@gtrabanco/pi-nan-provider 0.8.0 → 0.10.2

package/AGENTS.md

@@ -51,7 +51,7 @@ maintainer's quota, so they are **permission-gated**:
   gate covers ad-hoc diagnosis only, never `bun test`.
 - **Default to `qwen3.6` — it is unlimited.**
 - **For massive/bulk probes prefer a model the maintainer uses less with a large
-  token budget, e.g. `mimo-v2.5`** (1M context).
+  token budget, e.g. `mimo-v2.6-flash`** (1M context, 1.0B monthly quota).
 - **When the model under investigation is the point** (e.g. reproducing a
   model-specific 400), use it, but minimize tokens: smallest viable prompt,
   lowest `max_tokens`, stop at the first decisive response.
@@ -77,7 +77,7 @@ Every PR that changes code MUST bump `package.json` version in the same PR; CI p
 
 ## Verified API facts (do not re-derive from stale docs)
 
-- **Extension-side pi-ai imports + streaming-API instance binding (v0.6.10; verified on pi-ai 0.83.0–0.84.4 and pi 0.87.0):**
+- **Extension-side pi-ai imports + streaming-API instance binding (v0.6.10; verified on pi-ai 0.83.0–0.87.1 and pi 0.87.1):**
   statically import ONLY the bare `@earendil-works/pi-ai` root from `src/`. pi's
   extension loader maps that specifier to the compat entrypoint on the bundled
   CLI, Node-mode jiti aliases and compiled-binary virtualModules, and compat
@@ -107,7 +107,7 @@ Every PR that changes code MUST bump `package.json` version in the same PR; CI p
   `openAICompletionsApi`; `createProvider` and `envApiKeyAuth(name, envVars)` are
   on the root. `envApiKeyAuth` implements exactly: stored credential key wins →
   first set env var → unconfigured; `login()` prompts with `{ type: "secret" }`.
-- pi awaits extension factories (`await factory(api)`) on 0.83.0 and 0.84.4
+- pi awaits extension factories (`await factory(api)`) on 0.83.0 through 0.87.1
   alike, so the extension entrypoint may be async (v0.5.0: streaming-API
   resolution needs it).
 - pi-ai 0.83.0 runtime surface verified identical for this package's needs:
@@ -116,10 +116,48 @@ Every PR that changes code MUST bump `package.json` version in the same PR; CI p
   `fetchModels(context)`, `filterModels(models, credential)`, `api`) and
   `RefreshModelsContext.credential` match 0.84.4; `registerProvider` has both
   the full-`Provider` and `(name, config)` overloads in 0.83's ExtensionAPI.
+  (pi-ai 0.86 changed provider stream entry points to a branded `TranscriptContext`
+  — see next bullet — but this package's runtime is untouched.)
+- **pi-ai 0.86.0 `TranscriptContext` migration (verified 2026-09-29):** provider
+  stream entry points (`Provider.stream`, `ProviderStreams.stream/streamSimple`,
+  `StreamFunction`) now take a branded `TranscriptContext = { messages: Message[] }`
+  instead of the public `Context = { systemPrompt?: string; messages; tools? }`.
+  `normalizeContext(context: Context): TranscriptContext` (exported from the pi-ai
+  root) folds prompt + tools into a leading system message
+  `{ role: "system", content: <prompt>, toolsAdded: <tools>, timestamp: 0 }`.
+  `Context` itself still exists and is what the public entry points accept, so raw
+  `Context` literals only fail where a *provider-level* stream is called directly
+  (the 6 test call sites). Tests now build contexts with `normalizeContext(...)`.
+  This package's `src/` files never hand-build a provider context, so no `src/`
+  code changed: `wrapApiForStrictSanitization` forwards `context` untouched, and
+  `withContextOverflowClassification` only *reads* it for the estimate, which still
+  counts prompt + tools because they ride inside `messages[0]`.
+- pi 0.87 rejects extension tools that declare no parameter schema at registration;
+  all this package's MCP tools declare TypeBox `parameters`, so they register fine.
+  pi 0.87 also changed `pi.on()` to return an unsubscribe function (unused by
+  this package).
 - `pi.registerProvider(provider)` accepts a complete pi-ai `Provider`; pi's Models
   runtime then drives `fetchModels` refreshes (network refresh at interactive
   startup and periodically, cache-only at registration) and persists the overlay.
   A `fetchModels` rejection never blocks startup.
+- **pi 0.99 virtual models do not break the provider, but they do interact with the
+  cross-model thinking guard** (verified pi 0.99.1/0.99.2). A virtual model
+  (`pi.registerVirtualModel`, api `pi-virtual`) can be registered under `nan` and
+  route to physical `nan` models via `ctx.modelRegistry.find("nan", id)` — no
+  provider-side change needed. Inside the extension `context` hook, though,
+  `ctx.model` is the **selection** (the virtual entry), never the routed physical
+  model (`core/virtual-models.js`; `agent-session.js` "The selection stays in
+  agent state; only this request uses the routed model"), and the routed target is
+  not exposed to the hook. `stripCrossModelThinking` therefore skips virtual
+  selections (`src/cross-model-thinking-guard.ts`, `VIRTUAL_MODEL_API`); otherwise
+  its same-model check would strip reasoning on every continuation. Registered
+  virtual models are otherwise transparent to the catalog/factory.
