Files
openclaw/src/agents/command/run-embedded-attempt.ts
T
Peter Steinberger 4273ca9dbd refactor(sessions): remove file-era transcript runtime (#113233)
* refactor(sessions): keep helper transcripts in memory

* refactor(sessions): remove file-era transcript storage

* test(sessions): use SQLite identity in attempt persistence

* test(codex): isolate legacy transcript fixtures

* fix(sessions): preserve SQLite transcript identity

* fix(sessions): harden transcript lifecycle invariants

* fix(sessions): validate transcript identities

* fix(sessions): close identity compatibility gaps

* fix(sessions): preserve leaf and plugin identities

* fix(sessions): retain dispatch transcript targets

* fix(sessions): preserve active transcript context

* fix(sessions): isolate artifact accounting

* fix(sessions): bound SQLite usage accounting

* fix(sessions): retain bounded latest usage

* fix(sessions): align rebased transcript targets

* test(sessions): align accessor scope fixture

* fix(telegram): derive SQLite transcript identity

* refactor(sessions): remove file-era compaction residue

* chore(sessions): lower max-lines baseline

* fix(sessions): preserve structured transcript identity

* test(sessions): align doctor identity assertions

* fix(sessions): isolate default SDK database

* refactor(sessions): remove dead file-era exports

* fix(sessions): reconcile SQLite transcript identity

* fix(sessions): pass checkpoint identity explicitly

* test(sessions): make entry field probe explicit

* test(sessions): satisfy transcript cleanup lint

* test(sessions): align diagnostics identity proof

* fix(sessions): finish transcript runtime teardown

* fix(sessions): preserve transcript identity invariants

* fix(sessions): harden transcript compatibility edges

* fix(sessions): preserve checkpoint transcript anchors

* fix(sessions): preserve SQLite lifecycle invariants

* fix(sessions): retarget compaction successors

* test(sessions): preserve transcript fixture semantics

* feat(plugin-sdk): add command transcript targets

* fix(sessions): serialize transcript rewrites

* fix(sessions): validate legacy successor identity

* fix(sessions): normalize compaction ownership

* fix(sessions): validate successor identity before adoption

* fix(sessions): preserve plugin transcript ownership

* fix(sessions): carry transcript identity through commands

* fix(sessions): import legacy checkpoint artifacts into SQLite

* fix(sessions): preserve successor transcript ownership

* fix(sessions): align transcript consumers with target identity

* fix(sessions): scope transcript token estimates

* fix(sessions): retain agent identity across lifecycle hooks

* fix(sessions): resolve scoped SQLite targets

* fix(sessions): isolate lifecycle transcript targets

* fix(sessions): validate compaction agent ownership

* fix(sessions): preserve reset and cleanup lifecycle

* fix(sessions): serialize prompt cleanup lifecycle

* fix(sessions): remove stale lock import

* fix(sessions): preserve reset target context

* fix(sessions): fence prompt reload takeover

* fix(sessions): unblock abort and default lifecycle reads

* fix(sessions): validate legacy successor scope

* fix(sessions): reject metadata-only runtime rows

* fix(sessions): propagate custom transcript stores

* fix(sessions): preserve adopted retry targets

* fix(sessions): allow unkeyed usage reads

* fix(sessions): harden runtime target boundaries

* fix(sessions): serialize retry transcript writes

* fix(sessions): bound prompt reload disposal

* fix(sessions): complete retry marker identity

* fix(sessions): keep legacy marker identity minimal

* test(sessions): tighten teardown fixture types

* fix(sessions): preserve compatibility target identity

* test(sessions): persist post-checkpoint boundary turn

* test(sessions): align runtime store mock contracts

* style(sessions): simplify persisted identity guard

* fix(sessions): prefer complete typed targets

