pi-subagents 0.74.0 → 0.76.0

package/src/runs/background/subagent-runner.js

@@ -50,6 +50,7 @@ import { persistRunnerStartupFailure } from "./runner-startup-failure.js";
 import { currentPidNamespaceScope } from "./pid-namespace.js";
 import { createSteeringStatus, recordSteeringRequest, steeringStatus, terminalSteeringNoticeState, unconsumedSteerReason, updateSteeringTarget } from "./steering.js";
 import { PROMPT_REDACTED, detectSubagentError, extractTextFromContent, extractToolArgsPreview, formatEmptyTerminalAssistantResponseError, getAgentDir, hasEmptyTerminalAssistantResponse, readStatus } from "../../shared/utils.js";
+import { planBackgroundRunHistory, recordRun } from "../shared/run-history.js";
 import { planAbortRecovery } from "../shared/abort-recovery.js";
 import { createMutatingFailureState, didMutatingToolFail, isMutatingTool, nextLongRunningTrigger, recordMutatingFailure, resetMutatingFailureState, resolveCurrentPath, shouldEscalateMutatingFailures, summarizeRecentMutatingFailures, } from "../shared/long-running-guard.js";
 import { parseSessionTokens } from "../../shared/session-tokens.js";
@@ -618,7 +619,7 @@ export async function runSingleStepInner(step, ctx) {
         const adapterLaunch = step.runner.adapter === "codex-exec" || step.runner.adapter === "codex-exec-writer"
             ? resolveCodexExecLaunch({ adapter: step.runner.adapter, command: step.runner.command, asyncDir: path.dirname(ctx.outputFile), stepIndex: ctx.flatIndex })
             : step.runner.adapter === "claude-code" || step.runner.adapter === "claude-code-writer"
-                ? resolveClaudeCodeLaunch({ adapter: step.runner.adapter, command: step.runner.command })
+                ? resolveClaudeCodeLaunch({ adapter: step.runner.adapter, command: step.runner.command, overrideArgs: step.claudeCodeOverrideArgs })
                 : step.runner.adapter === "cursor-agent" || step.runner.adapter === "cursor-agent-writer"
                     ? resolveCursorAgentLaunch({ adapter: step.runner.adapter, command: step.runner.command, cwd: externalCwd, asyncDir: path.dirname(ctx.outputFile), stepIndex: ctx.flatIndex })
                     : undefined;
@@ -1095,12 +1096,14 @@ export async function runSingleStepInner(step, ctx) {
         break singleLaunch;
     }
     const rawOutput = finalResult?.finalOutput ?? "";
-    let outputForPersistence = stripAcceptanceReport(rawOutput);
-    if (!outputForPersistence.trim() && finalResult?.structuredOutput !== undefined)
-        outputForPersistence = JSON.stringify(finalResult.structuredOutput, null, 2);
+    const structuredText = finalResult?.structuredOutput === undefined ? undefined : JSON.stringify(finalResult.structuredOutput, null, 2);
+    let replyOutput = stripAcceptanceReport(rawOutput);
+    if (!replyOutput.trim() && structuredText !== undefined)
+        replyOutput = structuredText;
+    // The schema is the caller's contract: a bound output file holds the structured result, not closing prose.
     const resolvedOutput = step.outputPath && finalResult?.exitCode === 0
-        ? resolveSingleOutput(step.outputPath, outputForPersistence, finalOutputSnapshot, step.outputClaimPath)
-        : { fullOutput: outputForPersistence };
+        ? resolveSingleOutput(step.outputPath, structuredText ?? replyOutput, finalOutputSnapshot, step.outputClaimPath)
+        : { fullOutput: replyOutput };
     if (resolvedOutput.fatalError) {
         if (finalResult) {
             finalResult.exitCode = 1;
@@ -1113,7 +1116,8 @@ export async function runSingleStepInner(step, ctx) {
     if (finalResult?.stopped && !outputForSummary.trim()) {
         outputForSummary = ctx.stopMessage ?? "Subagent stopped by user.";
     }
-    const outputForAcceptance = rawOutput;
+    // Unfinished streamed text never stands in for the child's completed reply.
