pilotswarm-sdk 0.3.1 → 0.3.2

package/dist/managed-session.js

@@ -111,7 +111,7 @@ function acknowledgeTurnBoundary(action) {
     return `[SYSTEM: ${action} acknowledged. The runtime will suspend at the end of this turn. ` +
         `Finish any remaining tool results for the current step, then stop.]`;
 }
-const TERMINAL_TURN_BOUNDARY_ACTIONS = new Set(["wait", "input_required", "wait_for_agents", "list_sessions", "check_agents"]);
+const TERMINAL_TURN_BOUNDARY_ACTIONS = new Set(["completed", "wait", "input_required", "wait_for_agents", "list_sessions", "check_agents"]);
 function hasTerminalTurnBoundary(turnState) {
     return turnState.pendingActions.some((action) => TERMINAL_TURN_BOUNDARY_ACTIONS.has(action.type));
 }
@@ -119,6 +119,77 @@ function blockedAfterTurnBoundary(toolName) {
     return `[SYSTEM: ${toolName} was not executed because a previous control tool already scheduled this turn to suspend. ` +
         `Stop now; the runtime will resume with the control-tool result.]`;
 }
+function splitQualifiedModel(model) {
+    const configured = String(model || "").trim();
+    if (!configured)
+        return { provider: "(default)", model: "(default)" };
+    const separator = configured.indexOf(":");
+    if (separator <= 0)
+        return { provider: "(unqualified)", model: configured };
+    return {
+        provider: configured.slice(0, separator),
+        model: configured.slice(separator + 1),
+    };
+}
+function formatCurrentModelConfig(config) {
+    const configured = String(config.model || "").trim() || "(default)";
+    const { provider, model } = splitQualifiedModel(config.model);
+    const reasoningEffort = config.reasoningEffort ?? "(default)";
+    return [
+        "Current session configured model (this turn):",
+        `- provider: ${provider}`,
+        `- model: ${model}`,
+        `- qualified_model: ${configured}`,
+        `- reasoning_effort: ${reasoningEffort}`,
+    ].join("\n");
+}
+// ── Tool-call-as-text guard ──────────────────────────────────────
+// Some models (observed on claude-opus-4.8, especially on repetitive keepalive
+// cron cycles) intermittently emit a tool call as literal
+// `<invoke name="...">`/`<parameter>` text inside the assistant message instead
+// of a real tool_use block. That text is never executed, so a consequential
+// call (store_fact, complete_agent, an ADO write, etc.) would be silently
+// dropped while the transcript implies it happened. We detect the malformed
+// text and re-prompt the model — bounded — to actually invoke the tool.
+const MAX_TEXT_TOOL_CALL_CORRECTIONS = 2;
+const TEXT_TOOL_CALL_INVOKE_RE = /<(?:antml:)?invoke\s+name\s*=\s*"([^"]+)"/i;
+const TEXT_TOOL_CALL_STRUCTURE_RE = /<\/(?:antml:)?invoke\s*>|<(?:antml:)?parameter\b/i;
+const FENCED_CODE_BLOCK_RE = /```[\s\S]*?```/g;
+/**
+ * Detect a tool call the model emitted as literal text instead of a real
+ * tool_use block. Requires both the `<invoke name="...">` opener and a closing
+ * `</invoke>` or a `<parameter>` tag so prose that merely mentions the word
+ * "invoke" does not trip the guard. Returns the tool name, or null.
+ */
+function detectTextEmittedToolCall(content) {
+    if (typeof content !== "string" || content.length === 0)
+        return null;
+    const withoutExamples = content.replace(FENCED_CODE_BLOCK_RE, "");
+    const nameMatch = withoutExamples.match(TEXT_TOOL_CALL_INVOKE_RE);
+    if (!nameMatch)
+        return null;
+    if (!TEXT_TOOL_CALL_STRUCTURE_RE.test(withoutExamples))
+        return null;
+    const beforeInvoke = withoutExamples.slice(0, nameMatch.index ?? 0).trim();
+    const afterClose = withoutExamples.replace(/[\s\S]*<\/(?:antml:)?invoke\s*>/i, "").trim();
+    if (afterClose)
+        return null;
+    // Allow the common one-token junk prefix observed from claude-opus-4.8
+    // (for example "court\n<invoke ...>") but do not flag explanatory prose,
+    // markdown docs, or examples that happen to contain Anthropic XML syntax.
