@flowdesk/opencode-plugin 0.1.13 → 0.1.14

package/dist/agent-task-runner.js

@@ -1,9 +1,21 @@
 import { createHash } from "node:crypto";
 import { applyFlowDeskSessionEvidenceWriteIntentsV1, prepareFlowDeskSessionEvidenceWriteIntentV1, } from "@flowdesk/core";
 import { launchFlowDeskInjectedSdkRuntimeLaneFromPlanV1, materializeFlowDeskRuntimeLaneLaunchLifecycleEvidenceV1, } from "./managed-dispatch-adapter.js";
+import { observeFlowDeskAgentTaskOutputV1 } from "./agent-task-output.js";
+import { refreshFlowDeskCompletionUiCachesV1 } from "./completion-ui-cache.js";
 import { recordFlowDeskLaneHeartbeatV1 } from "./lane-heartbeat-writer.js";
 const TASK_RESULT_MAX_TEXT = 32_768;
 const AGENT_TASK_CONTEXT_MAX_PROMPT_TEXT = 32_768;
+const INVALID_PARENT_SESSION_REF = "ses-invalid-parent-session-binding";
+/** Schema version for async child session tracking evidence */
+export const AGENT_TASK_CHILD_SESSION_SCHEMA_VERSION = "flowdesk.agent_task_child_session.v1";
+export function sanitizeFlowDeskTaskResultTextV1(text) {
+    return {
+        text: text.length > TASK_RESULT_MAX_TEXT ? text.slice(0, TASK_RESULT_MAX_TEXT) : text,
+        changed: false,
+        truncated: text.length > TASK_RESULT_MAX_TEXT,
+    };
+}
 function agentTaskLaunchPlan(input) {
     return {
         schema_version: "flowdesk.runtime_lane_launch_plan.v1",
@@ -32,66 +44,162 @@ function agentTaskLaunchPlan(input) {
         runtimeExecution: false,
     };
 }
-function extractAssistantTextFromResponse(client, childSessionId) {
-    // We extract response text via messages API
+function validateAgentTaskParentSessionId(parentSessionId) {
+    const value = parentSessionId.trim();
+    if (value.length === 0)
+        return { ok: false, redactedReason: "missing_parent_session_binding", parentSessionRef: INVALID_PARENT_SESSION_REF };
+    if (value.length > 128)
+        return { ok: false, redactedReason: "invalid_parent_session_binding", parentSessionRef: INVALID_PARENT_SESSION_REF };
+    // `ses-...` is FlowDesk's opaque session-ref wrapper, not the raw OpenCode
+    // session id expected by SDK `session.create({ parentID })`. Accepting it here
+    // causes evidence such as `ses-ses-flowdesk-coordinator` and can make the SDK
+    // wait on a non-existent synthetic parent session until launch timeout.
+    if (value.startsWith("ses-"))
+        return { ok: false, redactedReason: "invalid_parent_session_binding", parentSessionRef: INVALID_PARENT_SESSION_REF };
+    if (/\s/.test(value))
+        return { ok: false, redactedReason: "invalid_parent_session_binding", parentSessionRef: INVALID_PARENT_SESSION_REF };
+    if (!/^[A-Za-z0-9_.:-]+$/.test(value))
+        return { ok: false, redactedReason: "invalid_parent_session_binding", parentSessionRef: INVALID_PARENT_SESSION_REF };
+    return { ok: true, parentSessionRef: `ses-${value}` };
+}
+/** Bounded nudge text — versioned constant, never echoes user input */
+const AGENT_TASK_NUDGE_TEXT = "Please provide your final answer now. If you have completed your analysis, output your complete response.";
+/**
+ * Polls `session.messages` with a per-call 3-second cap so it works whether the SDK
+ * uses snapshot (returns immediately) or long-poll (blocks until output) semantics.
+ *
+ * Heartbeat: fires every `quietPeriodMs` of silence — only when inactive.
+ * Nudge:     after `quietPeriodMs` of silence, sends a bounded prompt to the child
+ *            session asking for the final answer. Max `maxNudges` nudges total.
+ *            After exhausting nudges with no response, returns undefined.
