muonroi-cli 1.3.4 → 1.4.1

package/dist/src/orchestrator/orchestrator.d.ts

@@ -1,10 +1,11 @@
 import type { ModelMessage } from "ai";
-import type { ProviderId } from "../providers/types.js";
+import { type ProviderId } from "../providers/types.js";
 import { type ScheduleDaemonStatus, type StoredSchedule } from "../tools/schedule.js";
 import type { AgentMode, ChatEntry, SessionInfo, SessionSnapshot, StreamChunk, SubagentStatus, TaskRequest, ToolResult, VerifyRecipe } from "../types/index.js";
 import { type SandboxMode, type SandboxSettings } from "../utils/settings.js";
 import { type SideQuestionResult } from "../utils/side-question.js";
 import { type AgentOptions, type ProcessMessageObserver } from "./agent-options.js";
+import type { ToolLoopCapAsk } from "./tool-loop-cap.js";
 export declare class Agent {
     private provider;
     private providerId;
@@ -127,6 +128,8 @@ export declare class Agent {
     askSideQuestion(question: string, signal?: AbortSignal): Promise<SideQuestionResult>;
     abort(): void;
     cleanup(): Promise<void>;
+    private _toolLoopCapHandler;
+    setToolLoopCapHandler(fn: ToolLoopCapAsk | null): void;
     respondToToolApproval(approvalId: string, approved: boolean): void;
     clearHistory(): Promise<void>;
     startNewSession(): SessionSnapshot | null;
@@ -177,8 +180,6 @@ export declare class Agent {
     private postTurnCompact;
     respondToCouncilQuestion(questionId: string, answer: string): void;
     respondToCouncilPreflight(preflightId: string, approved: boolean): void;
-    private _createQuestionResponder;
-    private _createPreflightResponder;
     runCouncilV2(topic: string, options?: {
         skipClarification?: boolean;
         observer?: ProcessMessageObserver;
@@ -192,10 +193,13 @@ export declare class Agent {
             maxCost: number;
             maxSprints: number;
             doneThreshold: number;
+            budgetTokens?: number;
             stack?: string;
             noCustomerDebate?: boolean;
             noPriorContext?: boolean;
             forceCouncil?: boolean;
+            mode?: "maintain" | "new";
+            ghPr?: boolean;
         };
     }, options?: {
         observer?: ProcessMessageObserver;

package/dist/src/orchestrator/orchestrator.js

@@ -1,17 +1,19 @@
 // Multi-provider wired — runtime dispatch via providers/runtime.ts.
 import { extractSession } from "../ee/extract-session.js";
-import { bootstrapEEClient, getDefaultEEClient, getLastSurfacedState } from "../ee/intercept.js";
+import { bootstrapEEClient, getDefaultEEClient, getLastSurfacedState, updateLastSurfacedState, } from "../ee/intercept.js";
 import { getTenantId } from "../ee/tenant.js";
+import { emitTranscriptToDisk } from "../ee/transcript-emit.js";
 import { createRun, getActiveRunId, setActiveRunId } from "../flow/run-manager.js";
 import { ensureFlowDir } from "../flow/scaffold.js";
 import { executeEventHooks } from "../hooks/index.js";
 import { shutdownWorkspaceLspManager } from "../lsp/runtime.js";
 import { ensureDefaultMcpServers } from "../mcp/auto-setup.js";
-import { getModelByTier, getModelInfo, normalizeModelId } from "../models/registry.js";
+import { getModelInfo, normalizeModelId } from "../models/registry.js";
 import { getProviderCapabilities } from "../providers/capabilities.js";
 import { apiBaseFor } from "../providers/endpoints.js";
 import { loadKeyForProvider } from "../providers/keychain.js";
-import { createProviderFactory, createProviderFactoryAsync, detectProviderForModel, resolveModelRuntime as resolveRuntime, } from "../providers/runtime.js";
+import { createProviderFactory, createProviderFactoryAsync, detectProviderForModel, requireRuntimeProvider, resolveModelRuntime as resolveRuntime, } from "../providers/runtime.js";
+import { ALL_PROVIDER_IDS } from "../providers/types.js";
 import { appendCompaction, appendMessages, appendSystemMessage, buildChatEntries, getNextMessageSequence, getSessionTotalTokens, loadTranscript, loadTranscriptState, logInteraction, markMessageCompleted, recordUsageEvent, SessionStore, } from "../storage/index.js";
 import { BashTool } from "../tools/bash.js";
 import { createBuiltinTools } from "../tools/registry.js";
@@ -38,6 +40,7 @@ import { buildSystemPrompt, MAX_TOOL_ROUNDS } from "./prompts.js";
 import { getReadPathBudgetCap, ReadPathBudget } from "./read-path-budget.js";
 import { withStreamRetry } from "./retry-stream.js";
 import { StreamRunner } from "./stream-runner.js";
+import { resolveModelForTask } from "./sub-agent-model-tier.js";
 import { setProviderHint } from "./token-counter.js";
 import { firstLine, formatSubagentActivity, toToolResult } from "./tool-utils.js";
 // ---------------------------------------------------------------------------
@@ -49,6 +52,20 @@ import { firstLine, formatSubagentActivity, toToolResult } from "./tool-utils.js
 function createProvider(providerId, apiKey, baseURL) {
     return createProviderFactory(providerId, { apiKey, baseURL }).factory;
 }
+/**
+ * True iff `url` equals the default apiBase of ANY registered provider.
