@counterposition/skill-pi 0.87.1 → 0.99.1

package/skills/pi/references/mcp.md

@@ -0,0 +1,122 @@
+# MCP Servers
+
+Pi 0.99 connects to MCP servers through a built-in extension. There is no package to install. Checked against Pi v0.99.1: [mcp.md](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/mcp.md), [cli.md#mcp-commands](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/cli.md#mcp-commands), [extensions.md#mcp-servers](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/extensions.md#mcp-servers).
+
+## Configure
+
+| File | Scope |
+|------|-------|
+| `~/.pi/agent/mcp.json` | Global. Put personal servers and servers with credentials here |
+| `.pi/mcp.json` | Project. Read only after the project is trusted, because stdio servers run commands |
+
+A project entry replaces a global entry with the same name. The format is the usual `mcpServers` object, so Claude Desktop, Claude Code, and Cursor entries copy over unchanged:
+
+```json
+{
+  "mcpServers": {
+    "filesystem": {
+      "command": "npx",
+      "args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
+    },
+    "docs": {
+      "url": "https://example.com/mcp",
+      "headers": { "Authorization": "Bearer ${DOCS_TOKEN}" },
+      "exposure": "direct"
+    }
+  }
+}
+```
+
+- stdio: `command` (one executable, not a shell string), `args`, `env`, `cwd`. Relative `cwd` resolves against the session directory; a leading `~/` works in `command`, args, and `cwd`.
+- HTTP: `url`, `headers`, `oauth`. Only streamable HTTP; `"type": "sse"` is rejected (try the server's `/mcp` endpoint instead of `/sse`).
+- `type` is optional (`stdio`, `http`, or `streamable-http`); `command` implies stdio and `url` implies HTTP.
+- `env` and `headers` values accept `${NAME}` and `!command`. A command must produce the whole value: `"Authorization": "!echo Bearer $(gh auth token)"`.
+- `timeout` is seconds per request (default 60); progress notifications reset it. `enabled: false` keeps the entry without connecting.
+- Server names use letters, digits, `_`, `-`. Tools are named `mcp__<server>__<tool>`.
+- Invalid entries are skipped and reported; other servers still connect.
+
+Converting other clients: VS Code uses `servers` + `inputs` (move under `mcpServers`, replace `${input:...}` with `${NAME}`); Codex uses TOML `[mcp_servers.<name>]` (same fields as JSON); opencode uses `"type": "local"` with an array `command` (split into `command`/`args`), `environment` for `env`, and `{env:NAME}` for `${NAME}`.
+
+## Set Up From a Shell
+
+These commands work outside a session, so an agent can run them through `bash`:
+
+```bash
+pi mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem .
+pi mcp add docs --url https://example.com/mcp --bearer-token-env-var DOCS_TOKEN --exposure direct
+pi mcp add -l tools --env API_KEY='${TOOLS_KEY}' -- uvx tools-mcp   # -l writes .pi/mcp.json
+pi mcp list            # connects to every enabled server; exits 1 while anything is wrong
+pi mcp login sentry    # OAuth: opens the browser and waits for the user to approve
+pi mcp remove docs     # keeps stored OAuth credentials
+pi mcp logout sentry
+```
+
+`pi mcp add` does not connect; verify with `pi mcp list`. A running session picks up added or changed servers after `/reload` or a new session.
+
+In a session, `/mcp` opens the server manager (state, tools, errors, sign in/out, reconnect, exposure, enable/disable). Changes are written back to the `mcp.json` that defines the server. Outside the TUI, `/mcp` prints status and `/mcp login|logout|reconnect <server>` run directly.
+
+Pi connects on session start; the first prompt waits up to 10 seconds for startup connections. OAuth tokens live in `~/.pi/agent/mcp-auth.json`; server log notifications go to `~/.pi/agent/mcp.log`. OAuth is used only for HTTP servers whose config has no `Authorization` header. The optional `oauth: { clientId?, clientSecret?, callbackPort?, callbackUrl?, scope? }` sets a pre-registered client for servers without dynamic client registration, the loopback redirect (`callbackPort` alone gives `http://127.0.0.1:<port>/callback`; a `callbackUrl` without a port gets `callbackPort` or a free one), and `scope`, space-separated scopes to request when the server does not advertise the ones it needs.
+
+## Exposure
+
+`exposure` decides how the model reaches a server's tools:
+
+| Value | Declared to the model | Callable from codemode | Notes |
+|-------|----------------------|------------------------|-------|
+| `codemode` (default) | No | Yes, listed in the `codemode` description | Pi activates `codemode` when such a server connects |
+| `codemode-deferred` | No | Yes, found with `searchTools()` / `ALL_TOOLS` | Description names only the server and tool count; for large, rarely used servers |
+| `deferred` | After `tool_search` loads them | Yes | Pi activates `tool_search`; for large servers without codemode |
+| `direct` | Yes, like built-in tools | Yes | |
+| `hidden` | No | No | Registered but unreachable |
+
+`toolExposure` overrides single tools by name or `*` pattern (exact name wins, then first matching pattern):
+
+```json
+{
+  "mcpServers": {
+    "github": {
+      "url": "https://api.githubcopilot.com/mcp/",
+      "exposure": "deferred",
+      "toolExposure": { "search_code": "direct", "get_*": "codemode", "delete_*": "hidden" }
+    }
+  }
+}
+```
+
+Undeclared tools (`codemode`, `codemode-deferred`, `deferred`) are reachable through either tool: codemode scripts can call a `deferred` server's tools, and `tool_search` can load a `codemode` server's tools. Set top-level `"autoEnableCodemode": false` in `mcp.json` to stop Pi from activating `codemode`. Pi warns once when neither `codemode` nor `tool_search` is active, because the tools are then unreachable.
+
+## Results
+
+- Text results over 20KB reach the model with the middle cut out; the full text goes to a temp file named in the result.
+- Codemode scripts receive the whole `CallToolResult` (`content`, `structuredContent`, `isError`). A result with `isError` resolves in scripts and is reported to the model as an error for direct calls. `image(result.content[0])` forwards an image block.
+- Servers with resources add `list_mcp_resources`, `list_mcp_resource_templates`, and `read_mcp_resource` (marked read-only).
+- Tool calls are never retried (the server may have run them); resource reads retry once on 408/429/5xx.
+
+## Permissions
+
+Every MCP call goes through Pi's tool pipeline, so `tool_call` and `tool_result` handlers (permission gates included) see MCP tools. Calls from codemode scripts are nested calls; see [Nested Calls and Permissions](codemode.md#nested-calls-and-permissions). `pi.getAllTools()` reports the server's `readOnlyHint`, `destructiveHint`, `idempotentHint`, and `openWorldHint` annotations. They are unverified hints; see [Tool Exposure](extensions.md#tool-exposure--structured-results).
+
+## Servers From Extensions
+
+```typescript
+pi.registerMcpServer("jira", { url: "https://mcp.example.com/jira", exposure: "codemode" });
+pi.unregisterMcpServer("jira");
+```
+
+- Same config shape as an `mcpServers` entry. Registrations are not saved: register on every load.
+- Registered during load: connects on `session_start` with the `mcp.json` servers. Registered later: connects right away.
+- An `mcp.json` server with the same name wins. Another extension's name, an invalid name, or an invalid config throws.
+- `pi mcp` shell commands do not load extensions and only see `mcp.json` servers.
+- Alternative MCP extensions read registrations with `pi.getMcpServers()` on `session_start` and handle `mcp_servers_change`.
+
+## Turning It Off or Replacing It
+
+- `"extensions": ["-builtin:mcp"]` in settings, or `pi config` → Built-in, disables the built-in support. `pi mcp` shell commands still work.
+- `--no-extensions` disables built-ins too; `pi -ne -e builtin:mcp` keeps only MCP.
+- An extension that registers the `/mcp` command (for example `pi-mcp-adapter`) replaces the built-in support: `mcp.json` is then not read in sessions.
+
+## SDK
+
+SDK sessions do not load built-in extensions. Add `createMcpExtension()`, plus `createCodemodeExtension()` for `codemode`/`codemode-deferred` servers and `createToolSearchExtension()` for `deferred` servers, then call `session.bindExtensions()` so `session_start` connects the servers. See `references/sdk.md`.
+
+The standalone client is published as `@earendil-works/pi-mcp` (stdio and streamable HTTP transports, OAuth, an in-memory testing transport). It has no Pi session dependency.