+ */
+async function extractAssistantTextFromResponse(client, childSessionId, opts) {
     const messages = client.session.messages;
     if (messages === undefined)
         return undefined;
-    return (async () => {
-        try {
-            const method = messages;
+    const quietPeriodMs = opts?.quietPeriodMs ?? 30_000;
+    const maxNudges = opts?.maxNudges ?? 2;
+    const MESSAGES_TIMEOUT_MS = opts?.messagesTimeoutMs ?? 3_000; // per-call cap — handles both snapshot and long-poll
+    const method = messages;
+    /**
+     * Call session.messages with a ceiling timeout so we can check inactivity periodically.
+     * This handles both snapshot APIs (return immediately) and long-poll APIs
+     * (block until LLM produces output). With the timeout, a long-poll call that
+     * hasn't returned after MESSAGES_TIMEOUT_MS resolves as null so we can
+     * check the inactivity clock and possibly send a nudge.
+     */
+    const callMessages = () => {
+        const messagePromise = (async () => {
             const current = await method.call(client.session, { sessionID: childSessionId });
-            const response = isSdkErrorResponse(current)
-                ? await method.call(client.session, { path: { id: childSessionId } })
-                : current;
-            const data = asResponseData(response);
-            const record = asRecord(data);
-            const items = Array.isArray(data)
-                ? data
-                : Array.isArray(record?.items)
-                    ? record.items
-                    : Array.isArray(record?.messages)
-                        ? record.messages
-                        : [];
-            for (let index = items.length - 1; index >= 0; index -= 1) {
-                const message = items[index];
-                const record = asRecord(message);
-                const info = asRecord(record?.info) ?? record;
-                if (info?.role !== "assistant")
-                    continue;
-                const parts = Array.isArray(record?.parts)
-                    ? record.parts
-                    : Array.isArray(info?.parts)
-                        ? info.parts
-                        : [];
-                for (const part of parts) {
-                    const partRecord = asRecord(part);
-                    const text = typeof partRecord?.text === "string"
-                        ? partRecord.text
-                        : typeof partRecord?.content === "string"
-                            ? partRecord.content
-                            : undefined;
-                    if (typeof text === "string" && text.trim().length > 0)
-                        return text;
-                }
-            }
-            return undefined;
+            if (isSdkErrorResponse(current))
+                return method.call(client.session, { path: { id: childSessionId } });
+            return current;
+        })();
+        // Only race against timeout when the API might block (MESSAGES_TIMEOUT_MS > 0)
+        if (MESSAGES_TIMEOUT_MS <= 0)
+            return messagePromise;
+        return Promise.race([
+            messagePromise,
+            new Promise(resolve => setTimeout(() => resolve(null), MESSAGES_TIMEOUT_MS)),
+        ]);
+    };
+    /** Send a nudge to the child session with a hard timeout to prevent blocking.
+     * Uses noReply: true so the child does not generate a spurious second assistant turn.
