@cjhyy/code-shell 0.1.7 → 0.2.0

package/dist/cli/onboarding.d.ts

@@ -23,9 +23,14 @@ export interface ProviderDef {
     envKey: string;
     provider: string;
     baseUrl: string;
-    defaultModel: string;
     keyUrl: string;
     keyPrefix: string;
+    /**
+     * The catalog of model IDs to register for this provider. The first
+     * entry doubles as the zero-config default — auto-activated when the
+     * user boots with just an env key (no onboarding). Users override by
+     * picking in the wizard or running `/model`.
+     */
     models: string[];
     /** When true, skip API key prompt (e.g. local providers like Ollama). */
     noKey?: boolean;
@@ -77,7 +82,7 @@ export declare function findSavedKeyForProvider(provider: ProviderDef): {
  * Return the list of model IDs previously saved under this provider
  * (matched by baseUrl). Used by the onboarding wizard to pre-populate the
  * "model pool" step with the user's prior choices instead of resetting to
- * just provider.defaultModel.
+ * just the provider's first model.
  */
 export declare function loadSavedModelsForProvider(provider: ProviderDef): string[];
 export declare function hasApiKey(): boolean;

package/dist/client-base-F5FCRSB3.js

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

package/dist/compaction-L3Y6NG4N.js

@@ -0,0 +1 @@
+import {p}from'./chunk-M7VU4YVY.js';export{e as COMPACTABLE_TOOL_NAMES,b as adjustIndexToPreserveAPIInvariants,k as applySummaryCompaction,g as applyToolResultBudget,i as buildSummarizationPrompt,m as dropOldestRounds,a as estimateTokens,j as extractAnchoredSummary,n as extractReferencedFilePaths,l as groupMessagesByApiRound,f as microcompact,c as snipCompact,h as truncateToolResult,d as windowCompact}from'./chunk-M7VU4YVY.js';import'./chunk-3QP5BQ3L.js';p();

package/dist/context/compaction.d.ts

@@ -89,7 +89,8 @@ export declare function truncateToolResult(result: string, maxChars?: number): s
  * Build a structured summarization prompt from messages to be compacted.
  * Produces a 9-section summary preserving key details.
  */
-export declare function buildSummarizationPrompt(messagesToSummarize: Message[]): string;
+export declare function buildSummarizationPrompt(messagesToSummarize: Message[], priorSummary?: string): string;
+export declare function extractAnchoredSummary(messages: Message[]): string | undefined;
 /**
  * Apply LLM-generated summary compaction (hybrid mode).
  *

package/dist/data/static-catalogs.d.ts

@@ -0,0 +1,25 @@
+/**
+ * Hand-maintained static model catalogs for providers whose /v1/models
+ * endpoint returns only ids — no context window, max output, or pricing.
+ *
+ * At runtime we still call the live /models endpoint to discover ids
+ * (so newly-released models work without a code update), then left-join
+ * this table on `id` to fill in metadata. Edit the JSONs in src/data/
+ * when a vendor ships a new model.
+ *
+ * OpenRouter is NOT here — its catalog is generated at build time by
+ * scripts/sync-models.ts because OpenRouter exposes full metadata.
+ */
+import type { ProviderKindName } from "../llm/provider-kinds.js";
+export interface StaticModel {
+    id: string;
+    name: string;
+    contextLength: number;
+    maxOutputTokens: number;
+    inputPricePerMillion: number;
+    outputPricePerMillion: number;
+    modalities: string[];
+    thinking?: boolean;
+}
+export declare function listStaticModels(kind: ProviderKindName): StaticModel[];
+export declare function hasStaticCatalog(kind: ProviderKindName): boolean;