package/skills/pi/references/packages.md

@@ -94,16 +94,21 @@ If there is no `pi` manifest, Pi auto-discovers:
 
 ## Dependencies
 
+Checked against Pi v0.99.1 [packages.md](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/packages.md#declare-dependencies).
+
 - Normal runtime dependencies belong in `dependencies`
-- Pi core libraries should be peer deps with `"*"` ranges:
+- Pi supplies these to extensions; declare them as peer deps with `"*"` ranges and never bundle them:
   - `@earendil-works/pi-ai`
   - `@earendil-works/pi-agent-core`
   - `@earendil-works/pi-coding-agent`
   - `@earendil-works/pi-tui`
   - `typebox`
-- Pin dev dependencies to a current Pi line (e.g. `^0.84.2`) so typecheck/tests validate against what users run; keep peers `"*"` so installs stay compatible across Pi versions
-- Other Pi packages must be bundled explicitly
-- Git package installs run `npm install --omit=dev` — runtime needs go in `dependencies`, not `devDependencies`
+- Do not list those host-provided packages in `dependencies`. A physical copy can bypass Pi's extension module mapping and create duplicate classes and registries; since Pi 0.99.0 Pi reports an extension warning for this manifest.
+- Pi suppresses automatic peer installation for managed npm packages and for git packages installed with npm (`--omit=dev --legacy-peer-deps`), pnpm (`--prod`, no auto-installed peers), or Bun (`--omit=dev --omit=peer`). Local path packages are not installed or modified; their dependency tree is the author's responsibility.
+- Git package installs omit dev dependencies — runtime needs go in `dependencies`, not `devDependencies`
+- Installed packages load with separate module roots. Do not rely on two packages sharing one dependency instance, or on one package resolving another package's undeclared dependency.
+- Other Pi packages used as dependencies must be included in the published tarball and referenced through their `node_modules` resource paths.
+- Pin dev dependencies to the Pi line you test against (e.g. `^0.99.1`) so typecheck/tests validate against what users run; keep peers `"*"` so installs stay compatible across Pi versions. `"*"` does not prove compatibility: document the minimum supported Pi version and feature-check newer APIs (for example `ctx.executeTool`, `pi.registerMcpServer`) if you support older hosts.
 
