@pentoshi/clai 3.11.31 → 3.11.34

package/dist/agent/runner.js

@@ -63,14 +63,14 @@ import { CompactionAttemptLedger, compactionAttemptKey, } from "./compaction-att
 import { loadPlan, mutatePlan, markTask, appendPlanTask, readyPlanTasks, foregroundRemaining, responderOpenTasks, isPlanTerminal, isPlanSuccessful, } from "../store/plan.js";
 import { stat } from "node:fs/promises";
 import { isOutsideWorkingDirectory, resolveFsToolPath, } from "../tools/fs.js";
-import { stripSentinelTokens, parseToolCall, recognizeBareToolJson, looksLikeTruncatedToolCall, salvageTruncatedWrite, salvageTruncatedWriteFromNative, countToolFences, parseAllToolCalls, groupToolCallsForExecution, buildTurnHistory, collapseRepeatedText, textBeforeToolCall, formatToolArgs, looksLikePentestTask, looksLikeBuildTask, looksLikeInformationalQuery, looksLikeIdleOrSocialPrompt, looksLikeActionNarration, looksLikeWebActionNarration, localHttpProbeIsFailure, localHttpProbeIsSuccess, requiresFreshWebSearch, freshnessGuardMessage, buildWorkflowDirective, narrowNmapOperationDirective, pentestWorkflowDirective, pentestNoLocalServerDirective, shouldDimToolChatter, looksLikePromptLeak, } from "./tool-call-parser.js";
+import { stripSentinelTokens, parseToolCall, recognizeBareToolJson, looksLikeTruncatedToolCall, salvageTruncatedWrite, salvageTruncatedWriteFromNative, countToolFences, parseAllToolCalls, groupToolCallsForExecution, buildTurnHistory, collapseRepeatedText, textBeforeToolCall, formatToolArgs, looksLikePentestTask, looksLikeBuildTask, looksLikeInformationalQuery, looksLikeIdleOrSocialPrompt, looksLikeActionNarration, looksLikeWebActionNarration, localHttpProbeIsFailure, localHttpProbeIsSuccess, buildWorkflowDirective, narrowNmapOperationDirective, pentestWorkflowDirective, pentestNoLocalServerDirective, shouldDimToolChatter, looksLikePromptLeak, } from "./tool-call-parser.js";
 import { createSessionPolicy, isPreApprovalAllowedTool, isPlanModeAllowedShellCommand, isPlanModeAllowedTool, isPlanApprovedByStatus, planHasOpenWork, isAbortError, shouldEnableImageOcr, } from "./session-policy.js";
 import { saveToolOutput, summarizeOutput, formatToolContext, } from "./tool-output-formatting.js";
 import { codingSessionFromContext, isProtocolPlaceholderOutput, progressPauseMode, } from "./progress-pause-policy.js";
 import { renderPlanForTerminal, planContextMessage, upsertPlanContextMessage, handlePlanTool, resolvePlanTaskId, } from "./plan-tool.js";
 import { readDeclaredParentTaskId, resolveResponderParent, isExplicitResponderDelegation, delegationTaskTitle, } from "./responder-parent.js";
 import { readTaskUpdateArgs, distinctAdvancingTaskIds, isSimultaneousTaskAdvance, batchUpdateSignature, buildMultiUpdateReminder, buildMultiOpenRejection, multiUpdateToast, multiOpenToast, openingTaskIds, } from "./task-sync.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 { absorbLooseWorkIntoLedger, applyDestinationCwd, canMarkTaskDone, hasLocalRuntimeProof, hasRemoteWorkProof, isDevServerCall, isEvidenceWorkTool, isFeatureImplementationCall, isPackageInstallCommand, isPlanPreflightTool, isPortListeningOutput, isReadOnlyReconTool, isRemoteActiveTestCall, isRemoteReconToolCall, 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, looksLikeContinueOrResumePrompt, } from "./continue-orient.js";
 import { detectPackageManager } from "./workspace-orient.js";
@@ -442,19 +442,12 @@ export async function runAgentTurn(prompt, options = {}) {
         // brand-new plan (the exact failure where "what do u know till now"
         // triggered explore→plan and created an unrelated "Enhance clai" plan).
         const informationalQuery = looksLikeInformationalQuery(prompt);
-        // Greetings / thanks / short acks must never force tools, plans, or
-        // freshness retries — a false "act don't narrate" path burned tokens on
-        // web.search recovery loops after a simple "hi".