+- **Factory-time ExtensionAPI has NO `modelRegistry`** (verified pi 0.99.1):
+  The `ExtensionAPI` interface does not include `modelRegistry`; it exists only on
+  `ExtensionContext` (the `ctx` passed to command/event handlers). At factory time,
+  stored credentials are resolved via `readStoredCredential("nan")` from
+  `@earendil-works/pi-coding-agent`. Command-time paths (`/nan-mcp`, `/nan-usage`) use
+  `ctx.modelRegistry`.
 - models.json overrides compose **above** registered native providers.
 - Capability values that diverge from models.dev are recorded as build-time
   `MANUAL_OVERRIDES` (mandatory provenance note) in `scripts/manual-overrides.ts`,
@@ -132,20 +170,37 @@ Every PR that changes code MUST bump `package.json` version in the same PR; CI p
   override (maintainer-confirmed 2026-09-05) was withdrawn 2026-09-07 — the
   updated NaN docs still say 262K "the model's native window" and models.dev
   agrees at 262,144.
-- NaN's official chat model list (https://nan.builders/openapi.json `model`
-  param description + https://nan.builders/docs/models, checked 2026-09-07):
-  community `deepseek-v4-flash`, `mimo-v2.5`, `qwen3.8-flash`, `glm5.3-flash`,
-  `qwen3.6`, `gemma4` (all text+image vision) + premium-tier `glm5.3`
+- **Which models exist**: https://nan.builders/docs/models is the source of
+  truth (maintainer instruction, reaffirmed 2026-09-29), cross-checked against
+  the `model` param description in https://nan.builders/openapi.json; models.dev
+  (provider `nan`) supplies the numeric limits for the ids it documents, and
+  every divergence or exclusion is recorded with provenance.
+  Official chat models (checked 2026-09-29): community `deepseek-v4-flash`,
+  `mimo-v2.6-flash`, `qwen3.8-flash`, `glm5.3-flash`, `qwen3.6`, `gemma4`
+  (all text+image vision) + premium-tier `glm5.3`
   (~753B MoE, text-only input, 1M context, 400M tokens/rolling 4h window).
   glm5.3 is now documented by models.dev too (1M context / 131,072 max output,
   checked 2026-09-13) but stays out of the static catalog via
   `LIVE_ONLY_MODEL_IDS` (premium tier) so a non-premium key never sees a
   model it cannot call when the live `/models` fetch is unavailable; premium
   keys still receive it live with conservative placeholder limits.
-  `glm5.2` was removed by the provider (2026-09-05); models.dev may still list
-  it — the generator excludes it via `PROVIDER_REMOVED_MODEL_IDS`. Non-chat
-  endpoints: qwen3-embedding, rerank, kokoro (TTS), whisper (STT),
-  flux-2-klein (images) — MCP-bridge territory, not chat catalog models.
+  Removed by the provider: `glm5.2` (2026-09-05) and `mimo-v2.5` (checked
+  2026-09-29 — zero mentions in the docs and in openapi.json, and models.dev
+  dropped it too; superseded by `mimo-v2.6-flash`). models.dev may re-list a
+  removed id, so the generator excludes them via `PROVIDER_REMOVED_MODEL_IDS`,
+  and the four community ids the docs require (`qwen3.6`, `gemma4`,
+  `deepseek-v4-flash`, `mimo-v2.6-flash`) are `REQUIRED_MODEL_IDS` — if one
+  never reaches the catalog (from models.dev or from a manual-only entry),
+  generation exits non-zero instead of inventing data.
+  `mimo-v2.6-flash` entered through `MANUAL_ONLY_MODEL_IDS` while models.dev
+  lacked it; models.dev started listing it on 2026-09-29, and the generator
+  then skips the manual entry (attaching its note to the models.dev-derived
+  entry) so the id is never emitted twice.
+  Non-chat endpoints: qwen3-embedding, rerank, kokoro (TTS), whisper (STT),
+  flux-2-klein and qwen-image-2.1 (images) — MCP-bridge territory, not chat
+  catalog models; `qwen-image-2.1` also sits on models.dev with
+  `limit.output: 0`, so generation excludes it via `NON_CHAT_MODEL_IDS`
+  instead of flagging it "needs manual verification" forever.
 - NaN can close an SSE stream **before** `finish_reason`. The catalog sets
   `supportsFinishReason: true` so pi-ai raises the retryable
   `Stream ended without finish_reason` (pi-ai's `RETRYABLE_PROVIDER_ERROR_PATTERN`
@@ -165,6 +220,7 @@ Every PR that changes code MUST bump `package.json` version in the same PR; CI p
   `test/issue-2-truncated-stream.test.ts`, `test/issue-4-token-usage.test.ts`,
   `test/issue-7-streaming-usage-default.test.ts`;
   issues #2, #4 and #7.
+- **Catalog now declares `thinkingLevelMap` for reasoning suppression** (issue #16, measured 2026-09-27 against live gateway, repro script in the issue). `deepseek-v4-flash` → `{ off: "none" }` (0 reasoning tokens, reproducible); `glm5.3-flash` → `{ off: "minimal" }` (38 reasoning tokens + answer; `none` does NOT work on glm5.3 — 13,382 tokens); `qwen3.6` and `gemma4` → `{ off: "none" }` (gap left by the measured-only fix: the NaN docs state that with no parameter both reason by default with a 16,384-token budget and that `none`/`minimal` skip the reasoning phase entirely — https://nan.builders/docs/models#controlling-reasoning, checked 2026-10-01 — so `off` must send `none` explicitly). `reasoningEffortValues` corrected: deepseek-v4-flash from `[]` to `["none"]`; glm5.3-flash from `["low","medium","high","max"]` to include `"minimal"`. README limits table notes that `maxTokens` does NOT bound the reasoning phase (measured: max_tokens=512 still yielded 13,376 reasoning tokens). pi-ai's `buildRequest` (0.99.1) maps `reasoning: "off"` → `reasoningEffort: undefined` → fires the off-branch that checks `model.thinkingLevelMap?.off` (verified in pi-ai source ~line 718-722). A model whose reasoning depth is genuinely not adjustable (`qwen3.8-flash`, `mimo-v2.6-flash`) keeps no `thinkingLevelMap`: `off` sends nothing and the model uses its own default.