* fix(sessions): recover legacy marker targets

* test(sessions): align marker lookup fixture scope

* fix(sessions): validate partial transcript targets

* fix(sessions): reconcile partial transcript identities

* fix(sessions): canonicalize compatibility identities

* test(sessions): cover compatibility aliases

* fix(sessions): adopt legacy successor identity

* fix(sessions): preserve usage read identity

* fix(sessions): preserve partial marker compatibility

* fix(sessions): validate legacy successor mappings

* fix(sessions): reconcile marker store mappings

* fix(sessions): preserve legacy fallback identity

* fix(sessions): harden marker alias resolution

* fix(sessions): prefer verified successor aliases

* fix(sessions): resolve preferred marker aliases

* fix(sessions): serialize cleanup admission

* fix(sessions): align marker lookup scopes

* fix(codex): type marker alias summaries

* style(sessions): satisfy changed lint

* test(sessions): align structured target assertions

* fix(sessions): reconcile latest identity contracts

* fix(sessions): validate transcript identity boundaries

* docs(sessions): explain stable registry keys

* fix(sessions): harden compatibility target round trips

* fix(sessions): port usage identity to split modules

* test(sessions): align subagent transcript identity

* fix(sessions): finish transcript identity migration

* fix(agents): route subagent completion capture through transcript targets

* fix(agents): settle SQLite prompt handoff during cleanup

* chore: shrink max-lines baseline after teardown

* fix(sessions): port teardown across split runtime owners

* fix(sessions): carry transcript targets through split owners

* test(agents): use SQLite compaction target in abort coverage

* chore: retain unrelated max-lines suppressions

* chore: shrink max-lines baseline after main splits

* style(agents): const compaction checkpoint locals

* fix(sessions): harden SQLite teardown boundaries

* test(sessions): use typed metadata in predicate isolation fixture

* test(agents): cover malformed settlement rejections lint-safely

* fix(sessions): close remaining SQLite identity races

* fix(agents): fail closed on incomplete successor targets

* fix(sessions): preserve transcript identity fallbacks

* fix(agents): preserve session-key abort admission

* fix(trajectory): validate incomplete export targets

* test(sessions): drop retired pricing cache imports

* fix(sessions): validate partial transcript identities

* fix(sessions): close transcript identity edge cases

* fix(plugins): reserve retired transcript locator slot

* fix(sessions): scope transcript locks by target

* style(sessions): simplify SDK initialization error

* fix(sessions): preserve initialized transcript state

* fix(codex): verify mirrored history session keys

* fix(sessions): reject stale transcript ownership

* fix(sessions): anchor asynchronous transcript ownership

* fix(sessions): measure active transcript state

* fix(sessions): preserve scoped transcript compaction

* fix(sessions): harden transcript identity and lifecycle

* fix(sessions): resolve scoped command transcript stores

* fix(sessions): make transcript appends failure-atomic

* fix(sessions): enforce scoped transcript ownership

* fix(sessions): reject cross-owner transcript handoffs

* fix(sessions): fence cleanup transcript ownership

* fix(sessions): retire stale write ownership contexts

* fix(sessions): preserve pending session migration state

* fix(sessions): validate migrated transcript ownership

* fix(sessions): validate usage transcript targets

* fix(sessions): clear predecessor transcript metadata

* fix(sessions): align durable session event targets

* fix(sessions): fence late prompt handoffs

* fix(sessions): fence lifecycle transcript fallbacks

* fix(sessions): bound zero-length memory capture

* fix(sessions): preserve transcript teardown ownership

* fix(sessions): reject duplicate cleanup ownership

