@selesai/code 0.13.26 → 0.13.28

package/CHANGELOG.md

@@ -2,6 +2,109 @@
 
 All notable changes to `@selesai/code` will be documented in this file.
 
+## [0.13.28] - 2026-09-22
+
+### Automatic model routing
+
+A new bundled `auto-model` extension routes each idle, top-level prompt to one of four configured
+tiers — `simple`, `medium`, `complex`, and `reasoning` — by asking the Jev decisions model
+(`jev-1.13`) through the Token-In gateway which tier fits the request.
+
+- **Four-tier routing.** `settings.json` gains an `autoModel` block: a classifier (`provider` and
+  `model`, default `tokenin` / `jev-1.13`) plus one model per tier. The extension builds Jev's
+  `{state, questions}` decisions payload from the current ask, a bounded window of prior user turns,
+  and the active system prompt, reads the answered tier back, and calls `pi.setModel()` on the
+  matching model before the turn starts. Disabled by default.
+- **Tier model syntax.** Each tier value is `provider/modelId` with an optional `:thinkingLevel`
+  suffix (`tokenin/celestial-max:max`) applied through `pi.setThinkingLevel()`.
+- **Safe scope.** Only idle, top-level prompts are routed. Queued steering/follow-up input and
+  extension-injected messages are skipped because `pi.setModel()` is session-global, and `/`
+  commands are left untouched. A manual `/model` choice suspends routing for the rest of the session.
+- **Graceful failures.** A classifier timeout, error, malformed or low-confidence answer, missing or
+  out-of-scope model, or unavailable credentials keeps the current model or falls back to
+  `autoModel.fallbackTier` (default `medium`).
+
+### Graft
+
+- **Deep-build model default for existing installs.** `bootstrapAgentDir` now seeds `graft.deepModel`
+  (`deepseek-v4.1-flash`) into an existing user `settings.json` when the key is absent. It merges
+  into a present `graft` object, keeps the original formatting when it inserts a new section, and
+  never overwrites a user-chosen model.
+
+## [0.13.27] - 2026-09-21
+
+### Upstream sync: Pi v0.86.1
+
+Ported the coding-agent source from Pi v0.85.1 to v0.86.1 per file, and bumped
+`@earendil-works/pi-ai`, `pi-agent-core` and `pi-tui` to 0.86.1. The package now declares
+`engines.node >= 22.19.0`, matching upstream's floor.
+
+**Adopted from upstream**
+
+- **Custom-provider request contract.** Stream handlers now receive a normalized transcript:
+  the prompt and tool declarations live in its system messages and are read with the platform's
+  accessors, and handlers invoke the outgoing-payload hook before sending and the response hook
+  before consuming the body. Selesai's own providers never read the request context, so this
+  landed as a signature-and-type change at the forwarding boundary rather than a rewrite — your
+  `models.json` configuration and extension-registered providers are unaffected.
+- **Sectioned system prompt with a transcript-backed record.** A session now records the prompt
+  and tool changes it ran under, so resuming a session or navigating to an earlier branch
+  restores the instructions and tools that were actually in effect instead of applying today's.
+  Your prompt cache stays warm across resume and branch navigation.
+- **Whole-prompt replacement for one turn.** An extension can now replace the instruction for a
+  turn rather than only appending to it, so a preset or persona can take effect.
+- **Per-model compaction budgets.** The reserved and kept-recent token budgets can be set per
+  model, so a large-context model is no longer compacted with a small model's budget.
+- **Session runtime rebase.** Prompt and tool loadout are prepared as an explicit step with the
+  resulting state retained for the run; one shared abort controller per operation expresses
+  cancellation by throwing, so a cancellation arriving while compaction is still obtaining
+  credentials is no longer ignored; provider request settings (timeouts, retries, headers) come
+  from one builder and apply uniformly to a main turn, a compaction request and an extension's
+  model call; and the agent-level retry-backoff cap is configurable, so a transient provider
+  outage cannot hold you for minutes.
+- **Strict JSON-schema sampling by default.** Built-in tools prefer strict sampling without you
+  setting an environment variable, and a tool whose schema cannot be expressed strictly can opt out.
+- **Wider default tool set.** The default active set includes the search and listing tools, so a
+  fresh session can explore a repository without enabling tools by hand; each is still
+  individually disableable.
+- **Extension authoring.** Event subscriptions now return an unsubscribe handle, so a reload or
+  session replacement cannot leave a stale handler behind; tool-call arguments and tool-result
+  details are restricted to JSON-compatible values so they survive session persistence; shell
+  hooks now fail loudly instead of silently when a handler throws or returns something malformed;
+  and the embedded-module map moved into its own module that branches on runtime kind (compiled
+  binary, TypeScript source, unbundled Node).
+- **Interactive workspace.** Deferred extension-catalogue refresh, the new settings surfaces, and
+  the crash notice and bug report.
+
+**Fork fixes carried in this release**
+
+- **Clipboard failures are now surfaced.** A failed paste reports a warning instead of being
+  swallowed.
+- **Overflow recovery no longer loses its guard.** A response that stopped at the output-token
+  limit no longer clears the overflow-recovery attempt flag, so recovery is not re-armed into a
+  loop.
+- **Upstream path leak in a doc comment.** The session-manager documentation named
+  `~/.pi/agent/sessions`; runtime resolution was already correct, and the comment now names
+  `.selesai/agent/sessions`.
+
+**Dependency removed**
+
+- **`@mariozechner/clipboard` is gone.** The optional native clipboard dependency was retired in
+  favour of the clipboard helpers bundled with `@earendil-works/pi-tui`, so an install is smaller
+  and has one fewer native supply-chain boundary. Paste, copy-on-select and image paste are
+  unchanged.
+
+**Deliberately not adopted**
+
+- **Mermaid rendering.** Upstream's optional Mermaid path is not adopted, so no rendering
+  dependency or settings item was added.
+- **Fork behaviour kept.** Automatic handoff, the bounded length-continuation of an
+  output-limit-truncated response, the vision caption relay, the compaction-failure event and the
+  capability gateway all still behave as before, re-planted on the new upstream shapes.
+
+**Verification.** 53 test files / 1008 tests pass, the tree typechecks against Pi 0.86.1, the build
+succeeds, and a 37-row delta inventory proves each Selesai-owned behaviour survived the port.
+
 ## [0.13.26] - 2026-09-18
 