+     */
+    const sendNudge = async () => {
+        const promptFn = client.session.prompt ?? client.session.promptAsync;
+        if (promptFn === undefined)
+            return "skipped";
+        const NUDGE_TIMEOUT_MS = 5_000;
+        try {
+            await Promise.race([
+                promptFn.call(client.session, {
+                    sessionID: childSessionId,
+                    noReply: true,
+                    ...(opts?.runtimeModel !== undefined ? { model: opts.runtimeModel } : {}),
+                    ...(opts?.agentName !== undefined ? { agent: opts.agentName } : {}),
+                    parts: [{ type: "text", text: AGENT_TASK_NUDGE_TEXT }],
+                }),
+                new Promise((_, reject) => setTimeout(() => reject(new Error("nudge timeout")), NUDGE_TIMEOUT_MS)),
+            ]);
+            return "sent";
         }
         catch {
-            return undefined;
+            return "timeout";
         }
-    })();
-}
-function isProcessOnlyAssistantOutput(text) {
-    const normalized = text.trim().toLowerCase();
-    return normalized.length === 0 || [
-        "working",
-        "thinking",
-        "i'll take a look",
-        "i will take a look",
-        "let me inspect",
-        "i'm going to inspect",
-    ].some((fragment) => normalized.includes(fragment));
+    };
+    const observe = (response) => {
+        if (response === null)
+            return undefined; // timed-out poll cycle
+        return observeFlowDeskAgentTaskOutputV1(response);
+    };
+    const startMs = Date.now();
+    let lastActivityMs = startMs;
+    let lastSignature = "";
+    let lastHeartbeatMs = startMs;
+    let nudgeCount = 0;
+    let latestCandidate;
+    try {
+        while (true) {
+            const response = await callMessages();
+            const nowMs = Date.now();
+            // Build signature (null response = timeout, no change)
+            const sig = response === null ? lastSignature : (() => {
+                const data = asResponseData(response);
+                const record = asRecord(data);
+                const items = Array.isArray(data) ? data
+                    : Array.isArray(record?.items) ? record.items
+                        : Array.isArray(record?.messages) ? record.messages : [];
+                const observed = observe(response);
+                return `${items.length}:${observed?.latestText?.length ?? 0}:${observed?.terminalObserved === true ? "terminal" : "open"}`;
+            })();
+            if (sig !== lastSignature) {
+                // New activity — reset all inactivity clocks
+                lastSignature = sig;
+                lastActivityMs = nowMs;
+                lastHeartbeatMs = nowMs;
+            }
+            const observed = observe(response);
+            if (observed?.latestText !== undefined && observed.latestText.trim().length > 0)
+                latestCandidate = observed;
+            if (observed?.terminalObserved === true && observed.latestText !== undefined && observed.latestText.trim().length > 0) {
+                return { text: observed.latestText, completionStatus: "final", outputKind: observed.outputKind, usableForSynthesis: observed.usableForSynthesis };
+            }
+            const silenceMs = nowMs - lastActivityMs;
+            if (silenceMs >= quietPeriodMs) {
+                // Emit heartbeat on first quiet-period expiry of each silence window
+                if (nowMs - lastHeartbeatMs >= quietPeriodMs) {
+                    lastHeartbeatMs = nowMs;
+                    opts?.heartbeatFn?.(nowMs - startMs);
+                }
+                // Send nudge after quiet period
+                if (nudgeCount < maxNudges) {
+                    nudgeCount++;
+                    await sendNudge();
+                    // Reset activity clock after nudge — give a fresh quiet window
+                    lastActivityMs = Date.now();
+                    lastHeartbeatMs = lastActivityMs;
+                }
+                else {
+                    // Exhausted all nudges. Preserve usable candidate text as partial output.
+                    if (latestCandidate?.latestText !== undefined && latestCandidate.latestText.trim().length > 0) {
+                        return { text: latestCandidate.latestText, completionStatus: "partial", outputKind: latestCandidate.outputKind, usableForSynthesis: latestCandidate.usableForSynthesis };
+                    }
+                    return undefined;
+                }
+            }
+            else {
+                // No activity and not yet at quiet period — yield to event loop before next poll.
+                // Sleep for up to 1s or quietPeriodMs/10, whichever is smaller, to avoid tight loops
+                // while still being responsive when messages arrive quickly (snapshot mode).
