@chrok/pi-braid 0.3.6 → 0.5.0

package/dist/runner.js

@@ -1,10 +1,28 @@
 import { resolve } from "node:path";
-import { StringEnum, Type, validateToolCall, } from "@earendil-works/pi-ai";
-import { formatBudgetReminder, gitToolDefinition, finishMergeToolDefinition, mergeInstructions, parseGitToolArguments, parseFinishMergeArguments, } from "@chrok/braid";
+import { StringEnum, Type, isRetryableAssistantError, retryDelayMs, validateToolCall, } from "@earendil-works/pi-ai";
+import { formatBudgetReminder, interruptedToolResult, isStepInterrupted, notRunToolResult, gitToolDefinition, finishMergeToolDefinition, mergeInstructions, parseGitToolArguments, parseFinishMergeArguments, } from "@chrok/braid";
 import { createWorktreeWriteTools } from "./write-tools.js";
 import { createAvailableReadTools } from "./read-tools.js";
 import { createWorkspaceShellTools } from "./shell-tools.js";
 import { formatHumanInput } from "./human-input.js";
+const DEFAULT_MODEL_RETRIES = 3;
+/** Caps both Braid's backoff and a provider's Retry-After; longer requested waits fail immediately. */
+const MAX_MODEL_RETRY_DELAY_MS = 60_000;
+function abortableDelay(ms, signal) {
+    return new Promise((resolve, reject) => {
+        if (signal.aborted)
+            return reject(signal.reason);
+        const onAbort = () => {
+            clearTimeout(timer);
+            reject(signal.reason);
+        };
+        const timer = setTimeout(() => {
+            signal.removeEventListener("abort", onAbort);
+            resolve();
+        }, ms);
+        signal.addEventListener("abort", onAbort, { once: true });
+    });
+}
 /**
  * Present a Git result the way Pi's bash tool presents command output. Output
  * stays byte-exact, so a diff can be passed back to git apply.
@@ -30,6 +48,14 @@ export function createPiRunner(registry, onUsageOrOptions, cwd = process.cwd())
             throw new TypeError(`${name} must be Infinity or a positive safe integer`);
         }
     }
+    const maxModelRetries = options.maxModelRetries ?? DEFAULT_MODEL_RETRIES;
+    const modelRetryDelayMs = options.modelRetryDelayMs ?? 2000;
+    if (!Number.isSafeInteger(maxModelRetries) || maxModelRetries < 0) {
+        throw new TypeError("maxModelRetries must be a non-negative safe integer");
+    }
+    if (!Number.isFinite(modelRetryDelayMs) || modelRetryDelayMs < 0) {
+        throw new TypeError("modelRetryDelayMs must be a non-negative finite number");
+    }
     const onUsage = options.onUsage;
     const onProgress = options.onProgress;
     const reportProgress = (progress) => {
@@ -45,7 +71,7 @@ export function createPiRunner(registry, onUsageOrOptions, cwd = process.cwd())
         messages: context.messages,
         tools: context.tools,
     }).length / 4);
-    return async (request) => {
+    const runner = async (request) => {
         const name = request.model;
         const slash = name?.indexOf("/") ?? -1;
         if (!name || slash < 1 || slash === name.length - 1) {
@@ -182,8 +208,10 @@ export function createPiRunner(registry, onUsageOrOptions, cwd = process.cwd())
         let decided = false;
         let actualModel = name;
         let reportedContextTokens;
-        const complete = async () => {
+        const prepareRequest = async () => {
             request.signal.throwIfAborted();
+            // A pause takes effect here; messages sent while suspended are delivered after resume.
+            await request.checkpoint?.();
             deliverHumanInputs();
             const budgets = [];
             if (Number.isFinite(maxToolRounds)) {
@@ -195,6 +223,7 @@ export function createPiRunner(registry, onUsageOrOptions, cwd = process.cwd())
             if (budgets.length > 0) {
                 budgets.push("When a tool budget reaches zero, make no further tool calls and return your final answer. Reserve budget for decide or finish_merge if required.");
             }
+            // Rebuilt per attempt: a suspension moves the node deadline.
             context.systemPrompt = systemPrompt + formatBudgetReminder(request, budgets);
             reportProgress({
                 nodeId: request.node.id,
@@ -206,82 +235,125 @@ export function createPiRunner(registry, onUsageOrOptions, cwd = process.cwd())
                 toolRounds,
                 phase: "model",
             });
-            // The registry resolves stored API keys, OAuth, custom headers and endpoints.