package/dist/engine/engine.d.ts

@@ -78,6 +78,22 @@ export declare class Engine {
     private lastMessages;
     private lastSessionId;
     private compactedMessagesBySession;
+    /**
+     * SIDs whose ctx-bar seed we've already emitted in this process. The seed
+     * is a rough char/4 estimate; only useful before the first real
+     * usage_update arrives (cold start or cross-process resume). On subsequent
+     * turns the UI already shows the previous turn's accurate ctx — re-seeding
+     * would visibly drop the bar on every submit.
+     */
+    private ctxSeedSent;
+    /**
+     * Per-sid cache of "non-messages overhead" (system prompt + tool defs, in
+     * tokens). Survives across turns so each fresh TurnLoop instance can seed
+     * its first pre-llm emit with the right offset — without this the ctx bar
+     * visibly drops on every user submit (e.g. from ~20k → 3k) until the next
+     * LLM response arrives.
+     */
+    private ctxOverheadBySid;
     private activePermission;
     /** Public accessor so UI/clients can read the resolved per-model window. */
     get maxContextTokens(): number;

package/dist/engine/turn-loop.d.ts

@@ -23,6 +23,11 @@ export interface TurnLoopConfig {
      */
     onTurnBoundary?: (turnCount: number) => void;
 }
+export interface CtxOverheadStore {
+    /** Tokens for system prompt + tool defs, derived from provider's promptTokens. */
+    get(sid: string): number;
+    set(sid: string, tokens: number): void;
+}
 export interface TurnLoopDeps {
     model: ModelFacade;
     toolExecutor: ToolExecutor;
@@ -31,6 +36,10 @@ export interface TurnLoopDeps {
     transcript: Transcript;
     systemPrompt: string;
     tools: import("../types.js").ToolDefinition[];
+    /** Per-sid overhead cache so the ctx bar doesn't drop between turns. */
+    ctxOverheadStore: CtxOverheadStore;
+    /** Current session id, used to key the overhead store. */
+    sessionId: string;
 }
 export interface TurnLoopResult {
     text: string;
@@ -50,7 +59,18 @@ export declare class TurnLoop {
      * every line they write is tagged with the current turn/turnId.
      */
     private currentTurnLog;
+    /** Last emitted ctx token estimate; used to skip no-op usage_update events. */
+    private lastCtxEmit;
     constructor(deps: TurnLoopDeps, config: TurnLoopConfig);
+    /**
+     * Emit a usage_update so the UI ctx bar reflects current message-array
+     * size. Called at every point where messages mutate: after a tool_result
+     * is appended, after context management (which may shrink), and after an
+     * LLM response (which we also feed through with the provider's authoritative
+     * promptTokens to override our estimate).
+     */
+    private emitCtxFromMessages;
+    private emitCtxFromUsage;
     /**
      * Run the multi-turn agent loop until completion.
      */

package/dist/index.d.ts

@@ -3,7 +3,7 @@
  *
  * Public API exports.
  */
-export declare const VERSION = "0.1.7";
+export declare const VERSION = "0.2.0";
 export type { Message, ContentBlock, ToolDefinition, ToolCall, ToolResult, RegisteredTool, TranscriptEvent, TranscriptEventType, SessionState, TokenUsage, CompiledInput, PermissionDecision, PermissionMode, PermissionRule, TurnPhase, TurnResult, TerminalReason, StreamEvent, StreamCallback, LLMConfig, LLMResponse, Settings, MCPServerConfig, } from "./types.js";
 export { FrameworkError, LLMError, LLMRateLimitError, ContextLimitError, ToolError, ToolNotFoundError, ToolExecutionError, ToolTimeoutError, PermissionDeniedError, SessionError, TranscriptError, ConfigError, } from "./exceptions.js";
 export { Engine } from "./engine/engine.js";