 ## Package Filtering
 
@@ -134,4 +139,5 @@ Rules:
 
 - Project settings override global settings for the same package identity — unless the project entry has `autoload: false`, in which case it is applied as a delta over the global entry (Pi 0.80.4)
 - Identity is package name for npm, repo URL for git, and resolved absolute path for local sources
-- `pi config` can enable or disable package resources after installation; since Pi 0.80.4 it manages global vs project-local scopes (Tab to switch, `pi config -l` to start in project mode with inherited global resources dimmed)
+- `pi config` can enable or disable package resources after installation; since Pi 0.80.4 it manages global vs project-local scopes (Tab to switch, `pi config -l` to start in project mode with inherited global resources dimmed). Since Pi 0.99.0 its Built-in section also disables the built-in `mcp`, `llama.cpp`, `codemode`, and `tool-search` extensions, stored as `-builtin:<name>` in `extensions`
+- The `codemode`, `tool-search`, and `mcp` built-ins are replaceable: an extension that registers one of their tools, commands, or flags (for example `/mcp`, `codemode`, or `tool_search`) replaces that built-in, and Pi warns

package/skills/pi/references/patterns.md

@@ -19,6 +19,11 @@ Practical recipes for extending Pi. Each pattern is a self-contained example.
 13. [File Watcher / Trigger](#file-watcher--trigger)
 14. [Input Transform](#input-transform)
 15. [Tool Override](#tool-override)
+16. [Structured Tool](#structured-tool)
+17. [Nested Tool Calls](#nested-tool-calls)
+18. [Classifier Gate](#classifier-gate)
+
+Patterns 16–18 are full files under `examples/` that type-check against Pi 0.99.1 in this skill's package; read them rather than retyping.
 
 ---
 
@@ -43,6 +48,8 @@ export default function (pi: ExtensionAPI) {
 }
 ```
 
+The handler also runs for `bash` calls made by codemode scripts; see [Nested Calls and Permissions](codemode.md#nested-calls-and-permissions).
+
 ## Protected Paths
 
 Block writes to sensitive files:
@@ -313,3 +320,30 @@ pi.on("tool_call", async (event, ctx) => {
   }
 });
 ```
+
+## Structured Tool
+
+`examples/structured-tool.ts` registers a read-only data tool:
+
+- `outputSchema` plus `structuredContent` on every result, so codemode scripts get `{ files, missing }` as an object while the model reads a text summary.
+- `annotations: { readOnlyHint: true, openWorldHint: false }` for permission extensions, and a `namespace` for grouping in the `codemode` description.
+- `isError: true` when no file was found: the model sees an error, scripts still get the data.
+- Only `ENOENT` is treated as "missing"; other errors are rethrown.
+
+## Nested Tool Calls
+
+`examples/nested-tools.ts` registers `read_many`, which reads several files in parallel through `ctx.executeTool("read", ...)`:
+
+- `exposure: "model-only"`: the model can call it, other tools and codemode cannot, so it cannot recurse.
+- Checks `ctx.tools` first: an inactive `direct` tool is not callable.
+- Each nested call runs every `tool_call` gate (see [Nested Calls and Permissions](codemode.md#nested-calls-and-permissions)). `executeTool()` never rejects, and aborted calls come back as `isError`, so the tool calls `signal?.throwIfAborted()` after the batch to report the abort as an abort rather than as failed reads.
+- Streams progress through `onUpdate`.
+
+## Classifier Gate
+
+`examples/classifier-gate.ts` asks a classifier model whether a `bash` command is destructive before it runs:
+
+- Finds Jev with `ctx.modelRegistry.findOfType("classifier", "typesafe", "jev-latest")` and calls `ctx.modelRegistry.classify(..., { signal: ctx.signal })`.
+- `classify()` never rejects: it checks `stopReason` (`"aborted"` blocks, `"error"` falls through to confirmation) and validates the `bool` answer's probability.
+- Fails closed: a missing model, an error, an out-of-range answer, or a probability above the strict `MAX_ALLOWED_PROBABILITY` (0.2; tune it) asks the user, and blocks when `ctx.hasUI` is false.
+- Uses `satisfies ClassifierContext["questions"]` so the question literal keeps its narrow types.

package/skills/pi/references/providers.md

@@ -7,7 +7,8 @@ Pi supports subscription-based providers via OAuth, API-key providers via env va
 Use `/login` in interactive mode, then select a provider. The `/login` selector is fuzzy-searchable and shows where each entry's auth comes from (`--api-key`, env var, custom provider) without leaking the secret.
 
