL_OVERRIDES,
+	MANUAL_ONLY_MODEL_IDS,
+	REASONING_EFFORT_VALUES,
+} from "./manual-overrides.ts";
+
+/**
+ * Reasoning effort values sourced from the NaN docs (https://nan.builders/docs/models
+ * #controlling-reasoning, checked 2026-09-25). models.dev has no reasoning_effort_values
+ * field — only reasoning_options which is [{type:"toggle"}] or [] — so the actual
+ * effort levels must be hand-maintained from the docs. An empty array means the
+ * parameter is accepted but depth is not adjustable by the user.
+ */
+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-v2.5": [],
+};
 
 const MODELS_DEV_API_URL = "https://models.dev/api.json";
 const SOURCE_PROVIDER_ID = "nan";
@@ -158,6 +179,43 @@ function normalizeInput(modalitiesInput: string[] | undefined, modelId: string):
 	return input;
 }
 
+/**
+ * Build a GeneratedModelEntry for a model that is not yet on models.dev
+ * (MANUAL_ONLY_MODEL_IDS). Uses conservative but accurate defaults from the
+ * NaN docs so that the live /models refresh does not hand them
+ * UNKNOWN_MODEL_LIMITS.
+ */
+function buildManualOnlyModelEntry(
+	modelId: string,
+	detail: string,
+	override: ManualModelOverride | undefined,
+): GeneratedModel {
+	const reasoningEffortValues =
+		override?.reasoningEffortValues ??
+		REASONING_EFFORT_VALUES_FROM_DOCS[modelId] ??
+		undefined;
+
+	return {
+		entry: {
+			id: modelId,
+			name: override?.name ?? modelId,
+			reasoning: override?.reasoning ?? true,
+			input: override?.input ?? ["text", "image"],
+			cost: override?.cost ?? { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+			contextWindow: override?.contextWindow ?? 1_048_576,
+			maxTokens: override?.maxTokens ?? 131_072,
+			reasoningEffortValues,
+			compat: { ...NAN_COMPAT },
+			notes: [
+				NAN_COMPAT_NOTE,
+				`manual-only: "${modelId}" not yet on models.dev provider nan (${detail}).`,
+				...(override ? [override.note] : []),
+			],
+			extras: {},
+		},
+	};
+}
+
 function convertModel(modelId: string, m: ModelsDevModel): GeneratedModel | { skip: string } {
 	const removedReason = PROVIDER_REMOVED_MODEL_IDS[modelId];
 	if (removedReason) {
@@ -184,6 +242,13 @@ function convertModel(modelId: string, m: ModelsDevModel): GeneratedModel | { sk
 		console.log(`generate-models: manual override for "${modelId}" (${overridden.join(", ")})`);
 	}
 
+	// Reasoning effort values sourced from the NaN docs (models.dev does not
+	// expose them — only reasoning_options: [{type:"toggle"}] or []).
+	// An empty array means the parameter is accepted but depth is not
+	// adjustable by the user; the model manages its own reasoning depth.
+	const reasoningEffortValues =
+		override?.reasoningEffortValues ?? REASONING_EFFORT_VALUES_FROM_DOCS[modelId] ?? undefined;
+
 	return {
 		entry: {
 			id: modelId,
@@ -198,6 +263,7 @@ function convertModel(modelId: string, m: ModelsDevModel): GeneratedModel | { sk
 			},
 			contextWindow: override?.contextWindow ?? contextWindow,
 			maxTokens: override?.maxTokens ?? maxTokens,
+			reasoningEffortValues,
 			compat: { ...NAN_COMPAT },
 			notes: [
 				NAN_COMPAT_NOTE,
@@ -235,6 +301,17 @@ async function main(): Promise<void> {
 		}
 	}
 
+	// Add manual-only models (not yet on models.dev).
+	for (const [modelId, detail] of Object.entries(MANUAL_ONLY_MODEL_IDS)) {
+		const override = MANUAL_OVERRIDES[modelId];
+		if (override) {
+			console.log(`generate-models: manual override for "${modelId}" (${Object.keys(override).filter((k) => k !== "note").join(", ")})`);
+		}
+		const result = buildManualOnlyModelEntry(modelId, detail, MANUAL_OVERRIDES[modelId]);
+		entries.push(result.entry);
+		console.log(`generate-models: added manual-only model "${modelId}"`);
+	}
+
 	entries.sort((a, b) => a.id.localeCompare(b.id));
 