+    if (beforeInvoke && beforeInvoke.split(/\s+/).length > 1)
+        return null;
+    return { toolName: (nameMatch[1] || "").trim() || "the requested tool", rawContent: content };
+}
+function buildTextEmittedToolCallCorrection(toolName) {
+    const named = toolName && toolName !== "the requested tool" ? `the "${toolName}" tool` : "the intended tool";
+    return `[SYSTEM: Tool-call protocol error. Your previous message contained a tool call written as literal text ` +
+        `(for example \`<invoke name="${toolName}">\` with <parameter> tags). Text formatted like that is NOT executed — ` +
+        `the tool did not run and produced no result, so anything you implied there has NOT actually happened. ` +
+        `Come to your senses and actually invoke ${named} now using the real tool-calling mechanism, not text. ` +
+        `Do not write <invoke> or <parameter> tags as message content. ` +
+        `If you did not actually need a tool this turn, reply with plain prose only and no tool-call markup.]`;
+}
 function failureToolResult(error) {
     const message = error instanceof Error ? error.message : String(error ?? "Tool failed");
     return {
@@ -285,9 +356,9 @@ export class ManagedSession {
             handler: async () => "stub",
         });
         const reportCycleTool = defineTool("report_cycle", {
-            description: "Report the outcome of the current recurring cron/cron_at watcher cycle. " +
-                "Use status='quiet' when nothing material changed, status='material' when the parent should be notified, " +
-                "and status='blocked' when the cycle found a blocker or failure that needs parent attention. " +
+            description: "Report the outcome of the current recurring cron/cron_at watcher cycle when something material happened. " +
+                "Use status='material' when the parent should be notified, and status='blocked' when the cycle found a blocker or failure that needs parent attention. " +
+                "On an uneventful cycle, prefer NOT calling this tool at all — just end the turn silently; status='quiet' is accepted but unnecessary. " +
                 "This tool does not end the turn; after calling it, finish normally. It is ignored outside recurring watcher cycles.",
             parameters: {
                 type: "object",
@@ -303,6 +374,7 @@ export class ManagedSession {
         const listModelsTool = defineTool("list_available_models", {
             description: "List all available LLM models across all configured providers. " +
                 "Returns each model's exact qualified name (provider:model), description, and cost tier. " +
+                "Also returns this session's current configured provider, model, and reasoning effort for the current turn. " +
                 "This output is the authoritative source for model selection. " +
                 "Use this when choosing the best model for a sub-agent task, or when the user asks about available models. " +
                 "If you plan to pass spawn_agent(model=...), you must choose an exact provider:model value from this list and must not invent or shorten names. " +
@@ -315,13 +387,28 @@ export class ManagedSession {
             },
             handler: async () => "stub",
         });
+        const setSessionModelTool = defineTool("set_session_model", {
+            description: "Switch this session's model for the next turn boundary. " +
+                "Call list_available_models first and pass an exact provider:model value returned there. " +
+                "This ends the current turn. After it succeeds, stop; the runtime will continue on the selected model.",
+            parameters: {
+                type: "object",
+                properties: {
+                    model: { type: "string", description: "Exact provider:model value from list_available_models." },
+                    reasoning_effort: { type: "string", enum: ["low", "medium", "high", "xhigh"], description: "Optional reasoning effort supported by the selected model." },
+                },
+                required: ["model"],
+            },
+            handler: async () => "stub",
+        });
         const updateSessionSummaryTool = defineTool("update_session_summary", {
-            description: "Update this session's short live summary for session lists, discovery, and the Summary tab. " +
+            description: "Update this session's short live summary and optionally set this session's sticky title for session lists, discovery, and the Summary tab. " +
                 "Call it automatically after first meaningful work and after each notable update: changed intent, tangible progress toward the user's goal, received cross-session replies, delivered outputs, blockers, open questions, next actions, key links, schedule/delegate changes, or terminal state. " +
