@pentoshi/clai 3.8.19 → 3.8.20

package/dist/agent/runner.js

@@ -14,7 +14,7 @@ import { availableToolNames, normalizeToolCall, runToolCall, BATCH_SAFE_TOOLS, }
 import { getToolDefinitions, getCompactToolDefinitions, PLAN_TOOL_NAMES, } from "../tools/definitions.js";
 import { appendAssistantWithTools, appendToolResult, assertValidToolProtocol, fillMissingToolResults, } from "./tool-history.js";
 import { formatViewportHint, registerViewport } from "../ui/output-pane.js";
-import { compactMessagesWithSummary, estimateMessagesTokens, AUTO_COMPACT_TOKEN_BUDGET, COMPACTION_MEMORY_PREFIX, PLAN_IMPLEMENT_MEMORY_PREFIX, isCompactionMemoryMessage, } from "./context-manager.js";
+import { compactMessagesWithSummary, estimateMessagesTokens, AUTO_COMPACT_TOKEN_BUDGET, shouldApplyAutoCompact, COMPACTION_MEMORY_PREFIX, PLAN_IMPLEMENT_MEMORY_PREFIX, isCompactionMemoryMessage, } from "./context-manager.js";
 import { auditLog } from "../store/logs.js";
 import { loadProjectContext } from "../store/project.js";
 import { loadScope, isScopeActive } from "../store/scope.js";
@@ -37,6 +37,7 @@ import { saveToolOutput, summarizeOutput, formatToolContext, } from "./tool-outp
 import { renderPlanForTerminal, planContextMessage, handlePlanTool, resolvePlanTaskId, } from "./plan-tool.js";
 import { absorbLooseWorkIntoLedger, applyDestinationCwd, canMarkTaskDone, hasLocalRuntimeProof, hasRemoteWorkProof, isDevServerCall, isEvidenceWorkTool, isFeatureImplementationCall, isPackageInstallCommand, isPlanPreflightTool, isPortListeningOutput, isReadOnlyReconTool, isRemoteActiveTestCall, isRemoteObservationTask, isRemoteReconToolCall, isRuntimeObservationTask, isScaffoldCreateCommand, isServerReadyOutput, ledgerFromTaskEvidence, pickPendingTaskForToolCall, recordTaskWorkSuccess, resolveUserDestinationHint, taskEvidenceFromLedger, TOOL_ABORT_GRACE_MS, toolHardBudgetMs, toolStallBudgetMs, userAskedForFeatureApp, } from "./task-evidence.js";
 import { buildSessionStateBlock, inferNextHint, upsertSessionStateMessage, } from "./session-state.js";
+import { buildContinueOrientation } from "./continue-orient.js";
 import { detectPackageManager } from "./workspace-orient.js";
 import { budgetRemaining, consumeBudget, createRecoveryBudgets, freestyleClaimsAppReady, looksLikeShallowPentestReport, recoveryForErrorDiagnosis, recoveryForFailedProbe, recoveryForFreshness, recoveryForMissingFeature, recoveryForMissingPlan, recoveryForNarration, recoveryForPrematureComplete, recoveryForRuntimeVerify, recoveryForShallowPentest, } from "./must-continue.js";
 import { scopeContextMessage } from "./scope-context.js";
@@ -380,6 +381,22 @@ export async function runAgentTurn(prompt, options = {}) {
         if (activePlan) {
             systemSections.push(planContextMessage(activePlan, session.planApproved.value));
         }
+        // Soft mid-work recovery (any domain): re-attach to jobs / open tasks /
+        // last tools after interrupt or "continue" — not a hard gate.