 	const entryIds = new Set(entries.map((entry) => entry.id));

package/scripts/manual-overrides.ts

@@ -35,13 +35,69 @@ export interface ManualModelOverride {
 	contextWindow?: number;
 	/** Max output tokens override. */
 	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[];
 	/** 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).
+ *
+ * models.dev has NO `reasoning_effort_values` field — only `reasoning_options`
+ * (which is [{type:"toggle"}] or []), so this mapping must be hand-maintained
+ * from the docs. The values flow through MANUAL_OVERRIDES.reasoningEffortValues
+ * and appear in GeneratedModelEntry.reasoningEffortValues, where the pi
+ * model-selector can read them to show the correct granularity.
+ *
+ * Per-model contract from the docs:
+ * - glm5.3, glm5.3-flash: fully controllable (low/medium/high/max)
+ * - 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
+ *
+ * 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"],
+	"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.
+ */
+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).",
+};
+
+/**
+ * Manual-only model entries with the full override shape for test assertions.
+ * Mirrors MANUAL_ONLY_MODEL_IDS so the generator and the tests share the
+ * same set of models.
+ */
+export const MANUAL_ONLY_MODELS: Record<string, ManualModelOverride> = {
+	"mimo-v2.6-flash": {
+		input: ["text", "image"],
+		reasoning: true,
+		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).",
+	},
+};
+
 export const MANUAL_OVERRIDES: Record<string, ManualModelOverride> = {
 	"deepseek-v4-flash": {
 		input: ["text", "image"],
-		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.",
+		// 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).",
 	},
 };

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 2026-09-21T23:54:45.052Z
+// Source: https://models.dev/api.json (provider "nan"), fetched 2026-09-25T22:43:00.930Z
 // 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.
@@ -28,6 +28,7 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 		},
 		"contextWindow": 1000000,
 		"maxTokens": 384000,
+		"reasoningEffortValues": [],
 		"compat": {
 			"supportsDeveloperRole": false,
 			"supportsReasoningEffort": true,
@@ -37,7 +38,7 @@ 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."
+			"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)."
 		],
 		"extras": {
 			"id": "deepseek-v4-flash",
@@ -89,6 +90,14 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 		},
 		"contextWindow": 262144,
 		"maxTokens": 32768,
+		"reasoningEffortValues": [
+			"none",
+			"minimal",
+			"low",
+			"medium",
+			"high",
+			"max"
+		],
 		"compat": {
 			"supportsDeveloperRole": false,
 			"supportsReasoningEffort": true,
@@ -152,6 +161,12 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 		},
 		"contextWindow": 1000000,
 		"maxTokens": 131072,
+		"reasoningEffortValues": [
+			"low",
+			"medium",
+			"high",
+			"max"
+		],
 		"compat": {
 			"supportsDeveloperRole": false,
 			"supportsReasoningEffort": true,
@@ -211,6 +226,7 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 		},
 		"contextWindow": 1048576,
 		"maxTokens": 131072,
+		"reasoningEffortValues": [],
 		"compat": {
 			"supportsDeveloperRole": false,
 			"supportsReasoningEffort": true,
@@ -255,6 +271,35 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 			}
 		}
 	},
+	{
+		"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",
@@ -271,6 +316,14 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 		},
 		"contextWindow": 262144,
 		"maxTokens": 65536,
+		"reasoningEffortValues": [
+			"none",
+			"minimal",
+			"low",
+			"medium",
+			"high",
+			"max"
+		],
 		"compat": {
 			"supportsDeveloperRole": false,
 			"supportsReasoningEffort": true,
@@ -334,6 +387,7 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 		},
 		"contextWindow": 262144,
 		"maxTokens": 131072,
+		"reasoningEffortValues": [],
 		"compat": {
 			"supportsDeveloperRole": false,
 			"supportsReasoningEffort": true,
@@ -382,14 +436,15 @@ export const NAN_GENERATED_MODELS: readonly GeneratedModelEntry[] = [
 export const GENERATED_CATALOG_META = {
 	source: "https://models.dev/api.json",
 	modelsDevProvider: "nan",
-	fetchedAt: "2026-09-21T23:54:45.052Z",
-	modelCount: 6,
-	models: ["deepseek-v4-flash","gemma4","glm5.3-flash","mimo-v2.5","qwen3.6","qwen3.8-flash"],
+	fetchedAt: "2026-09-25T22:43:00.930Z",
+	modelCount: 7,
+	models: ["deepseek-v4-flash","gemma4","glm5.3-flash","mimo-v2.5","mimo-v2.6-flash","qwen3.6","qwen3.8-flash"],
 	notes: [
 	"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)",
 	"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.",
+	"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).).",
 	"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;

