@xzzpig/pi-subagents 0.16.0 → 0.18.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +164 -1
- package/README.md +5 -3
- package/agents/claude-code-writer.md +1 -0
- package/agents/codex-exec-writer.md +1 -0
- package/agents/cursor-agent-writer.md +1 -0
- package/agents/reviewer.md +3 -2
- package/agents/worker.md +3 -2
- package/docs/agents.md +17 -10
- package/docs/configuration.md +21 -3
- package/docs/extension-api.md +70 -4
- package/docs/models.md +5 -3
- package/docs/observability.md +23 -13
- package/docs/standalone-background.md +4 -4
- package/docs/tool-reference.md +40 -19
- package/docs/watchdog.md +2 -0
- package/docs/workflows.md +57 -3
- package/inspector-runner.mjs +1 -1
- package/package.json +14 -8
- package/prompts/council.md +6 -3
- package/runner-peer-loader.mjs +4 -7
- package/runner-peer-preload.mjs +5 -7
- package/skills/pi-subagents/SKILL.md +1 -0
- package/skills/pi-subagents/references/constraints-and-recipes.md +5 -2
- package/skills/pi-subagents/references/execution-controls.md +6 -4
- package/skills/pi-subagents/references/management-authoring-rpc.md +3 -3
- package/skills/pi-subagents/references/multi-lane-orchestration.md +8 -0
- package/skills/pi-subagents/references/prompting-and-roles.md +16 -15
- package/src/agents/agent-eject.ts +3 -3
- package/src/agents/agent-management.ts +61 -52
- package/src/agents/agent-memory.ts +90 -3
- package/src/agents/agent-serializer.ts +4 -4
- package/src/agents/agents.ts +125 -51
- package/src/agents/global-npm-root.ts +58 -0
- package/src/agents/proactive-skills.ts +4 -1
- package/src/agents/runtime-agent-registry.ts +14 -11
- package/src/agents/skills.ts +42 -9
- package/src/api/delegation.ts +15 -0
- package/src/api/inspectors.ts +27 -0
- package/src/api/preflight.ts +60 -60
- package/src/extension/doctor.ts +9 -4
- package/src/extension/fanout-child.ts +8 -0
- package/src/extension/index.ts +101 -48
- package/src/extension/rpc.ts +10 -1
- package/src/extension/schemas.ts +12 -6
- package/src/extension/subagent-guide.ts +12 -4
- package/src/extension/tool-activation.ts +197 -0
- package/src/extension/tool-description.ts +1 -1
- package/src/inspectors/ghostty/plugin.ts +13 -1
- package/src/inspectors/plugins.ts +70 -3
- package/src/intercom/intercom-bridge.ts +2 -0
- package/src/intercom/native-supervisor-channel.ts +22 -15
- package/src/intercom/result-intercom.ts +4 -0
- package/src/runs/background/active-async-capacity.ts +3 -25
- package/src/runs/background/async-execution.ts +100 -140
- package/src/runs/background/async-job-tracker.ts +28 -34
- package/src/runs/background/async-resume.ts +16 -8
- package/src/runs/background/async-status.ts +2 -0
- package/src/runs/background/binary-bootstrap.ts +34 -6
- package/src/runs/background/fleet-view.ts +51 -7
- package/src/runs/background/notify.ts +93 -7
- package/src/runs/background/pid-namespace.ts +18 -0
- package/src/runs/background/process-terminal-candidate.ts +98 -0
- package/src/runs/background/process-terminal.ts +17 -96
- package/src/runs/background/result-watcher.ts +1 -1
- package/src/runs/background/retained-children.ts +3 -2
- package/src/runs/background/run-child-session.ts +16 -5
- package/src/runs/background/run-id-resolver.ts +10 -5
- package/src/runs/background/run-status.ts +10 -4
- package/src/runs/background/runner-aliases.ts +25 -15
- package/src/runs/background/runner-child-launch.ts +1 -2
- package/src/runs/background/runner-startup-failure.ts +56 -0
- package/src/runs/background/stale-run-reconciler.ts +19 -3
- package/src/runs/background/subagent-runner-bootstrap.ts +276 -0
- package/src/runs/background/subagent-runner.ts +96 -287
- package/src/runs/background/wait-tool.ts +2 -0
- package/src/runs/background/workflow-terminal-proof.ts +68 -0
- package/src/runs/foreground/execution.ts +217 -250
- package/src/runs/foreground/prompt-audit.ts +3 -9
- package/src/runs/foreground/subagent-executor.ts +377 -119
- package/src/runs/shared/acceptance.ts +167 -68
- package/src/runs/shared/capability-ceiling.ts +4 -1
- package/src/runs/shared/child-hooks.ts +1 -16
- package/src/runs/shared/child-launch.ts +22 -14
- package/src/runs/shared/child-runtime-config.ts +3 -0
- package/src/runs/shared/child-session.ts +52 -2
- package/src/runs/shared/child-tool-plan.ts +23 -113
- package/src/runs/shared/dynamic-fanout.ts +1 -1
- package/src/runs/shared/external-cli-runner.ts +3 -1
- package/src/runs/shared/external-job-bridge.ts +54 -0
- package/src/runs/shared/git-environment.ts +29 -0
- package/src/runs/shared/mcp-config-sources.ts +3 -0
- package/src/runs/shared/mcp-direct-tool-allowlist.ts +1 -1
- package/src/runs/shared/mcp-direct-tool-grant.ts +22 -6
- package/src/runs/shared/model-resolution.ts +6 -1
- package/src/runs/shared/nested-events.ts +67 -55
- package/src/runs/shared/orca-progress-tabs.ts +3 -2
- package/src/runs/shared/parallel-utils.ts +2 -1
- package/src/runs/shared/pi-spawn.ts +92 -6
- package/src/runs/shared/single-output.ts +11 -2
- package/src/runs/shared/structured-output.ts +227 -7
- package/src/runs/shared/subagent-control.ts +2 -16
- package/src/runs/shared/subagent-prompt-runtime.ts +27 -5
- package/src/runs/shared/tool-availability.ts +2 -2
- package/src/runs/shared/tool-budget.ts +21 -0
- package/src/runs/shared/usage-reconciliation.ts +73 -0
- package/src/runs/shared/worktree.ts +35 -8
- package/src/shared/child-session-name.ts +4 -2
- package/src/shared/fork-context.ts +11 -4
- package/src/shared/launch-contract.ts +1 -1
- package/src/shared/pruned-fork.ts +69 -56
- package/src/shared/types.ts +52 -16
- package/src/slash/delegation-adapters.ts +25 -1
- package/src/slash/prompt-template-bridge.ts +16 -1