+        // Greetings / thanks / short acks must never force tools or plans.
         const idleOrSocialPrompt = looksLikeIdleOrSocialPrompt(prompt);
         suppressOutcomeDiagnostics =
             informationalQuery ||
                 idleOrSocialPrompt ||
                 looksLikeContinueOrResumePrompt(prompt);
-        const freshWebSearchRequired = !buildLikeTurn &&
-            !pentestLikeTurn &&
-            !idleOrSocialPrompt &&
-            toolNames.includes("web.search") &&
-            requiresFreshWebSearch(prompt);
         let provider = initialProvider;
         await ensureProviderConfigured(provider);
         let model = initialModel;
@@ -608,9 +601,6 @@ export async function runAgentTurn(prompt, options = {}) {
                 orientInput.candidateProject = candidate;
             systemSections.push(buildWorkspaceOrientation(orientInput));
         }
-        if (freshWebSearchRequired) {
-            systemSections.push(freshnessGuardMessage());
-        }
         // The live plan is mutable state: it is injected once as a keyed request
         // suffix (upsertPlanContextMessage) instead of being frozen into the stable
         // system prefix, so the model never sees a stale and a fresh plan together.
@@ -719,7 +709,6 @@ export async function runAgentTurn(prompt, options = {}) {
                     content.startsWith("ACTIVE PROJECT ROOT:") ||
                     content.startsWith("USER DESTINATION:") ||
                     content.startsWith("WORKSPACE STATUS") ||
-                    content.startsWith("Freshness guard for this turn:") ||
                     content.includes("MODE") ||
                     content.includes("OUTCOME"),
             }));
@@ -939,7 +928,6 @@ export async function runAgentTurn(prompt, options = {}) {
         // (e.g. malformed extra/missing braces that are NOT simple truncation). We
         // retry instead of leaking the raw block as the final answer.
         let malformedFenceRetries = 0;
-        let sawFreshWebSearch = false;
         const recovery = createRecoveryBudgets();
         let sawServerStart = false;
         let sawPlanCreateOk = false;
@@ -1011,70 +999,6 @@ export async function runAgentTurn(prompt, options = {}) {
                 remoteWorkVerified: planHasVerifiedRemoteWork(plan),
             });
         };
-        const reconcileOpenTaskBeforeFinalizing = async () => {
-            let plan = await loadPlan(session.sessionId).catch(() => undefined);
-            const open = plan?.tasks.find((task) => task.state === "in_progress" && !task.responderOwned);
-            if (!plan || !open)
-                return plan;
-            const gate = completionGateForTask(plan, open.id);
-            if (!gate.ok)
-                return plan;
-            const reconciledTaskIds = [open.id];
-            const working = plan;
-            markTask(working, open.id, "done", "Completion reconciled from verified task evidence.");
-            while (true) {
-                const observation = readyPlanTasks(working).find((task) => isRuntimeObservationTask(task.title) ||
-                    (working.kind === "pentest" && isRemoteObservationTask(task.title)));
-                if (!observation)
-                    break;
-                const observationGate = completionGateForTask(working, observation.id);
-                if (!observationGate.ok)
-                    break;
-                markTask(working, observation.id, "done", working.kind === "pentest"
-                    ? "Satisfied by verified remote evidence from the preceding task."