* fix(sessions): serialize runtime writes with sqlite leases

* fix(sessions): close sqlite teardown concurrency gaps

* fix(sessions): preserve nested lifecycle failures

* fix(sessions): canonicalize sqlite transcript ownership

* fix(sessions): settle disposed prompt handoffs

* fix(sessions): resolve canonical attempt lock targets

* test(sessions): align canonical target fixtures

* test(sessions): retire redundant jsonl parser coverage

* refactor(sessions): split active transcript cursors

* test(memory): retire legacy marker fixture

* fix(sessions): preserve canonical transcript access after rebase

* fix(sessions): fence prompt lease and return transcript targets

* fix(sessions): colocate transcript leases with target store

* fix(sessions): canonicalize transcript lease and worker targets

* fix(sessions): preserve plugin and fork identity markers

* fix(sessions): complete sqlite transcript target migration

* fix(sessions): integrate canonical followup identity

* fix(sessions): preserve bounded transcript topology

* fix(sessions): validate transcript identity boundaries

* fix(context): separate caller and successor targets

* test(sessions): split persistence compatibility coverage

* test(sessions): preserve fixture topology efficiently

* chore(sdk): refresh plugin api baseline

* test(agents): align compaction lock target mocks

* test(sessions): seed malformed transcript fixtures directly

* fix(agents): canonicalize transcript compatibility inputs

* fix(agents): type optional tool result ids

* test(ci): stabilize loaded process timing

* test(tui): wait for collect queue admission
2026-07-27 22:33:24 -04:00

637 lines
23 KiB
TypeScript

import { sanitizeForLog } from "../../../packages/terminal-core/src/ansi.js";
import type { SessionEntry } from "../../config/sessions/types.js";
import { emitAgentEvent } from "../../infra/agent-events.js";
import { createSubsystemLogger } from "../../logging/subsystem.js";
import {
MODEL_SELECTION_LOCKED_MESSAGE,
ModelSelectionLockedError,
isModelSelectionLocked,
} from "../../sessions/model-overrides.js";
import { createUserTurnTranscriptRecorder } from "../../sessions/user-turn-transcript.js";
import {
getGeneratedMediaTaskIdsForSessionKey,
hasNewGeneratedMediaTaskForSessionKey,
} from "../../tasks/task-status-access.js";
import { createTrajectoryRuntimeRecorder } from "../../trajectory/runtime.js";
import { resolveMessageChannel } from "../../utils/message-channel.js";
import {
clearAutoFallbackPrimaryProbeSelection,
entryMatchesAutoFallbackPrimaryProbe,
markAutoFallbackPrimaryProbe,
resolveEffectiveModelFallbacks,
} from "../agent-scope.js";
import {
runEmbeddedAgentEntry,
type EmbeddedAgentRunEntryTerminal,
} from "../embedded-agent-runner/run-entry.js";
import { resolveFastModeState } from "../fast-mode.js";
import { runAgentHarnessBeforeMessageWriteHook } from "../harness/hook-helpers.js";
import { prepareInternalSessionEffectsSession } from "../internal-session-effects.js";
import { LiveSessionModelSwitchError } from "../live-model-switch.js";
import { modelKey, resolveThinkingDefault } from "../model-selection.js";
import type { AgentRunSessionTarget } from "../run-session-target.js";
import {
isAgentRunDirectAbortReason,
isAgentRunRestartAbortReason,
resolveAgentRunErrorLifecycleFields,
} from "../run-termination.js";
import { resolveSessionRuntimeOverrideForProvider } from "../session-runtime-compat.js";
import {
resolveCandidateThinkingLevel,
resolveEffectiveAgentRuntime,
} from "../thinking-runtime.js";
import {
createAgentAttemptLifecycleCallbacks,
type AgentAttemptLifecycleState,
} from "./attempt-callbacks.js";
import { createAgentCommandLifecycle } from "./lifecycle.js";
import { normalizeAgentCommandModelRef } from "./model-ref.js";
import type { EmbeddedModelSelection } from "./model-selection.js";
import type { PreparedAgentCommandExecution } from "./prepare.js";