- package/src/slash/prompt-workflows.ts +12 -1
- package/src/slash/slash-commands.ts +7 -241
- package/src/slash/subagent-cost.ts +313 -0
- package/src/slash/subagents-admin.ts +28 -29
- package/src/tui/fleet-status.ts +24 -2
- package/src/tui/fleet.ts +7 -4
- package/src/tui/render.ts +53 -19
- package/src/watchdog/diff-tool.ts +24 -3
- package/src/watchdog/permission-arbiter.ts +13 -15
- package/src/watchdog/register-child.ts +10 -2
- package/src/watchdog/review.ts +14 -14
- package/src/watchdog/runtime.ts +2 -1
- package/src/watchdog/turn-delta.ts +11 -5
- package/src/workflows/scripted-workflow.ts +109 -23
- package/src/workflows/workflow-child-summary.ts +4 -2
- package/src/workflows/workflow-settlement.ts +4 -1
- package/src/runs/shared/completion-evidence.ts +0 -89
- package/src/runs/shared/completion-guard.ts +0 -255
- package/src/runs/shared/llm-intent-arbiter.ts +0 -321
- package/src/runs/shared/task-intent.ts +0 -244
|
@@ -53,9 +53,6 @@ import {
|
|
|
53
53
|
} from "../../shared/utils.ts";
|
|
54
54
|
import { resolveSkillsWithFallback } from "../../agents/skills.ts";
|
|
55
55
|
import { effectiveToolTimeoutMs, formatToolTimeoutMessage, resolveToolTimeoutMs, toolTimeoutCallKey, toolTimeoutFromEnv } from "../shared/tool-timeout.ts";
|
|
56
|
-
import { evaluateCompletionMutationGuard, expectsImplementationMutation, hasMutationToolCapability, validateImplementationToolContract } from "../shared/completion-guard.ts";
|
|
57
|
-
import { planCompletionEvidence } from "../shared/completion-evidence.ts";
|
|
58
|
-
import { arbitrateCompletionGuardRescue } from "../shared/llm-intent-arbiter.ts";
|
|
59
56
|
import { preflightLaunchCwd } from "../shared/launch-cwd.ts";
|
|
60
57
|
import { createJsonlWriter } from "../../shared/jsonl-writer.ts";
|
|
61
58
|
import { createOrcaProgressTab, type OrcaProgressTab } from "../shared/orca-progress-tabs.ts";
|
|
@@ -66,7 +63,7 @@ import { deriveChildSessionName } from "../../shared/child-session-name.ts";
|
|
|
66
63
|
import { assertAgentAllowedByCapabilityCeiling, intersectSubagentCapabilityCeilings, resolveCurrentSubagentCapabilityCeiling } from "../shared/capability-ceiling.ts";
|
|
67
64
|
import { resolveEffectiveThinking } from "../../shared/model-info.ts";
|
|
68
65
|
import { assertThinkingWithinCeiling, intersectThinkingCeilings } from "../../shared/thinking-ceiling.ts";
|
|
69
|
-
import { MISSING_STRUCTURED_ACCEPTANCE_REPORT_ERROR, MISSING_STRUCTURED_OUTPUT_CALL_ERROR } from "../shared/structured-output.ts";
|
|
66
|
+
import { formatStructuredOutputRejectionError, MISSING_STRUCTURED_ACCEPTANCE_REPORT_ERROR, MISSING_STRUCTURED_OUTPUT_CALL_ERROR } from "../shared/structured-output.ts";
|
|
70
67
|
import { formatMidToolExitError, isOrdinaryToolForMidToolExit } from "../shared/process-signal.ts";
|
|
71
68
|
import { formatChildToolDiagnostic } from "../shared/tool-availability.ts";
|
|
72
69
|
import { formatChildModelResolutionDiagnostic, isChildModelResolutionFailure } from "../shared/model-resolution-diagnostic.ts";
|
|
@@ -89,10 +86,10 @@ import {
|
|
|
89
86
|
shouldEscalateMutatingFailures,
|
|
90
87
|
summarizeRecentMutatingFailures,
|
|
91
88
|
} from "../shared/long-running-guard.ts";
|
|
92
|
-
import { acceptanceFailureMessage, buildSkippedAcceptanceLedger, evaluateAcceptance, formatAcceptancePrompt, resolveEffectiveAcceptance, stripAcceptanceReport, validateAcceptanceInput } from "../shared/acceptance.ts";
|
|
89
|
+
import { acceptanceFailureMessage, buildSkippedAcceptanceLedger, captureStagedIndexBaseline, evaluateAcceptance, formatAcceptancePrompt, resolveEffectiveAcceptance, stripAcceptanceReport, validateAcceptanceInput, typedVerifyOutput } from "../shared/acceptance.ts";
|
|
93
90
|
import { PROMPT_REDACTED } from "../../shared/utils.ts";
|
|
94
91
|
import { attachContractProjections, isAgentContract } from "../shared/agent-contract.ts";
|
|
95
|
-
import { initialToolBudgetState, toolBudgetState } from "../shared/tool-budget.ts";
|
|
92
|
+
import { initialToolBudgetState, isToolBudgetBlockedMessage, toolBudgetState } from "../shared/tool-budget.ts";
|
|
96
93
|
import { resolveWatchdogConfig } from "../../watchdog/settings.ts";
|
|
97
94
|
import { resolveLaunchBinding } from "../../shared/launch-contract.ts";
|
|
98
95
|
import { consumeWorkflowChildPermit } from "../../shared/workflow-child-permit.ts";
|
|
@@ -108,6 +105,7 @@ import {
|
|
|
108
105
|
} from "../../watchdog/child-status.ts";
|
|
109
106
|
import { buildInProcessChildLaunch, createReportedChildSessionInput } from "../shared/child-launch.ts";
|
|
110
107
|
import { childSessionFactory, childSessionHasQueuedMessages, projectChildSessionEventForJson, type ChildSession, type ChildSessionEvent } from "../shared/child-session.ts";
|
|
108
|
+
import { reconcileAttemptUsage } from "../shared/usage-reconciliation.ts";
|
|
111
109
|
|
|
112
110
|
const artifactOutputByResult = new WeakMap<SingleResult, string>();
|
|
113
111
|
const acceptanceOutputByResult = new WeakMap<SingleResult, string>();
|
|
@@ -160,7 +158,7 @@ function persistSingleResultMetadata(input: {
|
|
|
160
158
|
toolCount: target.progressSummary?.toolCount,
|
|
161
159
|
error: target.error,
|
|
162
160
|
agentContract: target.agentContract,
|
|
163
|
-
|
|
161
|
+
sandbox: target.sandbox,
|
|
164
162
|
launchContractDigest: target.launchContractDigest,
|
|
165
163
|
launchResolvedExtensions: target.launchResolvedExtensions,
|
|
166
164
|
runtimeAcknowledgedExtensions: target.runtimeAcknowledgedExtensions,
|
|
@@ -288,6 +286,7 @@ interface StructuredDelegationProgressState {
|
|
|
288
286
|
model?: string;
|
|
289
287
|
toolCount: number;
|
|
290
288
|
