@cjhyy/code-shell 0.1.6 → 0.2.0

package/dist/llm/capabilities/types.d.ts

@@ -0,0 +1,108 @@
+/**
+ * Per-(provider, model) capability descriptor.
+ *
+ * Captures the subset of request-shape divergence we've actually observed
+ * cause HTTP 400s or behavior changes. Each rule in `rules.ts` produces
+ * one of these; clients spread it into the outgoing request.
+ *
+ * Sources for every field are documented next to the matching rule in
+ * `rules.ts` — vendor docs, not folklore.
+ */
+import type { ProviderKindName } from "../provider-kinds.js";
+/** OpenAI-style reasoning effort levels — shared by several vendors. */
+export type ReasoningEffort = "minimal" | "low" | "medium" | "high";
+/** A binary thinking switch, like DeepSeek V4 and Z.AI GLM use. */
+export type ThinkingSwitch = "enabled" | "disabled";
+/**
+ * How the vendor exposes a "think harder / think less" knob (or none).
+ * Each kind has different field names and value shapes — we keep them
+ * separate so the engine can't accidentally cross-pollinate.
+ */
+export type ReasoningShape = {
+    kind: "none";
+}
+/** DeepSeek V4, Z.AI GLM-4.5+ — `{thinking: {type: "enabled"|"disabled"}}` at top level. */
+ | {
+    kind: "deepseek-thinking";
+}
+/**
+ * OpenAI o-series & gpt-5+ — `reasoning_effort: "minimal"|"low"|"medium"|"high"`.
+ * `disabledEffort` is the value we send when the caller asks for thinking
+ * "disabled" — defaults to `"minimal"` (OpenAI), but vendors with a
+ * narrower vocabulary override it (e.g. xAI Grok 4.3 has no `"minimal"`;
+ * Mistral Magistral only accepts `"high"` and `"none"`).
+ */
+ | {
+    kind: "openai-effort";
+    disabledEffort?: ReasoningEffort | "none";
+}
+/** Anthropic Claude 4.x ≤ 4.5 — `{thinking: {type: "enabled", budget_tokens: N}}`. */
+ | {
+    kind: "anthropic-budget";
+    minBudgetTokens: number;
+}
+/** Anthropic Claude 4.6+ — no opt-in, thinking is adaptive. Sending `type: "enabled"` 400s. */
+ | {
+    kind: "anthropic-adaptive";
+}
+/** OpenRouter normalized — `{reasoning: {effort, max_tokens, exclude, enabled}}`. */
+ | {
+    kind: "openrouter-reasoning";
+};
+/**
+ * How prior `reasoning_content` (or `thinking` blocks) must be threaded
+ * back into the next request.
+ */
+export type EchoReasoning = 
+/** DeepSeek `deepseek-reasoner`: 400 if you echo it back at all. */
+"never"
+/** DeepSeek V4 + tools, Claude 4.x + tools: must echo or 400. */
+ | "when-tools"
+/** Optional — echo if you have it, but absence is fine. */
+ | "optional";
+/**
+ * Parallel-tool-call shape. OpenAI exposes `parallel_tool_calls: bool`.
+ * Anthropic exposes it inside `tool_choice.disable_parallel_tool_use`.
+ * Some endpoints don't support it at all.
+ */
+export type ParallelToolCallsShape = "openai-flag" | "anthropic-disable-flag" | "unsupported";
+/**
+ * Streaming usage signal. OpenAI-compat needs `stream_options:
+ * {include_usage: true}`; Anthropic always emits usage in
+ * `message_delta`; some endpoints reject the field.
+ */
+export type StreamUsageShape = "include-usage-flag" | "auto" | "none";
+export interface Capability {
+    /** Token-limit request field. */
+    tokenLimitField: "max_tokens" | "max_completion_tokens";
+    /**
+     * Request fields the model 400s on (or silently drops). We just don't
+     * send them. Names are in OpenAI-compat spelling; clients translate
+     * for native protocols.
+     */
+    rejectedParams: ReadonlySet<"temperature" | "top_p" | "presence_penalty" | "frequency_penalty" | "logit_bias" | "logprobs" | "top_logprobs">;
+    /** How thinking is exposed (or not). */
+    reasoning: ReasoningShape;
+    /** Echo-back contract for prior reasoning. */
+    echoReasoning: EchoReasoning;
+    /** Parallel-tool-calls flag shape. */
+    parallelToolCalls: ParallelToolCallsShape;
+    /** Streaming usage reporting shape. */
+    streamUsage: StreamUsageShape;
+}
+/**
+ * Conservative default — used when no rule matches. Picks "send everything
+ * the vanilla OpenAI Chat Completions spec accepts," because that's what
+ * every OpenAI-compat endpoint at minimum understands.
+ */
+export declare const DEFAULT_CAPABILITY: Capability;
+export interface CapabilityRule {
+    /** Which provider kind this rule applies to. */
+    kind: ProviderKindName;
+    /** Model-id matcher. First matching rule wins (per kind). */
+    match: RegExp;
+    /** Patch applied on top of DEFAULT_CAPABILITY. */
+    capability: Partial<Capability>;
+    /** Free-text reason — shown in logs, helps future-you. */
+    why: string;
+}