 - A NaN request that still exceeds the destination model's context window (the
   cross-model thinking guard is disabled with `NAN_THINKING_GUARD=0`, the
   inflation is not a `thinking` block, or the window is smaller) gets NaN's
@@ -178,19 +234,44 @@ Every PR that changes code MUST bump `package.json` version in the same PR; CI p
   `test/context-overflow-classifier.test.ts`; issue #3.
 - Relative imports inside this package use `.ts` extensions (pi's official
   extension examples do the same; pi transpiles extension sources).
-- pi intentionally has NO built-in MCP client (docs/usage.md). MCP integration
-  happens by bridging servers into pi custom tools via `pi.registerTool()`:
+- pi 0.87 breaking changes NOT used by this package: `ContextEditEntry` added to
+  the `SessionEntry` union, expanded `TurnEndEvent`, removed `shouldStopAfterTurn`,
+  `SessionManager` canonical for provider context. The package touches none of them.
+- Since 0.10.0, this package uses native MCP via `pi.registerMcpServer()`.
+  Both servers are session-scoped, visible in `/mcp` with source "extension".
+  Server names and tool IDs:
+  - `nan-search` → tools: `mcp__nan-search__web_search`
+    (official remote, HTTP, `Authorization: Bearer <key>` header)
+  - `nan-media` → tools: `mcp__nan-media__generate_image/edit_image/text_to_speech/list_voices/speech_to_text`
+    (community stdio, `npx -y nan-mcp-server@1.1.2`, 120s timeout)
+  - `exposure: "direct"` on both — tools are declared to the model.
+  - User `mcp.json` entries with the same name take precedence.
+  - `/nan-mcp enable/disable` calls `pi.registerMcpServer()` / `pi.unregisterMcpServer()` directly.
+  - Missing `NAN_API_KEY` at load: web-search server NOT registered (console.warn).
+  NaN integration facts remain:
   - NaN's official remote MCP server: `https://api.nan.builders/mcp` (host
     root, NOT /v1; JSON-RPC 2.0 over streamable HTTP, stateless; same `sk-`
     key, shared rate limit/quota). Spec: https://nan.builders/openapi.json
-    (tag "MCP"). Currently exposes `web_search` (same args as POST /v1/search);
-    "growing registry" — use tools/list to discover.
+    (tag "MCP"). Currently exposes `web_search`.
   - Community `nan-mcp-server` (https://github.com/luciferfran/nan-mcp-server):
-    stdio MCP server, spawned per tool call (lazy), opt-in NAN_MEDIA_MCP=1,
-    version-pinned via NAN_MEDIA_MCP_VERSION (default 1.0.8) or a full command
-    override via NAN_MEDIA_MCP_COMMAND. Tools: generate_image, edit_image,
-    text_to_speech, list_voices, speech_to_text, embed, rerank, list_models
-    (we bridge the audio/image/transcription scope). An automated check
-    (scripts/check-nan-mcp-server.ts + .github/workflows/check-nan-mcp-server-update.yml)
-    compares the npm registry against the pin weekly and files a `dependencies` issue
-    with a breaking/safe verdict from the live server tool surface (unpkg).
+    stdio MCP server, version-pinned via NAN_MEDIA_MCP_VERSION (default 1.1.2)
+    or full command override via NAN_MEDIA_MCP_COMMAND. Tools: generate_image,
+    edit_image, text_to_speech, list_voices, speech_to_text.
+  - Automated check (scripts/check-nan-mcp-server.ts) compares the npm registry
+    against the pin weekly and files a `dependencies` issue when a newer release
+    exists, with a breaking/safe verdict from the live server tool surface.
+- **pi reports unhandled MCP registrations right after `session_start`** (verified
+  on pi 0.99.1): `AgentSession.bindExtensions()` emits `session_start` and then
+  calls `ExtensionRunner.reportUnhandledMcpServers()`, which raises
+  `MCP server "<name>" is registered, but no loaded extension connects MCP
+  servers` whenever no loaded extension handles `mcp_servers_change`. A host
+  that builds `DefaultResourceLoader` WITHOUT `extensionFactories` loads no
+  `builtin:*` extension at all — PI WEB's `createAgentSessionServices` does
+  exactly that (pi-web 1.202609.1, verified 2026-09-30 with a faithful loader
+  run: packages load, `builtin:mcp` absent, no `/mcp` command), so its sessions
+  have no MCP connector and `mcp.json` servers never connect either.
+  `src/mcp/host-support.ts` detects the gap through the `/mcp` command
+  (`pi.getCommands()` missing → conservative "connector present"), claims
+  `mcp_servers_change` so pi's report stays silent, and warns once per session;
+  `/nan-mcp enable` persists the toggle without registering. Regression tests:
+  `test/mcp-host-connector.test.ts`.