tokens: number;
|
|
289
|
+
turnCount?: number;
|
|
291
290
|
}
|
|
292
291
|
|
|
293
292
|
function captureStructuredDelegationProgressState(progress: AgentProgress, result: SingleResult): StructuredDelegationProgressState {
|
|
@@ -300,6 +299,7 @@ function captureStructuredDelegationProgressState(progress: AgentProgress, resul
|
|
|
300
299
|
model: progress.model ?? result.model,
|
|
301
300
|
toolCount: progress.toolCount,
|
|
302
301
|
tokens: progress.tokens,
|
|
302
|
+
turnCount: progress.turnCount,
|
|
303
303
|
};
|
|
304
304
|
}
|
|
305
305
|
|
|
@@ -314,6 +314,7 @@ function structuredDelegationProgressChanged(
|
|
|
314
314
|
|| previous.model !== (progress.model ?? result.model)
|
|
315
315
|
|| previous.toolCount !== progress.toolCount
|
|
316
316
|
|| previous.tokens !== progress.tokens
|
|
317
|
+
|| previous.turnCount !== progress.turnCount
|
|
317
318
|
|| previous.recentOutput.length !== progress.recentOutput.length) return true;
|
|
318
319
|
for (let index = 0; index < progress.recentOutput.length; index++) {
|
|
319
320
|
if (previous.recentOutput[index] !== progress.recentOutput[index]) return true;
|
|
@@ -328,6 +329,9 @@ function structuredDelegationProgressChanged(
|
|
|
328
329
|
return false;
|
|
329
330
|
}
|
|
330
331
|
|
|
332
|
+
function isCompleteUsageCounter(value: unknown): value is number {
|
|
333
|
+
return typeof value === "number" && Number.isFinite(value) && value >= 0;
|
|
334
|
+
}
|
|
331
335
|
|
|
332
336
|
const STOPPED_BEFORE_COMPLETION_ERROR = "Subagent stopped before completion.";
|
|
333
337
|
const AFTER_COMPACTION_SETTLEMENT = Symbol("afterCompactionSettlement");
|
|
@@ -376,8 +380,8 @@ async function runSingleAttempt(
|
|
|
376
380
|
})
|
|
377
381
|
: undefined;
|
|
378
382
|
const permissionRules = resolvePermissionRules(options.permissions, agent.permissions);
|
|
379
|
-
|
|
380
|
-
|
|
383
|
+
const sandbox = options.sandbox ?? agent.sandbox;
|
|
384
|
+
const permissionProfile = options.permissionProfile ?? agent.permissionProfile;
|
|
381
385
|
const permissionAuditPath = permissionRules && options.artifactsDir
|
|
382
386
|
? path.join(options.artifactsDir, "permission-audit", `${options.runId}-${options.index ?? 0}.jsonl`)
|
|
383
387
|
: undefined;
|
|
@@ -400,6 +404,7 @@ async function runSingleAttempt(
|
|
|
400
404
|
tools: agent.tools,
|
|
401
405
|
excludeTools: agent.excludeTools,
|
|
402
406
|
allowNestedSubagents: agent.allowNestedSubagents,
|
|
407
|
+
descendantAllowedAgents: agent.allowedAgents,
|
|
403
408
|
extensions: agent.extensions,
|
|
404
409
|
subagentOnlyExtensions: agent.subagentOnlyExtensions,
|
|
405
410
|
// Exact terminal protocol belongs to session resources, not compactable history.
|
|
@@ -421,10 +426,10 @@ async function runSingleAttempt(
|
|
|
421
426
|
toolBudget: options.toolBudget,
|
|
422
427
|
permissionRules,
|
|
423
428
|
permissionAuditPath,
|
|
424
|
-
|
|
425
|
-
|
|
426
|
-
|
|
427
|
-
|
|
429
|
+
sandbox,
|
|
430
|
+
permissionProfile,
|
|
431
|
+
projectTrusted: options.projectTrusted,
|
|
432
|
+
trustedProjectCwd: options.trustedProjectCwd,
|
|
428
433
|
childWatchdog,
|
|
429
434
|
// registerChildWatchdog returns before reading the sink when no watchdog exists.
|
|
430
435
|
watchdogStatus: childWatchdog ? (event) => onWatchdogStatus?.(event) : undefined,
|
|
@@ -434,7 +439,6 @@ async function runSingleAttempt(
|
|
|
434
439
|
thinkingCeiling: options.thinkingCeiling,
|
|
435
440
|
maxSubagentDepth: options.maxSubagentDepth,
|
|
436
441
|
runtimeSnapshotHost: options.runtimeSnapshotHost,
|
|
437
|
-
hostAvailableBuiltins: options.hostAvailableBuiltins,
|
|
438
442
|
inherited: options.childRuntime,
|
|
439
443
|
host: "parent",
|
|
440
444
|
});
|
|
@@ -446,34 +450,6 @@ async function runSingleAttempt(
|
|
|
446
450
|
}
|
|
447
451
|
|
|
448
452
|
const effectiveSystemPrompt = shared.systemPrompt;
|
|
449
|
-
const contractTools = toolPlan.explicitToolAllowlist ? toolPlan.effectiveToolAllowlist : undefined;
|
|
450
|
-
const contractError = validateImplementationToolContract({
|
|
451
|
-
agent: agent.name,
|
|
452
|
-
task: shared.originalTask ?? task,
|
|
453
|
-
tools: contractTools,
|
|
454
|
-
mcpDirectTools: toolPlan.effectiveMcpTools,
|
|
455
|
-
configuredExtensions: toolPlan.configuredExtensions,
|
|
456
|
-
requestedTools: toolPlan.requestedBuiltinTools,
|
|
457
|
-
acceptanceRole: agent.acceptanceRole,
|
|
458
|
-
completionGuard: agent.completionGuard,
|
|
459
|
-
});
|
|
460
|
-
if (contractError) {
|
|
461
|
-
return {
|
|
462
|
-
index: options.index ?? 0,
|
|
463
|
-
agent: agent.name,
|
|
464
|
-
task,
|
|
465
|
-
...(childSessionName ? { sessionName: childSessionName } : {}),
|
|
466
|
-
messages: [],
|
|
467
|
-
finalOutput: "",
|
|
468
|
-
exitCode: 1,
|
|
469
|
-
error: contractError,
|
|
470
|
-
usage: emptyUsage(),
|
|
471
|
-
model: modelArg,
|
|
472
|
-
progressSummary: { status: "failed", toolCount: 0, tokens: 0, durationMs: 0 },
|
|
473
|
-
...(toolPlan.capabilityCeiling ? { capabilityCeiling: toolPlan.capabilityCeiling } : {}),
|
|
474
|
-
...(toolPlan.capabilityAudit ? { capabilityAudit: toolPlan.capabilityAudit } : {}),
|
|
475
|
-
};
|
|
476
|
-
}
|
|
477
453
|
const fast = options.fast ?? agent.fast;
|
|
478
454
|
const { launchContractDigest } = resolveLaunchBinding({
|
|
479
455
|
agent,
|
|
@@ -482,12 +458,12 @@ async function runSingleAttempt(
|
|
|
482
458
|
...(fast !== undefined ? { fast } : {}),
|
|
483
459
|
...(resolvedThinking ? { thinking: resolvedThinking } : {}),
|
|
484
460
|
systemPrompt: effectiveSystemPrompt,
|
|
485
|
-
|
|
486
|
-
|
|
487
|
-
|
|
488
|
-
|
|
489
|
-
|
|
490
|
-
|
|
461
|
+