-            // complete() is stream().result(); streaming only adds observation.
-            const requestOptions = {
-                signal: request.signal,
-                sessionId,
-                cacheRetention: "none",
-                transport: "sse",
-                maxRetries: 0,
-            };
-            activity?.modelRequest();
-            let response;
-            try {
-                if (registry.stream) {
-                    const stream = registry.stream(model, context, requestOptions);
-                    for await (const event of stream) {
-                        if (event.type !== "done" && event.type !== "error")
-                            transcripts?.streaming(executionId, event.partial);
-                        if (event.type === "text_delta")
-                            activity?.modelStream("text", event.delta);
-                        else if (event.type === "thinking_delta")
-                            activity?.modelStream("reasoning", event.delta);
-                        else if (event.type === "toolcall_start" || event.type === "toolcall_delta") {
-                            const block = event.partial.content[event.contentIndex];
-                            activity?.modelStream("tool_call", event.type === "toolcall_delta" ? event.delta : "", block?.type === "toolCall" ? block.name : undefined);
+        };
+        const complete = async () => {
+            let retries = 0;
+            while (true) {
+                // Retries pass the checkpoint too, so a pause during backoff stops before the resend.
+                await prepareRequest();
+                // A node timeout pause interrupts only this step; the request is sent again after resume.
+                const step = request.stepSignal?.() ?? request.signal;
+                // The registry resolves stored API keys, OAuth, custom headers and endpoints.
+                // complete() is stream().result(); streaming only adds observation.
+                // pi-ai retries rejected requests before a response starts, honoring Retry-After.
+                const requestOptions = {
+                    signal: step,
+                    sessionId,
+                    cacheRetention: "none",
+                    transport: "sse",
+                    maxRetries: maxModelRetries,
+                    maxRetryDelayMs: MAX_MODEL_RETRY_DELAY_MS,
+                };
+                activity?.modelRequest();
+                let response;
+                // Without a stream we cannot tell where a request failed, so Braid may retry it too.
+                let started = !registry.stream;
+                try {
+                    if (registry.stream) {
+                        const stream = registry.stream(model, context, requestOptions);
+                        for await (const event of stream) {
+                            if (event.type === "start")
+                                started = true;
+                            if (event.type !== "done" && event.type !== "error")
+                                transcripts?.streaming(executionId, event.partial);
+                            if (event.type === "text_delta")
+                                activity?.modelStream("text", event.delta);
+                            else if (event.type === "thinking_delta")
+                                activity?.modelStream("reasoning", event.delta);
+                            else if (event.type === "toolcall_start" || event.type === "toolcall_delta") {
+                                const block = event.partial.content[event.contentIndex];
+                                activity?.modelStream("tool_call", event.type === "toolcall_delta" ? event.delta : "", block?.type === "toolCall" ? block.name : undefined);
+                            }
                         }
+                        response = await stream.result();
+                    }
+                    else {
+                        response = await registry.complete(model, context, requestOptions);
                     }
-                    response = await stream.result();
                 }
-                else {
-                    response = await registry.complete(model, context, requestOptions);
+                catch (error) {
+                    if (isStepInterrupted(request, step)) {
+                        activity?.modelError("Model request interrupted by the node timeout");
+                        continue;
+                    }
+                    activity?.modelError(error instanceof Error ? error.message : String(error));
+                    throw error;
                 }
-            }
-            catch (error) {
-                activity?.modelError(error instanceof Error ? error.message : String(error));
-                throw error;
-            }
-            transcripts?.assistant(executionId, response);
-            activity?.modelResponse({
-                text: response.content.flatMap(block => block.type === "text" ? [block.text] : []).join("\n\n"),
-                reasoning: response.content.flatMap(block => block.type === "thinking" ? [block.thinking] : []).join("\n\n"),