+ * Used to detect stale carryover of one provider's default URL into another
+ * provider's factory after a /model switch (see setModel + setApiKey).
+ */
+function isAnyProviderApiBase(url) {
+    if (!url)
+        return false;
+    for (const id of ALL_PROVIDER_IDS) {
+        if (url === apiBaseFor(id))
+            return true;
+    }
+    return false;
+}
 /**
  * Generate a session title using the Anthropic provider.
  * Kept as a lightweight stub for Phase 0 — title generation ships in Phase 1.
@@ -57,7 +74,7 @@ function genTitle(_provider, userMessage) {
     // Phase 0 stub: return a truncated version of the first user message as title.
     // Phase 1 will replace this with a real LLM-based title generation call.
     const title = userMessage.slice(0, 60).trim() || "New session";
-    return Promise.resolve({ title, modelId: DEFAULT_MODEL });
+    return Promise.resolve({ title, modelId: getCurrentModel() });
 }
 /**
  * Resolve a model ID to a runnable AI SDK LanguageModel.
@@ -66,7 +83,6 @@ function genTitle(_provider, userMessage) {
 function resolveModelRuntime(provider, modelId) {
     return resolveRuntime(provider, modelId);
 }
-const DEFAULT_MODEL = "claude-sonnet-4-6";
 async function toolSetToBatchTools(_tools) {
     // Batch API not supported with Anthropic in Phase 0. Phase 1 may add this.
     throw new Error("Batch API not available in Phase 0. Use standard streaming mode.");
@@ -102,7 +118,7 @@ function createTools(_bash, _provider, _mode, _opts) {
 // ============================================================================
 export class Agent {
     provider = null;
-    providerId = "anthropic";
+    providerId = null;
     apiKey = null;
     baseURL = null;
     bash;
@@ -281,9 +297,22 @@ export class Agent {
         if (newProviderId !== this.providerId && this.apiKey) {
             this.providerId = newProviderId;
             setProviderHint(this.providerId);
-            const effectiveBaseURL = this.providerId !== "anthropic" && this.baseURL === apiBaseFor("anthropic")
-                ? undefined
-                : (this.baseURL ?? undefined);
+            // Drop this.baseURL when it points at a DIFFERENT provider's default
+            // apiBase — otherwise the rebuilt factory binds the new provider's
+            // strategy to the OLD provider's URL, sending requests to the wrong
+            // host. Evidence: session 2492d6579b1d — user switched defaultProvider
+            // siliconflow→ (via UI), this.baseURL was still api.deepseek.com from
+            // startup, SiliconflowStrategy.createFactory was created with that
+            // baseURL, requests landed at api.deepseek.com which rejected the SF-
+            // style model id ("deepseek-ai/DeepSeek-V4-Flash") with "supported API
+            // model names are deepseek-v4-pro or deepseek-v4-flash".
+            // A user-supplied custom baseURL is preserved only when it does NOT
+            // match any known provider's apiBase (i.e. it's a real override, not
+            // a stale default).