package/dist/index.js

@@ -1 +1 @@
-export{a as defineProduct}from'./chunk-KCXSARAA.js';export{a as createRunManager}from'./chunk-RP26WO4R.js';export{_ as ArtifactTracker,i as AutoApprovalBackend,I as BUILTIN_AGENT_PRESETS,e as BUILTIN_TOOLS,Z as CheckpointWriter,ca as CompositeEvaluator,w as ContextManager,J as DEFAULT_AGENT_PRESET,K as DEFAULT_CLI_PRESET,o as DEFAULT_PERSIST_THRESHOLD,T as Engine,da as EngineRunner,V as FileRunStore,h as HeadlessApprovalBackend,aa as Heartbeat,n as HookRegistry,ba as NoopEvaluator,p as PER_MESSAGE_AGGREGATE_CAP,q as PREVIEW_SIZE,m as PermissionClassifier,Q as PromptComposer,X as RunApprovalBackend,$ as RunLock,ea as RunManager,W as RunQueue,x as SectionCache,g as ToolExecutor,f as ToolRegistry,U as VALID_TRANSITIONS,u as applyToolResultPersistence,H as availableSections,O as buildPresetSystemPrompt,D as buildSkillListing,z as combineInstructions,r as createContentReplacementState,Y as createRunAskUserFn,v as isPersistedReplacement,M as listPresetNames,F as loadSection,G as loadSections,B as matchSkillsByInput,C as matchSkillsByTool,s as reconstructContentReplacementState,L as registerPreset,E as registerSection,N as resolveAgentPreset,P as resolveBuiltinToolNames,t as resolveToolResultsDir,y as scanInstructions,A as scanSkills,c as taskManager}from'./chunk-NMBYBGC7.js';export{a as ModelPool}from'./chunk-2NS3JMNK.js';export{a as FileHistory}from'./chunk-WI6FSWGU.js';import'./chunk-P6JSTIFM.js';import'./chunk-SNFAPSLA.js';export{a as MCPManager}from'./chunk-XDOWC7EG.js';import {o}from'./chunk-W47Z7MGS.js';export{e as COMPACTABLE_TOOL_NAMES,j as applySummaryCompaction,i as buildSummarizationPrompt,a as estimateTokens,f as microcompact,h as truncateToolResult,d as windowCompact}from'./chunk-W47Z7MGS.js';export{b as MemoryManager}from'./chunk-T222K6IV.js';export{b as SessionManager,a as Transcript}from'./chunk-QTWU7UNA.js';export{c as SettingsManager,a as SettingsSchema,b as validateSettings}from'./chunk-UDPUTFQT.js';export{j as IterativeArena,f as defaultIterateConvergence,i as getIterateFormat,g as iterateCodeFormat,e as iterateDiffRatio,h as iterateDocumentFormat}from'./chunk-Q4O2HGWS.js';export{J as Arena,I as ArenaLedger,e as DiscussionStrategy,Q as MODEL_PRESETS,f as PlanningStrategy,d as ReviewStrategy,z as applyReviewResult,u as buildDigest,n as collectEvidence,b as createLLMClient,v as formatDigest,R as getMaxOutputTokens,k as getStrategy,l as getStrategyForPlan,p as hasTools,B as isTerminal,y as markUnderReview,A as markUnresolved,m as planArena,s as registerClaims,a as registerProvider,x as resolveClaimStatus,t as selectClaimsForReview,o as selectTools,w as transitionClaim,C as validTransitions}from'./chunk-Y2NS6HPO.js';import'./chunk-NVAZJ5XS.js';export{a as AnthropicClient}from'./chunk-WOP6MJKW.js';export{a as OpenAIClient}from'./chunk-PGANL6F2.js';import'./chunk-O5HFCH2U.js';export{a as LLMClientBase}from'./chunk-HIJ2XXGX.js';export{l as ConfigError,d as ContextLimitError,a as FrameworkError,b as LLMError,c as LLMRateLimitError,i as PermissionDeniedError,j as SessionError,e as ToolError,g as ToolExecutionError,f as ToolNotFoundError,h as ToolTimeoutError,k as TranscriptError}from'./chunk-FU7AHZNN.js';export{b as logger}from'./chunk-B7K2OWKC.js';import {h,a}from'./chunk-3QP5BQ3L.js';h();o();h();var vt={review:[{patterns:[/\breview\b/i,/\baudit\b/i,/\binspect\b/i,/\bcheck\b/i],weight:2},{patterns:[/\bcode\s*quality\b/i,/\bbug\b/i,/\bvulnerab/i,/\bsecur/i]},{patterns:[/\bpr\b/i,/\bpull\s*request\b/i,/\bdiff\b/i,/\bchanges?