+                const yieldMs = Math.max(10, Math.min(1_000, Math.floor(quietPeriodMs / 10)));
+                await new Promise(resolve => setTimeout(resolve, yieldMs));
+            }
+        }
+    }
+    catch {
+        return undefined;
+    }
 }
 function asRecord(value) {
     return typeof value === "object" && value !== null && !Array.isArray(value)
@@ -117,8 +225,38 @@ function writeSessionEvidence(input) {
         record: input.record,
     });
     if (prepared.ok && prepared.writeIntent !== undefined) {
-        applyFlowDeskSessionEvidenceWriteIntentsV1(input.rootDir, [prepared.writeIntent]);
+        const applied = applyFlowDeskSessionEvidenceWriteIntentsV1(input.rootDir, [prepared.writeIntent]);
+        return applied.ok && applied.writtenPaths.length > 0;
     }
+    return false;
+}
+function progressLabel(value) {
+    const compact = value.replace(/\s+/g, " ").trim();
+    return compact.length > 120 ? `${compact.slice(0, 119)}…` : compact;
+}
+function writeAgentTaskProgress(input) {
+    const observedAt = input.observedAt ?? new Date().toISOString();
+    const record = {
+        schema_version: "flowdesk.agent_task_progress.v1",
+        workflow_id: input.workflowId,
+        lane_id: input.laneId,
+        task_id: input.taskId,
+        agent_ref: input.agentRef,
+        provider_qualified_model_id: input.providerQualifiedModelId,
+        progress_seq: input.progressSeq,
+        observed_at: observedAt,
+        phase: input.phase,
+        progress_label: progressLabel(input.progressLabel),
+        progress_ref: `progress-${input.laneId}-${input.progressSeq}`,
+        redaction_version: "v1",
+        dispatch_authority_enabled: false,
+    };
+    writeSessionEvidence({
+        rootDir: input.rootDir,
+        workflowId: input.workflowId,
+        evidenceId: `agent-task-progress-${input.laneId}-${input.progressSeq}`,
+        record: record,
+    });
 }
 function writeAgentTaskTerminalLifecycle(input) {
     const childSessionRef = input.childSessionRef === input.parentSessionRef ? undefined : input.childSessionRef;
@@ -154,6 +292,44 @@ function writeAgentTaskTerminalLifecycle(input) {
 export async function executeFlowDeskAgentTaskV1(input) {
     const observedAt = new Date().toISOString();
     const token = `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 8)}`;
+    const parentBinding = validateAgentTaskParentSessionId(input.parentSessionId);
+    const parentSessionRef = parentBinding.parentSessionRef;
+    const attemptId = `attempt-task-${token}`;
+    if (!parentBinding.ok) {
+        const taskFailedEvidenceId = `task-failed-${input.taskId}-${token}-invalid-parent`;
+        const redactedReason = parentBinding.redactedReason;
+        writeSessionEvidence({
+            rootDir: input.rootDir,
+            workflowId: input.workflowId,
+            evidenceId: taskFailedEvidenceId,
+            record: {
+                schema_version: "flowdesk.task_failed.v1",
+                workflow_id: input.workflowId,
+                lane_id: input.laneId,
+                task_id: input.taskId,
+                agent_ref: input.agentRef,
+                provider_qualified_model_id: input.providerQualifiedModelId,
+                failure_category: "sdk_create_failed",
+                redacted_reason: redactedReason,
+                created_at: observedAt,
+                dispatch_authority_enabled: false,
+            },
+        });
+        writeAgentTaskTerminalLifecycle({
+            rootDir: input.rootDir,
+            workflowId: input.workflowId,
+            laneId: input.laneId,
+            attemptId,
+            parentSessionRef,
+            agentRef: input.agentRef,
+            providerQualifiedModelId: input.providerQualifiedModelId,
+            state: "invocation_failed",
+            evidenceId: `lifecycle-task-terminal-${input.laneId}-${token}-invalid-parent`,
+            createdAt: observedAt,
+            updatedAt: observedAt,
+        });
+        return { status: "task_failed", failureCategory: "sdk_create_failed", redactedReason, laneId: input.laneId };
+    }
     const launchPlan = agentTaskLaunchPlan({
         workflowId: input.workflowId,
         laneId: input.laneId,
@@ -163,8 +339,6 @@ export async function executeFlowDeskAgentTaskV1(input) {
         token,
     });
     const runningLifecycleEvidenceId = `lifecycle-task-running-${input.laneId}-${token}`;
-    const attemptId = launchPlan.attempt_id ?? `attempt-task-${token}`;
-    const parentSessionRef = `ses-${input.parentSessionId}`;
     const promptTextTruncated = input.promptText.length > AGENT_TASK_CONTEXT_MAX_PROMPT_TEXT;
     const agentTaskContextRecord = {
         schema_version: "flowdesk.agent_task_context.v1",
@@ -189,17 +363,73 @@ export async function executeFlowDeskAgentTaskV1(input) {
         evidenceId: `agent-task-context-${input.taskId}-${token}`,
         record: agentTaskContextRecord,
     });
-    // Launch the lane
-    const launchResult = await launchFlowDeskInjectedSdkRuntimeLaneFromPlanV1({
-        client: input.client,
-        launchPlan,
-        request: {
-            allowActualLaneLaunch: true,
-            parentSessionId: input.parentSessionId,
-            promptText: input.promptText,
-            dispatchMethod: "prompt",
-        },
+    writeAgentTaskProgress({
+        rootDir: input.rootDir,
+        workflowId: input.workflowId,
+        laneId: input.laneId,
+        taskId: input.taskId,
+        agentRef: input.agentRef,
+        providerQualifiedModelId: input.providerQualifiedModelId,
+        phase: "started",
+        progressSeq: 1,
+        progressLabel: "agent task lane launch started",
+        observedAt,
     });
+    // Launch the lane — wrap in absolute timeout so session.prompt blocking doesn't hang forever.