-                    : "Satisfied by the verified runtime evidence from the preceding task.");
-                reconciledTaskIds.push(observation.id);
-            }
-            if (working.status === "draft" || working.status === "approved") {
-                working.status = "in_progress";
-            }
-            if (isPlanTerminal(working)) {
-                working.status = isPlanSuccessful(working) ? "completed" : "abandoned";
-            }
-            // Replay the same reconciliation as a reducer so a concurrent
-            // responder settlement is preserved instead of overwritten.
-            const reconciledNotes = new Map(reconciledTaskIds.map((id) => [
-                id,
-                working.tasks.find((task) => task.id === id)?.note,
-            ]));
-            const reconcileResult = await mutatePlan(session.sessionId, (draft) => {
-                let changed = false;
-                for (const [id, note] of reconciledNotes) {
-                    const task = draft.tasks.find((candidate) => candidate.id === id);
-                    if (!task || task.state === "done")
-                        continue;
-                    task.state = "done";
-                    if (note !== undefined)
-                        task.note = note;
-                    changed = true;
-                }
-                if (draft.status === "draft" || draft.status === "approved") {
-                    draft.status = "in_progress";
-                    changed = true;
-                }
-                if (isPlanTerminal(draft)) {
-                    draft.status = isPlanSuccessful(draft) ? "completed" : "abandoned";
-                    changed = true;
-                }
-                return changed;
-            }).catch(() => undefined);
-            if (reconcileResult?.ok && reconcileResult.plan)
-                plan = reconcileResult.plan;
-            writePlanUpdate(plan, renderPlanForTerminal(plan) + "\n");
-            writeNotice("info", `reconciled ${reconciledTaskIds.map((id) => `[${id}]`).join(", ")} from verified evidence`, chalk.dim(`  ℹ reconciled ${reconciledTaskIds.map((id) => `[${id}]`).join(", ")} from verified evidence — no duplicate verification\n`));
-            taskWorkLedger = null;
-            return plan;
-        };
         async function persistProjectRootOnPlan(root) {
             const pm = detectPackageManager(root);
             // Metadata patches go through the transactional boundary so a
@@ -1806,9 +1730,6 @@ export async function runAgentTurn(prompt, options = {}) {
                     };
                 }
             }
-            if (call.name === "web.search") {
-                sawFreshWebSearch = true;
-            }
             if (!alreadyPrintedIds.has(toolEventId)) {
                 const toolCallLine = chalk.cyan(`  ▶ ${call.name}`) + chalk.gray(` ${formatToolArgs(call)}`);
                 writeToolCall(toolEventId, call, styleToolChatter(call, toolCallLine) + "\n");
@@ -3274,9 +3195,7 @@ export async function runAgentTurn(prompt, options = {}) {
                             ...(toolsAttached
                                 ? {
                                     tools: turnTools,
-                                    toolChoice: freshWebSearchRequired && !sawFreshWebSearch
-                                        ? { type: "function", name: "web.search" }
-                                        : "auto",
+                                    toolChoice: "auto",
                                     parallelToolCalls: true,
                                     // P2-3: emit tool cards as soon as the function name arrives.