package/dist/llm/model-pool.d.ts

@@ -29,6 +29,13 @@ export interface ModelEntry {
     /** Optional reference into ProviderCatalog. When set, baseUrl/apiKey
      *  come from the catalog unless the entry overrides them. */
     providerKey?: string;
+    /**
+     * Per-model thinking override. Wins over the provider-level setting
+     * (ProviderCatalog entry's `thinking`). Useful when models in the same
+     * provider need different defaults — e.g. DeepSeek V4 Pro off but
+     * V4 Flash on.
+     */
+    thinking?: "enabled" | "disabled";
 }
 export declare class ModelPool {
     private models;

package/dist/llm/provider-catalog.d.ts

@@ -13,6 +13,8 @@ export interface ProviderConfig {
     apiKey?: string;
     protocol?: "openai-compat" | "anthropic-style";
     modelsPath?: string;
+    /** Default DeepSeek V4 thinking-mode for this provider. See LLMConfig. */
+    thinking?: "enabled" | "disabled";
 }
 export declare class ProviderCatalog {
     private byKey;

package/dist/llm/provider-kinds.d.ts

@@ -6,7 +6,7 @@
  * (chat-completion only — embeddings / TTS / image generation are
  * filtered out so they don't show up in the model picker).
  */
-export type ProviderKindName = "openai" | "anthropic" | "deepseek" | "xai" | "mistral" | "groq" | "google" | "openrouter" | "ollama" | "custom";
+export type ProviderKindName = "openai" | "anthropic" | "deepseek" | "zai" | "xai" | "mistral" | "groq" | "google" | "openrouter" | "ollama" | "custom";
 export type ProviderProtocol = "openai-compat" | "anthropic-style" | "gemini" | "ollama";
 export interface ProviderKindMeta {
     label: string;

package/dist/llm/providers/openai.d.ts

@@ -1,15 +1,37 @@
 /**
  * OpenAI GPT provider using openai SDK.
+ *
+ * Serves every OpenAI-compatible endpoint we support (OpenAI, DeepSeek,
+ * OpenRouter, Z.AI, xAI, Mistral, Groq, Gemini-OpenAI-compat, Ollama,
+ * custom). Per-(provider, model) request-shape divergence — token-limit
+ * field name, rejected sampling params, thinking knob, reasoning echo —
+ * is resolved through `capabilitiesFor()`, not hardcoded.
  */
 import type { LLMConfig, LLMResponse } from "../../types.js";
 import type { CreateMessageOptions } from "../types.js";
 import { LLMClientBase } from "../client-base.js";
 export declare class OpenAIClient extends LLMClientBase {
     private _client;
+    private _forceMaxCompletionTokens;
     constructor(config: LLMConfig);
     protected initClient(): void;
     private get client();
+    /**
+     * Resolve the capability descriptor for the current (provider kind, model)
+     * pair. Cached for the lifetime of the client — model never changes
+     * mid-client. `providerKind` defaults to "openai" because legacy LLMConfig
+     * paths (arena, env-derived) don't always populate it; for those, the
+     * built-in rules table treats them as plain OpenAI.
+     */
+    private _capability;
+    private get capability();
     createMessage(options: CreateMessageOptions): Promise<LLMResponse>;
+    /**
+     * Build the request body honoring the model's capability descriptor.
+     * Centralized so both streaming and non-streaming paths agree on the
+     * exact shape.
+     */
+    private buildRequestBody;
     private nonStreamMessage;
     private streamMessage;
     private processChoice;