-                toolCalls: response.content.flatMap(block => block.type === "toolCall" ? [{ id: block.id, name: block.name }] : []),
-                stopReason: response.stopReason,
-                inputTokens: response.usage.input + response.usage.cacheRead + response.usage.cacheWrite,
-                outputTokens: response.usage.output,
-            });
-            if (response.stopReason !== "stop" && response.stopReason !== "toolUse") {
-                activity?.modelError(response.errorMessage || `Pi model ended with '${response.stopReason}'`);
-            }
-            actualModel = `${response.provider}/${response.responseModel ?? response.model}`;
-            const providerContextTokens = response.usage.input +
-                response.usage.cacheRead +
-                response.usage.cacheWrite;
-            if (providerContextTokens)
-                reportedContextTokens = providerContextTokens;
-            reportProgress({
-                nodeId: request.node.id,
-                ...(request.execution.executionId ? { executionId: request.execution.executionId } : {}),
-                contextTokens: reportedContextTokens ?? estimateContextTokens(context),
-                contextWindow: model.contextWindow,
-                contextSource: reportedContextTokens !== undefined ? "reported" : "estimate",
-                toolCalls,
-                toolRounds,
-                phase: "model",
-            });
-            onUsage?.(structuredClone(response.usage));
-            usage.inputTokens +=
-                response.usage.input +
+                // A failed attempt stays visible in the session view but never enters the context.
+                transcripts?.assistant(executionId, response);
+                activity?.modelResponse({
+                    text: response.content.flatMap(block => block.type === "text" ? [block.text] : []).join("\n\n"),
+                    reasoning: response.content.flatMap(block => block.type === "thinking" ? [block.thinking] : []).join("\n\n"),
+                    toolCalls: response.content.flatMap(block => block.type === "toolCall" ? [{ id: block.id, name: block.name }] : []),
+                    stopReason: response.stopReason,
+                    inputTokens: response.usage.input + response.usage.cacheRead + response.usage.cacheWrite,
+                    outputTokens: response.usage.output,
+                });
+                if (response.stopReason !== "stop" && response.stopReason !== "toolUse") {
+                    activity?.modelError(response.errorMessage || `Pi model ended with '${response.stopReason}'`);
+                }
+                actualModel = `${response.provider}/${response.responseModel ?? response.model}`;
+                const providerContextTokens = response.usage.input +
                     response.usage.cacheRead +
                     response.usage.cacheWrite;
-            usage.outputTokens += response.usage.output;
-            request.signal.throwIfAborted();
-            if (response.stopReason !== "stop" && response.stopReason !== "toolUse") {
-                throw new Error(response.errorMessage ||
-                    `Pi model ended with '${response.stopReason}'`);
+                if (providerContextTokens)
+                    reportedContextTokens = providerContextTokens;
+                reportProgress({
+                    nodeId: request.node.id,
+                    ...(request.execution.executionId ? { executionId: request.execution.executionId } : {}),
+                    contextTokens: reportedContextTokens ?? estimateContextTokens(context),
+                    contextWindow: model.contextWindow,
+                    contextSource: reportedContextTokens !== undefined ? "reported" : "estimate",
+                    toolCalls,
+                    toolRounds,
+                    phase: "model",
+                });
+                // Failed attempts are billed too.
+                onUsage?.(structuredClone(response.usage));
+                usage.inputTokens +=
+                    response.usage.input +
+                        response.usage.cacheRead +
+                        response.usage.cacheWrite;
+                usage.outputTokens += response.usage.output;
+                request.signal.throwIfAborted();
+                // The partial response never enters the context.
+                if (isStepInterrupted(request, step))
+                    continue;
+                // A stream that failed after it started is restarted from the same context. Tools run
+                // only after a complete response, so nothing from the failed attempt has executed.
+                if (response.stopReason === "error" && started && retries < maxModelRetries && isRetryableAssistantError(response)) {
+                    const delayMs = retryDelayMs({ baseDelayMs: modelRetryDelayMs, maxAgentDelayMs: MAX_MODEL_RETRY_DELAY_MS }, retries + 1);
+                    const deadline = Math.min(request.deadlines?.node ?? Infinity, request.deadlines?.graph ?? Infinity);
+                    // Report the provider error rather than a timeout when the wait cannot finish in time.