 - Anthropic Claude Pro / Max — third-party usage draws from extra usage and is billed per token (suppress the warning via `warnings.anthropicExtraUsage`)
-- OpenAI ChatGPT Plus / Pro (Codex) — `/login` defaults to browser auth but can use a device-code flow for headless environments
+- OpenAI with a ChatGPT subscription (Pi 0.99.0) — `/login openai` offers **Sign in with ChatGPT**, an OAuth login that uses the subscription with the OpenAI API on the regular `openai` provider. Pi stores a stable `deviceId` for it in global settings. Subscription usage-limit errors are not retried
+- OpenAI Codex (legacy) — the older `openai-codex` provider, renamed in Pi 0.99.0 and superseded by Sign in with ChatGPT. Its `/login` still offers browser auth or a device-code flow; its default model is `gpt-6.1-sol` (Pi 0.99.1). Prefer Sign in with ChatGPT on `openai` for new setups
 - GitHub Copilot
 - OpenRouter (Pi 0.82.0) — `/login openrouter` runs a PKCE flow that mints a user-controlled API key billed from OpenRouter credits; on headless/SSH hosts paste the redirect URL or authorization code into the prompt (Pi 0.83.0). `OPENROUTER_API_KEY` still works via **Use an API key**
 - xAI (Grok/X subscription, Pi 0.80.8) — `/login xai` then **Use a subscription** (device-code OAuth); `XAI_API_KEY` remains available via **Use an API key**
@@ -43,6 +44,7 @@ Since Pi 0.80.8, built-in catalogs are complemented by dynamic ones: `/model` re
 | OpenCode Go | `OPENCODE_API_KEY` | `opencode-go` |
 | Amazon Bedrock | `AWS_BEARER_TOKEN_BEDROCK` | `amazon-bedrock` |
 | Radius | `RADIUS_API_KEY` | `radius` |
+| TypeSafe (classifier models only) | `TYPESAFE_API_KEY` | `typesafe` |
 | Hugging Face | `HF_TOKEN` | `huggingface` |
 | Fireworks | `FIREWORKS_API_KEY` | `fireworks` |
 | Together AI | `TOGETHER_API_KEY` | `together` |
@@ -86,10 +88,12 @@ Entries may include an `env` object (Pi 0.79.5) for provider-scoped environment
 
 ## Credential Resolution Order
 
-1. `--api-key`
+1. `--api-key` (runtime override)
 2. `auth.json`
-3. Environment variables
-4. Custom provider keys from `models.json`
+3. `apiKey` from `models.json`
+4. Environment variables or ambient cloud credentials
+
+Provider extensions can define their own authentication behavior.
 
 ## Cloud Providers
 
@@ -151,12 +155,11 @@ Use `models.json` for OpenAI-compatible, Anthropic-compatible, Google-compatible
 
 ```json
 {
-  "providers": [
-    {
-      "name": "ollama",
+  "providers": {
+    "ollama": {
       "baseUrl": "http://localhost:11434/v1",
       "api": "openai-completions",
-      "apiKey": "unused",
+      "apiKey": "ollama",
       "models": [
         {
           "id": "llama3.2",
@@ -175,10 +178,12 @@ Use `models.json` for OpenAI-compatible, Anthropic-compatible, Google-compatible
         }
       ]
     }
-  ]
+  }
 }
 ```
 
+`providers` is an object keyed by provider ID, not an array; the optional `name` field is only a display label. A `models` entry adds or replaces a model with the same ID on that provider; use `modelOverrides` to change metadata of an existing built-in or extension model.
+
 `apiKey` is optional (Pi 0.80.0): omit it when auth comes from `/login`, `auth.json`, or CLI `--api-key`. Like `auth.json` keys, `apiKey` and `headers` values support `!command` execution and `$ENV_VAR` interpolation; plain uppercase strings are literals. Keyless local servers (e.g. Ollama) should keep a dummy value so their models appear in `/model`.
 
 Common `api` values:
@@ -206,7 +211,7 @@ Common `api` values:
 - `compat.thinkingFormat` supports OpenAI-compatible reasoning variants: `openrouter` sends `reasoning: { effort }`, `together` sends `reasoning: { enabled }` plus `reasoning_effort` when supported, `qwen-chat-template` targets local Qwen-compatible servers that read `chat_template_kwargs.enable_thinking`, `chat-template` (Pi 0.79.9) sends configurable `chat_template_kwargs` via `compat.chatTemplateKwargs` — e.g. `{ "thinking": { "$var": "thinking.enabled" } }` for DeepSeek models behind vLLM/Hugging Face chat templates — and `baseten` (Pi 0.84.0) sends `chat_template_args` via `compat.chatTemplateArgs`.
 - **Breaking (Pi 0.80.7):** `compat.sendSessionIdHeader` was removed. Session affinity is now controlled by `compat.sessionAffinityFormat` (`"openai"` sends `session_id`/`x-client-request-id`, `"openai-nosession"` omits the underscore-containing `session_id` header, `"openrouter"` sends `x-session-id`; default auto-detected). Replace `sendSessionIdHeader: false` with `sessionAffinityFormat: "openai-nosession"`. `sendSessionAffinityHeaders` still gates the behavior for `openai-completions`.
 - `compat.deferredToolsMode: "kimi"` (Pi 0.80.9) enables Kimi's deferred tool serialization; `compat.supportsToolReferences` / `compat.supportsToolSearch` enable native dynamic tool loading on verified custom endpoints, and `compat.supportsStrictTools` / `compat.supportsOpenAIGrammarTools` / per-model `constrainedSampling` gate constrained sampling (see `references/extensions.md`). `compat.supportsFinishReason: false` (Pi 0.84.0) tells pi to infer stop/toolUse for OpenAI-compatible streams that omit `finish_reason`.
-- Advanced `compat` flags exist for proxy quirks (`cacheControlFormat` — now also covering tool-result text content, `supportsReasoningEffort`, `supportsLongCacheRetention`, `supportsEagerToolInputStreaming`, `supportsStrictMode`). See [Pi `docs/models.md`](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/models.md) when a proxy rejects pi's defaults.
+- Advanced `compat` flags exist for proxy quirks (`cacheControlFormat` — now also covering tool-result text content, `supportsReasoningEffort`, `supportsLongCacheRetention`, `supportsEagerToolInputStreaming`, `supportsStrictMode`). See [Pi `docs/models.md`](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/models.md) when a proxy rejects pi's defaults.
 