 ### Added

package/dist/cli/args.js

@@ -395,6 +395,7 @@ ${chalk.bold("Environment Variables:")}
   MOONSHOT_API_KEY                 - Moonshot AI API key
   OPENCODE_API_KEY                 - OpenCode Zen/OpenCode Go API key
   KIMI_API_KEY                     - Kimi For Coding API key
+  META_API_KEY                     - Meta Model API key
   CLOUDFLARE_API_KEY               - Cloudflare API token (Workers AI and AI Gateway)
   CLOUDFLARE_ACCOUNT_ID            - Cloudflare account id (required for both)
   CLOUDFLARE_GATEWAY_ID            - Cloudflare AI Gateway slug (required for AI Gateway)

package/dist/cli/session-picker.d.ts

@@ -3,7 +3,7 @@
  */
 import type { SessionInfo, SessionListProgress } from "../core/session-manager.ts";
 import type { SettingsManager } from "../core/settings-manager.ts";
-type SessionsLoader = (onProgress?: SessionListProgress) => Promise<SessionInfo[]>;
+type SessionsLoader = (onProgress?: SessionListProgress, signal?: AbortSignal) => Promise<SessionInfo[]>;
 /** Show TUI session selector and return selected session path or null if cancelled */
 export declare function selectSession(currentSessionsLoader: SessionsLoader, allSessionsLoader: SessionsLoader, settingsManager: SettingsManager): Promise<string | null>;
 export {};

package/dist/config.d.ts

@@ -147,6 +147,15 @@ export declare function seedDefaultConfigFile(destPath: string, bundledName: str
  * file that already configures `subagents`, is left untouched.
  */
 export declare function seedMissingSubagentSettings(destPath: string, bundledDefaultsDir?: string): boolean;
+/**
+ * Add the bundled `graft.deepModel` default to an existing user settings file.
+ *
+ * The `graft` namespace may already exist without that leaf (graft is configured
+ * field-by-field), so a present `graft` object is merged rather than skipped. A
+ * user-provided `deepModel` is never overwritten. Returns true when the file was
+ * rewritten.
+ */
+export declare function seedMissingGraftSettings(destPath: string, bundledDefaultsDir?: string): boolean;
 /**
  * Bundled extensions load directly from the installed package. Do not copy them
  * into the user's agent dir, otherwise startup discovers both copies and tool