package/README.es.md

@@ -1,11 +1,11 @@
 # @gtrabanco/pi-nan-provider
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
-[![Version](https://img.shields.io/badge/version-0.8.0-blue)](https://github.com/gtrabanco/pi-nan-provider/releases)
+[![Version](https://img.shields.io/badge/version-0.10.0-blue)](https://github.com/gtrabanco/pi-nan-provider/releases)
 
-[NaN Builders](https://nan.builders) model provider + MCP bridges para [pi](https://github.com/earendil-works/pi). 
+[NaN Builders](https://nan.builders) model provider + MCP nativo para [pi](https://github.com/earendil-works/pi). 
 
-Registra el proveedor `nan` vía `pi.registerProvider()` usando la API OpenAI-compatible de NaN (`https://api.nan.builders/v1`), y conecta las herramientas MCP de NaN en pi con `pi.registerTool()`.
+Registra el proveedor `nan` vía `pi.registerProvider()` usando la API OpenAI-compatible de NaN (`https://api.nan.builders/v1`), y registra servidores MCP nativos con `pi.registerMcpServer()`.
 
 ---
 
@@ -97,40 +97,72 @@ Puedes sobrescribir el `compat` de cualquier modelo en `~/.pi/agent/models.json`
 
 Consigue una clave en la [plataforma NaN](https://cloud.nan.builders/r/7GK06FX8) (ajustes de usuario → API Keys; enlace de referidos).
 
-## 🔌 Puentes MCP
+## 🔌 MCP nativo
 
-Dado que [pi no incluye un cliente MCP integrado](https://github.com/earendil-works/pi/blob/main/docs/usage.md), este paquete conecta los servidores MCP como **herramientas nativas de pi**.
+Dado que [pi 0.99.0 incluye un cliente MCP integrado](https://github.com/earendil-works/pi/blob/main/docs/usage.md), este paquete migra sus puentes a MCP nativo (requiere `pi >=0.99` — cambio rotundo).
 
-Ambos puentes están **activados y son perezosos (lazy) por defecto**. Usa `/nan-mcp` para gestionarlos.
+Ambos puentes están **activados por defecto** (por sesión, visibles en `/mcp`). Usa `/nan-mcp` para gestionarlos.
 
 ### 🛠️ Comando de Gestión: `/nan-mcp`
 
 | Comando | Efecto |
 | :--- | :--- |
-| `/nan-mcp status` | Muestra el estado actual de ambos puentes. |
+| `/nan-mcp status` | Muestra el estado actual de ambos puentes y la resolución de la clave API. |
 | `/nan-mcp enable [target]` | Activa `web-search` o `nan-mcp-server` (persiste). |
 | `/nan-mcp disable [target]` | Desactiva un puente de forma persistente. |
 