+            const staleBaseURL = isAnyProviderApiBase(this.baseURL) && this.baseURL !== apiBaseFor(this.providerId);
+            const effectiveBaseURL = staleBaseURL ? undefined : (this.baseURL ?? undefined);
+            if (staleBaseURL)
+                this.baseURL = null;
             this.provider = createProvider(this.providerId, this.apiKey, effectiveBaseURL);
         }
         if (this.sessionStore && this.session) {
@@ -331,10 +360,14 @@ export class Agent {
     }
     setApiKey(apiKey, baseURL) {
         this.apiKey = apiKey;
-        this.baseURL = baseURL || null;
-        // Only pass baseURL to provider factory if it's an explicit override,
-        // not the default Anthropic URL (which would break non-Anthropic providers).
-        const effectiveBaseURL = this.providerId !== "anthropic" && baseURL === apiBaseFor("anthropic") ? undefined : baseURL;
+        // Drop baseURL when it points at a DIFFERENT provider's default apiBase
+        // (e.g. caller passed the legacy anthropic URL while providerId is
+        // siliconflow — without this we'd send siliconflow requests to
+        // api.anthropic.com or similar). User-supplied custom URLs that don't
+        // match any known provider's apiBase are preserved as real overrides.
+        const stale = isAnyProviderApiBase(baseURL) && baseURL !== apiBaseFor(this.providerId);
+        const effectiveBaseURL = stale ? undefined : baseURL;
+        this.baseURL = stale ? null : baseURL || null;
         this.provider = createProvider(this.providerId, apiKey, effectiveBaseURL);
     }
     setProviderAndKey(providerId, apiKey, baseURL) {
@@ -424,12 +457,29 @@ export class Agent {
         this.emitSubagentStatus(null);
     }
     async cleanup() {
+        // Slow-path sidecar — write transcript to disk SYNCHRONOUSLY before
+        // racing the async HTTP extract. Survives X-close + EE-offline.
+        try {
+            emitTranscriptToDisk(this.messages, this.getSessionId(), "cli-exit", this.bash.getCwd());
+        }
+        catch {
+            /* fail-open */
+        }
         await Promise.allSettled([
             this.bash.cleanup(),
             shutdownWorkspaceLspManager(this.bash.getCwd()),
             extractSession(this.messages, this.bash.getCwd(), "cli-exit", this.getSessionId()),
         ]);
     }
+    // Tool-loop cap handler — set by the UI (app.tsx) at startup. Invoked from
+    // the message-processor streamText loop when stepCount reaches the current
+    // cap. The UI surfaces an askcard ("Continue +50? / Stop and answer") and
+    // resolves with the verdict. When unset, the loop stops gracefully — no
+    // hard-throw, matches the user-friendly behaviour we promised.
+    _toolLoopCapHandler = null;
+    setToolLoopCapHandler(fn) {
+        this._toolLoopCapHandler = fn;
+    }
     respondToToolApproval(approvalId, approved) {
         const toolApprovalResponse = {
             role: "tool",
@@ -446,6 +496,12 @@ export class Agent {
     }
     async clearHistory() {
         // D-09: Extract messages accumulated since last clear BEFORE reset
+        try {
+            emitTranscriptToDisk(this.messages, this.getSessionId(), "cli-clear", this.bash.getCwd());
+        }
+        catch {
+            /* fail-open */
+        }
         await extractSession(this.messages, this.bash.getCwd(), "cli-clear", this.getSessionId()).catch(() => { }); // D-05: redundant safety — extractSession already swallows
         this.startNewSession();
     }
@@ -478,6 +534,21 @@ export class Agent {
         }
         this.sessionStore = new SessionStore(this.bash.getCwd());
         this.workspace = this.sessionStore.getWorkspace();
+        // Collapse double startNewSession() calls into a single row. The current
+        // session is REUSED (not orphaned with a fresh createSession) when it is
+        // brand-new and empty — no persisted messages in memory AND no title.
+        // Root cause of ~60% orphaned session rows (1187/1966): the /clear slash
+        // path calls clearHistory() — which already starts a new session — and then
+        // resetToNewSession(), which started ANOTHER. The first session never
+        // received any work and was left title-less and empty. The first call here
+        // still sees the prior conversation in `this.messages` (non-empty → new
+        // row, correct); the immediate second call sees the just-cleared empty
+        // session and reuses it instead of creating a twin.