+        if (!informationalQuery && !idleOrSocialPrompt) {
+            const continueBrief = buildContinueOrientation({
+                prompt,
+                history: options.history,
+                plan: activePlan,
+                runningJobs: jobManager.getRunningJobs(),
+                recentJobs: jobManager.getRecentJobs(12),
+                informationalQuery,
+                idleOrSocial: idleOrSocialPrompt,
+            });
+            if (continueBrief) {
+                systemSections.push(continueBrief);
+            }
+        }
         if (isPlanMode) {
             systemSections.push(planModeDirective());
         }
@@ -699,9 +716,14 @@ export async function runAgentTurn(prompt, options = {}) {
         refreshSessionState = (plan) => {
             if (idleOrSocialPrompt || informationalQuery)
                 return;
-            if (!buildLikeTurn && !pentestLikeTurn && !plan && !activePlan)
-                return;
+            const runningJobs = jobManager.getRunningJobs();
             const p = plan ?? activePlan;
+            if (!buildLikeTurn &&
+                !pentestLikeTurn &&
+                !p &&
+                runningJobs.length === 0) {
+                return;
+            }
             const root = getActiveProjectRoot() ?? p?.meta?.projectRoot;
             const pm = p?.meta?.packageManager ??
                 (root ? detectPackageManager(root) : undefined);
@@ -712,6 +734,10 @@ export async function runAgentTurn(prompt, options = {}) {
             const done = p?.tasks
                 .filter((t) => t.state === "done" || t.state === "skipped")
                 .map((t) => t.id);
+            const jobBits = runningJobs.slice(0, 4).map((j) => {
+                const cmd = (j.commandDisplay || j.command).replace(/\s+/g, " ").trim();
+                return `${j.id} ${j.status} ${cmd.slice(0, 40)}`;
+            });
             const snap = {
                 goal: p?.goal ?? prompt.slice(0, 160),
                 projectRoot: root,
@@ -728,6 +754,9 @@ export async function runAgentTurn(prompt, options = {}) {
                 serverProbedOk: sawLocalHttpProbe,
                 lastProbeFailed: sawFailedLocalHttpProbe,
                 lastOkTool: taskWorkLedger?.lastOkTool,
+                backgroundJobs: jobBits.length > 0
+                    ? `${runningJobs.length} running: ${jobBits.join("; ")}`
+                    : undefined,
                 engagementNote: pentestSession
                     ? "remote/security engagement — no local dev server as completion"
                     : undefined,
@@ -1879,7 +1908,8 @@ export async function runAgentTurn(prompt, options = {}) {
             }
             return { ok: result.ok, call, result, contextOutput };
         }
-        const AUTO_COMPACT_KEEP_RECENT = 6;
+        // Align with /compact default: small recency + dense memory (not keepRecent=6 fat tails).
+        const AUTO_COMPACT_KEEP_RECENT = 2;
         let lastCompactionMsgCount = 0;
         const summarizeForCompaction = async (summaryPrompt) => {
             const response = await completeWithProvider({
@@ -1890,7 +1920,8 @@ export async function runAgentTurn(prompt, options = {}) {
                     { role: "user", content: summaryPrompt },
                 ],
                 temperature: 0.1,
-                maxTokens: 4_096,
+                // Prefer dense memory; enough room for findings without forcing a tiny stub.
+                maxTokens: 2_048,
                 signal: options.signal,
             });
             return response.text;
@@ -1907,8 +1938,17 @@ export async function runAgentTurn(prompt, options = {}) {
                 return;
             try {
                 const result = await compactMessagesWithSummary(messages, summarizeForCompaction, { budgetTokens: 0, keepRecent: AUTO_COMPACT_KEEP_RECENT });
-                if (!result.summarized || result.afterTokens >= beforeTokens)
+                const summaryBody = result.messages.find((m) => isCompactionMemoryMessage(m))?.content ??
+                    "";
+                if (!shouldApplyAutoCompact({
+                    summarized: result.summarized,
+                    summaryBody,
+                    beforeTokens: result.beforeTokens,
+                    afterTokens: result.afterTokens,
+                    afterMessages: result.messages,
+                })) {
                     return;
+                }
                 messages.splice(0, messages.length, ...result.messages);
                 loopGuard.resetReadOnly();
                 // Token stats BEFORE plan re-injection so the reduction is accurate.