+Ambos puentes resuelven la credencial almacenada de `/login nan` (almacenada → env). Si ninguna se resuelve, el puente no se registra y se muestra un aviso en la salida de estado.
+
+> [!NOTE]
+> El registro ocurre cuando se carga la extensión. Si ejecutas `/login nan` **después** de que la sesión haya empezado, la factoría ya se ejecutó — usa `/reload` (o `/nan-mcp enable`) para registrar los puentes sin reiniciar pi.
+
+### 🚧 Hosts sin la extensión MCP integrada de pi (PI WEB)
+
+pi conecta los servidores registrados desde la extensión que maneja `mcp_servers_change` — la
+extensión integrada `mcp`, que además registra `/mcp`. Los hosts que construyen el resource loader
+de pi sin `extensionFactories` no cargan **ninguna** extensión integrada (verificado:
+`createAgentSessionServices` de PI WEB, pi 0.99.1 / pi-web 1.202609.1), así que nada conecta los
+servidores y pi reporta `MCP server "nan-media" is registered, but no loaded extension connects MCP
+servers`.
+
+Ahí los puentes no pueden exponer herramientas, así que este paquete reemplaza ese error críptico
+con un único aviso por sesión, mantiene los registros inactivos y deja que `/nan-mcp enable` persista
+el toggle sin registrar un servidor que nunca podrá conectarse. Para tener MCP en la CLI, activa el
+conector con `pi config` → Built-in extensions → `mcp`.
+
+### Funciones de pi 0.99 y modelos NaN
+
+- **Modelos virtuales** — registra un router bajo `nan` (`pi.registerVirtualModel({ provider: "nan", ... })`);
+  enruta a los modelos físicos de NaN sin configuración por parte del proveedor. El guard de
+  razonamiento entre modelos se salta las selecciones virtuales, porque el modelo físico enrutado no
+  es visible para el hook `context`.
+- **Codemode** — funciona con cualquier modelo NaN que soporte tool calling. Las herramientas MCP de
+  este paquete usan `exposure: "direct"`, así que los scripts de codemode también pueden llamarlas.
+  Requiere la extensión integrada `codemode` de pi (ver la nota de PI WEB arriba).
+- **Modelos clasificadores** — no aplica: NaN no expone ninguna API de clasificación (solo chat y
+  endpoints de embedding/rerank/audio/imagen), así que este proveedor no registra modelos
+  `type: "classifier"`.
+
 ---
 
 ### 1. Servidor MCP oficial de NaN
 *Puente oficial para herramientas remotas vía [https://api.nan.builders/mcp](https://nan.builders/docs/api).*
 
-- **`nan_web_search(query, ...)`**: Realiza búsquedas web a través del gateway de NaN.
+- **`mcp__nan-search__web_search(query, ...)`**: Realiza búsquedas web a través del gateway de NaN.
 
 ### 2. Servidor MCP de Media (Comunidad)
 *Conecta [`nan-mcp-server`](https://github.com/luciferfran/nan-mcp-server) mediante un cliente stdio local mínimo.*
 
-- **Carga Perezosa (Lazy)**: El proceso del servidor se lanza **solo** cuando se invoca una herramienta y se cierra inmediatamente después.
+- **Por sesión**: El servidor se conecta al registrarse (eager). Aparece en `/mcp` con fuente "extension".
 - **Configuración**: Los archivos se guardan en `~/nan-mcp-output/`.
 
 | Herramienta | Propósito |
 | :--- | :--- |
-| `nan_generate_image` | Generación de imágenes (flux-2-klein) |
-| `nan_edit_image` | Edición imagen→imagen (flux-2-klein) |
-| `nan_text_to_speech` | Síntesis de audio (kokoro) |
-| `nan_list_voices` | Listar voces disponibles |
-| `nan_speech_to_text` | Transcripción de audio (whisper) |
+| `mcp__nan-media__generate_image` | Generación de imágenes (flux-2-klein) |
+| `mcp__nan-media__edit_image` | Edición imagen→imagen (flux-2-klein) |
+| `mcp__nan-media__text_to_speech` | Síntesis de audio (kokoro) |
+| `mcp__nan-media__list_voices` | Listar voces disponibles |
+| `mcp__nan-media__speech_to_text` | Transcripción de audio (whisper) |
 
 #### 🔧 Configuración del Puente de Media
 
@@ -260,12 +292,33 @@ Catálogo base (verificado contra [docs de NaN](https://nan.builders/docs/models
 | `qwen3.6` | 262,144 | 65,536 | texto, imagen | ✅ |
 | `gemma4` | 262,144 | 32,768 | texto, imagen | ✅ |
 | `deepseek-v4-flash` | 1,000,000 | 384,000 | texto, imagen | ✅ |
-| `mimo-v2.5` | 1,048,576 | 131,072 | texto, imagen | ✅ |
+| `mimo-v2.6-flash` | 1,048,576 | 131,072 | texto, imagen | ✅ |
 | `glm5.3-flash` | 1,000,000 | 131,072 | texto, imagen | ✅ |
 | `qwen3.8-flash` | 262,144 | 131,072 | texto, imagen | ✅ |
 