+        const cur = this.session;
+        if (cur && this.messages.length === 0 && !cur.title) {
+            this.messageSeqs = [];
+            return this.getSessionSnapshot();
+        }
         this.session = this.sessionStore.createSession(this.modelId, this.mode, this.bash.getCwd());
         this.messages = [];
         this.messageSeqs = [];
@@ -567,7 +638,17 @@ export class Agent {
             }
         }
     }
-    recordUsage(usage, source = "message", model = this.modelId) {
+    recordUsage(usage, source = "message", model = this.modelId, 
+    /**
+     * Phase O1 — the providerOptions shape of the call that produced THIS event,
+     * threaded explicitly per call. Multi-step turns emit one event per step and
+     * a `task` sub-agent can run mid-turn, so a single mutable
+     * `_lastProviderOptionsShape` corrupted later events (the clear nulled it
+     * after step 1; an interleaved task overwrote it). When omitted (title /
+     * other one-shot calls that set no shape), fall back to the mutable field,
+     * which is cleared after each message event so they record null, not stale.
+     */
+    providerOptionsShape) {
         if (!usage)
             return;
         if (this.session) {
@@ -578,10 +659,12 @@ export class Agent {
             const lastSeq = lastPersistedSeq(this.messageSeqs);
             // Phase O1 — providerOptions shape (types only, no values) attached
             // to every usage event so post-mortem can answer "what provider
-            // options did this billed call carry?". Cleared below for "message"
-            // sources so non-message calls don't reuse stale data.
-            const providerOptionsShape = this._lastProviderOptionsShape;
-            recordUsageEvent(this.session.id, source, model, usage, lastSeq, pilActive, enrichmentDelta, providerOptionsShape);
+            // options did this billed call carry?". Prefer the explicitly-threaded
+            // shape (correct per-event, immune to the clear + task interleaving);
+            // fall back to the mutable field only when the caller passed nothing
+            // (title / other), which is cleared below so they record null, not stale.
+            const resolvedShape = providerOptionsShape !== undefined ? providerOptionsShape : this._lastProviderOptionsShape;
+            recordUsageEvent(this.session.id, source, model, usage, lastSeq, pilActive, enrichmentDelta, resolvedShape);
             if (source === "message") {
                 this._pilActive = false;
                 this._pilEnrichmentDelta = 0;
@@ -625,6 +708,12 @@ export class Agent {
         // API inputTokens includes cacheRead — subtract to get non-cached portion
         const nonCachedInput = Math.max(0, totalInput - cacheRead - cacheCreate);
         const turnCostMicros = nonCachedInput * priceIn + cacheRead * priceCached + cacheCreate * priceIn + output * priceOut;
+        // F5 — ctx_tokens reflects the CURRENT call's input size (≈ context
+        // window usage), not cumulative. Lets the user see "how full is my
+        // window" instead of "how much have I billed in total this session".
+        // Pair with context-fill % derived from model contextWindow.
+        const ctxWindow = info?.contextWindow ?? 0;
+        const ctxPct = ctxWindow > 0 ? Math.min(100, Math.round((totalInput / ctxWindow) * 100)) : undefined;
         statusBarStore.setState({
             in_tokens: prev.in_tokens + totalInput,
             out_tokens: prev.out_tokens + output,
@@ -633,6 +722,8 @@ export class Agent {
             session_usd: prev.session_usd + turnCostMicros / 1_000_000,
             provider: this.providerId,
             model,
+            ctx_tokens: totalInput,
+            ctx_pct: ctxPct,
         });
         // Append to cost-log JSONL so `usage report --by callsite` can surface
         // where orchestrator/task/title traffic is actually spending.