+                    if (performance.now() + delayMs < deadline) {
+                        retries++;
+                        activity?.modelError(`Retrying model request in ${Math.ceil(delayMs / 1000)}s (${retries}/${maxModelRetries})`);
+                        try {
+                            await abortableDelay(delayMs, step);
+                        }
+                        catch (error) {
+                            if (!isStepInterrupted(request, step))
+                                throw error;
+                        }
+                        continue;
+                    }
+                }
+                if (response.stopReason !== "stop" && response.stopReason !== "toolUse") {
+                    throw new Error(response.errorMessage ||
+                        `Pi model ended with '${response.stopReason}'`);
+                }
+                return response;
             }
-            return response;
         };
         const toolResult = (call, content, isError) => ({
             role: "toolResult",
@@ -291,11 +363,12 @@ export function createPiRunner(registry, onUsageOrOptions, cwd = process.cwd())
             isError,
             timestamp: Date.now(),
         });
-        const executeFileTool = async (call) => {
+        const executeFileTool = async (call, signal) => {
             const tool = toolByName.get(call.name);
             if (!tool && !(call.name === "git" && request.git) && !(call.name === "finish_merge" && request.merge)) {
                 return toolResult(call, `Tool '${call.name}' is unavailable in this Braid node. Available tools: ${allToolDefinitions.map(tool => tool.name).join(", ")}.`, true);
             }
+            let partialOutput = "";
             try {
                 request.signal.throwIfAborted();
                 if (call.name === "git" && request.git) {
@@ -309,17 +382,19 @@ export function createPiRunner(registry, onUsageOrOptions, cwd = process.cwd())
                     return toolResult(call, "Merge dispositions recorded. Return your final answer.", false);
                 }
                 const args = validateToolCall(allToolDefinitions, call);
-                const execute = () => tool.execute(call.id, args, request.signal, 
+                const execute = () => tool.execute(call.id, args, signal, 
                 // Bash sends cumulative, throttled output snapshots.
                 (partial) => {
-                    activity?.toolOutput(call.id, partial.content
-                        .flatMap(block => block.type === "text" ? [block.text] : []).join(""), "replace");
+                    partialOutput = partial.content.flatMap(block => block.type === "text" ? [block.text] : []).join("");
+                    activity?.toolOutput(call.id, partialOutput, "replace");
                     transcripts?.toolUpdate(executionId, call.id, { content: partial.content, details: partial.details });
                 });
                 const result = ["write", "edit", "bash", "powershell"].includes(call.name) && request.withWorkspaceWrite
                     ? await request.withWorkspaceWrite(execute)
                     : await execute();
                 request.signal.throwIfAborted();
+                if (isStepInterrupted(request, signal))
+                    return toolResult(call, interruptedToolResult(partialOutput), true);
                 return {
                     role: "toolResult",
                     toolCallId: call.id,
@@ -332,12 +407,14 @@ export function createPiRunner(registry, onUsageOrOptions, cwd = process.cwd())
             }
             catch (error) {
                 request.signal.throwIfAborted();
+                if (isStepInterrupted(request, signal))
+                    return toolResult(call, interruptedToolResult(partialOutput), true);
                 return toolResult(call, error instanceof Error ? error.message : String(error), true);
             }
         };
-        const executeCall = async (call) => {
+        const executeCall = async (call, signal) => {
             if (call.name !== "decide")
-                return executeFileTool(call);
+                return executeFileTool(call, signal);
             if (request.node.type !== "decision") {
                 return toolResult(call, "decide is available only on decision nodes", true);
             }
@@ -391,11 +468,19 @@ export function createPiRunner(registry, onUsageOrOptions, cwd = process.cwd())
             const results = [];
             for (const call of calls) {
                 request.signal.throwIfAborted();
+                const step = request.stepSignal?.() ?? request.signal;
+                // After an interruption the rest of the batch is not run, but every call still gets a result.
