open-multi-agent-kit 0.98.3 → 0.98.5

package/dist/core/agent-session.js

@@ -24,6 +24,7 @@ import { stripFrontmatter } from "../utils/frontmatter.js";
 import { resolvePath } from "../utils/paths.js";
 import { getShellConfig } from "../utils/shell.js";
 import { sleep } from "../utils/sleep.js";
+import { addActiveSkills, createActiveSkillState } from "./active-skill-state.js";
 import { createAdaptorchBridge } from "./adaptorch-bridge.js";
 import { applyCategoryTimeoutDefaults, resolveAgentToolSettings, resolveToolTimeoutCategory, } from "./agent-tool-settings.js";
 import { formatNoApiKeyFoundMessage, formatNoModelSelectedMessage } from "./auth-guidance.js";
@@ -43,7 +44,7 @@ import { exportSessionToHtml } from "./export-html/index.js";
 import { createToolHtmlRenderer } from "./export-html/tool-renderer.js";
 import { ExtensionRunner, wrapRegisteredTools, } from "./extensions/index.js";
 import { emitSessionShutdownEvent } from "./extensions/runner.js";
-import { assertTextChatModelForCompletion } from "./grok-harness.js";
+import { assertTextChatModelForCompletion, grokHarnessAutoApplyEnabled, isGrokOAuthProvider, selectGrokHarnessSkills, } from "./grok-harness.js";
 import { grokPlaybookAppendForProvider } from "./grok-playbook.js";
 import { captureHostResourceSnapshot } from "./host-resource-snapshot.js";
 import { decideLoadoutAccess } from "./loadout-access-policy.js";
@@ -102,14 +103,18 @@ import { resolveResourceGovernorSettings } from "./resource-governor-settings.js
 import { admissionObservationFacts, classificationObservationFacts, ResourceObservationJournal, settledObservationFacts, snapshotObservationFacts, soundObservationFacts, } from "./resource-observation-journal.js";
 import { decideResourceSafetyGate, RESOURCE_PRESSURE_REQUIRED_ACTION } from "./resource-safety-gate.js";
 import { appendRouterFeedbackRecord, ROUTER_FEEDBACK_LEVELS, } from "./router-feedback-collector.js";
+import { RunBudgetExceededError, RunBudgetPolicyError } from "./run-budget-policy.js";
 import { RunJournalStore } from "./run-journal-store.js";
 import { RunResourceLeaseController } from "./run-resource-lease.js";
 import { SessionBashRuntime } from "./session-bash-runtime.js";
 import { SessionBashService } from "./session-bash-service.js";
 import { SessionCompactionService } from "./session-compaction-service.js";
-import { preflightFailureCause, providerFailureCause, runtimeFailureCause, terminationMessage, } from "./session-failure-cause.js";
+import { preflightFailureCause, runtimeFailureCause, terminationMessage } from "./session-failure-cause.js";
 import { CURRENT_SESSION_VERSION, getLatestCompactionEntry } from "./session-manager.js";
 import { acquireSessionOwnerLeaseSync } from "./session-owner-lease.js";
+import { PromptExecutionBusyError, SessionPromptLifecycle } from "./session-prompt-lifecycle.js";
+import { SessionRunBudget } from "./session-run-budget.js";
+import { classifyRunTermination } from "./session-run-termination.js";
 import { assembleSessionSystemPrompt } from "./session-system-prompt.js";
 import { classifySessionTermination, } from "./session-termination.js";
 import { createSyntheticSourceInfo } from "./source-info.js";
@@ -136,17 +141,6 @@ export function parseSkillBlock(text) {
         userMessage: match[4]?.trim() || undefined,
     };
 }
-function mergePromptActiveSkillNames(first, second) {
-    const names = [];
-    const seen = new Set();
-    for (const name of [...first, ...second]) {
-        if (!seen.has(name)) {
-            seen.add(name);
-            names.push(name);
-        }
-    }
-    return names;
-}
 const PENDING_TOOL_RESULT_TOKENS = {
     text: 1024,
     image: 4096,
@@ -226,6 +220,11 @@ export class AgentSession {
     // Event subscription state
     _unsubscribeAgent;
     _eventListeners = [];
+    _runBudget;
+    _promptLifecycle = new SessionPromptLifecycle({
+        canSettle: () => !this.isStreaming && !this.agent.hasQueuedMessages(),
+        auditsLateSettlement: () => this.agent.toolExecutionPolicy?.lateSettlement !== "ignore",
+    });
     _messageEndReplacements = new WeakMap();
     /** Tracks pending steering messages for UI display. Removed when delivered. */
     _steeringMessages = [];
@@ -322,6 +321,21 @@ export class AgentSession {
     _promptCacheLastBreakReason;
     constructor(config) {