+                "Pass title when the user asks you to rename this session or when a durable human-readable title should stick; title updates lock the title against future automatic title summarization. " +
                 "Keep it concise and scannable; use compact bullets or short Markdown tables for structured progress, comparisons, rankings, decisions, or result sets instead of prose blobs. " +
                 "Do not paste long transcripts, raw logs, or bulky JSON into summary fields. " +
                 "Do not call it for no-op heartbeats, timer wakes, or unchanged cron cycles. " +
-                "Do not pass a string for summary_state. " + SESSION_SUMMARY_STATE_TEMPLATE,
+                "Do not pass a string for summary_state. summary_state is optional only when title is provided. " + SESSION_SUMMARY_STATE_TEMPLATE,
             parameters: {
                 type: "object",
                 properties: {
@@ -330,8 +417,8 @@ export class ManagedSession {
                         description: "Structured live summary state. Must be an object, not a string. Missing arrays should be [].",
                     },
                     short_summary: { type: "string", description: "Optional concise summary for session lists. If omitted, summary_state.summary is used." },
+                    title: { type: "string", description: "Optional sticky session title. When set, it behaves like a manual rename and prevents future automatic title changes." },
                 },
-                required: ["summary_state"],
             },
             handler: async () => "stub",
         });
@@ -368,7 +455,7 @@ export class ManagedSession {
             },
             handler: async () => "stub",
         });
-        return [waitTool, waitOnWorkerTool, cronTool, cronAtTool, askUserTool, reportCycleTool, listModelsTool, updateSessionSummaryTool, sendSessionMessageTool, replySessionMessageTool];
+        return [waitTool, waitOnWorkerTool, cronTool, cronAtTool, askUserTool, reportCycleTool, listModelsTool, setSessionModelTool, updateSessionSummaryTool, sendSessionMessageTool, replySessionMessageTool];
     }
     /**
      * Sub-agent tool definitions.
@@ -637,9 +724,9 @@ export class ManagedSession {
             },
         });
         const reportCycleTool = defineTool("report_cycle", {
-            description: "Report the outcome of the current recurring cron/cron_at watcher cycle. " +
-                "Use status='quiet' when nothing material changed, status='material' when the parent should be notified, " +
-                "and status='blocked' when the cycle found a blocker or failure that needs parent attention. " +
+            description: "Report the outcome of the current recurring cron/cron_at watcher cycle when something material happened. " +
+                "Use status='material' when the parent should be notified, and status='blocked' when the cycle found a blocker or failure that needs parent attention. " +
+                "On an uneventful cycle, prefer NOT calling this tool at all — just end the turn silently; status='quiet' is accepted but unnecessary. " +
                 "This tool does not end the turn; after calling it, finish normally. It is ignored outside recurring watcher cycles.",
             parameters: {
                 type: "object",
@@ -892,6 +979,7 @@ export class ManagedSession {
         const listModelsTool = defineTool("list_available_models", {
             description: "List all available LLM models across all configured providers. " +
                 "Returns each model's exact qualified name (provider:model), description, and cost tier. " +
+                "Also returns this session's current configured provider, model, and reasoning effort for the current turn. " +
                 "This output is the authoritative source for model selection. " +
                 "Use this when choosing the best model for a sub-agent task, or when the user asks about available models. " +
                 "If you plan to pass spawn_agent(model=...), you must choose an exact provider:model value from this list and must not invent or shorten names. " +
@@ -903,16 +991,62 @@ export class ManagedSession {
                 properties: {},
             },
             handler: async () => {
-                return opts?.modelSummary || "No model providers configured.";
+                return [
+                    formatCurrentModelConfig(this.config),
+                    opts?.modelSummary || "No model providers configured.",
+                ].join("\n\n");
+            },
+        });
+        const setSessionModelTool = defineTool("set_session_model", {
+            description: "Switch this session's model for the next turn boundary. " +
+                "Call list_available_models first and pass an exact provider:model value returned there. " +