+    // The launch phase timeout is longer (5 min) since promptAsync may queue work before responding.
+    // 1 min default — if session.prompt blocks for more than 1 min with no activity, give up
+    const LAUNCH_TIMEOUT_MS = input._launchTimeoutMs ?? 60_000;
+    const launchTimeoutHandle = setTimeout(() => { }, LAUNCH_TIMEOUT_MS);
+    const dispatchMethod = input.client.session.promptAsync !== undefined ? "promptAsync" : "prompt";
+    const launchResult = await Promise.race([
+        launchFlowDeskInjectedSdkRuntimeLaneFromPlanV1({
+            client: input.client,
+            launchPlan,
+            request: {
+                allowActualLaneLaunch: true,
+                parentSessionId: input.parentSessionId,
+                promptText: input.promptText,
+                dispatchMethod,
+            },
+        }),
+        new Promise(resolve => setTimeout(() => resolve({ status: "launch_timeout" }), LAUNCH_TIMEOUT_MS)),
+    ]);
+    clearTimeout(launchTimeoutHandle);
+    if ("status" in launchResult && launchResult.status === "launch_timeout") {
+        // session.prompt blocked for too long — treat as invocation failure
+        const failedEvidenceId = `task-failed-${input.taskId}-${token}-launch-timeout`;
+        writeSessionEvidence({
+            rootDir: input.rootDir,
+            workflowId: input.workflowId,
+            evidenceId: failedEvidenceId,
+            record: {
+                schema_version: "flowdesk.task_failed.v1",
+                workflow_id: input.workflowId,
+                lane_id: input.laneId,
+                task_id: input.taskId,
+                agent_ref: input.agentRef,
+                provider_qualified_model_id: input.providerQualifiedModelId,
+                failure_category: "sdk_create_failed",
+                redacted_reason: "lane launch timed out: session.prompt did not respond",
+                created_at: observedAt,
+                dispatch_authority_enabled: false,
+            },
+        });
+        writeAgentTaskTerminalLifecycle({
+            rootDir: input.rootDir,
+            workflowId: input.workflowId,
+            laneId: input.laneId,
+            attemptId,
+            parentSessionRef,
+            agentRef: input.agentRef,
+            providerQualifiedModelId: input.providerQualifiedModelId,
+            state: "invocation_failed",
+            evidenceId: `lifecycle-task-terminal-${input.laneId}-${token}-launch-timeout`,
+            createdAt: observedAt,
+            updatedAt: new Date().toISOString(),
+        });
+        return { status: "task_failed", failureCategory: "sdk_create_failed", redactedReason: "launch timeout: session.prompt did not respond within the allowed window", laneId: input.laneId };
+    }
     // Write running lifecycle evidence
     materializeFlowDeskRuntimeLaneLaunchLifecycleEvidenceV1({
         rootDir: input.rootDir,
@@ -278,22 +508,81 @@ export async function executeFlowDeskAgentTaskV1(input) {
         observedAt,
         progressSummaryLabel: `agent task lane launch heartbeat`,
     });
-    // Extract child session ID and get response text
+    // Extract child session ID
     const childSessionId = launchResult.childSessionRef?.startsWith("ses-")
         ? launchResult.childSessionRef.slice("ses-".length)
         : undefined;
-    let resultText;
+    // ── Async mode: return immediately, watchdog handles polling/nudging/abort ──
+    if (input.asyncMode === true) {
+        const resolvedChildId = childSessionId ?? "";
+        // Write child session evidence so watchdog can find it
+        writeSessionEvidence({
+            rootDir: input.rootDir,
+            workflowId: input.workflowId,
+            evidenceId: `agent-task-child-session-${input.laneId}-${token}`,
+            record: {
+                schema_version: AGENT_TASK_CHILD_SESSION_SCHEMA_VERSION,
+                workflow_id: input.workflowId,
+                lane_id: input.laneId,
+                task_id: input.taskId,