@@ -742,7 +833,7 @@ export class Agent {
     }
     async runTaskRequestBatch(args) {
         const { request, childMessages, childSystem, childRuntime, childTools, maxSteps, initialDetail, onActivity, signal, } = args;
-        const childCaps = getProviderCapabilities(childRuntime.modelInfo?.provider ?? "anthropic");
+        const childCaps = getProviderCapabilities(requireRuntimeProvider(childRuntime));
         if (childCaps.usesResponsesAPI(childRuntime.modelInfo)) {
             throw new Error("Batch mode currently supports chat-completions models only.");
         }
@@ -871,13 +962,14 @@ export class Agent {
             isBatchApiEnabled: () => this.batchApi,
             getCrossTurnDedup: () => this._crossTurnDedup,
             getReadBudget: () => this._readBudget,
-            recordUsage: (usage, source, model) => this.recordUsage(usage, source, model),
+            recordUsage: (usage, source, model, shape) => this.recordUsage(usage, source, model, shape),
             setCurrentCallId: (id) => {
                 this._currentCallId = id;
             },
             setLastProviderOptionsShape: (shape) => {
                 this._lastProviderOptionsShape = shape;
             },
+            getSessionId: () => this.session?.id,
             runTaskRequestBatch: (args) => this.runTaskRequestBatch(args),
         };
         const runner = new StreamRunner(deps);
@@ -1053,18 +1145,7 @@ export class Agent {
         return this._resolveModelForTask("compact");
     }
     _resolveModelForTask(task) {
-        const tierPrefs = {
-            compact: ["fast", "balanced"],
-            title: ["fast", "balanced"],
-            explore: ["balanced", "fast"],
-            general: ["premium", "balanced"],
-        };
-        for (const tier of tierPrefs[task] ?? ["balanced"]) {
-            const m = getModelByTier(tier, this.providerId);
-            if (m?.provider === this.providerId)
-                return m.id;
-        }
-        return this.modelId;
+        return resolveModelForTask(task, this.providerId, this.modelId);
     }
     async compactForContext(provider, system, contextWindow, signal, settings = this.getCompactionSettings(contextWindow), force = false) {
         if (!this.session)
@@ -1086,6 +1167,13 @@ export class Agent {
             })
                 .catch(() => { });
         }
+        // Emit pre-compact transcript snapshot so lessons survive the rewrite.
+        try {
+            emitTranscriptToDisk(this.messages, this.session?.id ?? null, "cli-compact", this.bash.getCwd());
+        }
+        catch {
+            /* fail-open */
+        }
         const preCompactInput = {
             hook_event_name: "PreCompact",
             trigger,
@@ -1143,6 +1231,24 @@ export class Agent {
         }
         this.messages = [createCompactionSummaryMessage(summary), ...pinnedReinjections, ...preparation.keptMessages];
         this.messageSeqs = [null, ...pinnedReinjectionSeqs, ...keptSeqs];
+        // EE anti-mù (Phase 1 of docs/ee-anti-mu-compaction-plan.md): immediately extract the fresh structured checkpoint summary
+        // so pilContext / layer3 search / ee.query can recall exact prior ✔ DONE items + progress for the rest of this long session
+        // (and sub-agents) even after further B3/B4 rewrites. Uses same client + fail-open pattern as promptStale above (1449).
+        // Transcript is the summary itself (not full history) to keep it small and focused on task state.
+        getDefaultEEClient()
+            .extract({
+            transcript: `[Context checkpoint summary]\n${summary}`,
+            projectPath: this.bash.getCwd(),
+            meta: {
+                source: "cli-compact-checkpoint",
+                sessionId: this.session?.id ?? undefined,
+                iteration: this._compactionStats.count + 1,
+                tokensBefore: preparation.tokensBefore,
+            },
+        }, AbortSignal.timeout(1500))
+            .catch(() => { });
+        // Mark as surfaced for prompt-stale reconciliation (per plan Phase 1).
+        updateLastSurfacedState([`compact-checkpoint-${this._compactionStats.count + 1}`]);
         // Track compaction stats — net of the tokens spent ON compaction itself.
         const tokensAfter = estimateConversationTokens(system, this.messages);
         const grossSaved = Math.max(0, preparation.tokensBefore - tokensAfter);
@@ -1232,15 +1338,6 @@ export class Agent {
     respondToCouncilPreflight(preflightId, approved) {
         this.councilManager.respondToPreflight(preflightId, approved);
     }
-    // Internal hooks used by orchestrator.agent.test.ts to exercise buffered
-    // question / preflight delivery through CouncilManager. They are reached
-    // through `as unknown as` casts in the test; keep them private but stable.