import { loadAttemptExecutionRuntime, type AgentAttemptResult } from "./runtime-loaders.js";
import { persistSessionEntry, resolveInternalSessionEffectsSource } from "./session-helpers.js";
import type { EmbeddedSessionState } from "./session-preparation.js";
import type { AgentCommandOpts } from "./types.js";
const log = createSubsystemLogger("agents/agent-command");
const MAX_LIVE_SWITCH_RETRIES = 5;
export async function runEmbeddedAgentAttempt(params: {
prepared: PreparedAgentCommandExecution;
opts: AgentCommandOpts;
sessionEntry?: SessionEntry;
lifecycleGeneration: string;
onLifecycleGenerationChanged: (lifecycleGeneration: string) => void;
suppressVisibleSessionEffects: boolean;
preserveUserFacingSessionModelState: boolean;
modelSelection: EmbeddedModelSelection;
embeddedSessionState: EmbeddedSessionState;
trackInternalModelRunTarget: (target: AgentRunSessionTarget | undefined) => void;
}) {
const {
cfg,
body,
transcriptBody,
sessionId,
sessionKey,
sessionStore,
storePath,
sessionAgentId,
workspaceDir,
cwd,
agentDir,
runId,
pluginsEnabled,
manifestMetadataSnapshot,
modelManifestContext,
normalizedSpawned,
isNewSession,
timeoutMs,
runTimeoutOverrideMs,
} = params.prepared;
const { runContext, skillsSnapshot, resolvedVerboseLevel } = params.embeddedSessionState;
const {
defaultProvider,
defaultModel,
configuredDefaultAuthProfileId,
visibilityPolicy,
hasExplicitRunOverride,
storedProviderOverride,
hasStoredAutoFallbackProvenance,
autoFallbackPrimaryProbe,
thinkingCatalog,
immutableThinkLevel,
sessionFile,
} = params.modelSelection;
let {
provider,
model,
providerForAuthProfileValidation,
sessionEntryForAttempt,
storedModelOverride,
storedModelOverrideSource,
effectiveTurnThinkLevel,
} = params.modelSelection;
let sessionEntry = params.sessionEntry;
let lifecycleGeneration = params.lifecycleGeneration;
const sessionEffectsSource = resolveInternalSessionEffectsSource({
agentId: sessionAgentId,
sessionId,
sessionKey,
storePath,
});
const internalSessionTarget = params.suppressVisibleSessionEffects
? await prepareInternalSessionEffectsSession({
agentId: sessionAgentId,
cwd: cwd ?? workspaceDir,
runId,
source: sessionEffectsSource,
storePath,
})
: undefined;
params.trackInternalModelRunTarget(internalSessionTarget);
const attemptSessionTarget =
internalSessionTarget ??
(sessionKey && storePath
? {
agentId: sessionAgentId,
sessionId,
sessionKey,
storePath,
}
: undefined);
const attemptSessionFile = internalSessionTarget?.sessionFile ?? sessionFile;
const startedAt = Date.now();
const attemptLifecycleState: AgentAttemptLifecycleState = {
currentTurnUserMessagePersisted: false,
lifecycleFinishing: false,
lifecycleEnded: false,
};
const attemptLifecycleCallbacks = createAgentAttemptLifecycleCallbacks(attemptLifecycleState);
const transcriptMedia = params.opts.transcriptMedia ?? [];
const hasTranscriptMedia = transcriptMedia.length > 0;
const suppressUserTurnPersistence =
params.opts.suppressPromptPersistence === true ||
(params.opts.transcriptMessage === "" && !hasTranscriptMedia);
const recorderTranscriptText = transcriptBody || undefined;
const userTurnTranscriptRecorder = createUserTurnTranscriptRecorder({
...(!suppressUserTurnPersistence && (recorderTranscriptText || hasTranscriptMedia)
? {
input: {
text: recorderTranscriptText,
...(hasTranscriptMedia ? { media: transcriptMedia } : {}),
},
}
: {}),
target: {
sessionId: internalSessionTarget?.sessionId ?? sessionId,
agentId: internalSessionTarget?.agentId ?? sessionAgentId,
sessionKey: internalSessionTarget?.sessionKey ?? sessionKey ?? sessionId,
sessionEntry: internalSessionTarget?.sessionEntry ?? sessionEntry,
sessionStore: params.suppressVisibleSessionEffects ? undefined : sessionStore,
storePath: internalSessionTarget?.storePath ?? storePath,
cwd: cwd ?? workspaceDir,
config: cfg,
},
beforeMessageWrite: runAgentHarnessBeforeMessageWriteHook,
errorContext: "agent command user turn transcript",
});
if (suppressUserTurnPersistence) {
userTurnTranscriptRecorder.markBlocked();
}