+                child_session_id: resolvedChildId,
+                parent_session_ref: parentSessionRef,
+                provider_qualified_model_id: input.providerQualifiedModelId,
+                agent_ref: input.agentRef,
+                nudge_count: 0,
+                last_nudge_at: null,
+                created_at: observedAt,
+                dispatch_authority_enabled: false,
+            },
+        });
+        writeAgentTaskProgress({
+            rootDir: input.rootDir,
+            workflowId: input.workflowId,
+            laneId: input.laneId,
+            taskId: input.taskId,
+            agentRef: input.agentRef,
+            providerQualifiedModelId: input.providerQualifiedModelId,
+            phase: "waiting",
+            progressSeq: 2,
+            progressLabel: "agent task waiting for async child result",
+        });
+        return { status: "task_launched", laneId: input.laneId, childSessionId: resolvedChildId };
+    }
+    let resultObservation;
     if (childSessionId !== undefined) {
-        resultText = await extractAssistantTextFromResponse(input.client, childSessionId);
+        const runtimeModel = launchResult.status === "lane_launch_started" && typeof launchResult.model === "string"
+            ? launchResult.model : undefined;
+        const agentName = launchResult.status === "lane_launch_started" && typeof launchResult.agent === "string"
+            ? launchResult.agent : undefined;
+        resultObservation = await extractAssistantTextFromResponse(input.client, childSessionId, {
+            quietPeriodMs: input._nudgeQuietPeriodMs ?? 20_000, // default 20s per policy
+            maxNudges: 2,
+            runtimeModel,
+            agentName,
+            messagesTimeoutMs: input._messagesTimeoutMs,
+            heartbeatFn: (elapsedMs) => {
+                recordFlowDeskLaneHeartbeatV1({
+                    rootDir: input.rootDir,
+                    workflowId: input.workflowId,
+                    attemptId,
+                    laneId: input.laneId,
+                    parentSessionRef,
+                    agentRef: input.agentRef,
+                    providerQualifiedModelId: input.providerQualifiedModelId,
+                    state: "running",
+                    observedAt: new Date().toISOString(),
+                    progressSummaryLabel: `agent task waiting for response elapsed=${Math.floor(elapsedMs / 1000)}s`,
+                });
+            },
+        });
     }
-    if (resultText === undefined || (input.outputContract === "final_assistant_text" && isProcessOnlyAssistantOutput(resultText))) {
+    const resultText = resultObservation?.text;
+    if (resultText === undefined) {
         // No response text - write task_failed
         const taskFailedEvidenceId = `task-failed-${input.taskId}-${token}`;
-        const failureCategory = resultText === undefined ? "no_response" : "contract_not_satisfied";
-        const evidenceFailureCategory = resultText === undefined ? "no_response" : "unknown";
-        const redactedReason = resultText === undefined
-            ? "lane launched but no assistant response text found"
-            : "lane launched but final assistant response did not satisfy requested output contract";
+        const failureCategory = "no_response";
+        const evidenceFailureCategory = "no_response";
+        const redactedReason = "lane launched but no assistant response text found";
         const taskFailedRecord = {
             schema_version: "flowdesk.task_failed.v1",
             workflow_id: input.workflowId,
@@ -312,6 +601,17 @@ export async function executeFlowDeskAgentTaskV1(input) {
             evidenceId: taskFailedEvidenceId,
             record: taskFailedRecord,
         });
+        writeAgentTaskProgress({
+            rootDir: input.rootDir,
+            workflowId: input.workflowId,
+            laneId: input.laneId,
+            taskId: input.taskId,
+            agentRef: input.agentRef,
