@bubblebrain-ai/bubble 0.0.49 → 0.0.51

package/dist/main.js

@@ -128,10 +128,6 @@ async function main() {
         return createProvider(providerId, target.apiKey, target.baseURL);
     };
     let agentRef;
-    const todoStore = {
-        getTodos: () => agentRef?.getTodos() ?? [],
-        setTodos: (todos) => agentRef?.setTodos(todos),
-    };
     const planHandlerRef = {};
     const planController = {
         getMode: () => agentRef?.mode ?? "default",
@@ -181,7 +177,6 @@ async function main() {
     const allowBackgroundTasks = !printMode;
     const promotionChannel = new PromotionChannel();
     const tools = createAllTools(args.cwd, skillRegistry, {
-        todoStore,
         planController,
         approvalController,
         questionController: printMode ? undefined : questionController,
@@ -338,7 +333,6 @@ async function main() {
                     ? normalizeThinkingLevel(configuredThinkingLevel, getAvailableThinkingLevels(activeProviderId, effectiveModelId))
                     : getDefaultThinkingLevel(activeProviderId, effectiveModelId)
         : (args.thinkingLevel ?? sessionThinkingLevel ?? configuredThinkingLevel ?? "off");
-    const restoredTodos = sessionManager?.getTodos() ?? [];
     const initialMode = args.mode ?? "default";
     const skillSummaries = skillRegistry.summaries();
     // This synchronous provenance gate must run before buildMemoryPrompt: the
@@ -401,7 +395,6 @@ async function main() {
         temperature: 0.2,
         thinkingLevel: initialThinkingLevel,
         mode: initialMode,
-        todos: restoredTodos,
         onMessageAppend: (message) => {
             if (!sessionManager)
                 return;
@@ -435,9 +428,6 @@ async function main() {
                 sessionManager.appendMarker("skill_activated", match[1].trim());
             }
         },
-        onTodosUpdate: (todos) => {
-            sessionManager?.appendTodosSnapshot(todos);
-        },
         onModeUpdate: (mode) => {
             sessionManager?.appendMarker("mode_switch", mode);
         },
@@ -703,7 +693,6 @@ async function main() {
                 // mirroring the /rewind history-replacement pattern.
                 const head = agent.messages.filter((m) => m.role === "system" || m.role === "meta");
                 const nextMessages = [...head, ...history];
-                const nextTodos = next.getTodos();
                 // Commit only after every file read/write and reconstruction step has
                 // succeeded. Callers can safely prepare candidate sessions without a
                 // failed switch rebinding persistence or replacing the live history.
@@ -714,10 +703,6 @@ async function main() {
                 // These update only live in-memory/UI mirrors after the persistence
                 // commit. A best-effort callback failure must not report the switch as
                 // rolled back after the outer session binding has already changed.
-                try {
-                    agent.setTodos(nextTodos);
-                }
-                catch { /* candidate already committed */ }
                 try {
                     agent.resetContextUsageAnchor();
                 }

package/dist/memory/phase1.js

@@ -160,8 +160,6 @@ function serializeSessionEntry(entry) {
             return `[tool_call:${entry.toolCall.name}] ${truncate(entry.toolCall.arguments, 1_500)}`;
         case "tool_result":
             return `[tool_result${entry.message.isError ? " error=true" : ""}] ${truncate(entry.message.content, 2_000)}`;
-        case "todos_snapshot":
-            return `[todos] ${truncate(JSON.stringify(entry.todos), 2_000)}`;
     }
 }
 function countMeaningfulEntries(entries) {