package/dist/llm/token-counter.d.ts

@@ -0,0 +1,17 @@
+/**
+ * Token counter — best-effort live token count for streaming output.
+ *
+ * Uses gpt-tokenizer's cl100k_base encoding. It's exact for GPT-4 family and
+ * close enough (within ~10%) for Claude/DeepSeek/Qwen for the purpose of a
+ * live counter in the UI. The authoritative number is still the provider's
+ * `usage` field, captured at end-of-stream.
+ *
+ * The encoder is lazy-initialized on first use so we don't pay the cost on
+ * startup for sessions that never stream.
+ */
+/**
+ * Count tokens in `text`. Synchronous. If the encoder isn't loaded yet,
+ * falls back to chars/4 — only happens for the very first stream chunk in a
+ * cold process.
+ */
+export declare function countTokens(text: string): number;

package/dist/llm/types.d.ts

@@ -19,6 +19,14 @@ export interface CreateMessageOptions {
      * @default true
      */
     recordUsage?: boolean;
+    /**
+     * DeepSeek V4 thinking-mode switch. "enabled" is the endpoint default;
+     * "disabled" turns reasoning off (faster, cheaper, no `reasoning_content`
+     * on the reply). Ignored by every provider except DeepSeek V4 — the OpenAI
+     * SDK forwards unknown top-level fields verbatim, so it's harmless on
+     * other compatible endpoints.
+     */
+    thinking?: "enabled" | "disabled";
 }
 export interface LLMUsageTracker {
     records: TokenUsage[];

package/dist/logging/logger.d.ts

@@ -1,7 +1,15 @@
 /**
  * Unified file logger for code-shell.
  *
- * Single sink: ~/.code-shell/logs/YYYY-MM-DD.log (JSON Lines).
+ * Sinks (JSON Lines), routed by `cat`:
+ *   ~/.code-shell/logs/ui-ink-YYYY-MM-DD.log
+ *     — UI/ink chatter: stream events, ctx render, ink screen diffs.
+ *   ~/.code-shell/logs/engine-YYYY-MM-DD.log
+ *     — everything else: engine, llm, tool, context, mcp, sandbox, ...
+ *
+ * Routing is by entry `cat`: cat ∈ {ui, ink, render, stream} → ui-ink bucket;
+ * anything else → engine bucket. Run scripts/logs.sh to query both at once
+ * (merge sorted by timestamp, filter by sid, etc).
  *
  * Level defaults:
  *   - Local dev (CODE_SHELL_DEV=1, or running from src/, or --debug)   → "debug"
@@ -9,7 +17,7 @@
  *   - CODE_SHELL_LOG_LEVEL overrides everything.
  *
  * Category filter (--debug=mcp,api or CODE_SHELL_DEBUG=mcp,api):
- *   When set, only logs whose `cat` matches are written.
+ *   When set, only debug logs whose `cat` matches are written.
  *
  * Disable entirely: CODE_SHELL_LOG=0
  */