systemPromptMode: agent.systemPromptMode,
|
|
462
|
+
inheritProjectContext: agent.inheritProjectContext,
|
|
463
|
+
inheritGlobalContext: agent.inheritGlobalContext,
|
|
464
|
+
inheritSkills: agent.inheritSkills,
|
|
465
|
+
sandbox,
|
|
466
|
+
permissionProfile,
|
|
491
467
|
skills: shared.resolvedSkillNames ?? [],
|
|
492
468
|
toolPlan,
|
|
493
469
|
...(options.outputPath ? { outputPath: options.outputPath } : {}),
|
|
@@ -501,8 +477,8 @@ async function runSingleAttempt(
|
|
|
501
477
|
task: shared.originalTask ?? task,
|
|
502
478
|
...(childSessionName ? { sessionName: childSessionName } : {}),
|
|
503
479
|
...(options.agentContract ? { agentContract: options.agentContract } : {}),
|
|
504
|
-
|
|
505
|
-
|
|
480
|
+
...(sandbox ? { sandbox } : {}),
|
|
481
|
+
...(permissionProfile ? { permissionProfile } : {}),
|
|
506
482
|
launchContractDigest,
|
|
507
483
|
launchResolvedExtensions,
|
|
508
484
|
exitCode: 0,
|
|
@@ -532,6 +508,10 @@ async function runSingleAttempt(
|
|
|
532
508
|
pendingControlEvents.push(event);
|
|
533
509
|
options.onControlEvent?.(event);
|
|
534
510
|
};
|
|
511
|
+
let inputUsageComplete = true;
|
|
512
|
+
let outputUsageComplete = true;
|
|
513
|
+
let cacheReadUsageComplete = true;
|
|
514
|
+
let cacheWriteUsageComplete = true;
|
|
535
515
|
|
|
536
516
|
const progress: AgentProgress = {
|
|
537
517
|
index: options.index ?? 0,
|
|
@@ -546,12 +526,20 @@ async function runSingleAttempt(
|
|
|
546
526
|
tokens: 0,
|
|
547
527
|
...(modelArg ? { model: modelArg } : {}),
|
|
548
528
|
...(resolvedThinking ? { thinking: resolvedThinking } : {}),
|
|
549
|
-
inputTokens: 0,
|
|
550
|
-
outputTokens: 0,
|
|
551
529
|
durationMs: 0,
|
|
552
530
|
lastActivityAt: startTime,
|
|
553
531
|
};
|
|
554
532
|
result.progress = progress;
|
|
533
|
+
const projectCompleteUsage = () => {
|
|
534
|
+
if (inputUsageComplete) progress.inputTokens = result.usage.input;
|
|
535
|
+
else delete progress.inputTokens;
|
|
536
|
+
if (outputUsageComplete) progress.outputTokens = result.usage.output;
|
|
537
|
+
else delete progress.outputTokens;
|
|
538
|
+
if (cacheReadUsageComplete) progress.cacheRead = result.usage.cacheRead;
|
|
539
|
+
else delete progress.cacheRead;
|
|
540
|
+
if (cacheWriteUsageComplete) progress.cacheWrite = result.usage.cacheWrite;
|
|
541
|
+
else delete progress.cacheWrite;
|
|
542
|
+
};
|
|
555
543
|
const attemptTimeout = resolveAttemptTimeout(options);
|
|
556
544
|
if (attemptTimeout?.remainingMs === 0) {
|
|
557
545
|
result.exitCode = 1;
|
|
@@ -568,7 +556,6 @@ async function runSingleAttempt(
|
|
|
568
556
|
return result;
|
|
569
557
|
}
|
|
570
558
|
const mutationSnapshot = options.machine ? { source: "tracked-files" as const, trackedOnly: true as const, cwd: options.cwd ?? runtimeCwd, dirtyFiles: [], fingerprints: {}, unavailable: "Local Git evidence is not authoritative for a pane-native remote run." } : snapshotTrackedMutations(options.cwd ?? runtimeCwd);
|
|
571
|
-
let observedMutationAttempt = false;
|
|
572
559
|
let structuredOutputToolInvoked = false;
|
|
573
560
|
let structuredOutputMessageStartIndex: number | undefined;
|
|
574
561
|
let toolAvailabilityError: string | undefined;
|
|
@@ -597,6 +584,7 @@ async function runSingleAttempt(
|
|
|
597
584
|
const exitCode = await new Promise<number>((resolve) => {
|
|
598
585
|
const jsonlWriter = createJsonlWriter(shared.jsonlPath, { pause() {}, resume() {} });
|
|
599
586
|
let session: ChildSession | undefined;
|
|
587
|
+
let messageBaseline: number | undefined;
|
|
600
588
|
let unsubscribe: (() => void) | undefined;
|
|
601
589
|
let sessionSettled = false;
|
|
602
590
|
let lifecycleFinished = false;
|
|
@@ -813,11 +801,11 @@ async function runSingleAttempt(
|
|
|
813
801
|
|
|
814
802
|
let activeLongRunningNotified = false;
|
|
815
803
|
let pendingToolResult: { tool: string; path?: string; mutates: boolean; startedAt?: number } | undefined;
|
|
816
|
-
|
|
817
|
-
|
|
818
|
-
|
|
819
|
-
|
|
820
|
-
|
|
804
|
+
// Pi host versions may represent one tool result as either a dedicated
|
|
805
|
+
// tool_result_end event or a message_end whose message role is toolResult.
|
|
806
|
+
// Remember correlated ids so normalizing both forms cannot duplicate output.
|
|
807
|
+
const handledToolResultIds = new Set<string>();
|
|
808
|
+
const MAX_HANDLED_TOOL_RESULT_IDS = 1024;
|
|
821
809
|
type ActiveToolCall = { key: string; tool: string; args: string; startedAt: number; path?: string };
|
|
822
810
|
let activeToolSequence = 0;
|
|
823
811
|
const activeToolCalls = new Map<string, ActiveToolCall>();
|
|
@@ -995,52 +983,61 @@ async function runSingleAttempt(
|
|
|
995
983
|
emitUpdateSnapshot(output || "(running...)");
|
|
996
984
|
};
|
|
997
985
|
|
|
998
|
-
|
|
999
|
-
|
|
1000
|
-
|
|
1001
|
-
|
|
1002
|
-
|
|
1003
|
-
|
|
1004
|
-
|
|
1005
|
-
|
|
1006
|
-
|
|
1007
|
-
|
|
1008
|
-
|
|
1009
|
-
|
|
1010
|
-
|
|
1011
|
-
|
|
1012
|
-
|
|
1013
|
-
|
|
1014
|
-
|
|
1015
|
-
|
|
1016
|
-
|
|
1017
|
-
|
|
1018
|
-
|
|
1019
|
-
|
|
1020
|
-
|
|
1021
|
-
|
|
1022
|
-
|
|
1023
|
-
|
|
1024
|
-
|
|
1025
|
-
|
|
1026
|
-
|
|
1027
|
-
|
|
1028
|
-
|
|
1029
|
-
|
|
1030
|
-
|
|
1031
|
-
|
|
1032
|
-
|
|
1033
|
-
|
|
1034
|
-
|
|
1035
|
-
|
|
1036
|
-
|
|
1037
|
-
|
|
1038
|
-
|
|
1039
|
-
|
|
1040
|
-
|
|
1041
|
-
|
|
1042
|
-
|
|
1043
|
-
|
|
986
|
+
// Fork: shared, id-deduplicated tool-result completion. Pi hosts may
|
|
987
|
+
// deliver the same result either as a tool_result_end event or as a
|
|
988
|
+
// message_end whose message role is toolResult; both funnel through here
|
|
989
|
+
// so one result updates budget, output, and mutating-failure state once.