package/dist/model-catalog.js

@@ -41,13 +41,12 @@ const OPENAI_CHAT_LEVELS = ["off"];
 // Internal representation for APIs that expose thinking as enabled/disabled.
 // UI code must render supported toggle models as on/off, not as "medium" effort.
 const TOGGLE_THINKING_LEVELS = ["off", "medium"];
-// GLM-5.2 is the first GLM to accept OpenAI-style `reasoning_effort`. The API
-// enum is none/minimal/low/medium/high/xhigh/max; we expose high and max (the
-// two effort tiers worth offering a coding agent) plus "off", which disables
-// thinking outright via `thinking: {type: "disabled"}`. Order matters: "high"
-// is first so it is the default (getDefaultThinkingLevel falls back to levels[0]
-// when "medium" is absent), since GLM-5.2 is a thinking-on-by-default model.
+// GLM-5.2 accepts OpenAI-style `reasoning_effort`. We expose high and max plus
+// "off", which disables thinking outright via `thinking: {type: "disabled"}`.
 const GLM_5_2_LEVELS = ["high", "max", "off"];
+// GLM-5.3 Coding Plan normalizes efforts into exactly three real tiers. Its
+// default is max, and disabling thinking maps to low rather than turning it off.
+const GLM_5_3_LEVELS = ["low", "high", "max"];
 // Kimi K2.7 Code variants only support thinking mode (disabling it errors), so
 // "off" is not offered — the model is always in its thinking variant.
 const KIMI_THINKING_ONLY_LEVELS = ["medium"];
@@ -124,6 +123,9 @@ export const BUILTIN_MODELS = [
     { id: "glm-5.1", name: "GLM-5.1", providerId: "zhipuai", tier: "balanced", reasoningLevels: TOGGLE_THINKING_LEVELS, contextWindow: 200000 },
     { id: "glm-4.7", name: "GLM-4.7", providerId: "zhipuai", reasoningLevels: TOGGLE_THINKING_LEVELS, contextWindow: 204800 },
     { id: "glm-4.6", name: "GLM-4.6", providerId: "zhipuai", reasoningLevels: TOGGLE_THINKING_LEVELS, contextWindow: 204800 },
+    // Coding Plan exposes GLM-5.3 before GET /models lists it (verified live
+    // 2026-08-14). The standard API is still marked "coming soon" in the docs.
+    { id: "glm-5.3", name: "GLM-5.3", providerId: "zhipuai-coding-plan", tier: "strong", reasoningLevels: GLM_5_3_LEVELS, defaultReasoningLevel: "max", contextWindow: 1000000 },
     { id: "glm-5.2", name: "GLM-5.2", providerId: "zhipuai-coding-plan", tier: "strong", reasoningLevels: GLM_5_2_LEVELS, contextWindow: 1000000 },
     { id: "glm-5.1", name: "GLM-5.1", providerId: "zhipuai-coding-plan", tier: "balanced", reasoningLevels: TOGGLE_THINKING_LEVELS, contextWindow: 200000 },
     { id: "glm-4.7", name: "GLM-4.7", providerId: "zhipuai-coding-plan", reasoningLevels: TOGGLE_THINKING_LEVELS, contextWindow: 204800 },

package/dist/permissions/settings.d.ts

@@ -43,7 +43,6 @@ export interface SettingsManagerOptions {
     bubbleHome?: string;
 }
 export declare class SettingsManager {
-    private readonly cwd;
     private readonly paths;
     private raw;
     private fileDiagnostics;

package/dist/permissions/settings.js

@@ -25,7 +25,6 @@ const KNOWN_MODES = new Set([
     "bypassPermissions",
 ]);
 export class SettingsManager {
-    cwd;
     paths;
     raw = {
         user: null,
@@ -34,7 +33,6 @@ export class SettingsManager {
     };
     fileDiagnostics = [];
     constructor(cwd, options = {}) {
-        this.cwd = cwd;
         const bubbleHome = options.bubbleHome ?? getBubbleHome();
         this.paths = {
             user: join(bubbleHome, "settings.json"),

package/dist/prompt/compose.js

@@ -86,9 +86,6 @@ function buildGuidelines(tools, extraGuidelines) {
     if (tools.includes("skill_search") && tools.includes("skill")) {
         add("Skills may provide specialized workflows. When a task appears to match a specialized workflow, call skill_search to find relevant skills, then call skill with the exact name to load the selected skill before applying it");
     }
-    if (tools.includes("todo_write")) {
-        add("Default to just doing the work. Reach for todo_write only when tracking progress would genuinely help — multi-phase work with real dependencies, a long task spanning several areas, an explicit user request, or steps you discover mid-task. Don't make a list to pad a task you could just do; when in doubt, skip it. Mark each item completed as soon as it is done; do not batch updates");
-    }
     for (const item of extraGuidelines) {
         add(item);
     }