+                if (isStepInterrupted(request, step)) {
+                    const result = toolResult(call, notRunToolResult(), true);
+                    transcripts?.toolResult(executionId, result);
+                    results.push(result);
+                    continue;
+                }
                 activity?.toolStart(call.id, call.name, call.arguments);
                 transcripts?.toolStart(executionId, call.id);
                 let result;
                 try {
-                    result = await executeCall(call);
+                    result = await executeCall(call, step);
                 }
                 catch (error) {
                     activity?.toolEnd(call.id, error instanceof Error ? error.message : String(error), true);
@@ -421,6 +506,7 @@ export function createPiRunner(registry, onUsageOrOptions, cwd = process.cwd())
             usage,
         };
     };
+    return Object.assign(runner, { pausable: true, interruptible: true });
 }
 /** Includes cache/cost fields for Pi's tool-usage accounting, separate from core token totals. */
 export function sumPiUsage(reports) {

package/dist/status-text.js

@@ -58,11 +58,23 @@ function workspaceLine(workspace) {
         parts.push(`${item.executionId} ${item.disposition}: ${item.reason}`);
     return parts.join(" · ");
 }
+/** Who paused an execution; a node timeout pause also says how long it ran. */
+export function pausedBy(suspension) {
+    if (suspension.source !== "node_timeout")
+        return `paused by ${suspension.source}`;
+    return suspension.timeout ? `paused by node timeout after ${duration(suspension.timeout.activeMs)} (budget ${duration(suspension.timeout.nodeTimeoutMs)})` : "paused by node timeout";
+}
+/** A node timeout pause waits for the parent's decision. */
+const TIMEOUT_PAUSE_NOTE = "Its node timeout ran out. Resume it with braid_resume (optionally with nodeTimeoutMs; the default is the job's nodeTimeoutMs again), or fail only this execution with braid_cancel and executionIds. Its successors wait until then.";
 function executionLine(execution) {
     const parts = [execution.executionId ?? "(no execution)", `node ${execution.id}`];
     if (execution.iteration)
         parts.push(`iteration ${execution.iteration}`);
     parts.push(execution.skipReason ? `${execution.status} (${execution.skipReason})` : execution.status);
+    if (execution.suspension)
+        parts.push(`${execution.suspension.state}, ${pausedBy(execution.suspension)}`);
+    if (pendingHold(execution))
+        parts.push(`holds when finished, set by ${holdSource(execution.holdAfter)}`);
     if (execution.latencyMs !== undefined)
         parts.push(duration(execution.latencyMs));
     if (execution.decision)
@@ -77,8 +89,27 @@ export function formatJobList(jobs) {
     return [`${jobs.length} Braid job${jobs.length === 1 ? "" : "s"} in this session, newest first:`,
         ...jobs.map(job => `- ${job.handle} · ${job.status} · started ${iso(job.createdAt)} · id ${job.jobId}\n  Goal: ${job.goal}`)].join("\n");
 }
+/** Held (pauseAfter) and suspended executions are resumed the same way; a job pause has its own resume. */
 function control(execution) {
-    return execution ? `Graph ${execution.status} · revision ${execution.revision} · paused executions: ${ids(execution.pausedExecutionIds)}` : "Graph not started";
+    if (!execution)
+        return "Graph not started";
+    const parts = [`Graph ${execution.status}`, `revision ${execution.revision}`, `paused executions: ${ids(execution.pausedExecutionIds)}`];
+    if (execution.suspendedExecutionIds?.length)
+        parts.push(`suspended executions: ${ids(execution.suspendedExecutionIds)}`);
+    if (execution.pause)
+        parts.push(`job paused by ${execution.pause.source} at ${iso(execution.pause.pausedAt)}`);
+    return parts.join(" · ");
+}
+/** Paused-by-user work is the user's call: say so wherever the parent sees it. */
+const USER_PAUSE_NOTE = "The user paused this; do not resume it unless the user asks.";
+const USER_HOLD_NOTE = "The user set this hold; do not remove it or release the result unless the user asks.";
+/** A hold that applies once a still-unsettled execution finishes. Settled holds are paused executions. */