\b/i]},{patterns:[/\brefactor/i,/\bclean\s*up\b/i,/\bimprove\b/i]}],discussion:[{patterns:[/\bdiscuss\b/i,/\bdebate\b/i,/\bcompare\b/i],weight:2},{patterns:[/\bpros?\s*(and|&|\/)\s*cons?\b/i,/\btrade\s*-?\s*offs?\b/i]},{patterns:[/\bshould\s+we\b/i,/\bwhich\s+(is|approach|way)\b/i,/\bvs\.?\b/i]},{patterns:[/\bopinion\b/i,/\bthoughts?\b/i,/\badvice\b/i]}],planning:[{patterns:[/\bplan\b/i,/\bplanning\b/i,/\broadmap\b/i,/\bstrategy\b/i],weight:2},{patterns:[/\barchitect/i,/\bdesign\b/i,/\bproposal\b/i]},{patterns:[/\bimplement(ation)?\s+(plan|strategy|approach)\b/i],weight:2},{patterns:[/\bphase\b/i,/\bmilestone\b/i,/\btimeline\b/i,/\bprioritiz/i]},{patterns:[/\bbuild\b/i,/\bcreate\b/i,/\bsetup\b/i,/\bbootstrap\b/i]}]},kt=3,Mt=2;function Pt(m){let a={review:0,discussion:0,planning:0};for(let[n,i]of Object.entries(vt))for(let p of i){let u=p.weight??1;for(let d of p.patterns)d.test(m)&&(a[n]+=u);}let s=Object.entries(a).sort((n,i)=>i[1]-n[1]),[t,e]=s[0],[,l]=s[1],r=e-l;return e>=kt&&r>=Mt?{mode:t,confidence:"high",reason:`Topic strongly matches "${t}" (score: ${e}, lead: +${r})`}:e>0&&r>0?{mode:t,confidence:"low",reason:`Topic weakly matches "${t}" (score: ${e}, lead: +${r})`}:{mode:"review",confidence:"low",reason:"No strong mode signal detected, defaulting to review"}}a(Pt,"detectArenaMode");var wt="0.1.7";export{wt as VERSION,Pt as detectArenaMode};
+export{a as defineProduct}from'./chunk-Q3RSX66I.js';export{a as createRunManager}from'./chunk-EU3J7TBP.js';export{$ as ArtifactTracker,j as AutoApprovalBackend,J as BUILTIN_AGENT_PRESETS,e as BUILTIN_TOOLS,_ as CheckpointWriter,da as CompositeEvaluator,x as ContextManager,K as DEFAULT_AGENT_PRESET,L as DEFAULT_CLI_PRESET,p as DEFAULT_PERSIST_THRESHOLD,U as Engine,ea as EngineRunner,W as FileRunStore,i as HeadlessApprovalBackend,ba as Heartbeat,o as HookRegistry,ca as NoopEvaluator,q as PER_MESSAGE_AGGREGATE_CAP,r as PREVIEW_SIZE,n as PermissionClassifier,R as PromptComposer,Y as RunApprovalBackend,aa as RunLock,fa as RunManager,X as RunQueue,y as SectionCache,h as ToolExecutor,f as ToolRegistry,V as VALID_TRANSITIONS,v as applyToolResultPersistence,I as availableSections,P as buildPresetSystemPrompt,E as buildSkillListing,A as combineInstructions,s as createContentReplacementState,Z as createRunAskUserFn,w as isPersistedReplacement,N as listPresetNames,G as loadSection,H as loadSections,C as matchSkillsByInput,D as matchSkillsByTool,t as reconstructContentReplacementState,M as registerPreset,F as registerSection,O as resolveAgentPreset,Q as resolveBuiltinToolNames,u as resolveToolResultsDir,z as scanInstructions,B as scanSkills,c as taskManager}from'./chunk-MGP35GKD.js';export{a as ModelPool}from'./chunk-MLCTC53I.js';export{a as FileHistory}from'./chunk-WI6FSWGU.js';import'./chunk-P6JSTIFM.js';import'./chunk-SNFAPSLA.js';export{a as MCPManager}from'./chunk-L7B2ORID.js';import {p}from'./chunk-M7VU4YVY.js';export{e as COMPACTABLE_TOOL_NAMES,k as applySummaryCompaction,i as