package/dist/prompt/environment.js

@@ -14,7 +14,6 @@ export const defaultToolSnippets = {
     question: "Ask the user structured questions when clarification or preference choices would materially improve the work",
     skill_search: "Search available skills by name, description, tags, and source",
     skill: "Load a named skill with specialized instructions and bundled resources",
-    todo_write: "Plan and track multi-step work. Mark each task completed as soon as it is done — do not batch.",
 };
 export const defaultToolNames = [
     "read",
@@ -31,7 +30,6 @@ export const defaultToolNames = [
     "question",
     "skill_search",
     "skill",
-    "todo_write",
 ];
 export function buildEnvironmentPrompt(options = {}) {
     const configuredProvider = options.configuredProvider ?? "unknown";

package/dist/prompt/reminders.js

@@ -29,7 +29,7 @@ const PLAN_MODE_ENTER = `
 Plan mode is now ACTIVE.
 
 Rules while in plan mode:
-- Only read-only tools are allowed, including read, glob, grep, web_search, web_fetch, spawn_agent, wait_agent, send_input, skill_search, skill, todo_write, tool_search, question, and exit_plan_mode.
+- Only read-only tools are allowed, including read, glob, grep, web_search, web_fetch, spawn_agent, wait_agent, send_input, skill_search, skill, tool_search, question, and exit_plan_mode.
 - Writes, edits, and shell commands WILL be rejected by the harness; do not try them.
 - Do not edit files or claim implementation is complete while plan mode is active.
 - Investigate the codebase, then use the question tool to clarify important ambiguities, tradeoffs, requirements, or preference choices that would materially change the plan.

package/dist/provider-registry.js

@@ -458,7 +458,19 @@ export class ProviderRegistry {
         if (!options.forceRefresh) {
             const cached = this.modelDiscoveryCache.get(key);
             if (cached && cached.expiresAt > now) {
-                const result = { ...cached.result, source: "cache" };
+                // A disk cache can outlive the application version that wrote it. Keep
+                // the current curated catalog authoritative over stale cached metadata,
+                // then retain cached remote-only ids for non-authoritative union results.
+                // This lets newly-added builtins surface immediately after an upgrade
+                // without waiting up to 24 hours or requiring a manual Ctrl+R refresh.
+                const local = this.localModelsForProvider(provider);
+                const localIds = new Set(local.map((model) => model.id));
+                const models = cached.result.source === "static"
+                    ? local
+                    : cached.result.authoritative
+                        ? cached.result.models
+                        : [...local, ...cached.result.models.filter((model) => !localIds.has(model.id))];
+                const result = { ...cached.result, models, source: "cache" };
                 const current = this.getConfigured().find((item) => item.id === provider.id);
                 if (!current || this.modelDiscoveryKey(current) === key) {
                     this.applyDynamicDiscoveryMetadata(provider, result);

package/dist/provider-transform.js

@@ -114,16 +114,15 @@ export function resolveProviderRequestConfig(providerId, modelId, requestedLevel
     // Zhipu/Z.AI OpenAI-compatible endpoints expose reasoning via a provider-specific
     // `thinking` block rather than OpenAI's `reasoning_effort` shape.
     if (["zhipuai", "zhipuai-coding-plan", "zai", "zai-coding-plan"].includes(providerId)) {
-        // GLM-5.2 is the only GLM that also accepts `reasoning_effort` (we expose
-        // high/max, which map 1:1 onto the API enum). "off" disables thinking via
-        // `thinking: {type: "disabled"}` — otherwise the server default (thinking
-        // on, effort max) would make "off" a no-op. The effort field rides inside
-        // the body alongside `thinking`, so it goes in extraBody, not the
-        // OpenRouter-style `reasoningEffort` config field.
-        if (modelId === "glm-5.2") {
+        // GLM-5.2+ also accept `reasoning_effort`. GLM-5.2 can disable thinking;
+        // GLM-5.3 requires thinking and Coding Plan normalizes its effort into
+        // low/high/max. The catalog clamps 5.3 choices before this serialization
+        // boundary. The effort field rides inside the body alongside `thinking`,
+        // so it goes in extraBody, not the OpenRouter-style config field.
+        if (modelId === "glm-5.2" || modelId === "glm-5.3") {
             return {
                 effectiveThinkingLevel,
-                extraBody: effectiveThinkingLevel === "off"
+                extraBody: modelId === "glm-5.2" && effectiveThinkingLevel === "off"
                     ? { thinking: { type: "disabled" } }
                     : {
                         thinking: { type: "enabled", clear_thinking: false },