|
|
990
|
+
const handleToolResult = (message: Message, completion: { toolCallId?: unknown; toolName?: unknown }, now: number): void => {
|
|
991
|
+
const rawCallId = completion.toolCallId ?? (message as { toolCallId?: unknown }).toolCallId;
|
|
992
|
+
const callId = typeof rawCallId === "string" && rawCallId.length > 0 ? rawCallId : undefined;
|
|
993
|
+
if (callId) {
|
|
994
|
+
if (handledToolResultIds.has(callId)) return;
|
|
995
|
+
if (handledToolResultIds.size >= MAX_HANDLED_TOOL_RESULT_IDS) {
|
|
996
|
+
const oldest = handledToolResultIds.values().next().value;
|
|
997
|
+
if (typeof oldest === "string") handledToolResultIds.delete(oldest);
|
|
998
|
+
}
|
|
999
|
+
handledToolResultIds.add(callId);
|
|
1000
|
+
}
|
|
1001
|
+
result.messages!.push(message);
|
|
1002
|
+
const resultText = extractTextFromContent(message.content);
|
|
1003
|
+
// The result event's own tool name is authoritative; the single pending slot can
|
|
1004
|
+
// describe a different, overlapping call and serves only as a fallback.
|
|
1005
|
+
const blockedTool = (typeof completion.toolName === "string" && completion.toolName.length > 0
|
|
1006
|
+
? completion.toolName
|
|
1007
|
+
: undefined) ?? pendingToolResult?.tool;
|
|
1008
|
+
if (options.toolBudget && isToolBudgetBlockedMessage(options.toolBudget, resultText, blockedTool)) {
|
|
1009
|
+
result.toolBudgetBlocked = true;
|
|
1010
|
+
result.toolBudget = toolBudgetState(
|
|
1011
|
+
options.toolBudget,
|
|
1012
|
+
Math.max(progress.toolCount, options.toolBudget.hard + 1),
|
|
1013
|
+
blockedTool,
|
|
1014
|
+
);
|
|
1015
|
+
}
|
|
1016
|
+
appendRecentOutput(progress, resultText.split("\n").slice(-10));
|
|
1017
|
+
const toolSnapshot = pendingToolResult;
|
|
1018
|
+
pendingToolResult = undefined;
|
|
1019
|
+
if (toolSnapshot?.mutates && didMutatingToolFail(resultText)) {
|
|
1020
|
+
recordMutatingFailure(mutatingFailures, {
|
|
1021
|
+
tool: toolSnapshot.tool,
|
|
1022
|
+
path: toolSnapshot.path,
|
|
1023
|
+
error: resultText.split("\n").find((line) => line.trim())?.trim().slice(0, 180) ?? "mutating tool failed",
|
|
1024
|
+
ts: now,
|
|
1025
|
+
}, mutatingFailureWindowMs);
|
|
1026
|
+
if (shouldEscalateMutatingFailures(mutatingFailures, controlConfig.failedToolAttemptsBeforeAttention)) {
|
|
1027
|
+
emitNeedsAttention(now, {
|
|
1028
|
+
message: `${agent.name} needs attention after repeated mutating tool failures`,
|
|
1029
|
+
reason: "tool_failures",
|
|
1030
|
+
currentTool: toolSnapshot.tool,
|
|
1031
|
+
currentPath: toolSnapshot.path,
|
|
1032
|
+
currentToolDurationMs: toolSnapshot.startedAt ? Math.max(0, now - toolSnapshot.startedAt) : undefined,
|
|
1033
|
+
recentFailureSummary: summarizeRecentMutatingFailures(mutatingFailures),
|
|
1034
|
+
});
|
|
1035
|
+
}
|
|
1036
|
+
} else if (toolSnapshot?.mutates) {
|
|
1037
|
+
resetMutatingFailureState(mutatingFailures);
|
|
1038
|
+
}
|
|
1039
|
+
fireUpdate();
|
|
1040
|
+
};
|
|
1044
1041
|
|
|
1045
1042
|
const processEvent = (evt: ChildSessionEvent & { message?: Message; toolName?: string; toolCallId?: string; args?: unknown; willRetry?: unknown }) => {
|
|
1046
1043
|
if (lifecycleFinished) return;
|
|
@@ -1125,7 +1122,6 @@ async function runSingleAttempt(
|
|
|
1125
1122
|
result.toolBudget = toolBudgetState(options.toolBudget, progress.toolCount);
|
|
1126
1123
|
}
|
|
1127
1124
|
const mutates = isMutatingTool(evt.toolName, toolArgs, agent.mutationTools);
|
|
1128
|
-
observedMutationAttempt = observedMutationAttempt || mutates;
|
|
1129
1125
|
pendingToolResult = { tool: evt.toolName ?? "tool", path: activeTool?.path, mutates, startedAt: now };
|
|
1130
1126
|
fireUpdate();
|
|
1131
1127
|
}
|
|
@@ -1145,68 +1141,71 @@ async function runSingleAttempt(
|
|
|
1145
1141
|
}
|
|
1146
1142
|
|
|
1147
1143
|
if (evt.type === "message_end" && evt.message) {
|
|
1148
|
-
|
|
1149
|
-
|
|
1150
|
-
|
|
1151
|
-
|
|
1152
|
-
|
|
1153
|
-
|
|
1154
|
-
|
|
1155
|
-
|
|
1156
|
-
|
|
1144
|
+
if (evt.message.role === "toolResult") {
|
|
1145
|
+
// Fork: hosts that deliver tool results as message_end instead of
|
|
1146
|
+
// tool_result_end reach the same deduplicated completion path.
|
|
1147
|
+
handleToolResult(evt.message, evt, now);
|
|
1148
|
+
} else {
|
|
1149
|
+
result.messages!.push(evt.message);
|
|
1150
|
+
if (childWatchdog) {
|
|
1151
|
+
const next = applyChildWatchdogMessage(childWatchdogState, evt.message);
|
|
1152
|
+
if (next) updateChildWatchdogState(next);
|
|
1153
|
+
}
|
|
1154
|
+
if (evt.message.role === "assistant") {
|
|
1155
|
+
result.usage.turns++;
|
|
1156
|
+
progress.turnCount = result.usage.turns;
|
|
1157
|
+
const toolCalls = Array.isArray(evt.message.content)
|
|
1158
|
+
? evt.message.content.filter((part) => (part as { type?: string }).type === "toolCall")
|
|
1159
|
+
: [];
|
|
1160
|
+
const hasToolCall = toolCalls.length > 0;
|
|
1161
|
+
const terminalAssistantStop = (evt.message as { stopReason?: string }).stopReason === "stop" && !hasToolCall;
|
|
1162
|
+
const u = evt.message.usage;
|
|
1163
|
+
inputUsageComplete &&= isCompleteUsageCounter(u?.input);
|
|
1164
|
+
outputUsageComplete &&= isCompleteUsageCounter(u?.output);
|
|
1165
|
+
cacheReadUsageComplete &&= isCompleteUsageCounter(u?.cacheRead);
|
|
1166
|
+
cacheWriteUsageComplete &&= isCompleteUsageCounter(u?.cacheWrite);
|
|
1167
|
+
if (u) {
|
|
1168
|
+
const window = (u.input || 0) + (u.cacheRead || 0);
|
|
1169
|
+
result.usage.input += u.input || 0;
|
|
1170
|
+
result.usage.output += u.output || 0;
|
|
1171
|
+
result.usage.cacheRead += u.cacheRead || 0;
|
|
1172
|
+
result.usage.cacheWrite += u.cacheWrite || 0;
|
|
1173
|
+
result.usage.cost += u.cost?.total || 0;
|
|
1174
|
+
progress.tokens = result.usage.input + result.usage.output;
|
|
1175
|
+
progress.window = window;
|
|
1176
|
+
progress.windowPeak = Math.max(progress.windowPeak ?? 0, window);
|
|
1157
1177
|
}
|
|
1158
|
-
|
|
1159
|
-
|
|
1160
|
-
progress.