+                "This ends the current turn. After it succeeds, stop; the runtime will continue on the selected model.",
+            parameters: {
+                type: "object",
+                properties: {
+                    model: { type: "string", description: "Exact provider:model value from list_available_models." },
+                    reasoning_effort: { type: "string", enum: ["low", "medium", "high", "xhigh"], description: "Optional reasoning effort supported by the selected model." },
+                },
+                required: ["model"],
+            },
+            handler: async (args) => {
+                if (hasTerminalTurnBoundary(turnState))
+                    return blockedAfterTurnBoundary("set_session_model");
+                const model = String(args.model || "").trim();
+                if (!model)
+                    return "Error: model is required.";
+                const reasoningEffort = args.reasoning_effort ? normalizeReasoningEffort(args.reasoning_effort) : undefined;
+                if (args.reasoning_effort && !reasoningEffort) {
+                    return "Error: reasoning_effort must be one of low, medium, high, xhigh.";
+                }
+                if (!controlBridge)
+                    return "Error: set_session_model is unavailable in this session.";
+                const result = await controlBridge.setSessionModel({ model, ...(reasoningEffort ? { reasoning_effort: reasoningEffort } : {}) });
+                if (/model switch accepted/i.test(String(result))) {
+                    turnState.pendingActions.push({
+                        type: "completed",
+                        content: "Model switch requested. Continuing on the selected model.",
+                    });
+                    return `${result}\n${acknowledgeTurnBoundary("set_session_model")}`;
+                }
+                if (/set_session_model failed/i.test(String(result)) || !/model switch accepted/i.test(String(result))) {
+                    turnState.pendingActions.push({
+                        type: "completed",
+                        content: "Model switch failed. Continuing on the unchanged model.",
+                    });
+                    return `${result}\n${acknowledgeTurnBoundary("set_session_model")}`;
+                }
+                return result;
             },
         });
         const updateSessionSummaryTool = defineTool("update_session_summary", {
-            description: "Update this session's short live summary for session lists, discovery, and the Summary tab. " +
+            description: "Update this session's short live summary and optionally set this session's sticky title for session lists, discovery, and the Summary tab. " +
                 "Call it automatically after first meaningful work and after each notable update: changed intent, tangible progress toward the user's goal, received cross-session replies, delivered outputs, blockers, open questions, next actions, key links, schedule/delegate changes, or terminal state. " +
+                "Pass title when the user asks you to rename this session or when a durable human-readable title should stick; title updates lock the title against future automatic title summarization. " +
                 "Keep it concise and scannable; use compact bullets or short Markdown tables for structured progress, comparisons, rankings, decisions, or result sets instead of prose blobs. " +
                 "Do not paste long transcripts, raw logs, or bulky JSON into summary fields. " +
                 "Do not call it for no-op heartbeats, timer wakes, or unchanged cron cycles. " +
-                "Do not pass a string for summary_state. " + SESSION_SUMMARY_STATE_TEMPLATE,
+                "Do not pass a string for summary_state. summary_state is optional only when title is provided. " + SESSION_SUMMARY_STATE_TEMPLATE,
             parameters: {
                 type: "object",
                 properties: {
@@ -921,8 +1055,8 @@ export class ManagedSession {
                         description: "Structured live summary state. Must be an object, not a string. Missing arrays should be [].",
                     },
                     short_summary: { type: "string", description: "Optional concise summary for session lists. If omitted, summary_state.summary is used." },
+                    title: { type: "string", description: "Optional sticky session title. When set, it behaves like a manual rename and prevents future automatic title changes." },
                 },
-                required: ["summary_state"],
             },
             handler: async (args) => {
                 if (hasTerminalTurnBoundary(turnState))
@@ -1271,7 +1405,7 @@ export class ManagedSession {
                 return acknowledgeTurnBoundary("delete_agent");