package/dist/sdk/index.d.ts

@@ -17,10 +17,11 @@ import { type SessionSummary } from "../session.js";
 import type { ApprovalDecision, ApprovalRequest } from "../approval/types.js";
 import { UserConfig } from "../config.js";
 import { ProviderRegistry } from "../provider-registry.js";
+import type { ResolvedSubagentRoute } from "../agent/categories.js";
 import { type QuestionAnswer, type QuestionRequest } from "../question/controller.js";
 import { SkillRegistry } from "../skills/registry.js";
 import type { SkillSummary } from "../skills/types.js";
-import type { AgentEvent, ContentPart, Message, PermissionMode, ThinkingLevel } from "../types.js";
+import type { AgentEvent, ContentPart, Message, PermissionMode, Provider, ThinkingLevel } from "../types.js";
 export interface BubbleSdkOptions {
     /** Fallback working directory for sessions created without an explicit cwd. */
     defaultCwd?: string;
@@ -66,6 +67,14 @@ export interface RunTurnOptions extends TurnHandlers {
      * harness) A/B prompt variants without forking the prompt builder.
      */
     appendSystemPrompt?: string;
+    /**
+     * Creates the provider instance for a subagent route whose provider differs
+     * from the turn's provider (e.g. spawn_agent with `model: "provider:model"`).
+     * When omitted the SDK builds a default factory over its own registry —
+     * same semantics as the TUI's createProviderForRoute (OAuth prepare, then
+     * configured+enabled+keyed profile lookup). Pass your own to override.
+     */
+    providerFactory?: (route: ResolvedSubagentRoute) => Provider | Promise<Provider>;
     signal?: AbortSignal;
 }
 export interface SdkSessionRef {
@@ -106,6 +115,12 @@ export declare class BubbleSdk {
     };
     runTurn(sessionId: string, options: RunTurnOptions): AsyncGenerator<AgentEvent>;
     private bashAllowlistFor;
+    /**
+     * Default cross-provider subagent factory over this SDK instance's own
+     * registry — the same semantics as the TUI's createProviderForRoute. The
+     * registry is reloaded first so a key added after construction works.
+     */
+    private defaultProviderFactory;
     /** MCP servers are started lazily, once per cwd (McpManager has no stop). */
     private mcpToolsFor;
     private resolveSession;
@@ -149,4 +164,4 @@ export { ExternalHookController } from "../hooks/controller.js";
 export { calculateUsageCost, getModelPricing, type ModelPricing, type PricingCurrency, } from "../model-pricing.js";
 export { registry as slashCommandRegistry } from "../slash-commands/index.js";
 export type { SlashCommandContext } from "../slash-commands/types.js";
-export type { AgentEvent, ContentPart, Message, PermissionMode, Provider, ThinkingLevel, Todo, TokenUsage, UsageCost, ToolRegistryEntry, ToolResult, ToolUpdate, } from "../types.js";
+export type { AgentEvent, ContentPart, Message, PermissionMode, Provider, ThinkingLevel, TokenUsage, UsageCost, ToolRegistryEntry, ToolResult, ToolUpdate, } from "../types.js";