package/dist/logging/session-recorder.d.ts

@@ -1,9 +1,16 @@
 /**
  * Per-session verbose recorder for local-dev runs.
  *
- * Writes a JSON-lines trace to <repo>/log/<YYYY-MM-DD>/session-<sid>.jsonl
- * containing every step: user input, system prompt, each LLM request+response
- * (full bodies), each tool call's args and result, and engine decision events.
+ * Layout (per day, split by bucket):
+ *   <repo>/log/<YYYY-MM-DD>/engine/session-<sid>.jsonl
+ *     — unified timeline: LLM, tools, session events, AND UI events
+ *       (UI events are dual-written so this file is the full story.)
+ *   <repo>/log/<YYYY-MM-DD>/ui/session-<sid>.jsonl
+ *     — UI-only events (stream events received, ctx renders, chatStore
+ *       mutations). Open when chasing a display bug.
+ *
+ * Same sid → same filename in both folders. Diff/merge by sid to align
+ * "what the model said" with "what the UI did about it."
  *
  * Why a separate sink from logger.ts?
  *   - logger.ts is the terse process-wide log (~/.code-shell/logs/) and stays
@@ -69,6 +76,12 @@ export declare function recordToolResult(sid: string, result: {
     error?: string;
 }): void;
 export declare function recordEvent(sid: string, name: string, data?: Record<string, unknown>): void;
+/**
+ * Record a UI-side event. Goes into both engine.jsonl (unified timeline) and
+ * ui.jsonl (focused view). Use from the UI when chasing display bugs — pass
+ * the current sessionId.
+ */
+export declare function recordUIEvent(sid: string | undefined, name: string, data?: Record<string, unknown>): void;
 export declare function recordSessionEnd(sid: string, info: {
     reason?: string;
     turns?: number;

package/dist/manager-J6FXEUEK.js

@@ -0,0 +1 @@
+export{c as SettingsManager}from'./chunk-DTPABBSV.js';import'./chunk-3QP5BQ3L.js';

package/dist/mcp-manager-27EVBKRT.js

@@ -0,0 +1 @@
+export{a as MCPManager}from'./chunk-L7B2ORID.js';import'./chunk-ZSAONDT6.js';import'./chunk-3QP5BQ3L.js';

package/dist/memory-BNI4WEWL.js

@@ -0,0 +1 @@
+export{b as MemoryManager,a as resolveMemoryBaseDir}from'./chunk-T222K6IV.js';import'./chunk-3QP5BQ3L.js';

package/dist/model-cache-HEDXI7UJ.js

@@ -0,0 +1 @@
+export{d as defaultCacheDir,c as isStale,a as readCache,b as writeCache}from'./chunk-P6JSTIFM.js';import'./chunk-3QP5BQ3L.js';

package/dist/model-fetcher-X5W5EASG.js

@@ -0,0 +1 @@
+export{b as fetchModelList}from'./chunk-2OD6VR4N.js';import'./chunk-P6JSTIFM.js';import'./chunk-3QP5BQ3L.js';

package/dist/off-7SDVMMXW.js

@@ -0,0 +1 @@
+export{a as createOffBackend}from'./chunk-SNFAPSLA.js';import'./chunk-3QP5BQ3L.js';

package/dist/openai-3QEADSYL.js

@@ -0,0 +1 @@
+export{a as OpenAIClient}from'./chunk-VKKU4LOM.js';import'./chunk-O5HFCH2U.js';import'./chunk-2EUIMKJA.js';import'./chunk-FU7AHZNN.js';import'./chunk-ZSAONDT6.js';import'./chunk-3QP5BQ3L.js';

package/dist/openrouter-models-GZVCRWD2.js

@@ -0,0 +1 @@
+export{d as findOpenRouterModel,c as getOpenRouterModels,b as getOpenRouterSnapshot,e as listOpenRouterModelsByVendor,a as setOpenRouterSnapshot}from'./chunk-A4M2Y42L.js';import'./chunk-3QP5BQ3L.js';

package/dist/openrouter-sync-WG4AG5NC.js

@@ -0,0 +1 @@
+export{a as syncOpenRouterCatalog}from'./chunk-XXCZE2U2.js';export{b as getOpenRouterSnapshot}from'./chunk-A4M2Y42L.js';import'./chunk-3QP5BQ3L.js';