             },
         });
-        const SYSTEM_TOOL_NAMES = new Set(["wait", "wait_on_worker", "cron", "cron_at", "ask_user", "report_cycle", "list_available_models", "update_session_summary", "send_session_message", "reply_session_message", "spawn_agent", "message_agent", "check_agents", "wait_for_agents", "list_sessions", "complete_agent", "cancel_agent", "delete_agent"]);
+        const SYSTEM_TOOL_NAMES = new Set(["wait", "wait_on_worker", "cron", "cron_at", "ask_user", "report_cycle", "list_available_models", "set_session_model", "update_session_summary", "send_session_message", "reply_session_message", "spawn_agent", "message_agent", "check_agents", "wait_for_agents", "list_sessions", "complete_agent", "cancel_agent", "delete_agent"]);
         // Merge user tools with system tools
         const userTools = this.config.tools ?? [];
         // Wrap user tool handlers to augment invocation with the PilotSwarm
@@ -1310,6 +1444,7 @@ export class ManagedSession {
             askUserTool,
             reportCycleTool,
             listModelsTool,
+            setSessionModelTool,
             updateSessionSummaryTool,
             sendSessionMessageTool,
             replySessionMessageTool,
@@ -1342,6 +1477,7 @@ export class ManagedSession {
         let lastPublishedReasoning = "";
         let lastReasoningPublishAt = 0;
         let deferredSessionError = null;
+        const textEmittedToolCallRef = { current: null };
         // Streaming progress + turn timing state.
         // Token-level deltas (`assistant.message_delta`,
         // `assistant.streaming_delta`, `assistant.reasoning_delta`) stay
@@ -1480,6 +1616,16 @@ export class ManagedSession {
                 if (isEmptyAssistantTranscriptEvent(eventType, eventData)) {
                     return;
                 }
+                if (eventType === "assistant.message") {
+                    const content = extractAssistantMessageContent({ data: eventData });
+                    const textToolCall = detectTextEmittedToolCall(content);
+                    if (textToolCall) {
+                        textEmittedToolCallRef.current = textToolCall;
+                        return;
+                    }
+                    finalContent = content ?? finalContent;
+                    publishReasoningSnapshot("assistant.message", true);
+                }
                 // Track turn boundaries so we can stamp turn_end with a
                 // durationMs and the streaming counters.
                 if (eventType === "assistant.turn_start") {
@@ -1547,11 +1693,6 @@ export class ManagedSession {
                     catch { }
                 }
             }));
-            // Capture the final assistant message
-            unsubscribers.push(this.copilotSession.on("assistant.message", (event) => {
-                finalContent = extractAssistantMessageContent(event) ?? finalContent;
-                publishReasoningSnapshot("assistant.message", true);
-            }));
             unsubscribers.push(this.copilotSession.on("assistant.reasoning", (event) => {
                 currentReasoning = String(extractReasoningText(event?.data ?? event) || "").trim();
                 if (currentReasoning) {
@@ -1594,6 +1735,18 @@ export class ManagedSession {
                 setTimeout(() => reject(new Error("Turn timed out")), TURN_TIMEOUT);
             })
             : null;
+        // Re-armable idle waiter used by the tool-call-as-text guard to wait for
+        // the model's response to a mid-turn correction without tearing down the
+        // event subscriptions set up above.
+        const waitForNextIdle = () => new Promise((resolve) => {
+            const unsub = this.copilotSession.on("session.idle", () => {
+                flushStreamingProgress(true);
+                publishReasoningSnapshot("session.idle", true);
+                unsub();
+                resolve();
+            });
+            unsubscribers.push(unsub);
+        });
         const effectivePrompt = opts?.requiredTool
             ? [
                 `[SYSTEM: For this request, you MUST invoke the tool "${opts.requiredTool}" before giving your answer.`,
@@ -1618,6 +1771,61 @@ export class ManagedSession {
             else {
                 await turnComplete;
             }
+            // ── Guard: tool call emitted as text instead of executed ──────────
+            // If the model typed a tool call as `<invoke .../>` text rather than
+            // calling it, that call did not run. Re-prompt it (bounded) to
+            // actually invoke the tool so a consequential call is never silently
+            // dropped. If a control tool already scheduled a turn boundary, we
+            // still return an error rather than accepting a transcript that
+            // implied another unexecuted tool side effect.