package/dist/sdk/index.js

@@ -171,10 +171,6 @@ export class BubbleSdk {
                 approvalController,
                 fileStateTracker,
                 planController,
-                todoStore: {
-                    getTodos: () => agentRef?.getTodos() ?? [],
-                    setTodos: (todos) => agentRef?.setTodos(todos),
-                },
                 questionController,
                 goalStore: new GoalStore(),
                 checkpoints: () => session.getCheckpoints(),
@@ -214,12 +210,17 @@ export class BubbleSdk {
                 temperature: 0.2,
                 thinkingLevel,
                 mode,
-                todos: session.getTodos(),
                 budgetLedger: new BudgetLedger(),
                 fileStateTracker,
                 skills: skillRegistry.summaries(),
                 memoryPrompt,
                 externalHooks: hookController,
+                // Cross-provider subagent routes (spawn_agent with "provider:model")
+                // need a factory; without one the child blocks with
+                // "no provider factory is configured". Default mirrors the TUI's
+                // createProviderForRoute (OAuth prepare + configured+enabled+keyed
+                // profile lookup); hosts may override per turn.
+                providerFactory: options.providerFactory ?? this.defaultProviderFactory(),
                 onMessageAppend: (message) => {
                     if (message.role === "system" || message.role === "meta")
                         return;
@@ -230,7 +231,6 @@ export class BubbleSdk {
                 onCompactionApplied: (summary) => {
                     session.applyLLMCompaction(summary);
                 },
-                onTodosUpdate: (todos) => session.appendTodosSnapshot(todos),
                 onModeUpdate: (m) => session.appendMarker("mode_switch", m),
             });
             agentRef = agent;
@@ -276,6 +276,37 @@ export class BubbleSdk {
         }
         return allowlist;
     }
+    /**
+     * Default cross-provider subagent factory over this SDK instance's own
+     * registry — the same semantics as the TUI's createProviderForRoute. The
+     * registry is reloaded first so a key added after construction works.
+     */
+    defaultProviderFactory() {
+        return async (route) => {
+            const providerId = route.providerId;
+            if (!providerId) {
+                throw new Error(`Subagent route for model "${route.model}" did not include a provider.`);
+            }
+            // getConfigured() reads config.json from disk every call, so keys
+            // added after SDK construction are picked up without a reload call.
+            if (this.registry.supportsOAuth(providerId) && this.registry.getAuthStorage().has(providerId)) {
+                await this.registry.prepareProvider(providerId);
+            }
+            const target = this.registry.getConfigured().find((item) => item.id === providerId);
+            if (!target?.enabled || !target.apiKey) {
+                throw new Error(`Subagent route requires provider "${providerId}", but it is not configured or has no active credentials.`);
+            }
+            return createProviderInstance({
+                providerId,
+                apiKey: target.apiKey,
+                baseURL: target.baseURL,
+                protocol: target.protocol,
+                headers: target.headers,
+                openAICodexAuth: this.registry.createOpenAICodexAuthAdapter(providerId),
+                grokAuth: this.registry.createGrokAuthAdapter(providerId),
+            });
+        };
+    }
     /** MCP servers are started lazily, once per cwd (McpManager has no stop). */
     mcpToolsFor(cwd) {
         if (!this.mcpEnabled)

package/dist/session-log.d.ts

@@ -1,5 +1,5 @@
-import type { Message, Todo } from "./types.js";
-import type { SessionLogEntry, SessionMarkerKind, SessionMetadata, SessionSummaryEntry, SessionTodosSnapshotEntry } from "./session-types.js";
+import type { Message } from "./types.js";
+import type { SessionLogEntry, SessionMarkerKind, SessionMetadata, SessionSummaryEntry } from "./session-types.js";
 export declare class SessionLog {
     private entries;
     load(lines: string[]): void;
@@ -9,8 +9,6 @@ export declare class SessionLog {
     setMetadata(metadata: SessionMetadata): SessionLogEntry[];
     appendMessage(message: Message): SessionLogEntry[];
     appendSummary(summary: string): SessionSummaryEntry;
-    appendTodosSnapshot(todos: Todo[]): SessionTodosSnapshotEntry;
-    getTodos(): Todo[];
     appendMarker(kind: SessionMarkerKind, value: string): SessionLogEntry;
     toMessages(): Message[];
 }

package/dist/session-log.js

@@ -60,28 +60,6 @@ export class SessionLog {
         this.entries.push(entry);
         return entry;
     }
-    appendTodosSnapshot(todos) {
-        const entry = {
-            id: nextEntryId(this.entries),
-            type: "todos_snapshot",
-            todos: todos.map((todo) => ({ ...todo })),
-            timestamp: Date.now(),
-        };
-        this.entries.push(entry);
-        return entry;
-    }
-    getTodos() {
-        for (let i = this.entries.length - 1; i >= 0; i--) {
-            const entry = this.entries[i];
-            if (entry.type === "marker" && entry.kind === "conversation_clear") {
-                return [];
-            }
-            if (entry.type === "todos_snapshot") {
-                return entry.todos.map((todo) => ({ ...todo }));
-            }
-        }
-        return [];
-    }
     appendMarker(kind, value) {
         const entry = {
             id: nextEntryId(this.entries),
@@ -239,7 +217,6 @@ function isSessionLogEntry(entry) {
         "assistant_message",
         "tool_call",
         "tool_result",
-        "todos_snapshot",
     ].includes(entry.type);
 }
 function nextEntryId(entries) {