package/dist/product/index.js

@@ -1,21 +1 @@
-import {
-  defineProduct
-} from "../chunk-ABRRONJ7.js";
-import "../chunk-PKGRTCCY.js";
-import "../chunk-BSW22MPA.js";
-import "../chunk-3OA5G4TM.js";
-import "../chunk-SIHHIQLO.js";
-import "../chunk-7Y6S3OCX.js";
-import "../chunk-VIAGTYYW.js";
-import "../chunk-76HYEPDM.js";
-import "../chunk-3SBSUOWZ.js";
-import "../chunk-T6VYG46N.js";
-import "../chunk-QARL3IDQ.js";
-import "../chunk-BZ5TNE7L.js";
-import "../chunk-3QO25N6E.js";
-import "../chunk-SFXHZSRF.js";
-import "../chunk-BXCY2FSE.js";
-import "../chunk-MD32R22I.js";
-export {
-  defineProduct
-};
+export{a as defineProduct}from'../chunk-Q3RSX66I.js';import'../chunk-MGP35GKD.js';import'../chunk-MLCTC53I.js';import'../chunk-WI6FSWGU.js';import'../chunk-P6JSTIFM.js';import'../chunk-SNFAPSLA.js';import'../chunk-L7B2ORID.js';import'../chunk-M7VU4YVY.js';import'../chunk-T222K6IV.js';import'../chunk-QTWU7UNA.js';import'../chunk-DTPABBSV.js';import'../chunk-RERBX75D.js';import'../chunk-2GJ7H5NC.js';import'../chunk-A4M2Y42L.js';import'../chunk-FU7AHZNN.js';import'../chunk-ZSAONDT6.js';import'../chunk-3QP5BQ3L.js';

package/dist/protocol/client.d.ts

@@ -45,7 +45,7 @@ export declare class AgentClient {
     /**
      * Query server state.
      */
-    query(type: QueryParams["type"], sessionId?: string, ...extra: unknown[]): Promise<QueryResult>;
+    query(type: QueryParams["type"], arg?: string | Record<string, unknown>, ...extra: unknown[]): Promise<QueryResult>;
     /**
      * Inject context into the session transcript without triggering a LLM turn.
      * Used to make arena results, tool outputs, etc. visible to subsequent LLM calls.

package/dist/run/index.js

@@ -1,48 +1 @@
-import {
-  createRunManager
-} from "../chunk-7QIM3BKD.js";
-import {
-  ArtifactTracker,
-  CheckpointWriter,
-  CompositeEvaluator,
-  EngineRunner,
-  FileRunStore,
-  Heartbeat,
-  NoopEvaluator,
-  RunApprovalBackend,
-  RunLock,
-  RunManager,
-  RunQueue,
-  VALID_TRANSITIONS,
-  createRunAskUserFn
-} from "../chunk-PKGRTCCY.js";
-import "../chunk-BSW22MPA.js";
-import "../chunk-3OA5G4TM.js";
-import "../chunk-SIHHIQLO.js";
-import "../chunk-7Y6S3OCX.js";
-import "../chunk-VIAGTYYW.js";
-import "../chunk-76HYEPDM.js";
-import "../chunk-3SBSUOWZ.js";
-import "../chunk-T6VYG46N.js";
-import "../chunk-QARL3IDQ.js";
-import "../chunk-BZ5TNE7L.js";
-import "../chunk-3QO25N6E.js";
-import "../chunk-SFXHZSRF.js";
-import "../chunk-BXCY2FSE.js";
-import "../chunk-MD32R22I.js";
-export {
-  ArtifactTracker,
-  CheckpointWriter,
-  CompositeEvaluator,
-  EngineRunner,
-  FileRunStore,
-  Heartbeat,
-  NoopEvaluator,
-  RunApprovalBackend,
-  RunLock,
-  RunManager,
-  RunQueue,
-  VALID_TRANSITIONS,
-  createRunAskUserFn,
-  createRunManager
-};
+export{a as createRunManager}from'../chunk-EU3J7TBP.js';export{$ as ArtifactTracker,_ as CheckpointWriter,da as CompositeEvaluator,ea as EngineRunner,W as FileRunStore,ba as Heartbeat,ca as NoopEvaluator,Y as RunApprovalBackend,aa as RunLock,fa as RunManager,X as RunQueue,V as VALID_TRANSITIONS,Z as createRunAskUserFn}from'../chunk-MGP35GKD.js';import'../chunk-MLCTC53I.js';import'../chunk-WI6FSWGU.js';import'../chunk-P6JSTIFM.js';import'../chunk-SNFAPSLA.js';import'../chunk-L7B2ORID.js';import'../chunk-M7VU4YVY.js';import'../chunk-T222K6IV.js';import'../chunk-QTWU7UNA.js';import'../chunk-DTPABBSV.js';import'../chunk-RERBX75D.js';import'../chunk-2GJ7H5NC.js';import'../chunk-A4M2Y42L.js';import'../chunk-FU7AHZNN.js';import'../chunk-ZSAONDT6.js';import'../chunk-3QP5BQ3L.js';