                                     onToolCallDelta: (delta) => {
@@ -3709,12 +3628,15 @@ export async function runAgentTurn(prompt, options = {}) {
                 }
                 if (nativeToolCalls.length) {
                     const first = nativeToolCalls[0];
-                    if (first.args?._parseError) {
-                        call = undefined;
-                    }
-                    else {
-                        call = normalizeToolCall({ name: first.name, args: first.args });
-                    }
+                    call = first.args?._parseError
+                        ? {
+                            name: first.name || "unknown",
+                            args: {
+                                __nativeParseError: true,
+                                _raw: first.args._raw,
+                            },
+                        }
+                        : normalizeToolCall({ name: first.name, args: first.args });
                 }
                 else {
                     call = parseToolCall(assistantText.visible, {
@@ -3809,6 +3731,18 @@ export async function runAgentTurn(prompt, options = {}) {
                     }
                 }
                 if (!canonicalAssistantVisible.trim() && !call) {
+                    const incompleteNativeStream = nativeToolCalls.length === 0 &&
+                        deferredToolCalls.some((entry) => entry.shown && entry.call.name !== "…");
+                    if (incompleteNativeStream) {
+                        const reason = "The provider began this native tool call but never completed it. Nothing ran; reissue a complete call.";
+                        for (const deferred of deferredToolCalls) {
+                            if (!deferred.shown || deferred.call.name === "…")
+                                continue;
+                            writeToolBlocked(deferred.eventId, deferred.call.name, reason, chalk.yellow(`  ⚠ ${reason}\n`));
+                        }
+                        markTextOnlyModel(provider, model);
+                        writeNotice("warn", "provider abandoned a native tool call — switching this model to the text tool protocol", chalk.yellow("  ⚠ provider abandoned a native tool call — switching this model to the text tool protocol\n"));
+                    }
                     emptyVisibleRetries += 1;
                     if (emptyVisibleRetries <= 3) {
                         if (assistantText.hasThinking) {
@@ -3821,10 +3755,8 @@ export async function runAgentTurn(prompt, options = {}) {
                             retryWithoutThinking = true;
                         commitAssistantRetry(assistantText.visible);
                         // Keep nudges SHORT — cheap models lose the key instruction in long text.
-                        const buildNudge = freshWebSearchRequired && !sawFreshWebSearch
-                            ? toolsAttached
-                                ? "No visible output. This is current or scheduled information: call web.search now. Do NOT answer from memory."
-                                : "No visible output. This is current or scheduled information: emit exactly one valid ```tool block for web.search now. Do NOT answer from memory or hide the tool call in <think> tags."
+                        const buildNudge = incompleteNativeStream
+                            ? "Your native tool call was incomplete, so nothing ran. Use exactly one complete fenced ```tool block now; do not repeat the incomplete native call."
                             : isPlanMode && !activePlan
                                 ? toolsAttached
                                     ? "No visible output. In plan mode: gather context or call plan.create when ready (do not only describe the plan)."
@@ -4029,15 +3961,12 @@ export async function runAgentTurn(prompt, options = {}) {
                     const cleaned = collapseRepeatedText(stripSentinelTokens(canonicalAssistantVisible));
                     const narratedAction = looksLikeActionNarration(cleaned);
                     const narratedWebAction = looksLikeWebActionNarration(cleaned);
-                    const reconciledPlanAtCompletion = await reconcileOpenTaskBeforeFinalizing();
-                    const livePlanAtCompletion = reconciledPlanAtCompletion ??
-                        (await loadPlan(session.sessionId).catch(() => undefined));
+                    const livePlanAtCompletion = await loadPlan(session.sessionId).catch(() => undefined);
                     const planStatusAtCompletion = livePlanAtCompletion?.status ?? activePlan?.status;
                     const completedPlanDuringThisTurn = activePlan?.status !== "completed" &&
                         planStatusAtCompletion === "completed";
                     const planHasOpenWorkNow = planHasOpenWork(planStatusAtCompletion);
-                    const userExpectsWork = freshWebSearchRequired ||
-                        (planHasOpenWorkNow && session.planApproved.value) ||
+                    const userExpectsWork = (planHasOpenWorkNow && session.planApproved.value) ||
                         (!informationalQuery &&
                             !idleOrSocialPrompt &&
                             (buildLikeTurn || pentestLikeTurn));
@@ -4105,11 +4034,6 @@ export async function runAgentTurn(prompt, options = {}) {
                         pentestSession,
                         informationalQuery,
                         idleOrSocialPrompt,
-                        freshWebSearchRequired,
-                        freshnessGuardText: freshWebSearchRequired
-                            ? freshnessGuardMessage()
-                            : "",
-                        sawFreshWebSearch,
                         sawPlanCreateOk,
                         sawFeatureImplWrite,
                         sawScaffoldOk,