package/dist/session-types.d.ts

@@ -1,4 +1,4 @@
-import type { AssistantMessage, Message, ThinkingLevel, Todo, ToolCall, ToolMessage, UserMessage } from "./types.js";
+import type { AssistantMessage, Message, ThinkingLevel, ToolCall, ToolMessage, UserMessage } from "./types.js";
 import type { GoalState } from "./goal/store.js";
 export interface ExternalRuntimeSessionMetadata {
     /** External agent runtime that owns the conversational state. */
@@ -65,11 +65,7 @@ export interface SessionToolResultEntry extends BaseSessionLogEntry {
     type: "tool_result";
     message: ToolMessage;
 }
-export interface SessionTodosSnapshotEntry extends BaseSessionLogEntry {
-    type: "todos_snapshot";
-    todos: Todo[];
-}
-export type SessionLogEntry = SessionMetadataEntry | SessionSummaryEntry | SessionMarkerEntry | SessionUserMessageEntry | SessionAssistantMessageEntry | SessionToolCallEntry | SessionToolResultEntry | SessionTodosSnapshotEntry;
+export type SessionLogEntry = SessionMetadataEntry | SessionSummaryEntry | SessionMarkerEntry | SessionUserMessageEntry | SessionAssistantMessageEntry | SessionToolCallEntry | SessionToolResultEntry;
 export interface LegacySessionEntry {
     id: string;
     type: "metadata" | "message" | "compaction";

package/dist/session.d.ts

@@ -3,7 +3,7 @@
  */
 import { CheckpointStore } from "./checkpoints.js";
 import { type CompactOptions, type CompactResult } from "./context/compact.js";
-import type { Message, Todo } from "./types.js";
+import type { Message } from "./types.js";
 import type { SessionLogEntry, SessionMarkerKind, SessionMetadata } from "./session-types.js";
 export interface SessionSummary {
     file: string;
@@ -55,8 +55,6 @@ export declare class SessionManager {
     appendMessage(message: Message): void;
     appendCompaction(summary: string): void;
     appendMarker(kind: SessionMarkerKind, value: string): void;
-    appendTodosSnapshot(todos: Todo[]): void;
-    getTodos(): Todo[];
     compact(options?: CompactOptions): CompactResult;
     /**
      * Inspect whether the session is large enough to compact and, if so, return

package/dist/session.js

@@ -152,14 +152,6 @@ export class SessionManager {
         const entry = this.log.appendMarker(kind, value);
         this.persist(entry);
     }
-    appendTodosSnapshot(todos) {
-        const entry = this.log.appendTodosSnapshot(todos);
-        this.persist(entry);
-        this.maybeAutoCompact();
-    }
-    getTodos() {
-        return this.log.getTodos();
-    }
     compact(options) {
         const result = compactSessionEntries(this.log.list(), options);
         if (result.compacted && result.entries) {