 ### Context Overflow Recovery
 
@@ -214,12 +219,96 @@ Pi can auto-compact and retry when a provider fails with a recognized context-wi
 
 ### llama.cpp
 
-Pi supports the llama.cpp router server (Pi 0.81.0): configure with `/login llama.cpp`, search/download Hugging Face models and load/unload with live progress via `/llama`, then select loaded models in `/model`. The model catalog persists across restarts (fixed in Pi 0.82.1). See [Pi `docs/llama-cpp.md`](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/llama-cpp.md).
+Pi supports the llama.cpp router server (Pi 0.81.0): configure with `/login llama.cpp`, search/download Hugging Face models and load/unload with live progress via `/llama`, then select loaded models in `/model`. The model catalog persists across restarts (fixed in Pi 0.82.1). Every llama.cpp chat model is also listed as a classifier model (see [Classifier Models](#classifier-models)). `"extensions": ["-builtin:llama.cpp"]` removes the provider and `/llama`. See [Pi `docs/llama-cpp.md`](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/llama-cpp.md).
+
+## Model Types
+
+Since Pi 0.99.0 catalogs hold typed models: `type` is `"chat"` (the default when omitted), `"image"`, or `"classifier"`. Chat-facing reads (`getModels()`, `getAvailable()`, `ctx.modelRegistry.getAll()`, `/model`) stay chat-only. `ModelRuntime`, `ctx.modelRegistry`, and pi-ai `Models` all have `getModelsOfType(type, provider?)`, `getModelOfType(type, provider, id)`, and async `getAvailableOfType(type, provider?)`. Only `ModelRuntime` and `Models` have `getAllModels()` and `getAllAvailable()` for every type; `ctx.modelRegistry` adds `findOfType(type, provider, id)` instead. pi-ai exports `isModelType()` and `getModelType()` for mixed lists. One upstream ID may have separate chat and image entries. Sources: [models.md](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/models.md#use-classifier-models), [pi-ai types](https://github.com/earendil-works/pi/blob/v0.99.1/packages/ai/src/types.ts).
+
+### Classifier Models
+
+Classifiers do not chat. They answer typed questions about JSON state with probabilities, and do not appear in `/model`.
+
+| Provider | Model IDs | Auth |
+|----------|-----------|------|
+| `typesafe` | `jev-latest` | `TYPESAFE_API_KEY` |
+| `openrouter` | `typesafe/jev-1.13`, `~typesafe/jev-latest` | `OPENROUTER_API_KEY` or `/login` |
+| `cloudflare-workers-ai` | `typesafe/jev` | `CLOUDFLARE_API_KEY` + `CLOUDFLARE_ACCOUNT_ID` |
+| `vercel-ai-gateway` | `typesafe-ai/jev` | `AI_GATEWAY_API_KEY` |
+| `opencode` | `jev-1.13`, `jev-1.13-free` | `OPENCODE_API_KEY` |
+| llama.cpp router | every loaded chat model ID | `/login llama.cpp` |
+
+```typescript
+const jev = ctx.modelRegistry.findOfType("classifier", "typesafe", "jev-latest");
+if (jev) {
+  const result = await ctx.modelRegistry.classify(
+    jev,
+    {
+      state: { message: "The change works, thanks." },
+      questions: {
+        approved: {
+          type: "bool",
+          instructions: "Does the user approve of the result?",
+          criteria: { true: "Approval", false: "No approval" },
+        },
+      },
+    },
+    { signal: ctx.signal },
+  );
+  const answer = result.stopReason === "stop" ? result.answers.approved : undefined;
+  if (answer?.type === "bool") console.log(answer.probability);
+}
+```
+
+- Question types: `choice` (`criteria` maps labels to descriptions; answer has `choice`, `probabilities`, `confidence`), `bool` (`criteria: { true, false }`; answer has `probability`), and `score` (`criteria` is an ordered list; answer has `score`, `confidence`). Pi translates `bool` to TypeSafe's wire format.
+- `classify()` never rejects. Failures and aborts return `stopReason: "error"` or `"aborted"` with `errorMessage`. Check `stopReason` and each answer's `type` before use, and validate probabilities yourself if a decision depends on them (for example finite and within `[0, 1]`).
+- `result.usage` carries token counts and catalog cost when the service reports them; TypeSafe's direct `jev-latest` has no catalog price.
+- `jev-latest` moves with the service. Pin a versioned ID such as `openrouter`'s `typesafe/jev-1.13` when thresholds were tuned against one model.
+- Codemode scripts reach classifiers through `models.classify()` (see `references/codemode.md`). The SDK equivalent is `modelRuntime.classify()`. A llama.cpp classifier accepts a per-request `temperature` to soften overconfident label probabilities.
+
+A gate that fails closed on classifier errors is in `examples/classifier-gate.ts`.
+
+### Image Generation
+
+`ModelRuntime.generateImages(model, { input }, options?)` (Pi 0.99.0) generates images with runtime-resolved auth; list image models with `getModelsOfType("image")`. OpenRouter image models appear under the `openrouter` provider with its credential. Like `classify()`, it returns an error result rather than rejecting.
+
+In pi-ai, image models live in the same catalog as chat models: `builtinModels()` includes them, `models.getModelOfType("image", provider, id)` finds one (type `ImageModel`), `models.generateImages()` runs it, and `createProvider({ models, images })` registers image implementations. The separate `*Images*` API was removed in pi-ai 0.99.0.
+
+## Virtual Models
+
+A virtual model (experimental, Pi 0.99.0) is a selectable entry that routes each request to a physical model. It is for routing by task, cost, or conversation state, not for permission decisions. Source: [virtual-models.md](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/virtual-models.md), [`jev-router.ts`](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/examples/extensions/jev-router.ts).
+
+```typescript
+pi.registerVirtualModel({
+  provider: "router",
+  id: "auto",
+  name: "Auto",
+  thinkingLevels: ["low", "high"],
+  route(request, ctx) {
+    const sticky = request.failed ?? request.previous;
+    if (request.reason !== "user" && sticky) {
+      return { model: sticky.model, thinkingLevel: sticky.thinkingLevel ?? "medium" };
+    }
+    const id = request.thinkingLevel === "high" ? "claude-sonnet-4-5" : "claude-haiku-4-5";
+    const model = ctx.modelRegistry.find("anthropic", id);
+    if (!model) throw new Error(`anthropic/${id} is not in the catalog`);
+    return { model, thinkingLevel: "medium" };
+  },
+});
+```
+
+- `route()` runs before every request with `reason` (`user`, `continuation`, `retry`, `direct`), `previous`, `failed`, `state`, `messages`, and `signal`. Returning `previous` for continuations and `failed` for retries keeps prompt caches and thinking signatures valid.
+- Returned `state` (JSON) is stored on the session branch and comes back as `request.state`; `direct` requests (compaction summaries, `ctx.modelRegistry.streamSimple()`) have none.
+- The route must be a physical model with credentials; throwing or returning a virtual model ends the request with an error.
+- Selection (`ctx.model`, `model_change`) names the virtual model; each assistant message records the physical `provider`/`model`/`thinkingLevel`. The footer shows both; `/session` lists cost per physical model.
+- `pi.unregisterVirtualModel(provider, id)` removes one (`unregisterProvider()` does not). SDK code can call `modelRuntime.registerVirtualModel(definition)`.
 