package/src/fetch-models.ts

@@ -39,6 +39,13 @@ export interface GeneratedModelEntry {
 	cost: { input: number; output: number; cacheRead: number; cacheWrite: number };
 	contextWindow: number;
 	maxTokens: number;
+	/**
+	 * Reasoning effort values as declared by NaN docs.
+	 * An empty array means the parameter is accepted but depth is not
+	 * adjustable by the user — the model manages its own reasoning depth.
+	 * An undefined array means the field was not set (legacy / uncatalogued).
+	 */
+	reasoningEffortValues?: string[];
 	/** Compat applied to every NaN-compatible model (LiteLLM-confirmed, see scripts/generate-models.ts). */
 	compat?: OpenAICompletionsCompat;
 	/** Provenance notes for values overriding models.dev or needing manual confirmation. */

package/src/pi-ai-loader.ts

@@ -18,13 +18,15 @@
  * `undefined is not an object (evaluating 'block.name.length')` — before the
  * request is ever sent, so it reads like a NaN/gateway failure.
  *
- * `import.meta.resolve("@earendil-works/pi-ai")` returns the *same instance*
- * the bare-root static import used in every environment measured (both the
- * host's 0.87 core on pi-web, and the extension-tree copy under plain
- * node/jiti). It is therefore the anchor: derive a **file URL** for the
- * sibling `api/openai-completions.lazy.js` (then `compat.js`) from that
- * root and dynamic-import the URL. A file URL bypasses package resolution
- * entirely, so the loaded module is guaranteed to be the host's instance.
+ * The anchor is the **host process entrypoint** (`process.argv[1]`), passed to
+ * `import.meta.resolve(specifier, parent)`. An extension-relative resolve is
+ * NOT enough: it returns whatever the extension's own tree holds, which under
+ * pi-web is exactly the stale 0.85.1 copy (that was the v0.6.10 regression:
+ * the "host-resolved root" was still the extension's copy). From the
+ * host-resolved root derive a **file URL** for the sibling
+ * `api/openai-completions.lazy.js` (then `compat.js`) and dynamic-import the
+ * URL. A file URL bypasses package resolution entirely, so the loaded module is
+ * guaranteed to be the host's instance.
  *
  * Under the bundled CLI / Node-mode aliases / compiled binary, the bare root
  * is the compat entrypoint and already exposes the factory, so the first
@@ -42,6 +44,7 @@
 import * as piAi from "@earendil-works/pi-ai";
 import type { ProviderStreams } from "@earendil-works/pi-ai";
 import { createRequire } from "node:module";
+import { pathToFileURL } from "node:url";
 
 /** The only pi-ai specifier this package may import. */
 export const PI_AI_PACKAGE_SPECIFIER = "@earendil-works/pi-ai";
@@ -106,22 +109,83 @@ export function openAICompletionsApiFrom(namespace: unknown): OpenAICompletionsA
  * `import.meta.resolve` is absent from bun-types' `ImportMeta`, so read it
  * through an explicit shape. Bun/Node expose it at runtime; when it is missing
  * or throws, `createRequire` resolves the same bare root from this module.
+ *
+ * Referenced directly (not via a variable or type cast) so pi's jiti loader
+ * can rewrite it. PR #11 (jiti compatibility) and PR #12 (host-anchored
+ * resolution) merge: the direct reference fixes jiti, the host anchor fixes
+ * pi-web.
  */
-type ImportMetaWithResolve = ImportMeta & { resolve?: (specifier: string) => string };
+function readImportMetaResolve(): ((specifier: string, parent?: string) => string) | undefined {
+	if (typeof import.meta.resolve !== "function") return undefined;
+	return (specifier, parent) =>
+		parent === undefined ? import.meta.resolve(specifier) : import.meta.resolve(specifier, parent);
+}
 