+> [!NOTE]  
+> `mimo-v2.6-flash` está servido por NaN y desde el 2026-09-29 aparece también en models.dev (proveedor `nan`); la entrada manual que lo sostenía se conserva como fallback y su nota de procedencia sigue unida a la entrada generada.
+>
+> `mimo-v2.5` fue eliminado por NaN: ya no aparece ni en [los docs de NaN](https://nan.builders/docs/models) ni en la [lista de modelos del OpenAPI](https://nan.builders/openapi.json) (comprobado 2026-09-29), así que sale del catálogo — lo sustituye `mimo-v2.6-flash` (mismos límites 1,048,576 / 131,072 y cuota mensual de 1.0B). El consumo histórico de `mimo-v2.5` en `/nan-usage` sigue apareciendo en la sección de modelos no documentados.
+
 ---
 
+---
+
+## 🧠 Controles de razonamiento
+
+El parámetro `reasoning_effort` de NaN controla cuánto piensa el modelo antes de responder — pero el grado de control varía por modelo:
+
+| Modelo | Effort de razonamiento | Cómo funciona |
+| :--- | :--- | :--- |
+| `glm5.3`, `glm5.3-flash` | `minimal` · `low` · `medium` · `high` · `max` | Completamente controlable — valores más altos permiten al modelo razonar más; `minimal` produce ~38 tokens de razonamiento (issue #16, medido 2026-09-27); `none` NO suprime el razonamiento en glm5.3-flash (13,382 tokens) |
+| `qwen3.6`, `gemma4` | `none` · `minimal` · `low` · `medium` · `high` · `max` | `none`/`minimal` omiten el razonamiento por completo; otros valores limitan a 2K / 8K / 16K / 32K tokens. `thinking: "off"` envía `reasoning_effort: "none"` automáticamente vía el `thinkingLevelMap` del catálogo — los docs de NaN dicen que sin parámetro estos modelos razonan por defecto (presupuesto de 16,384 tokens), así que `off` debe ser explícito |
+| `deepseek-v4-flash` | `none` · `minimal` · `low` · `medium` · `high` · `max` (el catálogo declara ["none"]) | `none` deshabilita el razonamiento de forma fiable (0 tokens de razonamiento, medido 2026-09-27); cualquier otro valor deja 13-14K tokens de razonamiento y causa razonamiento descontrolado (finish_reason:"length" con CERO texto de respuesta) — el catálogo solo declara `none` porque el resto es ruido. `thinking: "off"` ahora envía `reasoning_effort: "none"` automáticamente vía el `thinkingLevelMap` del catálogo |
+| `qwen3.8-flash`, `mimo-v2.6-flash` | *(aceptado pero no ajustable)* | El parámetro es aceptado y nunca rechazado, pero el modelo gestiona su propia profundidad de razonamiento — nunca es un error enviar un valor que estos modelos no ajustan |
+
+> [!NOTE]  
+> `maxTokens` NO limita la fase de razonamiento en los modelos de NaN. La fase de razonamiento corre hasta completarse (o se trunca a ~60,000 caracteres para deepseek-v4-flash) independientemente de `maxTokens` — medido: `max_tokens=512` en deepseek-v4-flash aún consumió 13,376 tokens de completación, todos de razonamiento (26x el techo). Este es un comportamiento del gateway confirmado por el issue #16 (medido 2026-09-27).
+
 ## 🚀 Desarrollo
 
 ```bash

package/README.md

@@ -1,11 +1,11 @@
 # @gtrabanco/pi-nan-provider
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
-[![Version](https://img.shields.io/badge/version-0.8.0-blue)](https://github.com/gtrabanco/pi-nan-provider/releases)
+[![Version](https://img.shields.io/badge/version-0.10.0-blue)](https://github.com/gtrabanco/pi-nan-provider/releases)
 
-[NaN Builders](https://nan.builders) model provider + MCP bridges for [pi](https://github.com/earendil-works/pi). 
+[NaN Builders](https://nan.builders) model provider + native MCP servers for [pi](https://github.com/earendil-works/pi). 
 
-Registers the `nan` provider via `pi.registerProvider()` using NaN's OpenAI-compatible API (`https://api.nan.builders/v1`), and bridges NaN's MCP tools into pi with `pi.registerTool()`.
+Registers the `nan` provider via `pi.registerProvider()` using NaN's OpenAI-compatible API (`https://api.nan.builders/v1`), and registers MCP servers via pi's native `pi.registerMcpServer()`.
 
 ---
 
@@ -97,40 +97,69 @@ You can override any model's `compat` per-model in `~/.pi/agent/models.json` (pi
 
 Get a key from the [NaN platform](https://cloud.nan.builders/r/7GK06FX8) (user settings → API Keys; referral link).
 
-## 🔌 MCP Bridges
+## 🔌 Native MCP Servers
 
-Since [pi does not include a built-in MCP client](https://github.com/earendil-works/pi/blob/main/docs/usage.md), this package bridges MCP servers as **native pi tools**.