package/dist/config.js

@@ -281,7 +281,7 @@ export function getSelfUpdateUnavailableInstruction(packageName, npmCommand, upd
     const method = detectInstallMethod();
     const target = normalizeSelfUpdatePackageTarget(updatePackageTarget);
     if (method === "bun-binary") {
-        return `Download from: https://github.com/earendil-works/pi-mono/releases/latest`;
+        return `Download from: https://github.com/earendil-works/pi/releases/latest`;
     }
     const command = getSelfUpdateCommandForMethod(method, packageName, target, npmCommand);
     if (command) {
@@ -676,6 +676,51 @@ export function seedMissingSubagentSettings(destPath, bundledDefaultsDir = getBu
         return false;
     }
 }
+/**
+ * Add the bundled `graft.deepModel` default to an existing user settings file.
+ *
+ * The `graft` namespace may already exist without that leaf (graft is configured
+ * field-by-field), so a present `graft` object is merged rather than skipped. A
+ * user-provided `deepModel` is never overwritten. Returns true when the file was
+ * rewritten.
+ */
+export function seedMissingGraftSettings(destPath, bundledDefaultsDir = getBundledDefaultsDir()) {
+    if (!existsSync(destPath))
+        return false;
+    try {
+        const raw = readFileSync(destPath, "utf-8");
+        const settings = JSON.parse(raw);
+        const defaults = JSON.parse(readFileSync(join(bundledDefaultsDir, "settings.json"), "utf-8"));
+        if (typeof settings !== "object" || settings === null || Array.isArray(settings))
+            return false;
+        if (typeof defaults !== "object" || defaults === null || Array.isArray(defaults))
+            return false;
+        const defaultGraft = defaults.graft;
+        const deepModel = typeof defaultGraft === "object" && defaultGraft !== null && !Array.isArray(defaultGraft)
+            ? defaultGraft.deepModel
+            : undefined;
+        if (typeof deepModel !== "string")
+            return false;
+        const graft = settings.graft;
+        if (graft === undefined) {
+            const injected = injectRootObjectKey(raw, "graft", defaultGraft);
+            if (injected === undefined)
+                return false;
+            writeFileSync(destPath, injected);
+            return true;
+        }
+        if (typeof graft !== "object" || graft === null || Array.isArray(graft))
+            return false;
+        if (Object.hasOwn(graft, "deepModel"))
+            return false;
+        graft.deepModel = deepModel;
+        writeFileSync(destPath, `${JSON.stringify(settings, null, 2)}\n`);
+        return true;
+    }
+    catch {
+        return false;
+    }
+}
 /**
  * Recursively copy a bundled directory tree into the user's agent dir,
  * overwriting any existing file so the bundled copy stays authoritative.
@@ -787,6 +832,7 @@ export function bootstrapAgentDir(agentDir = getAgentDir()) {
     const settingsPath = join(agentDir, "settings.json");
     seedDefaultConfigFile(settingsPath, "settings.json");
     seedMissingSubagentSettings(settingsPath);
+    seedMissingGraftSettings(settingsPath);
     seedDefaultExtensions(agentDir);
     seedDefaultSkills(agentDir);
     seedDefaultThemes(agentDir);

package/dist/core/agent-session-auto-handoff.test.js

@@ -2,6 +2,7 @@ import { describe, expect, it, vi } from "vitest";
 import { Type } from "typebox";
 import { AgentSession } from "./agent-session.js";
 import { AUTO_HANDOFF_GOAL } from "./handoff.js";
+import { SessionManager } from "./session-manager.js";
 import { SettingsManager } from "./settings-manager.js";
 function createMockSession({ enabled, threshold = 128_000, tokens, mode = "tui", command, customTools, allowedToolNames, }) {
     const settings = SettingsManager.inMemory({
@@ -45,7 +46,7 @@ function createMockSession({ enabled, threshold = 128_000, tokens, mode = "tui",
     };
     const session = new AgentSession({
         agent: agent,
-        sessionManager: { getBranch: () => [] },
+        sessionManager: SessionManager.inMemory(),
         settingsManager: settings,
         cwd: "/tmp",
         resourceLoader: resourceLoader,