         this.agent = config.agent;
+        this._runBudget = new SessionRunBudget(this.agent, {
+            assertIdle: () => {
+                this._promptLifecycle.assertIdle();
+                if (this.isRetrying || this.isCompacting || this._branchSummaryAbortController)
+                    throw new PromptExecutionBusyError();
+            },
+            stop: (error) => {
+                this._pendingRuntimeTerminationCause = { area: "budget", code: error.code };
+                this.agent.abort();
+                this.abortRetry();
+                this.abortCompaction();
+                this.abortBranchSummary();
+            },
+            reject: (error) => this._publishRuntimeFailure(error),
+        });
         this.sessionManager = config.sessionManager;
         this.settingsManager = config.settingsManager;
         this._scopedModels = config.scopedModels ?? [];
@@ -515,64 +529,6 @@ export class AgentSession {
         });
         this._appendReplayEvent(event, details);
     }
-    _classifyRunTermination(runId, event) {
-        const timestamp = new Date().toISOString();
-        let assistant;
-        for (let index = event.messages.length - 1; index >= 0; index -= 1) {
-            const message = event.messages[index];
-            if (message?.role === "assistant") {
-                assistant = message;
-                break;
-            }
-        }
-        let cause;
-        let message;
-        let sideEffects = this._sessionRiskLevel === "elevated" ? "possible" : "none";
-        let toolCallId;
-        let toolName;
-        if (this._activeRunToolTermination) {
-            cause = { area: "tool", code: "timeout" };
-            message = `Tool ${this._activeRunToolTermination.toolName} timed out.`;
-            toolCallId = this._activeRunToolTermination.toolCallId;
-            toolName = this._activeRunToolTermination.toolName;
-            sideEffects = this._activeRunToolTermination.executionStarted ? "possible" : sideEffects;
-        }
-        else if (!assistant) {
-            cause = { area: "internal", code: "unclassified" };
-            message = "Agent run ended without an assistant result.";
-        }
-        else if (assistant.stopReason === "aborted") {
-            cause = this._userAbortRequested ? { area: "user", code: "abort" } : { area: "provider", code: "abort" };
-            message = terminationMessage(assistant.errorMessage, this._userAbortRequested ? "The user aborted the run." : "The provider aborted the run.");
-        }
-        else if (assistant.stopReason === "error") {
-            cause = providerFailureCause(assistant, this.model?.contextWindow ?? 0);
-            message = terminationMessage(assistant.errorMessage, "The provider request failed.");
-        }
-        else {
-            cause = { area: "completed" };
-            message = "Run completed.";
-        }
-        let provider = this.model?.provider;
-        if (assistant?.provider)
-            provider = assistant.provider;
-        let model = this.model?.id;
-        if (assistant?.model)
-            model = assistant.model;
-        return classifySessionTermination({
-            sessionId: this.sessionId,
-            runId,
-            timestamp,
-            source: "observed",
-            message,
-            cause,
-            sideEffects,
-            ...(provider ? { provider } : {}),
-            ...(model ? { model } : {}),
-            ...(toolCallId ? { toolCallId } : {}),
-            ...(toolName ? { toolName } : {}),
-        });
-    }
     _publishTermination(termination) {
         this._lastTermination = termination;
         this._emit({ type: "session_termination", termination });
@@ -647,11 +603,23 @@ export class AgentSession {
             }
             return;
         }
+        if (event.type === "provider_denied" && event.deniedReason === "contract-violation") {
+            this._pendingRuntimeTerminationCause = { area: "configuration", code: "invalid" };
+        }
         if (event.type !== "agent_end")
             return;
         if (this._activeRunId === null)
             throw new Error("run journal received agent_end without run_started");
-        const termination = this._classifyRunTermination(this._activeRunId, event);