|
|
1161
|
-
|
|
1162
|
-
|
|
1163
|
-
|
|
1164
|
-
|
|
1165
|
-
const terminalAssistantStop = (evt.message as { stopReason?: string }).stopReason === "stop" && !hasToolCall;
|
|
1166
|
-
const u = evt.message.usage;
|
|
1167
|
-
if (u) {
|
|
1168
|
-
const window = (u.input || 0) + (u.cacheRead || 0);
|
|
1169
|
-
result.usage.input += u.input || 0;
|
|
1170
|
-
result.usage.output += u.output || 0;
|
|
1171
|
-
result.usage.cacheRead += u.cacheRead || 0;
|
|
1172
|
-
result.usage.cacheWrite += u.cacheWrite || 0;
|
|
1173
|
-
result.usage.cost += u.cost?.total || 0;
|
|
1174
|
-
progress.tokens = result.usage.input + result.usage.output;
|
|
1175
|
-
progress.inputTokens = result.usage.input;
|
|
1176
|
-
progress.outputTokens = result.usage.output;
|
|
1177
|
-
progress.window = window;
|
|
1178
|
-
progress.windowPeak = Math.max(progress.windowPeak ?? 0, window);
|
|
1179
|
-
}
|
|
1180
|
-
if (evt.message.model) {
|
|
1181
|
-
progress.model = evt.message.model;
|
|
1182
|
-
if (!result.model) result.model = evt.message.model;
|
|
1183
|
-
if (expectedModelForVerification && !hasToolCall) {
|
|
1184
|
-
const modelVerificationError = formatSubagentModelVerificationError(expectedModelForVerification, evt.message.model, options.availableModels, options.modelResponseAliases);
|
|
1185
|
-
if (modelVerificationError && !result.error) result.error = modelVerificationError;
|
|
1186
|
-
}
|
|
1187
|
-
}
|
|
1188
|
-
if (evt.message.errorMessage) assistantError = evt.message.errorMessage;
|
|
1189
|
-
else if (hasToolCall && evt.message.stopReason === "toolUse") {
|
|
1190
|
-
// A recovered request can finish via a terminating tool, without a text stop.
|
|
1191
|
-
assistantError = undefined;
|
|
1192
|
-
}
|
|
1193
|
-
const assistantText = extractTextFromContent(evt.message.content);
|
|
1194
|
-
appendRecentOutput(progress, assistantText.split("\n").slice(-10));
|
|
1195
|
-
// Final assistant message: start the settle drain window.
|
|
1196
|
-
if (terminalAssistantStop) {
|
|
1197
|
-
if (!evt.message.errorMessage && assistantText.trim()) assistantError = undefined;
|
|
1198
|
-
cleanTerminalAssistantStopReceived ||= !evt.message.errorMessage;
|
|
1199
|
-
clearAllToolTimeouts();
|
|
1200
|
-
activeToolCalls.clear();
|
|
1201
|
-
activeToolKeysByName.clear();
|
|
1202
|
-
refreshCurrentTool();
|
|
1203
|
-
applyChildLifecycle(projectChildLifecycle(evt, true, childLifecycleState));
|
|
1178
|
+
projectCompleteUsage();
|
|
1179
|
+
if (evt.message.model) {
|
|
1180
|
+
progress.model = evt.message.model;
|
|
1181
|
+
if (!result.model) result.model = evt.message.model;
|
|
1182
|
+
if (expectedModelForVerification && !hasToolCall) {
|
|
1183
|
+
const modelVerificationError = formatSubagentModelVerificationError(expectedModelForVerification, evt.message.model, options.availableModels, options.modelResponseAliases);
|
|
1184
|
+
if (modelVerificationError && !result.error) result.error = modelVerificationError;
|
|
1204
1185
|
}
|
|
1205
1186
|
}
|
|
1206
|
-
|
|
1207
|
-
|
|
1187
|
+
if (evt.message.errorMessage) assistantError = evt.message.errorMessage;
|
|
1188
|
+
else if (hasToolCall && evt.message.stopReason === "toolUse") {
|
|
1189
|
+
// A recovered request can finish via a terminating tool, without a text stop.
|
|
1190
|
+
assistantError = undefined;
|
|
1191
|
+
}
|
|
1192
|
+
const assistantText = extractTextFromContent(evt.message.content);
|
|
1193
|
+
appendRecentOutput(progress, assistantText.split("\n").slice(-10));
|
|
1194
|
+
// Final assistant message: start the settle drain window.
|
|
1195
|
+
if (terminalAssistantStop) {
|
|
1196
|
+
if (!evt.message.errorMessage && assistantText.trim()) assistantError = undefined;
|
|
1197
|
+
cleanTerminalAssistantStopReceived ||= !evt.message.errorMessage;
|
|
1198
|
+
clearAllToolTimeouts();
|
|
1199
|
+
activeToolCalls.clear();
|
|
1200
|
+
activeToolKeysByName.clear();
|
|
1201
|
+
refreshCurrentTool();
|
|
1202
|
+
applyChildLifecycle(projectChildLifecycle(evt, true, childLifecycleState));
|
|
1203
|
+
}
|
|
1208
1204
|
}
|
|
1205
|
+
updateActivityState(now);
|
|
1206
|
+
fireUpdate();
|
|
1209
1207
|
}
|
|
1208
|
+
}
|
|
1210
1209
|
|
|
1211
1210
|
if (evt.type === "tool_result_end" && evt.message) {
|
|
1212
1211
|
const toolResultCompletion = {
|
|
@@ -1223,16 +1222,16 @@ async function runSingleAttempt(
|
|
|
1223
1222
|
});
|
|
1224
1223
|
refreshCurrentTool();
|
|
1225
1224
|
}
|
|
1226
|
-
|
|
1227
|
-
|
|
1228
|
-
|
|
1229
|
-
|
|
1225
|
+
// Fork: the shared, id-deduplicated tail (message push, budget,
|
|
1226
|
+
// mutating-failure tracking, output) lives in handleToolResult; hosts
|
|
1227
|
+
// may deliver the same result via message_end(role=toolResult).
|
|
1228
|
+
handleToolResult(evt.message, evt, now);
|
|
1230
1229
|
}
|
|
1231
1230
|
};
|
|
1232
|
-
|
|
1233
|
-
|
|
1234
|
-
|
|
1235
|
-
|
|
1231
|
+
// SAFETY: this callback is only ever invoked with child-watchdog status events
|
|
1232
|
+
// (declared as ChildWatchdogStatusEvent above); those carry the session-event
|
|
1233
|
+
// fields `processEvent` needs, and `processEvent` re-validates the shape with
|
|
1234
|
+
// `isChildWatchdogStatusEvent` before touching watchdog-specific fields.