const lifecycle = createAgentCommandLifecycle({
runId,
lifecycleGeneration: () => lifecycleGeneration,
startedAt,
abortSignal: params.opts.abortSignal,
state: attemptLifecycleState,
});
const attemptExecutionRuntime = await loadAttemptExecutionRuntime();
const messageChannel = resolveMessageChannel(
runContext.messageChannel,
params.opts.replyChannel ?? params.opts.channel,
);
let result: AgentAttemptResult;
let fallbackProvider = provider;
let fallbackModel = model;
let fallbackExhausted = false;
let terminal: EmbeddedAgentRunEntryTerminal;
let liveSwitchRetries = 0;
let autoFallbackPrimaryProbeInterruptedByLiveSwitch = false;
const fastModeStartedAtMs = Date.now();
const fallbackTrajectoryRecorder = createTrajectoryRuntimeRecorder({
cfg,
runId,
sessionId,
sessionKey,
sessionFile: attemptSessionFile,
provider,
modelId: model,
workspaceDir,
});
let liveSwitchMediaTaskIds: ReadonlySet<string> = new Set();
for (;;) {
try {
liveSwitchMediaTaskIds = sessionKey
? getGeneratedMediaTaskIdsForSessionKey(sessionKey)
: new Set<string>();
const spawnedBy = normalizedSpawned.spawnedBy ?? sessionEntry?.spawnedBy;
const effectiveFallbacksOverride = isModelSelectionLocked(sessionEntry)
? []
: (params.opts.modelFallbacksOverride ??
resolveEffectiveModelFallbacks({
cfg,
agentId: sessionAgentId,
sessionKey,
hasSessionModelOverride:
hasExplicitRunOverride || Boolean(storedProviderOverride || storedModelOverride),
modelOverrideSource: hasExplicitRunOverride ? "user" : storedModelOverrideSource,
hasAutoFallbackProvenance: hasExplicitRunOverride
? false
: hasStoredAutoFallbackProvenance,
}));
const fallbackRuntimeState: { originRuntime?: "cli" | "embedded" } = {};
attemptLifecycleState.currentTurnUserMessagePersisted = false;
let attemptMediaTaskIds = liveSwitchMediaTaskIds;
const currentAttemptCommittedCronMedia = () =>
Boolean(
sessionKey && hasNewGeneratedMediaTaskForSessionKey(sessionKey, attemptMediaTaskIds),
);
const fallbackResult = await runEmbeddedAgentEntry<AgentAttemptResult>({
selection: {
cfg,
provider,
model,
requestedRouteResolution: params.modelSelection.requestedRouteResolution,
agentDir,
fallbacksOverride: effectiveFallbacksOverride,
...modelManifestContext,
},
identity: {
runId,
agentId: sessionAgentId,
sessionId,
sessionKey: sessionKey ?? sessionId,
},
harness: {
workspaceDir,
sessionKey,
preparation: { kind: "direct" },
resolveRuntimeOverride: (candidateProvider) =>
resolveSessionRuntimeOverrideForProvider({
provider: candidateProvider,
entry: sessionEntryForAttempt,
cfg,
}),
},
behavior: {
kind: "command-rpc",
hasCommittedSideEffect: currentAttemptCommittedCronMedia,
},
sessionOverride: {
kind: "reconcile-completed",
reconcile: async ({ provider: winnerProvider, model: winnerModel }) => {
if (
!autoFallbackPrimaryProbe ||
autoFallbackPrimaryProbeInterruptedByLiveSwitch ||
!sessionEntry ||
!sessionStore ||
!sessionKey ||
isModelSelectionLocked(sessionEntry) ||
params.suppressVisibleSessionEffects ||
params.preserveUserFacingSessionModelState ||
!entryMatchesAutoFallbackPrimaryProbe(sessionEntry, autoFallbackPrimaryProbe) ||
winnerProvider !== autoFallbackPrimaryProbe.provider ||
winnerModel !== autoFallbackPrimaryProbe.model
) {
return;
}
const nextSessionEntry = { ...sessionEntry };
clearAutoFallbackPrimaryProbeSelection(nextSessionEntry);
sessionEntry = await persistSessionEntry({
sessionStore,
sessionKey,
storePath,
initialEntry: sessionEntry,
entry: nextSessionEntry,
shouldPersist: (current) =>
Boolean(
current &&
entryMatchesAutoFallbackPrimaryProbe(current, autoFallbackPrimaryProbe),
),
});
},
},
abortSignal: params.opts.abortSignal,
onFallbackStep: (step) => {
fallbackTrajectoryRecorder?.recordEvent("model.fallback_step", step);
},
runCandidate: async (providerOverride, modelOverride, runOptions) => {
attemptMediaTaskIds = sessionKey
? getGeneratedMediaTaskIdsForSessionKey(sessionKey)
: new Set<string>();
attemptLifecycleState.lifecycleError = undefined;
attemptLifecycleState.lifecycleFinishing = false;