package/dist/core/agent-session-length-continuation.test.js

@@ -1,5 +1,6 @@
 import { describe, expect, it, vi } from "vitest";
 import { AgentSession } from "./agent-session.js";
+import { SessionManager } from "./session-manager.js";
 import { SettingsManager } from "./settings-manager.js";
 function createMockSession() {
     const settings = SettingsManager.inMemory({});
@@ -43,7 +44,7 @@ function createMockSession() {
     };
     const session = new AgentSession({
         agent: agent,
-        sessionManager: { getBranch: () => [], appendMessage: vi.fn() },
+        sessionManager: SessionManager.inMemory(),
         settingsManager: settings,
         cwd: "/tmp",
         resourceLoader: resourceLoader,

package/dist/core/agent-session.d.ts

@@ -15,6 +15,7 @@
 import type { Agent, AgentEvent, AgentMessage, AgentState, AgentTool, ThinkingLevel } from "@earendil-works/pi-agent-core";
 import type { ImageContent, Model, TextContent } from "@earendil-works/pi-ai/compat";
 import { type BashResult } from "./bash-executor.ts";
+import type { CacheWarmer, CacheWarmingStatus } from "./cache-warmer.ts";
 import { type CompactionResult } from "./compaction/index.ts";
 import { type ContextUsage, type ExtensionCommandContextActions, type ExtensionErrorListener, type ExtensionMode, ExtensionRunner, type ExtensionUIContext, type InputSource, type ReplacedSessionContext, type SessionStartEvent, type ShutdownHandler, type ToolDefinition, type ToolInfo } from "./extensions/index.ts";
 import type { CustomMessage } from "./messages.ts";
@@ -22,7 +23,7 @@ import type { ModelRuntime } from "./model-runtime.ts";
 import { type PromptTemplate } from "./prompt-templates.ts";
 import type { ResourceLoader } from "./resource-loader.ts";
 import type { BranchSummaryEntry, SessionEntry, SessionManager } from "./session-manager.ts";
-import type { SettingsManager } from "./settings-manager.ts";
+import type { CacheWarmingMode, SettingsManager } from "./settings-manager.ts";
 import { type BashOperations } from "./tools/bash.ts";
 /** Parsed skill block from a user message */
 export interface ParsedSkillBlock {
@@ -131,7 +132,9 @@ export interface AgentSessionConfig {
     customTools?: ToolDefinition[];
     /** Canonical model/auth runtime used by coding-agent internals. */
     modelRuntime: ModelRuntime;
-    /** Initial active built-in tool names. Default: [read, bash, edit, write, grep, find, ls] */
+    /** Keeps the prompt cache entry of the last session request warm. */
+    cacheWarmer?: Pick<CacheWarmer, "cancel" | "status" | "onAgentSettled" | "onModeChanged" | "onWarmed">;
+    /** Initial active built-in tool names. Default: [read, bash, edit, write] */
     initialActiveToolNames?: string[];
     /** Optional allowlist of tool names. When provided, only these tool names are exposed. */
     allowedToolNames?: string[];
@@ -211,6 +214,7 @@ export declare class AgentSession {
     private _unsubscribeAgent?;
     private _eventListeners;
     private _isAgentRunActive;
+    private _agentRunAbortRequested;
     private _idleWaitPromise;
     private _resolveIdleWait;
     /** Tracks pending steering messages for UI display. Removed when delivered. */
@@ -250,13 +254,14 @@ export declare class AgentSession {
     private _extensionErrorListener?;
     private _extensionErrorUnsubscriber?;
     private _modelRuntime;
+    private _cacheWarmer?;
     private _toolRegistry;
     private _toolDefinitions;
     private _toolPromptSnippets;
     private _toolPromptGuidelines;
-    private _baseSystemPrompt;
     private _baseSystemPromptOptions;
-    private _systemPromptOverride?;
+    /** Prompt options after before_agent_start mutations for the active run. */
+    private _runSystemPromptOptions?;
     constructor(config: AgentSessionConfig);
     get modelRuntime(): ModelRuntime;
     private _getRequiredRequestAuth;
@@ -314,6 +319,10 @@ export declare class AgentSession {
     dispose(): void;
     /** Full agent state */
     get state(): AgentState;
+    /** Current cache-warming state and the policy inputs that produced it. */
+    get cacheWarmingStatus(): CacheWarmingStatus | undefined;
+    /** Persist the cache-warming mode and immediately reconcile active warming. */
+    setCacheWarmingMode(mode: CacheWarmingMode): void;
     /** Current model (may be undefined if not yet selected) */
     get model(): Model<any> | undefined;
     /** Current thinking level */
@@ -322,7 +331,7 @@ export declare class AgentSession {
     get isStreaming(): boolean;
     /** Whether the session has no active agent run, compaction, branch summary, retry, or queued continuation. */
     get isIdle(): boolean;
-    /** Current effective system prompt (includes any per-turn extension modifications) */