|
|
1236
1235
|
onWatchdogStatus = (event) => processEvent(event as unknown as Parameters<typeof processEvent>[0]);
|
|
1237
1236
|
|
|
1238
1237
|
fireUpdate();
|
|
@@ -1370,6 +1369,12 @@ async function runSingleAttempt(
|
|
|
1370
1369
|
}
|
|
1371
1370
|
const finalCode = forcedDrainAfterFinalSuccess && !forcedDrainAfterEmptyTerminal ? 0 : closeError || promptError !== undefined ? 1 : 0;
|
|
1372
1371
|
if (!result.error && closeError) result.error = closeError;
|
|
1372
|
+
if (session && messageBaseline !== undefined) {
|
|
1373
|
+
result.usage = reconcileAttemptUsage(result.usage, session.messages, messageBaseline);
|
|
1374
|
+
progress.tokens = result.usage.input + result.usage.output;
|
|
1375
|
+
projectCompleteUsage();
|
|
1376
|
+
progress.turnCount = result.usage.turns;
|
|
1377
|
+
}
|
|
1373
1378
|
finish(finalCode);
|
|
1374
1379
|
};
|
|
1375
1380
|
|
|
@@ -1453,6 +1458,7 @@ async function runSingleAttempt(
|
|
|
1453
1458
|
abortChild();
|
|
1454
1459
|
}
|
|
1455
1460
|
options.onChildSession?.({ steer: (text) => created.steer(text), followUp: (text) => created.followUp(text) });
|
|
1461
|
+
messageBaseline = created.messages.length;
|
|
1456
1462
|
await created.prompt(`Task: ${task}`);
|
|
1457
1463
|
settle(undefined);
|
|
1458
1464
|
} catch (error) {
|
|
@@ -1492,15 +1498,11 @@ async function runSingleAttempt(
|
|
|
1492
1498
|
result.exitCode = 1;
|
|
1493
1499
|
}
|
|
1494
1500
|
let validatedStructuredOutput = false;
|
|
1495
|
-
if (options.structuredOutput
|
|
1501
|
+
if (options.structuredOutput) {
|
|
1496
1502
|
result.structuredOutputSchemaPath = options.structuredOutput.schemaPath;
|
|
1497
1503
|
result.structuredOutputPath = options.structuredOutput.outputPath;
|
|
1498
1504
|
const structured = capture.structuredOutput();
|
|
1499
|
-
if (
|
|
1500
|
-
result.exitCode = 1;
|
|
1501
|
-
result.error = MISSING_STRUCTURED_OUTPUT_CALL_ERROR;
|
|
1502
|
-
result.structuredOutputFailed = true;
|
|
1503
|
-
} else {
|
|
1505
|
+
if (structuredOutputToolInvoked && structured.called) {
|
|
1504
1506
|
result.structuredOutput = structured.value;
|
|
1505
1507
|
const acceptanceMode = options.structuredOutput.acceptanceReportPath
|
|
1506
1508
|
? options.structuredOutput.acceptanceReportRequired ? "required" : "optional"
|
|
@@ -1512,6 +1514,12 @@ async function runSingleAttempt(
|
|
|
1512
1514
|
(result as SingleResult & { structuredAcceptanceReport?: unknown; structuredAcceptanceReportError?: string }).structuredAcceptanceReportError = acceptanceReportError;
|
|
1513
1515
|
writeStructuredOutputArtifacts(options.structuredOutput, structured.value, acceptanceMode ? structured.acceptanceReport : undefined);
|
|
1514
1516
|
validatedStructuredOutput = true;
|
|
1517
|
+
} else if (result.exitCode === 0 && !result.error) {
|
|
1518
|
+
result.exitCode = 1;
|
|
1519
|
+
result.error = structuredOutputToolInvoked
|
|
1520
|
+
? formatStructuredOutputRejectionError(result.messages ?? [])
|
|
1521
|
+
: MISSING_STRUCTURED_OUTPUT_CALL_ERROR;
|
|
1522
|
+
result.structuredOutputFailed = true;
|
|
1515
1523
|
}
|
|
1516
1524
|
}
|
|
1517
1525
|
if (result.exitCode === 0 && !result.error) {
|
|
@@ -1526,17 +1534,17 @@ async function runSingleAttempt(
|
|
|
1526
1534
|
&& hasEmptyTerminalAssistantResponse(messages)
|
|
1527
1535
|
&& (progress.toolCount > 0 || Boolean(finalText?.trim()));
|
|
1528
1536
|
if ((missingOutput || terminalEmptyAfterUsefulWork) && (!errInfo.hasError || hasEmptyTerminalAssistantResponse(messages))) {
|
|
1529
|
-
|
|
1530
|
-
|
|
1531
|
-
|
|
1537
|
+
// A startup guard can block the first turn, which leaves no message to
|
|
1538
|
+
// inspect. Prefer the reason the blocked child published.
|
|
1539
|
+
const sandboxDiagnostic = readSandboxStartupDiagnostic(launch.sandboxDiagnosticsPath);
|
|
1532
1540
|
result.exitCode = 1;
|
|
1533
|
-
|
|
1534
|
-
|
|
1535
|
-
|
|
1536
|
-
|
|
1537
|
-
|
|
1538
|
-
|
|
1539
|
-
|
|
1541
|
+
if (sandboxDiagnostic) {
|
|
1542
|
+
result.error = sandboxDiagnostic.reason;
|
|
1543
|
+
result.startupBlocked = true;
|
|
1544
|
+
clearSandboxStartupDiagnostic(launch.sandboxDiagnosticsPath);
|
|
1545
|
+
} else {
|
|
1546
|
+
result.error = formatEmptyTerminalAssistantResponseError(messages);
|
|
1547
|
+
}
|
|
1540
1548
|
} else if (errInfo.hasError) {
|
|
1541
1549
|
result.exitCode = errInfo.exitCode ?? 1;
|
|
1542
1550
|
result.error = errInfo.details
|
|
@@ -1590,70 +1598,6 @@ async function runSingleAttempt(
|
|
|
1590
1598
|
fullOutput = fullOutput.trim()
|
|
1591
1599
|
? `${timeoutMessage}\n\n${result.timeoutRecovery.message}\n\nPartial output before timeout:\n${fullOutput}`
|
|
1592
1600
|
: `${timeoutMessage}\n\n${result.timeoutRecovery.message}`;
|
|
1593
|
-
}
|
|
1594
|
-
const completionGuardEnabled = isAgentContract(options.agentContract) ? agent.completionGuard === true : agent.completionGuard !== false;
|
|
1595
|
-
const completionGuard = ((result.exitCode === 0 && !result.error) || toolAvailabilityError) && completionGuardEnabled
|
|
1596
|
-
? evaluateCompletionMutationGuard({
|
|
1597
|
-
agent: agent.name,
|
|
1598
|
-
task: shared.originalTask ?? task,
|
|
1599
|
-
messages: result.messages ?? [],
|
|
1600
|
-
tools: contractTools,
|
|
1601
|
-
mcpDirectTools: toolPlan.effectiveMcpTools,
|
|
1602
|
-
mutationTools: agent.mutationTools,
|
|
1603
|
-
toolAvailabilityError,
|
|
1604
|
-
mutationEvidence,
|
|
1605
|
-
})
|
|
1606
|
-
: undefined;
|
|
1607
|
-
const mutationAttemptObserved = observedMutationAttempt || mutationEvidence.attemptedMutation;
|
|
1608
|
-
let completionGuardTriggered = completionGuard?.triggered === true && !mutationAttemptObserved;
|
|
1609
|
-
// The classifier is deliberately narrow, so a read-only review task can
|
|
1610
|
-
// still be misread as implementation. Arbitrate BEFORE any failure side
|
|
1611
|
-
// effect is published (effects, exit code, progress, notifications,
|
|
1612
|
-
// acceptance, output persistence): only a confident read-only verdict
|
|
1613
|
-
// rescues, and the task text alone is evidence — never the child's own
|
|
1614
|
-
// final message.