buildSummarizationPrompt,a as estimateTokens,f as microcompact,h as truncateToolResult,d as windowCompact}from'./chunk-M7VU4YVY.js';export{b as MemoryManager}from'./chunk-T222K6IV.js';export{b as SessionManager,a as Transcript}from'./chunk-QTWU7UNA.js';export{c as SettingsManager,a as SettingsSchema,b as validateSettings}from'./chunk-DTPABBSV.js';export{j as IterativeArena,f as defaultIterateConvergence,i as getIterateFormat,g as iterateCodeFormat,e as iterateDiffRatio,h as iterateDocumentFormat}from'./chunk-RERBX75D.js';export{J as Arena,I as ArenaLedger,e as DiscussionStrategy,Q as MODEL_PRESETS,f as PlanningStrategy,d as ReviewStrategy,z as applyReviewResult,u as buildDigest,n as collectEvidence,b as createLLMClient,v as formatDigest,R as getMaxOutputTokens,k as getStrategy,l as getStrategyForPlan,p as hasTools,B as isTerminal,y as markUnderReview,A as markUnresolved,m as planArena,s as registerClaims,a as registerProvider,x as resolveClaimStatus,t as selectClaimsForReview,o as selectTools,w as transitionClaim,C as validTransitions}from'./chunk-2GJ7H5NC.js';import'./chunk-A4M2Y42L.js';export{a as AnthropicClient}from'./chunk-DEYFK6FP.js';export{a as OpenAIClient}from'./chunk-VKKU4LOM.js';import'./chunk-O5HFCH2U.js';export{a as LLMClientBase}from'./chunk-2EUIMKJA.js';export{l as ConfigError,d as ContextLimitError,a as FrameworkError,b as LLMError,c as LLMRateLimitError,i as PermissionDeniedError,j as SessionError,e as ToolError,g as ToolExecutionError,f as ToolNotFoundError,h as ToolTimeoutError,k as TranscriptError}from'./chunk-FU7AHZNN.js';export{b as logger}from'./chunk-ZSAONDT6.js';import {h,a}from'./chunk-3QP5BQ3L.js';h();p();h();var vt={review:[{patterns:[/\breview\b/i,/\baudit\b/i,/\binspect\b/i,/\bcheck\b/i],weight:2},{patterns:[/\bcode\s*quality\b/i,/\bbug\b/i,/\bvulnerab/i,/\bsecur/i]},{patterns:[/\bpr\b/i,/\bpull\s*request\b/i,/\bdiff\b/i,/\bchanges?\b/i]},{patterns:[/\brefactor/i,/\bclean\s*up\b/i,/\bimprove\b/i]}],discussion:[{patterns:[/\bdiscuss\b/i,/\bdebate\b/i,/\bcompare\b/i],weight:2},{patterns:[/\bpros?\s*(and|&|\/)\s*cons?\b/i,/\btrade\s*-?\s*offs?\b/i]},{patterns:[/\bshould\s+we\b/i,/\bwhich\s+(is|approach|way)\b/i,/\bvs\.?\b/i]},{patterns:[/\bopinion\b/i,/\bthoughts?\b/i,/\badvice\b/i]}],planning:[{patterns:[/\bplan\b/i,/\bplanning\b/i,/\broadmap\b/i,/\bstrategy\b/i],weight:2},{patterns:[/\barchitect/i,/\bdesign\b/i,/\bproposal\b/i]},{patterns:[/\bimplement(ation)?\s+(plan|strategy|approach)\b/i],weight:2},{patterns:[/\bphase\b/i,/\bmilestone\b/i,/\btimeline\b/i,/\bprioritiz/i]},{patterns:[/\bbuild\b/i,/\bcreate\b/i,/\bsetup\b/i,/\bbootstrap\b/i]}]},kt=3,Mt=2;function Pt(m){let a={review:0,discussion:0,planning:0};for(let[n,i]of Object.entries(vt))for(let p of i){let u=p.weight??1;for(let d of p.patterns)d.test(m)&&(a[n]+=u);}let s=Object.entries(a).sort((n,i)=>i[1]-n[1]),[t,e]=s[0],[,l]=s[1],r=e-l;return e>=kt&&r>=Mt?{mode:t,confidence:"high",reason:`Topic strongly matches "${t}" (score: ${e}, lead: +${r})`}:e>0&&r>0?{mode:t,confidence:"low",reason:`Topic weakly matches "${t}" (score: ${e}, lead: +${r})`}:{mode:"review",confidence:"low",reason:"No strong mode signal detected, defaulting to review"}}a(Pt,"detectArenaMode");var wt="0.2.0";export{wt as VERSION,Pt as detectArenaMode};