+function pendingHold(execution) {
+    return Boolean(execution.holdAfter) && (execution.status === "running" || execution.status === "runnable");
+}
+const holdSource = (hold) => hold.source === "definition" ? "pauseAfter" : hold.source;
+function userPaused(execution, executions) {
+    return execution?.pause?.source === "user" || executions.some(item => item.suspension?.source === "user" || item.holdAfter?.source === "user");
 }
 /** The submission receipt. The caller already has the graph it submitted. */
 export function formatReceipt(job, message) {
@@ -86,10 +117,40 @@ export function formatReceipt(job, message) {
 }
 export function formatUpdate(job) {
     return [`Braid job ${job.handle} updated to revision ${job.execution.revision}.`,
-        `Paused executions: ${ids(job.execution.pausedExecutionIds)}`, "Current graph:", formatGraph(job.execution.graph)].join("\n");
+        `Paused executions: ${ids(job.execution.pausedExecutionIds)}`,
+        ...job.execution.suspendedExecutionIds?.length ? [`Suspended executions: ${ids(job.execution.suspendedExecutionIds)}`] : [],
+        "Current graph:", formatGraph(job.execution.graph)].join("\n");
 }
 export function formatResume(job) {
-    return `Braid job ${job.handle} resumed at revision ${job.execution.revision}. Still paused: ${ids(job.execution.pausedExecutionIds)}.`;
+    const execution = job.execution;
+    return `Braid job ${job.handle} resumed at revision ${execution.revision}. Still paused: ${ids(execution.pausedExecutionIds)}. ` +
+        `Still suspended: ${ids(execution.suspendedExecutionIds)}.${execution.pause ? ` The job itself is still paused by ${execution.pause.source}.` : ""}`;
+}
+/** The result of braid_pause or braid_resume with after=true. */
+export function formatHold(job, executionIds, hold) {
+    const execution = job.execution;
+    if (!hold)
+        return [`Braid job ${job.handle}: hold removed for ${ids(executionIds)}; their results go downstream when they finish.`, control(execution)].join("\n");
+    return [
+        `Braid job ${job.handle}: ${ids(executionIds)} keep running and will be held when they finish; downstream work waits until you release them.`,
+        control(execution),
+        "Each one appears in paused executions once it finishes, and you get a pause reminder. Review it with braid_status, then release it with " +
+            `braid_resume({"jobId":"${job.handle}","expectedRevision":<current>,"executionIds":[...]}), or atomically with braid_update resume. ` +
+            `To remove the hold before it finishes, use braid_resume({"jobId":"${job.handle}","executionIds":[...],"after":true}).`,
+    ].join("\n");
+}
+/** The result of braid_pause: what is pausing now and how to continue. */
+export function formatPause(job, executionIds) {
+    const execution = job.execution;
+    const pausing = Object.values(execution.executions).filter(item => item.suspension?.state === "pausing").map(item => item.executionId);
+    return [
+        executionIds ? `Braid job ${job.handle}: pause requested for ${ids(executionIds)}.` : `Braid job ${job.handle} paused: no new executions start, and running ones pause.`,
+        control(execution),
+        `Pausing (each stops before its next model request; a running tool finishes first): ${ids(pausing)}.`,
+        "Suspended executions keep their conversation and workspace and release their concurrency slot. The node timeout stops while suspended; the graph timeout keeps running. " +
+            (executionIds ? `Continue with braid_resume({"jobId":"${job.handle}","expectedRevision":${execution.revision},"executionIds":[...]}).`
+                : `Continue with braid_resume({"jobId":"${job.handle}"}).`),
+    ].join("\n");
 }
 export function formatActivity(activity, now, jobId) {
     const lines = [`Activity: ${activity.phase} for ${duration(now - activity.phaseStartedAt)}; last observed signal ${duration(now - activity.lastActivityAt)} ago (${activity.lastActivity}). Model requests ${activity.modelRequests}, tool calls ${activity.toolCalls}.`];
@@ -124,6 +185,8 @@ export function formatActivity(activity, now, jobId) {
 /** A focused read: the current control state, then one execution in full. */
 export function formatNodeStatus({ job, node, activity, now }) {