package/dist/slash-commands/commands.js

@@ -16,7 +16,7 @@ import { collectUsageStatsBundle, formatStatsText } from "../stats/usage.js";
 import { buildMemoryPrompt, getMemoryStatus, isMemoryDisabled, resetMemory, searchMemory, } from "../memory/index.js";
 import { feishuCommand } from "./feishu.js";
 import { GROK_LOCAL_COMMAND_HELP } from "../external-runtime/grok-input-policy.js";
-import { isGrokSubscriptionProviderId, } from "../external-runtime/grok-provider.js";
+import { isGrokSubscriptionProviderId } from "../external-runtime/grok-provider.js";
 import { classifyExternalRuntimeBinding } from "../external-runtime/session-policy.js";
 const VALID_SCOPES = ["user", "project", "local"];
 const VALID_LISTS = ["allow", "deny"];
@@ -107,7 +107,7 @@ function persistSelectedModel(model, ctx) {
     }
 }
 function syncSystemPrompt(ctx, model) {
-    const { providerId, modelId } = decodeModel(model);
+    const { providerId } = decodeModel(model);
     const toolPromptOptions = typeof ctx.agent.getSystemPromptToolOptions === "function"
         ? ctx.agent.getSystemPromptToolOptions()
         : {};
@@ -412,14 +412,14 @@ const builtinSlashCommandEntries = [
     {
         name: "skills",
         description: "Open the searchable skills picker",
-        async handler(args, ctx) {
+        async handler(_args, ctx) {
             ctx.openPicker("skill");
         },
     },
     {
         name: "help",
         description: "Show available slash commands",
-        async handler(args, ctx) {
+        async handler(_args, ctx) {
             if (hasExternalRuntimeSession(ctx)) {
                 const grok = isGrokSessionActive(ctx);
                 return [
@@ -448,14 +448,14 @@ const builtinSlashCommandEntries = [
     {
         name: "context",
         description: "Show current context window usage and breakdown",
-        async handler(args, ctx) {
+        async handler(_args, ctx) {
             return `${formatContextUsage(ctx.agent.getContextUsageSnapshot())}\n\n${formatMcpContextStatus(ctx)}`;
         },
     },
     {
         name: "quit",
         description: "Exit the application",
-        async handler(args, ctx) {
+        async handler(_args, ctx) {
             ctx.exit();
         },
     },
@@ -482,13 +482,14 @@ const builtinSlashCommandEntries = [
     {
         name: "clear",
         description: "Clear the current conversation history",
-        async handler(args, ctx) {
+        async handler(_args, ctx) {
             ctx.agent.messages = ctx.agent.messages.filter((m) => m.role === "system" || m.role === "meta");
+            // The resident history just shrank: drop the incremental usage anchor
+            // too (same as /compact and /rewind), or budget accounting keeps
+            // measuring against the pre-clear message count.
+            ctx.agent.resetContextUsageAnchor();
             ctx.sessionManager?.appendMarker("conversation_clear", "");
             ctx.sessionManager?.clearTitleMetadata?.();
-            if (ctx.agent.getTodos().length > 0) {
-                ctx.agent.setTodos([]);
-            }
             ctx.clearMessages();
         },
     },
@@ -559,7 +560,6 @@ const builtinSlashCommandEntries = [
                 session.rewindToEntry(target.id);
                 const head = ctx.agent.messages.filter((m) => m.role === "system" || m.role === "meta");
                 ctx.agent.messages = [...head, ...session.getMessages()];
-                ctx.agent.setTodos(session.getTodos());
                 ctx.agent.resetContextUsageAnchor();
                 if (ctx.fillComposer) {
                     // Put the rewound message back into the input box for re-editing.
@@ -845,7 +845,7 @@ const builtinSlashCommandEntries = [
     {
         name: "plan",
         description: "Toggle plan mode on/off (Tab switches Build/Plan)",
-        async handler(args, ctx) {
+        async handler(_args, ctx) {
             const next = ctx.agent.mode === "plan" ? "default" : "plan";
             ctx.agent.setMode(next);
             return next === "plan"
@@ -1069,7 +1069,7 @@ const builtinSlashCommandEntries = [
     {
         name: "compact",
         description: "Manually compact the current session context",
-        async handler(args, ctx) {
+        async handler(_args, ctx) {
             if (!ctx.sessionManager) {
                 return "Compaction requires session persistence. Start an interactive session first.";
             }

package/dist/tools/child-tools.d.ts

@@ -33,7 +33,7 @@ export declare class WorktreeApprovalController implements ApprovalController {
  * exposes a cloneForChild hook (the standard `read`, which carries a
  * FileStateTracker) is rebuilt as a fresh per-child instance, so concurrent
  * members of a fan-out never share mutable tool state. Stateless tools
- * (glob/grep, web/memory/skill/todo) and custom/mock tools without the hook are
+ * (glob/grep, web/memory/skill) and custom/mock tools without the hook are
  * passed through unchanged. Write children get full isolation via
  * createWorktreeChildTools instead.
  */

package/dist/tools/child-tools.js

@@ -97,7 +97,7 @@ const WORKTREE_TOOL_NAMES = new Set(["read", "glob", "grep", "edit", "write", "b
  * exposes a cloneForChild hook (the standard `read`, which carries a
  * FileStateTracker) is rebuilt as a fresh per-child instance, so concurrent
  * members of a fan-out never share mutable tool state. Stateless tools
- * (glob/grep, web/memory/skill/todo) and custom/mock tools without the hook are
+ * (glob/grep, web/memory/skill) and custom/mock tools without the hook are
  * passed through unchanged. Write children get full isolation via
  * createWorktreeChildTools instead.
  */