 ## Custom Providers via Extensions
 
 Use extensions when you need custom streaming, custom headers, OAuth/device-flow logic, or want to override an existing provider's `baseUrl`/`headers`:
 
+A config-form `models` list replaces the provider's models across chat, image, and classifier operations; image and classifier entries need an explicit `type` and implementations keyed by `api` through `images` and `classifiers`. Registering only `baseUrl` or `headers` keeps the built-in models. See [custom-provider.md](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/custom-provider.md).
+
 ```typescript
 // Full registration
 pi.registerProvider("my-provider", {

package/skills/pi/references/sdk.md

@@ -30,10 +30,15 @@ session.subscribe((event) => {
   }
 });
 
-await session.prompt("What files are in the current directory?");
+try {
+  await session.prompt("What files are in the current directory?");
+  console.log(session.getLastAssistantText());
+} finally {
+  session.dispose();
+}
 ```
 
-Run with `npx tsx my-script.ts`.
+Run with `npx tsx my-script.ts` (or Node's built-in type stripping). `prompt()` resolves when the run finishes, including automatic retries. `session.dispose()` aborts active work, invalidates extension contexts, and removes listeners. Checked against Pi v0.99.1: [sdk.md](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/sdk.md), [SDK examples](https://github.com/earendil-works/pi/tree/v0.99.1/packages/coding-agent/examples/sdk).
 
 **Breaking (Pi 0.80.8):** `ModelRuntime` replaced the old `AuthStorage`/`ModelRegistry` pair as the SDK model/auth facade. `CreateAgentSessionOptions.authStorage` and `modelRegistry` are gone — pass the async `modelRuntime` instead. `AuthStorage` is no longer exported; use `ModelRuntime` (or a custom pi-ai `CredentialStore`), or `readStoredCredential()` for one-off reads of `auth.json`. `ModelRegistry` still exists only as the synchronous extension-facing compatibility facade, and its `refresh()` is now `Promise<void>`.
 