     const lines = [`Braid job ${job.handle} · ${job.status} · ${control(job.execution)}`,
+        ...node.suspension ? [`Execution is ${node.suspension.state}, ${pausedBy(node.suspension)}${node.suspension.suspendedAt !== undefined ? ` since ${iso(node.suspension.suspendedAt)}` : ""}.${node.suspension.source === "user" ? ` ${USER_PAUSE_NOTE}` : node.suspension.source === "node_timeout" ? ` ${TIMEOUT_PAUSE_NOTE}` : ""}`] : [],
+        ...pendingHold(node) ? [`Its result will be held when it finishes (set by ${holdSource(node.holdAfter)}).${node.holdAfter.source === "user" ? ` ${USER_HOLD_NOTE}` : ""}`] : [],
         `Execution ${node.executionId ?? "(none yet)"} · node ${node.id}${node.node ? ` (${node.node.type})` : ""}${node.iteration ? ` · iteration ${node.iteration}` : ""} · ${node.skipReason ? `${node.status} (${node.skipReason})` : node.status}`];
     const timing = [node.startedAt !== undefined && `started ${iso(node.startedAt)}`, node.latencyMs !== undefined && `took ${duration(node.latencyMs)}`,
         node.model && `model ${node.model}`, node.usage && `usage input ${node.usage.inputTokens}, output ${node.usage.outputTokens}`].filter(Boolean);
@@ -148,6 +211,8 @@ export function formatJobStatus(job, now) {
     const state = job.result ?? job.execution;
     // Control fields lead: a long goal or output must not push them out of a truncated preview.
     const header = [`Braid job ${job.handle} · ${job.status}`, `ID: ${job.jobId}`, control(job.execution)];
+    if (job.status === "running" && userPaused(job.execution, Object.values(job.execution?.executions ?? {})))
+        header.push(USER_PAUSE_NOTE);
     const error = job.error ?? (job.result?.error ? failure(job.result.error) : job.execution?.error ? failure(job.execution.error) : undefined);
     if (error)
         header.push(`Error: ${error}`);
@@ -177,3 +242,24 @@ export function formatJobStatus(job, now) {
         sections.push(`Current graph (revision ${job.execution.revision}):\n${formatGraph(job.execution.graph)}`);
     return sections.join("\n\n");
 }
+/** The waking reminder for an execution its node timeout paused. */
+export function formatTimeoutPause(pause) {
+    const execution = `${pause.executionId}${pause.iteration ? ` (iteration ${pause.iteration})` : ""}`;
+    return [
+        `Braid job ${pause.handle} node ${JSON.stringify(pause.nodeId)} execution ${execution} was paused because it exceeded its node timeout:`,
+        `it ran ${duration(pause.activeMs)} against a budget of ${duration(pause.nodeTimeoutMs)}.`,
+        pause.activity ? `When the timeout fired it was: ${pause.activity}.` : "",
+        "Its conversation, workspace, and tool budgets are kept; the step that was interrupted is reported to it after resume.",
+        "Its successors wait until you decide. Check it with",
+        `braid_status(${JSON.stringify({ jobId: pause.handle, executionId: pause.executionId })}), then either`,
+        `resume it with braid_resume({"jobId":"${pause.handle}","executionIds":["${pause.executionId}"],"expectedRevision":<current>}) — it gets the job's nodeTimeoutMs again, or pass nodeTimeoutMs —`,
+        "or resume it in braid_update with resume and resumeOptions together with graph changes,",
+        `or fail only this execution with braid_cancel({"jobId":"${pause.handle}","executionIds":["${pause.executionId}"]}).`,
+        "The graph timeout keeps running while it is paused.",
+    ].filter(Boolean).join(" ");
+}
+/** The result of braid_cancel with executionIds. */
+export function formatCancelExecutions(job, executionIds) {
+    return [`Braid job ${job.handle}: cancellation requested for ${ids(executionIds)}. They fail; their successors follow the usual failure rules.`,
+        control(job.execution)].join("\n");
+}

package/package.json

@@ -1,12 +1,12 @@
 {
   "name": "@chrok/pi-braid",
-  "version": "0.3.6",
+  "version": "0.5.0",
   "license": "MIT",
   "description": "Pi extension for Braid: background agent graphs, bounded loops, live updates, and a flow panel",
   "type": "module",
   "dependencies": {
     "grok-mermaid": "^0.2.2",
-    "@chrok/braid": "0.3.6"
+    "@chrok/braid": "0.5.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-ai": "*",