package/dist/tools/grep.js

@@ -54,7 +54,7 @@ export function createGrepTool(cwd) {
             }
             rgArgs.push(searchPath);
             return new Promise((resolve) => {
-                execFile("rg", rgArgs, { cwd, maxBuffer: 10 * 1024 * 1024 }, (error, stdout, stderr) => {
+                execFile("rg", rgArgs, { cwd, maxBuffer: 10 * 1024 * 1024 }, (_error, stdout, _stderr) => {
                     // rg returns exit code 1 when no matches found, which is not an error for us
                     const lines = stdout.split("\n").filter((l) => l.trim() !== "");
                     const matches = [];

package/dist/tools/index.d.ts

@@ -15,7 +15,6 @@ export { createWebSearchTool } from "./web-search.js";
 export { createSkillTool } from "./skill.js";
 export { createSkillSearchTool } from "./skill-search.js";
 export { createAgentLifecycleTools, createCloseAgentTool, createSendInputTool, createSpawnAgentTool, createWaitAgentTool } from "./agent-lifecycle.js";
-export { createTodoTool, type TodoStore } from "./todo.js";
 export { createExitPlanModeTool, type PlanController } from "./exit-plan-mode.js";
 export { createToolSearchTool, type ToolSearchController } from "./tool-search.js";
 export { createQuestionTool } from "./question.js";
@@ -25,7 +24,6 @@ import type { ApprovalController } from "../approval/types.js";
 import type { SkillRegistry } from "../skills/registry.js";
 import { type PlanController } from "./exit-plan-mode.js";
 import { type LspService } from "../lsp/index.js";
-import { type TodoStore } from "./todo.js";
 import { type ToolSearchController } from "./tool-search.js";
 import type { QuestionController } from "../question/index.js";
 import type { CheckpointStore } from "../checkpoints.js";
@@ -34,7 +32,6 @@ import type { GoalStore } from "../goal/store.js";
 import type { ProcessManager } from "../tasks/manager.js";
 import type { PromotionChannel } from "../tasks/promotion.js";
 export interface CreateAllToolsOptions {
-    todoStore?: TodoStore;
     planController?: PlanController;
     approvalController?: ApprovalController;
     questionController?: QuestionController;

package/dist/tools/index.js

@@ -15,7 +15,6 @@ export { createWebSearchTool } from "./web-search.js";
 export { createSkillTool } from "./skill.js";
 export { createSkillSearchTool } from "./skill-search.js";
 export { createAgentLifecycleTools, createCloseAgentTool, createSendInputTool, createSpawnAgentTool, createWaitAgentTool } from "./agent-lifecycle.js";
-export { createTodoTool } from "./todo.js";
 export { createExitPlanModeTool } from "./exit-plan-mode.js";
 export { createToolSearchTool } from "./tool-search.js";
 export { createQuestionTool } from "./question.js";
@@ -32,7 +31,6 @@ import { createReadTool } from "./read.js";
 import { createSkillTool } from "./skill.js";
 import { createSkillSearchTool } from "./skill-search.js";
 import { createAgentLifecycleTools } from "./agent-lifecycle.js";
-import { createTodoTool } from "./todo.js";
 import { createToolSearchTool } from "./tool-search.js";
 import { createWebFetchTool } from "./web-fetch.js";
 import { createWebSearchTool } from "./web-search.js";
@@ -71,7 +69,6 @@ export function createAllTools(cwd, skillRegistry, options = {}) {
         createToolSearchTool(options.toolSearchController),
         ...(options.questionController ? [createQuestionTool(options.questionController)] : []),
         ...(skillRegistry ? [createSkillSearchTool(skillRegistry), createSkillTool(skillRegistry)] : []),
-        ...(options.todoStore ? [createTodoTool(options.todoStore)] : []),
         ...(options.planController ? [createExitPlanModeTool(options.planController)] : []),
         ...(options.goalStore ? createGoalTools(options.goalStore) : []),
     ];