+Since [pi 0.99.0 ships a built-in MCP client](https://github.com/earendil-works/pi/blob/main/docs/usage.md), this package migrates its bridges to native MCP (requires `pi >=0.99` — breaking).
 
-Both bridges are **enabled and lazy by default**. Use `/nan-mcp` to manage them.
+Both bridges are **enabled by default** (session-scoped, visible in `/mcp`). Use `/nan-mcp` to manage them.
 
 ### 🛠️ Management Command: `/nan-mcp`
 
 | Command | Effect |
 | :--- | :--- |
-| `/nan-mcp status` | Shows current state of both bridges. |
+| `/nan-mcp status` | Shows current state of both bridges and API key resolution. |
 | `/nan-mcp enable [target]` | Enables `web-search` or `nan-mcp-server` (persisted). |
 | `/nan-mcp disable [target]` | Disables a bridge persistently. |
 
+Both bridges resolve the stored `/login nan` credential (stored → env). If neither resolves, the bridge does not register and a warning is shown in the status output.
+
+> [!NOTE]
+> Registration happens when the extension loads. If you run `/login nan` **after** the session started, the factory has already run — run `/reload` (or `/nan-mcp enable`) to register the bridges without restarting pi.
+
+### 🚧 Hosts without pi's built-in MCP extension (PI WEB)
+
+pi connects registered servers from the extension that handles `mcp_servers_change` — the built-in
+`mcp` extension, which also registers `/mcp`. Hosts that build pi's resource loader without
+`extensionFactories` load **no** built-in extension at all (verified: PI WEB's
+`createAgentSessionServices`, pi 0.99.1 / pi-web 1.202609.1), so nothing connects the servers and pi
+reports `MCP server "nan-media" is registered, but no loaded extension connects MCP servers`.
+
+There the bridges cannot expose tools, so this package replaces that cryptic error with one warning
+per session, keeps the registrations idle, and lets `/nan-mcp enable` persist the toggle without
+registering a server that can never connect. To get MCP in the CLI, enable the connector with
+`pi config` → Built-in extensions → `mcp`.
+
+### pi 0.99 features and NaN models
+
+- **Virtual models** — register a router under `nan` (`pi.registerVirtualModel({ provider: "nan", ... })`);
+  it routes to the physical NaN models with no provider-side configuration. The cross-model thinking
+  guard skips virtual selections, because the routed physical model is not visible to the `context` hook.
+- **Codemode** — works with any tool-calling NaN model. This package's MCP tools use
+  `exposure: "direct"`, so codemode scripts can call them as well. Requires pi's built-in `codemode`
+  extension (see the PI WEB note above).
+- **Classifier models** — not applicable: NaN exposes no classifier API (chat plus
+  embedding/rerank/audio/image endpoints only), so this provider registers no `type: "classifier"` models.
+
 ---
 
 ### 1. Official NaN MCP Server
-*Official bridge for remote tools via [https://api.nan.builders/mcp](https://nan.builders/docs/api).*
+*Official remote MCP server registered natively via `pi.registerMcpServer()`.*
 
-- **`nan_web_search(query, ...)`**: Performs web searches through NaN's gateway.
+- **`mcp__nan-search__web_search(query, ...)`**: Performs web searches through NaN's gateway.
 
 ### 2. Community Media MCP Server
 *Bridges [`nan-mcp-server`](https://github.com/luciferfran/nan-mcp-server) via a minimal local stdio client.*
 
-- **Lazy Loading**: The server process is spawned **only** when a tool is invoked and terminated immediately after.
+- **Session-scoped**: The server connects at registration time (eager). Appears in `/mcp` with source "extension".
 - **Config**: Files land in `~/nan-mcp-output/`.
 
 | Tool | Purpose |
 | :--- | :--- |
-| `nan_generate_image` | Image generation (flux-2-klein) |
-| `nan_edit_image` | Image-to-image editing (flux-2-klein) |
-| `nan_text_to_speech` | Audio synthesis (kokoro) |
-| `nan_list_voices` | List available voices |
-| `nan_speech_to_text` | Audio transcription (whisper) |
+| `mcp__nan-media__generate_image` | Image generation (flux-2-klein) |
+| `mcp__nan-media__edit_image` | Image-to-image editing (flux-2-klein) |
+| `mcp__nan-media__text_to_speech` | Audio synthesis (kokoro) |
+| `mcp__nan-media__list_voices` | List available voices |
+| `mcp__nan-media__speech_to_text` | Audio transcription (whisper) |
 
 #### 🔧 Media Bridge Configuration
 
@@ -260,13 +289,14 @@ Baseline catalog (verified against [NaN docs](https://nan.builders/docs/models)