|
|
1615
|
-
let arbiterRescued = false;
|
|
1616
|
-
if (completionGuardTriggered) {
|
|
1617
|
-
const arbitration = await arbitrateCompletionGuardRescue({
|
|
1618
|
-
guardTriggered: true,
|
|
1619
|
-
task: shared.originalTask ?? task,
|
|
1620
|
-
arbiter: options.llmIntentArbiter,
|
|
1621
|
-
});
|
|
1622
|
-
completionGuardTriggered = arbitration.triggered;
|
|
1623
|
-
arbiterRescued = arbitration.rescued;
|
|
1624
|
-
}
|
|
1625
|
-
const completionEvidence = planCompletionEvidence({
|
|
1626
|
-
guard: completionGuard,
|
|
1627
|
-
guardTriggered: completionGuardTriggered,
|
|
1628
|
-
completionGuardEnabled,
|
|
1629
|
-
mutationCapable: hasMutationToolCapability(contractTools, toolPlan.effectiveMcpTools),
|
|
1630
|
-
implementationMutationExpected: expectsImplementationMutation(agent.name, shared.originalTask ?? task),
|
|
1631
|
-
mutationAttemptObserved,
|
|
1632
|
-
mutationEvidence,
|
|
1633
|
-
arbiterRescued,
|
|
1634
|
-
agentContractEnabled: isAgentContract(options.agentContract),
|
|
1635
|
-
});
|
|
1636
|
-
if (completionEvidence.fileMutation) {
|
|
1637
|
-
result.effects = {
|
|
1638
|
-
...(result.effects ?? {}),
|
|
1639
|
-
fileMutation: completionEvidence.fileMutation,
|
|
1640
|
-
};
|
|
1641
|
-
}
|
|
1642
|
-
if (completionEvidence.legacyFailureError) {
|
|
1643
|
-
result.exitCode = 1;
|
|
1644
|
-
result.error = completionEvidence.legacyFailureError;
|
|
1645
|
-
progress.status = "failed";
|
|
1646
|
-
progress.error = result.error;
|
|
1647
|
-
emitControlEvent(buildControlEvent({
|
|
1648
|
-
from: progress.activityState,
|
|
1649
|
-
to: "needs_attention",
|
|
1650
|
-
runId: options.runId ?? agent.name,
|
|
1651
|
-
agent: agent.name,
|
|
1652
|
-
index: options.index,
|
|
1653
|
-
ts: Date.now(),
|
|
1654
|
-
message: `${agent.name} completed without making edits for an implementation task`,
|
|
1655
|
-
reason: "completion_guard",
|
|
1656
|
-
}));
|
|
1657
1601
|
}
|
|
1658
1602
|
if (options.outputPath && result.exitCode === 0) {
|
|
1659
1603
|
const resolvedOutput = resolveSingleOutput(options.outputPath, fullOutput, shared.outputSnapshot, options.outputClaimPath);
|
|
@@ -1953,6 +1897,22 @@ async function runSyncCompletionInner(
|
|
|
1953
1897
|
const verifyModel = Boolean(candidate) && !options.modelOverrideFromParent;
|
|
1954
1898
|
let lastResult: SingleResult | undefined;
|
|
1955
1899
|
let recoveryPrompt = task;
|
|
1900
|
+
let stagedIndexBaseline: string | undefined;
|
|
1901
|
+
if (effectiveAcceptance.preserveStagedIndex) {
|
|
1902
|
+
try {
|
|
1903
|
+
stagedIndexBaseline = captureStagedIndexBaseline(options.cwd ?? runtimeCwd);
|
|
1904
|
+
} catch (error) {
|
|
1905
|
+
return redactResultPrompt(withRunContext({
|
|
1906
|
+
index: options.index ?? 0,
|
|
1907
|
+
agent: agentName,
|
|
1908
|
+
task,
|
|
1909
|
+
exitCode: 1,
|
|
1910
|
+
messages: [],
|
|
1911
|
+
usage: emptyUsage(),
|
|
1912
|
+
error: error instanceof Error ? error.message : String(error),
|
|
1913
|
+
}, options.context));
|
|
1914
|
+
}
|
|
1915
|
+
}
|
|
1956
1916
|
for (let attemptIndex = 0; attemptIndex < 2; attemptIndex++) {
|
|
1957
1917
|
const outputSnapshot = captureSingleOutputSnapshot(options.outputPath);
|
|
1958
1918
|
const attemptResult = await runSingleAttempt(runtimeCwd, agent, recoveryPrompt, candidate, attemptOptions, {
|
|
@@ -2083,6 +2043,7 @@ async function runSyncCompletionInner(
|
|
|
2083
2043
|
? { content: childWrittenOutput, path: options.outputPath, authoritative: options.outputMode === "file-only", durable: result.savedOutputPath !== undefined }
|
|
2084
2044
|
: undefined,
|
|
2085
2045
|
cwd: options.cwd ?? runtimeCwd,
|
|
2046
|
+
stagedIndexBaseline,
|
|
2086
2047
|
reportOptional: isAgentContract(options.agentContract),
|
|
2087
2048
|
artifactsDir: options.artifactsDir,
|
|
2088
2049
|
runId: options.runId,
|
|
@@ -2095,6 +2056,12 @@ async function runSyncCompletionInner(
|
|
|
2095
2056
|
}
|
|
2096
2057
|
const acceptanceFailure = acceptanceFailureMessage(result.acceptance);
|
|
2097
2058
|
stripAcceptanceReportsFromMessages(result.messages);
|
|
2059
|
+
// A passing typed gate supplies the structured output for runs that have no
|
|
2060
|
+
// outputSchema of their own; preflight rejects the combination.
|
|
2061
|
+
const typedGate = typedVerifyOutput(result.acceptance);
|
|
2062
|
+
if (typedGate && result.structuredOutput === undefined && !acceptanceFailure && result.exitCode === 0) {
|
|
2063
|
+
result.structuredOutput = typedGate.value;
|
|
2064
|
+
}
|
|
2098
2065
|
if (acceptanceFailure && result.acceptance.explicit && result.exitCode === 0 && !result.interrupted && !result.timedOut && !isAgentContract(options.agentContract)) {
|
|
2099
2066
|
result.exitCode = 1;
|
|
2100
2067
|
if (result.savedOutputPath) {
|