-    _createQuestionResponder() {
-        return this.councilManager.createQuestionResponder();
-    }
-    _createPreflightResponder() {
-        return this.councilManager.createPreflightResponder();
-    }
     // ========================================================================
     // Council v2 — Clarify → Confirm → Debate → Plan → Execute
     // ========================================================================
@@ -1314,6 +1411,7 @@ export class Agent {
                 maxCost: payload.flags.maxCost,
                 maxSprints: payload.flags.maxSprints,
                 doneThreshold: payload.flags.doneThreshold,
+                budgetTokens: payload.flags.budgetTokens,
                 stack: payload.flags.stack,
                 forceCouncil: payload.flags.forceCouncil,
             },
@@ -1321,9 +1419,14 @@ export class Agent {
             respondToPreflight: this.councilManager.createPreflightResponder(),
             cwd: this.bash.getCwd(),
             processMessageFn,
+            // Mode C — wire verify-recipe detector so runProductLoop auto-detect can probe cwd.
+            detectVerifyRecipe: () => this.detectVerifyRecipe(),
             skipPriorContext: payload.flags.noPriorContext === true,
             complexity,
             sufficiencyMissing,
+            // Mode C explicit override + gh pr create opt-in (see .planning/MAINTAIN-MODE.md).
+            mode: payload.flags.mode,
+            ghPr: payload.flags.ghPr === true,
             // Chat session id — used as the FK key for interaction_logs telemetry.
             // The /ideal runId is NOT a sessions.id and would silently fail FK insert.
             sessionId: this.session?.id,
@@ -1794,7 +1897,7 @@ export class Agent {
             executeBatchToolCall: (tools, toolCall, messages, signal) => self.executeBatchToolCall(tools, toolCall, messages, signal),
             appendCompletedTurn: (user, asst) => self.appendCompletedTurn(user, asst),
             discardAbortedTurn: (user) => self.discardAbortedTurn(user),
-            recordUsage: (usage, source, model) => self.recordUsage(usage, source, model),
+            recordUsage: (usage, source, model, shape) => self.recordUsage(usage, source, model, shape),
         };
     }
     appendCompletedTurn(userMessage, newMessages) {
@@ -1996,8 +2099,20 @@ export class Agent {
             listDelegations: () => self.listDelegations(),
             appendCompletedTurn: (user, asst) => self.appendCompletedTurn(user, asst),
             discardAbortedTurn: (user) => self.discardAbortedTurn(user),
-            recordUsage: (usage, source, model) => self.recordUsage(usage, source, model),
+            recordUsage: (usage, source, model, shape) => self.recordUsage(usage, source, model, shape),
             respondToToolApproval: (id, ok) => self.respondToToolApproval(id, ok),
+            askToolLoopContinue: async (info) => {
+                const h = self._toolLoopCapHandler;
+                if (!h)
+                    return "stop";
+                try {
+                    return await h(info);
+                }
+                catch (err) {
+                    console.error(`[Agent] askToolLoopContinue crashed: ${err?.message ?? err}`);
+                    return "stop";
+                }
+            },
             runCouncilV2: (msg, opts) => self.runCouncilV2(msg, opts),
             processMessage: (msg, obs, imgs) => self.processMessage(msg, obs, imgs),
             processMessageBatchTurn: (args) => self.processMessageBatchTurn(args),

package/dist/src/orchestrator/pending-calls.test.js

@@ -34,7 +34,7 @@ beforeEach(async () => {
 });
 afterEach(async () => {
     delete process.env.MUONROI_CLI_HOME;
-    await fs.rm(tmpBase, { recursive: true, force: true }).catch(() => {
+    await fs.rm(tmpBase, { recursive: true, force: true, maxRetries: 10, retryDelay: 50 }).catch(() => {
         /* best-effort cleanup */
     });
 });

package/dist/src/orchestrator/prompts.d.ts

@@ -1,6 +1,6 @@
 import type { AgentMode, TaskRequest } from "../types/index.js";
 import { type CustomSubagentConfig, type SandboxMode, type SandboxSettings } from "../utils/settings.js";
-export declare const MAX_TOOL_ROUNDS = 75;
+export declare const MAX_TOOL_ROUNDS: number;
 export declare const VISION_MODEL = "grok-4-1-fast-reasoning";
 export declare const COMPUTER_MODEL = "grok-4.20-0309-reasoning";
 export declare function findCustomSubagent(agent: string, subagents?: CustomSubagentConfig[]): CustomSubagentConfig | undefined;