+        const termination = classifyRunTermination(event.messages, {
+            sessionId: this.sessionId,
+            runId: this._activeRunId,
+            timestamp: new Date().toISOString(),
+            model: this.model,
+            elevatedRisk: this._sessionRiskLevel === "elevated",
+            userAbortRequested: this._userAbortRequested,
+            pendingCause: this._pendingRuntimeTerminationCause,
+            toolTimeout: this._activeRunToolTermination,
+        });
         try {
             this._runJournalStore.finish({
                 termination,
@@ -743,6 +711,7 @@ export class AgentSession {
         throw new Error(formatNoApiKeyFoundMessage(model.provider));
     }
     async _getCompactionRequestAuth(model) {
+        this._runBudget.assertAdmission();
         const options = { minRemainingMs: COMPACTION_MIN_TOKEN_VALIDITY_MS };
         if (isBuiltinStreamFn(this.agent.streamFn)) {
             return this._getRequiredRequestAuth(model, options);
@@ -897,6 +866,8 @@ export class AgentSession {
                 }
             }
         }
+        if (event.type === "tool_execution_late_settlement")
+            this._promptLifecycle.flush();
     };
     _willRetryAfterAgentEnd(event) {
         const settings = this.settingsManager.getRetrySettings();
@@ -1164,6 +1135,8 @@ export class AgentSession {
      * Call this when completely done with the session.
      */
     dispose() {
+        this._promptLifecycle.dispose();
+        this._runBudget.close();
         try {
             try {
                 this.abortRetry();
@@ -1290,7 +1263,7 @@ export class AgentSession {
         for (const name of desiredToolNames) {
             const tool = this._toolRegistry.get(name);
             if (tool) {
-                tools.push(tool);
+                tools.push(this._promptLifecycle.wrapTool(tool));
                 validToolNames.push(name);
             }
             else if (lockedToolNames?.includes(name)) {
@@ -1515,13 +1488,14 @@ export class AgentSession {
         // One identity per top-level run (§16.2): internal retries and
         // continuations inside this call share it, as does the resource lease.
         const promptRunId = `prompt-run-${randomUUID()}`;
-        const startedAtEpochMs = Date.now();
+        const ownedRun = this._promptLifecycle.begin(promptRunId);
         // Roadmap M2: the lease spans the whole run including internal retries
         // and continuations, so they share one admission decision (§8.3).
         const resourceLease = await this._beginResourceGovernedRun(promptRunId);
         const resourceObservations = this._resourceObservationJournals.get(promptRunId) ?? null;
         let outcome = "completed";
         try {
+            this._runBudget.assertActive();
             await this.agent.prompt(messages);
             while (await this._handlePostAgentRun()) {
                 await this.agent.continue();
@@ -1533,43 +1507,18 @@ export class AgentSession {
             throw error;
         }
         finally {
+            if (this._runBudget.failure && this._lastTermination?.kind !== "budget_exhausted")
+                this._publishRuntimeFailure(this._runBudget.failure);
             outcome = promptSettlement.resolvePromptSettlementOutcome(outcome, this._lastTermination?.kind);
-            if (resourceLease !== null) {
-                // Generation-safe: a stale release after another run acquired the
-                // lease is a no-op instead of clobbering the newer cap (§8.1).
-                this._resourceLeaseController?.release(resourceLease);
-                resourceObservations?.record("resource_lease_released_v1", { promptRunId });
-            }
             this._flushPendingBashMessages();
-            // §16.4 (M4): settle only after every in-run continuation drained.
-            this._emitPromptSettledIfReady({ promptRunId, startedAtEpochMs, outcome, resourceObservations });
-        }
-    }
-    /**
-     * Emit `prompt_settled` when the §16.4 conditions hold. Called exactly
-     * once per promptRunId (from `_runAgentPrompt`'s finally), so the same
-     * run can never emit twice; late-queued messages block emission
-     * conservatively — the continuation run settles under its own id.
-     * Settlement is a UX signal only and must never affect the run (§16.1).