+    /** Current effective system prompt, including changes not yet sent to the model. */
     get systemPrompt(): string;
     /** Current retry attempt (0 if not retrying) */
     get retryAttempt(): number;
@@ -372,8 +381,33 @@ export declare class AgentSession {
     private _normalizePromptSnippet;
     private _normalizePromptGuidelines;
     private _rebuildSystemPrompt;
+    /**
+     * Apply a prompt and tool loadout for the next request. Sets the executable tools and
+     * returns a system message patching the prompt sections the model currently has (replayed
+     * from `messages`), or undefined when the prompt is unchanged. Tool changes are declared by
+     * the agent loop before the request.
+     *
+     * A forced prompt does not affect the transcript: the structured sections are still diffed
+     * and persisted, and the forced text is projected onto the request by
+     * {@link _installAgentForcedPromptProjection}.
+     */
+    private _preparePromptAndToolLoadout;
+    /**
+     * Send a forced prompt as the provider's leading system prompt without recording it.
+     *
+     * A `before_agent_start` handler that returns `systemPrompt` needs that exact text at the
+     * head of the request; a mid-conversation system message would leave the original prompt
+     * in place. The forced text is a rendering of the current prompt, so the transcript keeps
+     * its structured sections and the request is projected instead: the system messages
+     * collapse into one head holding the forced text and the current tools. Runs after the
+     * `context` extension handlers.
+     */
+    private _installAgentForcedPromptProjection;
+    /** Restore the active tool loadout declared by the session transcript, if it declares one. */
+    private _restoreToolsFromTranscript;
     private _runAgentPrompt;
     private _handlePostAgentRun;
+    private _runInputHandlers;
     /**
      * Send a prompt to the agent.
      * - Handles extension commands (registered via pi.registerCommand) immediately, even during streaming
@@ -391,23 +425,30 @@ export declare class AgentSession {
      * Emits errors via extension runner if file read fails.
      */
     private _expandSkillCommand;
+    private _queueUserInput;
     /**
      * Queue a steering message while the agent is running.
      * Delivered after the current assistant turn finishes executing its tool calls,
      * before the next LLM call.
      * Expands skill commands and prompt templates. Errors on extension commands.
      * @param images Optional image attachments to include with the message
+     * @param options Input source; defaults to interactive
      * @throws Error if text is an extension command
      */
-    steer(text: string, images?: ImageContent[]): Promise<void>;
+    steer(text: string, images?: ImageContent[], options?: {
+        source?: InputSource;
+    }): Promise<void>;
     /**
      * Queue a follow-up message to be processed after the agent finishes.
      * Delivered only when agent has no more tool calls or steering messages.
      * Expands skill commands and prompt templates. Errors on extension commands.
      * @param images Optional image attachments to include with the message
+     * @param options Input source; defaults to interactive
      * @throws Error if text is an extension command
      */
-    followUp(text: string, images?: ImageContent[]): Promise<void>;
+    followUp(text: string, images?: ImageContent[], options?: {
+        source?: InputSource;
+    }): Promise<void>;
     private _queueSteer;
     private _queueFollowUp;
     /**
@@ -540,6 +581,7 @@ export declare class AgentSession {
      * Saves to settings.
      */
     setFollowUpMode(mode: "all" | "one-at-a-time"): void;
+    private _runDefaultCompaction;
     private _clearManualCompactionState;
     /**
      * Manually compact the session context.
@@ -589,6 +631,7 @@ export declare class AgentSession {
      * the TUI needs to render the retry and recreate the underlying indicator.
      */
     private _summarizationRetryCallbacks;
+    private _finishCancelledRetry;
     private _prepareRetry;
     /**
      * Cancel in-progress retry.
@@ -691,6 +734,14 @@ export declare class AgentSession {
      * @returns The resolved output file path.
      */
     exportToJsonl(outputPath?: string): string;
+    /**
+     * Ask the current model to describe what went wrong in this session for a bug report.
+     * Used when the user declines to share the transcript itself.
+     */
+    summarizeForBugReport(options: {
+        hint?: string;
+        signal: AbortSignal;
+    }): Promise<string>;
     /**
      * Get text content of last assistant message.
      * Useful for /copy command.