+    const outputForAcceptance = finalResult?.outputPartial ? "" : rawOutput;
     const childWrittenOutput = step.outputPath
         ? extractChildWrittenOutput(finalResult?.messages, step.outputPath, step.cwd ?? ctx.cwd)
         : undefined;
@@ -1243,6 +1247,7 @@ export async function runSingleStepInner(step, ctx) {
         launchContractDigest: actualLaunchContractDigest,
         output: outputForSummary,
         outputState,
+        outputPartial: finalResult?.outputPartial,
         exitCode: effectiveFinalExitCode,
         error: effectiveFinalError,
         sessionFile: step.sessionFile,
@@ -1335,7 +1340,13 @@ function markParallelGroupSetupFailure(input) {
             throw new Error(`Missing parallel task at index ${taskIndex}`);
         const stopped = statusStep.stopped || statusStep.stopRequested || input.statusPayload.stopped;
         const paused = !stopped && input.statusPayload.state === "paused";
+        const timedOut = !stopped && !paused && input.statusPayload.timedOut === true;
         statusStep.status = stopped ? "stopped" : paused ? "paused" : "failed";
+        // Mirror the run-level timeout onto the step so consumers (status readers,
+        // run-history recording) see timed_out rather than a bare failure — the
+        // StepResult below already carries it, the status step must too.
+        if (timedOut)
+            statusStep.timedOut = true;
         statusStep.startedAt = input.failedAt;
         statusStep.endedAt = input.failedAt;
         statusStep.durationMs = 0;
@@ -1506,6 +1517,11 @@ export async function runSubagent(config, childSessions) {
     let previousOutput = "";
     const outputs = {};
     const results = [];
+    // Flat indices of steps a child session was actually dispatched for — the
+    // ground truth for run-history: stopRunner/timeoutRunner/fail-fast/budget
+    // skips relabel never-launched steps to terminal statuses, so status alone
+    // cannot distinguish them from steps that really ran.
+    const launchedFlatIndices = new Set();
     const overallStartTime = Date.now();
     const shareEnabled = config.share === true;
     const asyncDir = config.asyncDir;
@@ -2392,7 +2408,7 @@ export async function runSubagent(config, childSessions) {
         return key ? removeActiveToolCallKey(flatIndex, key) : undefined;
     };
     const openToolAttentionTarget = (flatIndex, now) => [...(activeToolCalls[flatIndex]?.values() ?? [])]
-        .filter((active) => shouldEmitOpenToolAttention({ config: controlConfig, currentTool: active.tool, currentToolStartedAt: active.startedAt, now }))
+        .filter((active) => !active.attentionEmitted && shouldEmitOpenToolAttention({ config: controlConfig, currentTool: active.tool, currentToolStartedAt: active.startedAt, now }))
         .sort((left, right) => left.startedAt - right.startedAt)[0];
     const supervisorAttentionSteps = new Map();
     const mutatingFailureWindowMs = 5 * 60_000;
@@ -2446,11 +2462,15 @@ export async function runSubagent(config, childSessions) {
     };
     const maybeEmitOpenToolAttention = (flatIndex, now) => {
         const step = statusPayload.steps[flatIndex];
-        if (!step || step.status !== "running" || step.activityState === "needs_attention")
+        if (!step || step.status !== "running")
             return false;
         const target = openToolAttentionTarget(flatIndex, now);
         if (!target)
             return false;
+        target.attentionEmitted = true;
+        // Keep this attention when a concurrent supervisor request ends and restores its saved state.
+        if (supervisorAttentionSteps.has(flatIndex))
+            supervisorAttentionSteps.set(flatIndex, "needs_attention");
         const previous = step.activityState;
         step.activityState = "needs_attention";
         statusPayload.activityState = "needs_attention";
@@ -2469,6 +2489,7 @@ export async function runSubagent(config, childSessions) {
             tokens: step.tokens?.total,
             toolCount: step.toolCount,
             currentTool: target.tool,
+            toolCallId: target.key.startsWith("id:") ? target.key.slice(3) : undefined,
             currentToolDurationMs: toolDurationMs,
             currentPath: target.path,
         })));
@@ -3342,6 +3363,7 @@ export async function runSubagent(config, childSessions) {
                 return omitUndefinedProperties({
                     agent: task.agent,
                     ...(task.sessionName ? { sessionName: task.sessionName } : {}),
+                    ...(externalRunnerStatus(task.runner) ? { runner: externalRunnerStatus(task.runner) } : {}),
                     ...(statusStepDescription(task.task) ? { description: statusStepDescription(task.task) } : {}),
                     ...(task.context ? { context: task.context } : {}),
                     ...(task.phase ?? step.phase ? { phase: task.phase ?? step.phase } : {}),
@@ -3449,6 +3471,7 @@ export async function runSubagent(config, childSessions) {