@@ -44,6 +49,7 @@ import {
   createAgentSession,
   DefaultResourceLoader,
   defineTool,
+  getAgentDir,
   ModelRuntime,
   SessionManager,
 } from "@earendil-works/pi-coding-agent";
@@ -56,11 +62,14 @@ import { getBuiltinModel } from "@earendil-works/pi-ai/providers/all";
 
 ```typescript
 const cwd = "/path/to/project";
+const agentDir = getAgentDir();
 const modelRuntime = await ModelRuntime.create();
+const resourceLoader = new DefaultResourceLoader({ cwd, agentDir });
+await resourceLoader.reload();   // a loader you pass in is not reloaded for you
 
 const { session } = await createAgentSession({
   cwd,
-  agentDir: "~/.pi/agent",
+  agentDir,
   modelRuntime,
   model: getBuiltinModel("anthropic", "claude-opus-4-8"),
   thinkingLevel: "medium",  // off | minimal | low | medium | high | xhigh | max
@@ -69,7 +78,7 @@ const { session } = await createAgentSession({
   ],
   tools: ["read", "bash", "edit", "write"],
   customTools: [/* defineTool(...) entries */],
-  resourceLoader: new DefaultResourceLoader(),
+  resourceLoader,
   sessionManager: SessionManager.inMemory(),
 });
 ```
@@ -80,12 +89,61 @@ Notes:
 - Pi 0.68.0 changed the SDK `tools` option from `Tool[]` to a `string[]` allowlist of built-in, extension, and custom tool names. Use `noTools: "builtin"` to disable built-ins while keeping extension/custom tools enabled, or `noTools: "all"` for none.
 - `customTools` accepts `ToolDefinition[]`. Build them with `defineTool({...})` for full TypeScript inference.
 - The `create*Tool(cwd)` factories still exist for code that needs explicit `AgentTool` instances (e.g. when wiring tools into pi-agent-core directly), but they are no longer the value passed to `createAgentSession({ tools })`.
-- `DefaultResourceLoader` loads extensions, skills, prompt templates, themes, and context files. Replace it to drive resource discovery from custom sources (and it must implement `loadProjectContextFiles()` if you want `AGENTS.md`/`CLAUDE.md` discovery; that helper is also exported standalone).
+- `DefaultResourceLoader` requires `{ cwd, agentDir }` and loads extensions, skills, prompt templates, themes, and context files. Replace it to drive resource discovery from custom sources (and it must implement `loadProjectContextFiles()` if you want `AGENTS.md`/`CLAUDE.md` discovery; that helper is also exported standalone).
 - **Breaking (Pi 0.87.0):** pi-agent-core's `shouldStopAfterTurn` option was removed (it was never a `createAgentSession()` option). On a low-level `Agent`, set `finishTurn` instead: it runs after the assistant message and tool results are finalized but before `turn_end`, also for error/aborted responses. Return `{ action: "end" }` to stop, `{ action: "continue" }` to ensure one more request, or `undefined` for normal scheduling — guard hard exits first: `if (turn.message.stopReason === "error" || turn.message.stopReason === "aborted") return;`. Inside `createAgentSession()`, use an extension's actionable `turn_end` / `agent_before_settle` boundaries (see `references/extensions.md`).
 
+## Built-in Extensions: Codemode, Tool Search, MCP
+
+The CLI loads `codemode`, `tool_search`, MCP, and llama.cpp as built-in extensions. SDK sessions load none of them. Opt in through the resource loader (Pi 0.99.0, [example](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/examples/sdk/14-codemode-mcp.ts)):
+
+```typescript
+import {
+  createAgentSession,
+  createCodemodeExtension,
+  createMcpExtension,
+  createToolSearchExtension,
+  DefaultResourceLoader,
+  getAgentDir,
+  SessionManager,
+  SettingsManager,
+} from "@earendil-works/pi-coding-agent";
+
+const cwd = process.cwd();
+const resourceLoader = new DefaultResourceLoader({
+  cwd,
+  agentDir: getAgentDir(),
+  extensionFactories: [
+    createCodemodeExtension({ mode: "on" }),   // options override codemode.mode / inlineBudget
+    createToolSearchExtension(),
+    createMcpExtension(),                        // reads mcp.json like the CLI (project file only when trusted)
+  ],
+});
+await resourceLoader.reload();
+
+const settingsManager = SettingsManager.create(cwd);
+settingsManager.applyOverrides({ defaultTools: ["+codemode", "+tool_search"] });
+
+const { session } = await createAgentSession({
+  resourceLoader,
+  settingsManager,
+  sessionManager: SessionManager.inMemory(),
+});
+
+try {
+  await session.bindExtensions({});   // emits session_start; MCP connects here, in the background
+  await session.prompt("Use codemode to count the TypeScript files in src/.");
+} finally {
+  session.dispose();
+}
+```
+
+- `codemode` and `tool_search` are registered inactive. Enable them with `defaultTools` (`+name` keeps the other defaults), or let MCP activate them for servers that need them. Passing `tools: [...]` instead restricts the session to the named tools, which hides MCP tools.
+- Extension `session_start` handlers run only when the host calls `session.bindExtensions()`. Without it, MCP servers never connect.
+- Inline extensions can be named `{ name, factory, replaceable?: true, builtin?: true }`. `replaceable` drops the extension when another one registers the same tool, command, or flag name. `builtin` supplies the code for `builtin:<name>`: it loads by default, is listed in `pi config`, is disabled by `-builtin:<name>` or `noExtensions`, and loads after project trust (so it cannot handle `project_trust`).
+
 ## Models & Auth (`ModelRuntime`)
 
-`ModelRuntime` (Pi 0.80.8) implements the pi-ai `Models` interface and owns credential storage. Auth resolution priority: runtime overrides (`setRuntimeApiKey`, not persisted) → stored credentials in `auth.json` → environment variables → fallback resolver for `models.json` keys.
+`ModelRuntime` (Pi 0.80.8) implements the pi-ai `Models` interface and owns credential storage. Auth resolution priority: runtime overrides (`setRuntimeApiKey`, not persisted) → stored credentials in `auth.json` → `apiKey` from `models.json` or an extension registration → the provider's environment variables or ambient cloud credentials.
 
 ```typescript
 import { InMemoryCredentialStore } from "@earendil-works/pi-ai";
@@ -115,6 +173,8 @@ const inMemory = await ModelRuntime.create({ credentials: new InMemoryCredential
 
 Pi 0.84.x made catalog refresh cancellation-aware: `refresh(options?)` accepts `{ providers?, allowNetwork?, force?, signal? }` and returns a `ModelsRefreshResult` (`{ errors, aborted }`) instead of discarding provider errors; `create({ signal })`, `login()`, `logout()`, and the now-async `setRuntimeApiKey()/removeRuntimeApiKey()` accept optional abort signals too (unbounded when omitted — applications own deadline policy). Credential mutations resolve once local state is consistent, without waiting for remote freshness; if that synchronization fails they reject with the exported `CredentialSynchronizationError` (inspect `providerId`, `operation`, `credential`, `cause` rather than retrying blindly). A failed refresh never undoes a committed credential change, and each `refresh()` starts a fresh generation so stale refreshes cannot publish afterward.
 
+Non-chat models (Pi 0.99.0): `getModelsOfType(type, provider?)`, `getModelOfType(type, provider, id)`, `getAvailableOfType()`, `getAllModels()`, and `getAllAvailable()` cover `"chat"`, `"image"`, and `"classifier"` entries; `getModels()` stays chat-only. `modelRuntime.classify(model, context, options?)` and `modelRuntime.generateImages(model, context, options?)` resolve auth at request time and return error results (`stopReason: "error" | "aborted"`) instead of rejecting. `modelRuntime.registerVirtualModel(definition)` registers a routing model without an extension. See `references/providers.md#model-types`.
+
 To match CLI model parsing, use the exported resolver helpers (Pi 0.80.4):
 