-const defaultPiAiLoaderHost: PiAiLoaderHost = {
-	namespace: piAi as unknown as ModuleNamespace,
-	resolveSpecifier(specifier: string): string {
-		const meta = import.meta as ImportMetaWithResolve;
-		if (typeof meta.resolve === "function") {
+/**
+ * File URL of the host process entrypoint (pi-web's `sessiond.js`, the pi CLI,
+ * a test runner). Resolving FROM it pins the result to the host's module graph
+ * instead of the extension's own tree.
+ */
+export function hostAnchorUrl(entry: string | undefined): string | undefined {
+	if (typeof entry !== "string" || entry.length === 0) return undefined;
+	try {
+		return pathToFileURL(entry).href;
+	} catch {
+		return undefined;
+	}
+}
+
+/** Injection seam for {@link resolvePiAiSpecifier}. */
+export interface PiAiSpecifierResolution {
+	/** Anchor URL to resolve FROM; omitted means extension-relative resolution. */
+	anchorUrl?: string;
+	/** `import.meta.resolve`, when the runtime exposes it. */
+	resolve?: (specifier: string, parent?: string) => string;
+	/** Last-resort CJS resolver; only used when neither `resolve` call succeeds. */
+	fallback?: (specifier: string) => string;
+}
+
+/**
+ * Resolve a pi-ai specifier the way the host runtime does.
+ *
+ * The `anchorUrl` is the whole point. Under pi-web the extension's own tree
+ * holds a stale hoisted `@earendil-works/pi-ai` (0.85.1), so an
+ * extension-relative resolve returns a package the host never loaded, and the
+ * derived "host-resolved root" is still that stale copy. Passing the host
+ * entrypoint as the resolver's parent returns the instance the host itself
+ * uses (the v0.6.10 fix anchored on `import.meta.resolve` alone and therefore
+ * still selected the stale copy).
+ */
+export function resolvePiAiSpecifier(
+	specifier: string,
+	options: PiAiSpecifierResolution = {},
+): string {
+	const resolve = options.resolve ?? readImportMetaResolve();
+	const fallback =
+		options.fallback ?? ((value: string) => createRequire(import.meta.url).resolve(value));
+	if (resolve !== undefined) {
+		if (options.anchorUrl !== undefined) {
 			try {
-				const resolved = meta.resolve(specifier);
-				if (typeof resolved === "string" && resolved.length > 0) return resolved;
+				const anchored = resolve(specifier, options.anchorUrl);
+				if (typeof anchored === "string" && anchored.length > 0) return anchored;
 			} catch {
-				// Fall through to createRequire — same bare root, same instance.
+				// Runtimes without parent support fall through to the bare call.
 			}
 		}
-		return createRequire(import.meta.url).resolve(specifier);
+		try {
+			const resolved = resolve(specifier);
+			if (typeof resolved === "string" && resolved.length > 0) return resolved;
+		} catch {
+			// Fall through to createRequire — same bare root, same instance.
+		}
+	}
+	return fallback(specifier);
+}
+
+const defaultPiAiLoaderHost: PiAiLoaderHost = {
+	namespace: piAi as unknown as ModuleNamespace,
+	resolveSpecifier(specifier: string): string {
+		return resolvePiAiSpecifier(specifier, { anchorUrl: hostAnchorUrl(process.argv[1]) });
 	},
 	importModule: (url: string) => import(url) as Promise<ModuleNamespace>,
 };

package/src/usage.ts

@@ -37,6 +37,7 @@ export interface ModelQuota {
 export const MODEL_QUOTAS: readonly ModelQuota[] = [
 	{ model: "deepseek-v4-flash", label: "DeepSeek V4 Flash", monthlyCap: 3_000_000_000, rollingWindowCap: 0, rollingWindowHours: 0, premium: false },
 	{ model: "mimo-v2.5", label: "MiMo V2.5", monthlyCap: 1_000_000_000, rollingWindowCap: 0, rollingWindowHours: 0, premium: false },
+	{ model: "mimo-v2.6-flash", label: "MiMo V2.6 Flash", monthlyCap: 1_000_000_000, rollingWindowCap: 0, rollingWindowHours: 0, premium: false },
 	{ model: "qwen3.6", label: "Qwen 3.6", monthlyCap: 0, rollingWindowCap: 0, rollingWindowHours: 0, premium: false },
 	{ model: "gemma4", label: "Gemma 4", monthlyCap: 0, rollingWindowCap: 0, rollingWindowHours: 0, premium: false },
 	{ model: "qwen3.8-flash", label: "Qwen 3.8 Flash", monthlyCap: 500_000_000, rollingWindowCap: 0, rollingWindowHours: 0, premium: false },