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

@@ -0,0 +1,20 @@
+/**
+ * Capability layer — single entry point.
+ *
+ * `capabilitiesFor(kind, model)` walks the rules in `rules.ts`, returns
+ * a `Capability` describing what params the (provider, model) pair will
+ * accept. Clients spread the result into their request-build code.
+ *
+ * Pure function. No I/O, no caching beyond the rules array.
+ */
+import type { ProviderKindName } from "../provider-kinds.js";
+import { type Capability } from "./types.js";
+export { DEFAULT_CAPABILITY } from "./types.js";
+export type { Capability, ReasoningShape, ReasoningEffort, ThinkingSwitch, EchoReasoning, ParallelToolCallsShape, StreamUsageShape, } from "./types.js";
+/**
+ * Resolve the capability for a given (provider kind, model id).
+ *
+ * First matching rule for the kind wins — rules.ts is ordered so that
+ * more specific patterns come before catch-alls.
+ */
+export declare function capabilitiesFor(kind: ProviderKindName, model: string): Capability;

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

@@ -0,0 +1,26 @@
+/**
+ * Per-(provider kind, model family) capability overrides.
+ *
+ * Each rule patches the conservative DEFAULT_CAPABILITY. First match per
+ * kind wins — order matters for overlapping families.
+ *
+ * Authority for every entry is the vendor's own docs; the `why` field
+ * cites the specific page. When a model 400s on something not captured
+ * here, add a rule, don't patch the client.
+ *
+ * Sources audited 2026-05-15:
+ *   OpenAI:    https://platform.openai.com/docs/guides/reasoning
+ *              https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/reasoning
+ *   Anthropic: https://platform.claude.com/docs/en/build-with-claude/extended-thinking
+ *   DeepSeek:  https://api-docs.deepseek.com/guides/reasoning_model
+ *              https://api-docs.deepseek.com/guides/thinking_mode
+ *   Z.AI:      https://docs.z.ai/guides/llm/glm-4.6
+ *   Gemini:    https://ai.google.dev/gemini-api/docs/openai
+ *              https://ai.google.dev/gemini-api/docs/thinking
+ *   OpenRouter: https://openrouter.ai/docs/guides/best-practices/reasoning-tokens
+ *   xAI:       https://docs.x.ai/docs/guides/reasoning
+ *   Mistral:   https://docs.mistral.ai/api/
+ *   Groq:      https://console.groq.com/docs/reasoning
+ */
+import type { CapabilityRule } from "./types.js";
+export declare const RULES: ReadonlyArray<CapabilityRule>;