 ```typescript
@@ -138,11 +198,11 @@ await session.prompt("Review the current directory");
 await session.prompt("Stop and do this instead", { streamingBehavior: "steer" });
 await session.prompt("After you finish, also check tests", { streamingBehavior: "followUp" });
 
-await session.steer("Use a smaller diff");
+const disposition = await session.steer("Use a smaller diff");   // "queued" | "handled"
 await session.followUp("Summarize the changes afterward");
 ```
 
-`prompt()` expands file-based prompt templates. During active streaming, calling it without `streamingBehavior` throws.
+`prompt()` expands file-based prompt templates and resolves after an accepted run finishes. During active streaming, calling it without `streamingBehavior` throws. Since Pi 0.99.0 `steer()` and `followUp()` return a disposition: `"queued"` (also after an extension transformed the input) or `"handled"` (an extension consumed it, so nothing was queued). Both throw for extension commands (`/name`); send those through `prompt()`, which runs them immediately, even while streaming. Read state through `session.messages`, `session.model`, `session.thinkingLevel`, `session.systemPrompt`, and `session.getActiveToolNames()`.
 
 ## Session Runtime
 
@@ -230,7 +290,7 @@ session.subscribe((event) => {
         process.stdout.write(event.assistantMessageEvent.delta);
       }
       break;
-    case "tool_execution_start":
+    case "tool_execution_start":   // nested calls (ctx.executeTool, codemode) carry parentToolCallId
     case "tool_execution_update":
     case "tool_execution_end":
     case "agent_start":
@@ -261,8 +321,9 @@ Pi 0.84.0 added transport-neutral remote-session client APIs: `@earendil-works/p
 
 ## RPC Mode
 
-`pi --mode rpc` speaks newline-delimited JSON over stdio. Additions since Pi 0.79:
+`pi --mode rpc` speaks newline-delimited JSON over stdio ([rpc.md](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/rpc.md), [rpc-commands.md](https://github.com/earendil-works/pi/blob/v0.99.1/packages/coding-agent/docs/rpc-commands.md)). Additions since Pi 0.79:
 
+- Input dispositions (Pi 0.99.0): a successful `prompt` response carries `data.disposition` — `"started"` (a run started), `"queued"` (queued during a run), or `"handled"` (an extension command or input handler consumed it). `steer` and `follow_up` responses carry `"queued"` or `"handled"`. On `"handled"` no run started for that input: do not wait for `agent_settled`. Otherwise keep consuming events and wait for `agent_settled`, not `agent_end`. Failures after acceptance arrive as events, never as a second response. `RpcClient.prompt(message, images?, streamingBehavior?)`, `steer()`, and `followUp()` return the disposition.
 - `agent_settled` event (Pi 0.80.4) fires when a run is fully settled — no automatic retry, compaction retry, or queued continuation remains; `agent_end` now carries `willRetry`. `set_thinking_level` accepts `"max"` where the model supports it.
 - `clear_queue` (Pi 0.84.4) returns and removes queued steering and follow-up messages; direct RPC `steer`/`follow_up` commands pass through extension `input` handlers since Pi 0.86.0.
 - `get_entries` / `get_tree` (Pi 0.80.3) read session entries and tree snapshots over RPC; `get_available_thinking_levels` (Pi 0.81.0) lists the current model's supported thinking levels.
@@ -285,13 +346,18 @@ Pi 0.84.0 added transport-neutral remote-session client APIs: `@earendil-works/p
 To name an inline factory in the startup Extensions list (instead of `<inline:1>`), wrap it in an `InlineExtension` (Pi 0.80.4):
 
 ```typescript
-import type { InlineExtension } from "@earendil-works/pi-coding-agent";
+import { DefaultResourceLoader, getAgentDir, type InlineExtension } from "@earendil-works/pi-coding-agent";
 
 const myProvider: InlineExtension = {
   name: "my-provider",
   factory: (pi) => { /* ... */ },
 };
-const loader = new DefaultResourceLoader({ extensionFactories: [myProvider] });
+const loader = new DefaultResourceLoader({
+  cwd: process.cwd(),
+  agentDir: getAgentDir(),
+  extensionFactories: [myProvider],
+});
+await loader.reload();
 ```
 
 Bare factory functions are still accepted.