+            providerQualifiedModelId: input.providerQualifiedModelId,
+            phase: "failed",
+            progressSeq: 3,
+            progressLabel: failureCategory === "no_response" ? "agent task finished without response" : "agent task output contract not satisfied",
+        });
         writeAgentTaskTerminalLifecycle({
             rootDir: input.rootDir,
             workflowId: input.workflowId,
@@ -322,7 +622,7 @@ export async function executeFlowDeskAgentTaskV1(input) {
             messageRef: launchResult.messageRef?.startsWith("msg-") ? launchResult.messageRef : undefined,
             agentRef: input.agentRef,
             providerQualifiedModelId: input.providerQualifiedModelId,
-            state: resultText === undefined ? "no_output" : "incomplete",
+            state: "no_output",
             evidenceId: `lifecycle-task-terminal-${input.laneId}-${token}`,
             createdAt: observedAt,
             updatedAt: new Date().toISOString(),
@@ -335,10 +635,9 @@ export async function executeFlowDeskAgentTaskV1(input) {
             laneId: input.laneId,
         };
     }
-    // Truncate if needed
     const fullResultText = resultText;
-    const truncated = fullResultText.length > TASK_RESULT_MAX_TEXT;
-    const storedResultText = truncated ? fullResultText.slice(0, TASK_RESULT_MAX_TEXT) : fullResultText;
+    const sanitizedResult = sanitizeFlowDeskTaskResultTextV1(fullResultText);
+    const storedResultText = sanitizedResult.text;
     const promptSha256 = sha256Hex(input.promptText);
     const resultSha256 = sha256Hex(fullResultText);
     // Write task_result evidence
@@ -352,17 +651,42 @@ export async function executeFlowDeskAgentTaskV1(input) {
         provider_qualified_model_id: input.providerQualifiedModelId,
         task_prompt_sha256: promptSha256,
         result_text: storedResultText,
-        result_text_truncated: truncated,
+        result_text_truncated: sanitizedResult.truncated,
         result_text_sha256: resultSha256,
+        completion_status: resultObservation?.completionStatus ?? "final",
+        output_kind: resultObservation?.outputKind ?? "final_answer",
+        usable_for_synthesis: resultObservation?.usableForSynthesis ?? true,
+        missing_contract: input.outputContract === "final_assistant_text" &&
+            (resultObservation?.completionStatus !== "final" ||
+                ["empty", "process_notes", "tool_trace_only"].includes(String(resultObservation?.outputKind ?? ""))),
         created_at: observedAt,
         dispatch_authority_enabled: false,
     };
-    writeSessionEvidence({
+    const taskResultWritten = writeSessionEvidence({
         rootDir: input.rootDir,
         workflowId: input.workflowId,
         evidenceId: taskResultEvidenceId,
         record: taskResultRecord,
     });
+    if (!taskResultWritten) {
+        return {
+            status: "task_failed",
+            failureCategory: "unknown",
+            redactedReason: "task_result evidence persistence failed",
+            laneId: input.laneId,
+        };
+    }
+    writeAgentTaskProgress({
+        rootDir: input.rootDir,
+        workflowId: input.workflowId,
+        laneId: input.laneId,
+        taskId: input.taskId,
+        agentRef: input.agentRef,
+        providerQualifiedModelId: input.providerQualifiedModelId,
+        phase: "finalizing",
+        progressSeq: 3,
+        progressLabel: "agent task result captured",
+    });
     writeAgentTaskTerminalLifecycle({
         rootDir: input.rootDir,
         workflowId: input.workflowId,
@@ -380,6 +704,11 @@ export async function executeFlowDeskAgentTaskV1(input) {
         updatedAt: new Date().toISOString(),
         timeoutMs: input.timeoutMs,
     });
+    refreshFlowDeskCompletionUiCachesV1({
+        rootDir: input.rootDir,
+        workflowId: input.workflowId,
+        observedAt,
+    });
     return {
         status: "task_completed",
         resultText: fullResultText,