-     */
-    _emitPromptSettledIfReady(input) {
-        try {
-            let state = promptSettlement.createPromptSettlementState(input.promptRunId, input.startedAtEpochMs);
-            state = promptSettlement.reducePromptSettlement(state, { kind: "terminal", outcome: input.outcome });
-            if (this.isStreaming) {
-                state = promptSettlement.reducePromptSettlement(state, { kind: "tool", delta: 1 });
-            }
-            if (this.agent.hasQueuedMessages()) {
-                state = promptSettlement.reducePromptSettlement(state, { kind: "continuation", delta: 1 });
-            }
-            const settled = promptSettlement.settlePromptIfReady(state, Date.now());
-            if (settled.event !== null) {
-                input.resourceObservations?.record("prompt_settled_v1", settledObservationFacts(settled.event));
-                this._emit(settled.event);
-            }
-        }
-        catch {
-            // Never let settlement bookkeeping break a run.
+            ownedRun.finish(outcome, (event) => {
+                if (resourceLease !== null) {
+                    this._resourceLeaseController?.release(resourceLease);
+                    resourceObservations?.record("resource_lease_released_v1", { promptRunId });
+                }
+                resourceObservations?.record("prompt_settled_v1", settledObservationFacts(event));
+                this._emit(event);
+            });
         }
     }
     /**
@@ -1581,7 +1530,7 @@ export class AgentSession {
      *   delayed. The pending probe is bounded (~cpuSampleMs) and never rejects.
      * - `adaptive`/`strict`: bounded blocking probe (default 300 ms deadline),
      *   then a generation-safe lease applies the effective tool cap for this
-     *   run. The caller's `finally` releases it exactly once.
+     *   run. Prompt settlement releases it after owned tool promises terminate.
      *
      * Never throws: any probe or policy failure leaves this run ungoverned
      * (§2.1: probe failure must not crash or block the prompt).
@@ -1776,6 +1725,7 @@ export class AgentSession {
         }
     }
     async _handlePostAgentRun() {
+        this._runBudget.assertActive();
         const msg = this._lastAssistantMessage;
         this._lastAssistantMessage = undefined;
         if (!msg) {
@@ -1810,17 +1760,20 @@ export class AgentSession {
      * @throws Error if streaming and no streamingBehavior specified
      * @throws Error if no model selected or no API key available (when not streaming)
      */
-    async prompt(text, options) {
+    prompt(text, options) {
+        if (options?.runBudget === undefined && (this.isStreaming || this.isRetrying))
+            return this._prompt(text, options);
+        return this._runBudget.execute(options?.runBudget, () => this._prompt(text, options), options?.preflightResult);
+    }
+    getRunBudgetSnapshot() {
+        return this._runBudget.snapshot();
+    }
+    async _prompt(text, options) {
         const expandPromptTemplates = options?.expandPromptTemplates ?? true;
         const preflightResult = options?.preflightResult;
         let currentText = redactSensitiveText(text);
         const defaultActiveSkills = this._getDefaultActiveSkills();
-        let promptActiveSkillNames = mergePromptActiveSkillNames(defaultActiveSkills, options?.activeSkillNames ?? []);
-        let promptActiveSkillSource = defaultActiveSkills.length > 0
-            ? options?.activeSkillSource
-                ? `settings+${options.activeSkillSource}`
-                : "settings"
-            : options?.activeSkillSource;
+        let promptSkills = createActiveSkillState(defaultActiveSkills, options?.activeSkillNames ?? [], options?.activeSkillSource);
         let isBangSkillInvocation = false;
         if (expandPromptTemplates) {
             const bangInvocation = parseBangInvocation(text, {
@@ -1831,9 +1784,7 @@ export class AgentSession {
                 currentText = bangInvocation.prompt
                     ? `/skill:${bangInvocation.skillName} ${bangInvocation.prompt}`
                     : `/skill:${bangInvocation.skillName}`;
-                promptActiveSkillNames = mergePromptActiveSkillNames(promptActiveSkillNames, bangInvocation.activeSkillNames);
-                promptActiveSkillSource =
-                    defaultActiveSkills.length > 0 ? `settings+${bangInvocation.source}` : bangInvocation.source;
+                promptSkills = addActiveSkills(promptSkills, bangInvocation.activeSkillNames, bangInvocation.source);