package/dist/seatbelt-L54CNI2Y.js

@@ -0,0 +1,37 @@
+import {h,a}from'./chunk-3QP5BQ3L.js';import {mkdtempSync,writeFileSync,rmSync}from'fs';import {tmpdir}from'os';import {join}from'path';h();function y(e){return {name:"seatbelt",wrap(o,l){let n=b(e),t=mkdtempSync(join(tmpdir(),"codeshell-sandbox-")),a$1=join(t,"profile.sb");return writeFileSync(a$1,n,"utf-8"),{file:"/usr/bin/sandbox-exec",args:["-f",a$1,l.shell,"-c",o],cleanup:a(()=>{try{rmSync(t,{recursive:!0,force:!0});}catch{}},"cleanup")}},hintForBlockedOutput(o){if(/Operation not permitted|sandbox/.test(o))return `
+[sandbox:seatbelt] A syscall was blocked by the sandbox. If this path should be writable or this network call legitimate, ask the user to update sandbox.writableRoots / sandbox.network in settings.json.`}}}a(y,"createSeatbeltBackend");function b(e){let o=e.writableRoots.map(t=>`  (subpath ${d(t)})`).join(`
+`),l=e.deniedReads.map(t=>`  (subpath ${d(t)})`).join(`
+`),n=e.network==="deny"?"(deny network-outbound)":"(allow network*)";return `(version 1)
+(deny default)
+
+;; Process control
+(allow process-fork)
+(allow process-exec)
+(allow signal (target self))
+(allow signal (target children))
+
+;; Reads: broadly allowed, then explicit deny of sensitive paths
+(allow file-read*)
+${l?`(deny file-read*
+${l})`:""}
+
+;; Writes: workspace + listed roots only
+(allow file-write*
+${o})
+(allow file-write-data
+  (literal "/dev/null")
+  (literal "/dev/zero")
+  (literal "/dev/random")
+  (literal "/dev/urandom")
+  (literal "/dev/dtracehelper"))
+
+;; IPC & system services common tools need
+(allow mach-lookup)
+(allow ipc-posix-shm)
+(allow sysctl-read)
+(allow system-socket)
+(allow iokit-open)
+
+;; Network
+${n}
+`}a(b,"buildProfile");function d(e){return `"${e.replace(/"/g,'\\"')}"`}a(d,"quote");export{y as createSeatbeltBackend};

package/dist/session/session-manager.d.ts

@@ -23,4 +23,6 @@ export declare class SessionManager {
 }
 export type SessionListEntry = SessionState & {
     preview?: string;
+    /** Last activity time — transcript mtime, falling back to startedAt. */
+    lastActiveAt: number;
 };

package/dist/session-manager-AGWS5SMJ.js

@@ -0,0 +1 @@
+export{b as SessionManager}from'./chunk-QTWU7UNA.js';import'./chunk-FU7AHZNN.js';import'./chunk-3QP5BQ3L.js';