 | `qwen3.6` | 262,144 | 65,536 | text, image | ✅ |
 | `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.
+> `mimo-v2.6-flash` is served by NaN and has been listed by models.dev provider `nan` since 2026-09-29. Before that it entered the catalog through a manual-only entry; that entry is kept as a fallback (re-emitted automatically if models.dev drops the model again) and its provenance note stays attached to the generated entry.
+>
+> `mimo-v2.5` was removed by NaN: it no longer appears in the [NaN docs](https://nan.builders/docs/models) nor in the [OpenAPI model list](https://nan.builders/openapi.json) (checked 2026-09-29), so it is out of the catalog — `mimo-v2.6-flash` replaces it (same 1,048,576 / 131,072 limits, 1.0B monthly quota). Historical `/nan-usage` consumption for `mimo-v2.5` still shows under the undocumented-models section.
 
 ---
 
@@ -276,9 +306,13 @@ NaN's `reasoning_effort` parameter controls how much the model thinks before ans
 
 | 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 |
+| `glm5.3`, `glm5.3-flash` | `minimal` · `low` · `medium` · `high` · `max` | Fully controllable — higher values let the model reason longer; `minimal` yields ~38 reasoning tokens (issue #16, measured 2026-09-27); `none` does NOT suppress reasoning on glm5.3-flash (13,382 tokens) |
+| `qwen3.6`, `gemma4` | `none` · `minimal` · `low` · `medium` · `high` · `max` | `none`/`minimal` skip reasoning entirely; others cap at 2K / 8K / 16K / 32K tokens. `thinking: "off"` sends `reasoning_effort: "none"` automatically via the catalog's `thinkingLevelMap` — the NaN docs state that with no parameter these models reason by default (16,384-token budget), so off must be explicit |
+| `deepseek-v4-flash` | `none` · `minimal` · `low` · `medium` · `high` · `max` (catalog declares ["none"]) | `none` reliably disables reasoning (0 reasoning tokens, measured 2026-09-27); every other value leaves 13-14K reasoning tokens and causes runaway reasoning (finish_reason:"length" with ZERO answer text) — the catalog only declares `none` because the rest are useless. `thinking: "off"` now sends `reasoning_effort: "none"` automatically via the catalog's `thinkingLevelMap` |
+| `qwen3.8-flash`, `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 |
+> [!NOTE]  
+> `maxTokens` does NOT bound the reasoning phase on NaN models. The reasoning phase runs to completion (or truncates at ~60,000 chars for deepseek-v4-flash) regardless of `maxTokens` — measured: `max_tokens=512` on deepseek-v4-flash still consumed 13,376 completion tokens, all reasoning (26x the ceiling). This is gateway behavior confirmed by issue #16 (measured 2026-09-27).
+
 
 ---
 

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@gtrabanco/pi-nan-provider",
-	"version": "0.8.0",
+	"version": "0.10.2",
 	"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",
@@ -42,12 +42,12 @@
 		"check-pi-sdk-versions": "node scripts/check-pi-sdk-versions.mjs --report"
 	},
 	"peerDependencies": {
-		"@earendil-works/pi-ai": ">=0.83.0 <1",
-		"@earendil-works/pi-coding-agent": ">=0.83.0 <1"
+		"@earendil-works/pi-ai": ">=0.99.0 <1",
+		"@earendil-works/pi-coding-agent": ">=0.99.0 <1"
 	},
 	"devDependencies": {
-		"@earendil-works/pi-ai": "0.84.4",
-		"@earendil-works/pi-coding-agent": "0.84.4",
+		"@earendil-works/pi-ai": "0.99.1",
+		"@earendil-works/pi-coding-agent": "0.99.1",
 		"@types/bun": "1.4",
 		"typescript": "7"
 	},

package/scripts/check-nan-mcp-server.ts

@@ -42,8 +42,16 @@
 
 import {
 	DEFAULT_NAN_MEDIA_MCP_VERSION,
-	NAN_MEDIA_MCP_SERVER_TOOLS,
-} from "../src/mcp/nan-media.ts";
+} from "../src/mcp/media-server.ts";
+
+/** The audio/image/transcription scope we bridge from the community server. */
+const bridgedtools = [
+	"generate_image",
+	"edit_image",
+	"text_to_speech",
+	"list_voices",
+	"speech_to_text",
+] as const;
 
 /** npm package name of the community stdio MCP server we bridge. */
 export const NAN_MCP_SERVER_PACKAGE = "nan-mcp-server";
@@ -143,7 +151,7 @@ export function assessUpdate(latestTools: string[]): {
 	removedTools: string[];
 	reason: string;
 } {
-	const bridged: readonly string[] = [...NAN_MEDIA_MCP_SERVER_TOOLS];
+	const bridged: readonly string[] = [...bridgedtools];
 	const latest = new Set(latestTools.map((t) => t.trim()));
 
 	const removedTools = bridged.filter((tool) => !latest.has(tool));