                     return omitUndefinedProperties({ agent: task.agent, ...(task.sessionName ? { sessionName: task.sessionName } : {}), context: task.context, output: "(skipped — fail-fast)", exitCode: -1, skipped: true });
                 }
                 const taskStartTime = Date.now();
+                launchedFlatIndices.add(fi);
                 statusPayload.currentStep = fi;
                 requiredStatusStep(statusPayload, fi).status = "running";
                 delete requiredStatusStep(statusPayload, fi).error;
@@ -3576,6 +3599,7 @@ export async function runSubagent(config, childSessions) {
                     runtimeAcknowledgedExtensions: pr.runtimeAcknowledgedExtensions,
                     output: pr.output,
                     outputState: pr.outputState,
+                    outputPartial: pr.outputPartial,
                     error: pr.error,
                     success: pr.stopped !== true && pr.interrupted !== true && pr.exitCode === 0 && pr.execution?.status !== "partial",
                     exitCode: pr.interrupted === true ? 0 : pr.exitCode,
@@ -3865,6 +3889,7 @@ export async function runSubagent(config, childSessions) {
                         return omitUndefinedProperties({ agent: task.agent, ...(task.sessionName ? { sessionName: task.sessionName } : {}), context: task.context, output: "(skipped — fail-fast)", exitCode: -1, skipped: true });
                     }
                     const taskStartTime = Date.now();
+                    launchedFlatIndices.add(fi);
                     statusPayload.currentStep = fi;
                     requiredStatusStep(statusPayload, fi).status = "running";
                     delete requiredStatusStep(statusPayload, fi).error;
@@ -4027,6 +4052,7 @@ export async function runSubagent(config, childSessions) {
                         launchResolvedExtensions: pr.launchResolvedExtensions,
                         output: pr.output,
                         outputState: pr.outputState,
+                        outputPartial: pr.outputPartial,
                         error: pr.error,
                         success: pr.stopped !== true && pr.interrupted !== true && pr.exitCode === 0 && pr.execution?.status !== "partial",
                         exitCode: pr.interrupted === true ? 0 : pr.exitCode,
@@ -4256,6 +4282,7 @@ export async function runSubagent(config, childSessions) {
             }
             const singleCwd = singleWorktreeSetup?.worktrees[0]?.agentCwd ?? cwd;
             const stepStartTime = Date.now();
+            launchedFlatIndices.add(flatIndex);
             statusPayload.currentStep = flatIndex;
             requiredStatusStep(statusPayload, flatIndex).status = "running";
             delete requiredStatusStep(statusPayload, flatIndex).activityState;
@@ -4339,6 +4366,7 @@ export async function runSubagent(config, childSessions) {
                 runtimeAcknowledgedExtensions: singleResult.runtimeAcknowledgedExtensions,
                 output: stopped || childStopped ? stopMessage : timedOut ? singleResult.output || (timeoutMessage ?? "Subagent timed out.") : singleResult.output,
                 outputState: singleResult.outputState,
+                outputPartial: singleResult.outputPartial,
                 error: stopped || childStopped ? stopMessage : timedOut ? (timeoutMessage ?? "Subagent timed out.") : singleResult.error,
                 success: !stopped && !childStopped && !timedOut && singleResult.interrupted !== true && singleResult.exitCode === 0 && singleResult.execution?.status !== "partial",
                 exitCode: stopped || childStopped ? 1 : timedOut ? 1 : singleResult.interrupted === true ? 0 : singleResult.exitCode,
@@ -4743,6 +4771,7 @@ export async function runSubagent(config, childSessions) {
                 context: r.context,
                 output: r.output,
                 outputState: r.outputState,
+                outputPartial: r.outputPartial,
                 error: r.error,
                 success: r.success,
                 skipped: r.skipped || undefined,
@@ -4845,6 +4874,26 @@ export async function runSubagent(config, childSessions) {
         totalCost: finalTotalCost,
         usageBudget: statusPayload.usageBudget,
     }));
+    // Paused runs record an interrupted attempt; a later resume records another.
+    for (const historyEntry of planBackgroundRunHistory({
+        steps,
+        resultMode,
+        statusSteps: statusPayload.steps.map((step, index) => ({
+            agent: step.agent,
+            status: step.status,
+            durationMs: step.durationMs,
+            timedOut: step.timedOut,
+            stopped: step.stopped,
+            launched: launchedFlatIndices.has(index),
+        })),
+        stepResults: results,
+        runDurationMs: runEndedAt - overallStartTime,
+        stopped,
+        interrupted,
+        timedOut,
+    })) {
+        recordRun(historyEntry.agent, historyEntry.task, historyEntry.exitCode, historyEntry.durationMs, historyEntry.terminal);
+    }
     writeRunLog(logPath, omitUndefinedProperties({
         id,
         mode: statusPayload.mode,