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/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/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/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-AM6M52D6.js → openrouter-models-GZVCRWD2.js}

@@ -1 +1 @@
-export{d as findOpenRouterModel,c as getOpenRouterModels,b as getOpenRouterSnapshot,e as listOpenRouterModelsByVendor,a as setOpenRouterSnapshot}from'./chunk-NVAZJ5XS.js';import'./chunk-3QP5BQ3L.js';
+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 +1 @@
-export{a as defineProduct}from'../chunk-KCXSARAA.js';import'../chunk-NMBYBGC7.js';import'../chunk-2NS3JMNK.js';import'../chunk-WI6FSWGU.js';import'../chunk-P6JSTIFM.js';import'../chunk-SNFAPSLA.js';import'../chunk-XDOWC7EG.js';import'../chunk-W47Z7MGS.js';import'../chunk-T222K6IV.js';import'../chunk-QTWU7UNA.js';import'../chunk-UDPUTFQT.js';import'../chunk-Q4O2HGWS.js';import'../chunk-Y2NS6HPO.js';import'../chunk-NVAZJ5XS.js';import'../chunk-FU7AHZNN.js';import'../chunk-B7K2OWKC.js';import'../chunk-3QP5BQ3L.js';
+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 +1 @@
-export{a as createRunManager}from'../chunk-RP26WO4R.js';export{_ as ArtifactTracker,Z as CheckpointWriter,ca as CompositeEvaluator,da as EngineRunner,V as FileRunStore,aa as Heartbeat,ba as NoopEvaluator,X as RunApprovalBackend,$ as RunLock,ea as RunManager,W as RunQueue,U as VALID_TRANSITIONS,Y as createRunAskUserFn}from'../chunk-NMBYBGC7.js';import'../chunk-2NS3JMNK.js';import'../chunk-WI6FSWGU.js';import'../chunk-P6JSTIFM.js';import'../chunk-SNFAPSLA.js';import'../chunk-XDOWC7EG.js';import'../chunk-W47Z7MGS.js';import'../chunk-T222K6IV.js';import'../chunk-QTWU7UNA.js';import'../chunk-UDPUTFQT.js';import'../chunk-Q4O2HGWS.js';import'../chunk-Y2NS6HPO.js';import'../chunk-NVAZJ5XS.js';import'../chunk-FU7AHZNN.js';import'../chunk-B7K2OWKC.js';import'../chunk-3QP5BQ3L.js';
+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-WBYROQT6.js → seatbelt-L54CNI2Y.js}

@@ -1,7 +1,7 @@
-import {h,a}from'./chunk-3QP5BQ3L.js';import {mkdtempSync,writeFileSync}from'fs';import {tmpdir}from'os';import {join}from'path';h();function x(e){return {name:"seatbelt",wrap(o,l){let r=b(e),t=mkdtempSync(join(tmpdir(),"codeshell-sandbox-")),a=join(t,"profile.sb");return writeFileSync(a,r,"utf-8"),{file:"/usr/bin/sandbox-exec",args:["-f",a,l.shell,"-c",o]}},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(x,"createSeatbeltBackend");function b(e){let o=e.writableRoots.map(t=>`  (subpath ${d(t)})`).join(`
+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(`
-`),r=e.network==="deny"?"(deny network-outbound)":"(allow network*)";return `(version 1)
+`),n=e.network==="deny"?"(deny network-outbound)":"(allow network*)";return `(version 1)
 (deny default)
 
 ;; Process control
@@ -33,5 +33,5 @@ ${o})
 (allow iokit-open)
 
 ;; Network
-${r}
-`}a(b,"buildProfile");function d(e){return `"${e.replace(/"/g,'\\"')}"`}a(d,"quote");export{x as createSeatbeltBackend};
+${n}
+`}a(b,"buildProfile");function d(e){return `"${e.replace(/"/g,'\\"')}"`}a(d,"quote");export{y as createSeatbeltBackend};