             }
         }
         let messages;
@@ -1882,6 +1833,7 @@ export class AgentSession {
                 preflightResult?.(true);
                 return;
             }
+            this._promptLifecycle.assertIdle();
             // Flush any pending bash messages before the new prompt
             this._flushPendingBashMessages();
             // Validate model
@@ -1934,11 +1886,18 @@ export class AgentSession {
                 messages.push(msg);
             }
             this._pendingNextTurnMessages = [];
+            if (isGrokOAuthProvider(this.model?.provider) && grokHarnessAutoApplyEnabled(process.env)) {
+                const skillQuery = isBangSkillInvocation ? currentText : expandedText;
+                const grokSkills = selectGrokHarnessSkills(skillQuery, this._resourceLoader.getSkills().skills, {
+                    contextPressure: this._computePressureBucket(messages) > 0,
+                });
+                promptSkills = addActiveSkills(promptSkills, grokSkills, "grok-harness");
+            }
             const turnSystemPromptOptions = {
                 ...this._baseSystemPromptOptions,
                 contextBudget: this._getContextBudgetOptions(expandedText),
-                activeSkillNames: promptActiveSkillNames,
-                activeSkillSource: promptActiveSkillSource,
+                activeSkillNames: promptSkills.names,
+                activeSkillSource: promptSkills.source,
             };
             const turnSystemPrompt = buildSystemPromptPlan(turnSystemPromptOptions);
             // Emit before_agent_start extension event
@@ -1973,11 +1932,16 @@ export class AgentSession {
             }
             this._recordPromptCachePlan(result?.systemPrompt ? "extension-override" : "turn-plan");
             await this._checkProjectedCompaction(messages);
+            this._runBudget.assertActive();
         }
         catch (error) {
             preflightResult?.(false);
             const rawMessage = error instanceof Error ? error.message : String(error);
-            const cause = preflightFailureCause(rawMessage, Boolean(this.model));
+            const cause = error instanceof RunBudgetExceededError
+                ? { area: "budget", code: error.code }
+                : error instanceof RunBudgetPolicyError
+                    ? { area: "configuration", code: "invalid" }
+                    : preflightFailureCause(rawMessage, Boolean(this.model));
             const timestamp = new Date().toISOString();
             this._publishTermination(classifySessionTermination({
                 sessionId: this.sessionId,
@@ -2227,6 +2191,7 @@ export class AgentSession {
         this._followUpMessages = [];
         this.agent.clearAllQueues();
         this._emitQueueUpdate();
+        this._promptLifecycle.flush();
         return { steering, followUp };
     }
     /** Number of pending messages (includes both steering and follow-up) */
@@ -2573,11 +2538,11 @@ export class AgentSession {
      * 1..<0.75, 2..<0.9, 3..>=0.9. The released default pressure coefficient
      * is 1 and participates only after the classifier has prompt evidence.
      */
-    _computePressureBucket() {
+    _computePressureBucket(pendingMessages = []) {
         const contextWindow = this.model?.contextWindow ?? 0;
         if (contextWindow <= 0)
             return 0;
-        const estimate = estimateProjectedContextTokens(this.agent.state.messages, []);
+        const estimate = estimateProjectedContextTokens(this.agent.state.messages, pendingMessages);
         const pressure = estimate.tokens / contextWindow;
         if (pressure >= 0.9)
             return 3;
@@ -3993,7 +3958,7 @@ export class AgentSession {
         const failoverTo = await this._maybeFailoverFromSafetyStop(message);
         const switchedVia = failoverTo ? `failover ${failoverTo}` : rotatedTo ? `route ${rotatedTo}` : undefined;
         // Safety stops are usually immediate false positives — short delay after failover, full backoff otherwise.
-        const delayMs = computeRetryDelayMs(settings.baseDelayMs, attempt, switchedVia !== undefined);
+        const delayMs = Math.min(computeRetryDelayMs(settings.baseDelayMs, attempt, switchedVia !== undefined), this._runBudget.remainingMs ?? Infinity);
         const errorMessage = switchedVia
             ? `${message.errorMessage || "content/safety stop"} → ${switchedVia}`
             : message.errorMessage || "Unknown error";