+            let textToolCallCorrections = 0;
+            while (textToolCallCorrections < MAX_TEXT_TOOL_CALL_CORRECTIONS &&
+                !hasTerminalTurnBoundary(turnState) &&
+                textEmittedToolCallRef.current) {
+                const offendingTool = textEmittedToolCallRef.current.toolName;
+                const rawContent = textEmittedToolCallRef.current.rawContent;
+                textEmittedToolCallRef.current = null;
+                textToolCallCorrections++;
+                const diagnostic = {
+                    eventType: "runtime.tool_call_as_text",
+                    data: { toolName: offendingTool, rawContent, attempt: textToolCallCorrections, sessionId: this.sessionId },
+                };
+                collectedEvents.push(diagnostic);
+                if (opts?.onEvent) {
+                    try {
+                        opts.onEvent(diagnostic);
+                    }
+                    catch { }
+                }
+                finalContent = undefined;
+                const nextIdle = waitForNextIdle();
+                await this.copilotSession.send({ prompt: buildTextEmittedToolCallCorrection(offendingTool) });
+                if (timeoutPromise) {
+                    await Promise.race([nextIdle, timeoutPromise]);
+                }
+                else {
+                    await nextIdle;
+                }
+            }
+            if (textEmittedToolCallRef.current) {
+                const diagnostic = {
+                    eventType: "runtime.tool_call_as_text",
+                    data: { toolName: textEmittedToolCallRef.current.toolName, rawContent: textEmittedToolCallRef.current.rawContent, final: true, sessionId: this.sessionId },
+                };
+                collectedEvents.push(diagnostic);
+                if (opts?.onEvent) {
+                    try {
+                        opts.onEvent(diagnostic);
+                    }
+                    catch { }
+                }
+                const message = buildTextEmittedToolCallCorrection(textEmittedToolCallRef.current.toolName);
+                return {
+                    type: "error",
+                    message,
+                    events: collectedEvents,
+                };
+            }
         }
         catch (err) {
             // Timeout — kill it
@@ -1648,6 +1856,8 @@ export class ManagedSession {
             const combinedQueuedActions = [...turnState.queuedActions, ...remainingActions];
             const queuedActions = combinedQueuedActions.length > 0 ? combinedQueuedActions : undefined;
             switch (firstAction.type) {
+                case "completed":
+                    return { ...firstAction, events: collectedEvents, queuedActions };
                 case "input_required":
                     return { ...firstAction, events: collectedEvents, queuedActions };
                 case "wait":
@@ -1723,6 +1933,8 @@ export class ManagedSession {
     updateConfig(config) {
         if (config.model !== undefined)
             this.config.model = config.model;
+        if (config.reasoningEffort !== undefined)
+            this.config.reasoningEffort = config.reasoningEffort;
         if (config.tools !== undefined)
             this.config.tools = config.tools;
         if (config.systemMessage !== undefined)
@@ -1732,6 +1944,16 @@ export class ManagedSession {
         if (config.waitThreshold !== undefined)
             this.config.waitThreshold = config.waitThreshold;
     }
+    requiresModelRebind(config) {
+        const currentModel = this.config.model;
+        const nextModel = config.model ?? this.config.model;
+        const currentReasoningEffort = this.config.reasoningEffort ?? null;
+        const nextReasoningEffort = config.reasoningEffort !== undefined
+            ? config.reasoningEffort ?? null
+            : this.config.reasoningEffort ?? null;
+        return Boolean((currentModel || nextModel)
+            && (currentModel !== nextModel || currentReasoningEffort !== nextReasoningEffort));
+    }
     /** Get the underlying CopilotSession (for direct access when needed). */
     getCopilotSession() {
         return this.copilotSession;