attemptLifecycleState.lifecycleEnded = false;
const isAutoFallbackPrimaryProbeCandidate =
autoFallbackPrimaryProbe &&
providerOverride === autoFallbackPrimaryProbe.provider &&
modelOverride === autoFallbackPrimaryProbe.model;
const attemptSessionEntry =
autoFallbackPrimaryProbe &&
providerOverride === autoFallbackPrimaryProbe.fallbackProvider &&
!isAutoFallbackPrimaryProbeCandidate
? sessionEntry
: sessionEntryForAttempt;
if (isAutoFallbackPrimaryProbeCandidate) {
markAutoFallbackPrimaryProbe({ probe: autoFallbackPrimaryProbe, sessionKey });
}
await params.opts.onActiveModelSelected?.({
provider: providerOverride,
model: modelOverride,
});
const fastModeState = resolveFastModeState({
cfg,
provider: providerOverride,
model: modelOverride,
agentId: sessionAgentId,
sessionEntry,
});
const fastMode = params.opts.fastMode ?? fastModeState.mode;
const configuredAuthProfileId =
providerOverride === defaultProvider && modelOverride === defaultModel
? configuredDefaultAuthProfileId
: undefined;
const agentHarnessRuntimeOverride = resolveSessionRuntimeOverrideForProvider({
provider: providerOverride,
entry: attemptSessionEntry,
cfg,
});
const candidateRuntime = resolveEffectiveAgentRuntime({
cfg,
provider: providerOverride,
modelId: modelOverride,
agentId: sessionAgentId,
sessionKey,
sessionEntry: attemptSessionEntry,
});
const candidateRequestedThinkLevel =
immutableThinkLevel ??
resolveThinkingDefault({
cfg,
provider: providerOverride,
model: modelOverride,
catalog: thinkingCatalog,
agentRuntime: candidateRuntime,
});
const candidateThinkLevel =
resolveCandidateThinkingLevel({
cfg,
provider: providerOverride,
modelId: modelOverride,
level: candidateRequestedThinkLevel,
catalog: thinkingCatalog,
agentId: sessionAgentId,
sessionKey,
sessionEntry: attemptSessionEntry,
agentRuntime: candidateRuntime,
}) ?? candidateRequestedThinkLevel;
effectiveTurnThinkLevel = candidateThinkLevel;
return attemptExecutionRuntime.runAgentAttempt({
providerOverride,
modelOverride,
configuredAuthProfileId,
modelFallbacksOverride: effectiveFallbacksOverride,
originalProvider: provider,
cfg,
sessionEntry: attemptSessionEntry,
agentHarnessRuntimeOverride,
sessionId,
sessionKey,
...(attemptSessionTarget ? { sessionTarget: attemptSessionTarget } : {}),
sessionAgentId,
sessionFile: attemptSessionFile,
workspaceDir,
cwd,
body,
transcriptBody,
isFallbackRetry: runOptions.isFallbackRetry,
resolvedThinkLevel: candidateThinkLevel,
fastMode,
fastModeStartedAtMs,
fastModeAutoOnSeconds:
fastMode === "auto"
? (params.opts.fastModeAutoOnSeconds ?? fastModeState.fastAutoOnSeconds)
: fastModeState.fastAutoOnSeconds,
isFinalFallbackAttempt: runOptions?.isFinalFallbackAttempt,
timeoutMs,
runTimeoutOverrideMs,
runId,
lifecycleGeneration,
opts: params.opts,
runContext,
spawnedBy,
messageChannel,
skillsSnapshot,
resolvedVerboseLevel,
agentDir,
authProfileProvider: providerForAuthProfileValidation,
sessionStore: params.suppressVisibleSessionEffects ? undefined : sessionStore,
storePath: params.suppressVisibleSessionEffects ? undefined : storePath,
pluginsEnabled,
...(manifestMetadataSnapshot ? { metadataSnapshot: manifestMetadataSnapshot } : {}),
allowTransientCooldownProbe: runOptions?.allowTransientCooldownProbe,
sessionHasHistory:
!isNewSession ||
(await attemptExecutionRuntime.sessionFileHasContent(attemptSessionFile)),
fallbackRuntimeState,
suppressPromptPersistenceOnRetry:
suppressUserTurnPersistence ||
userTurnTranscriptRecorder.hasPersisted() ||
userTurnTranscriptRecorder.isBlocked() ||
(runOptions.isFallbackRetry && attemptLifecycleState.currentTurnUserMessagePersisted),
userTurnTranscriptRecorder,
onUserMessagePersisted: attemptLifecycleCallbacks.onUserMessagePersisted,
onLifecycleGenerationChanged: (nextLifecycleGeneration) => {
lifecycleGeneration = nextLifecycleGeneration;
// Outer cleanup owns the run context, so publish before the attempt can reject.
params.onLifecycleGenerationChanged(nextLifecycleGeneration);
},
onAgentEvent: attemptLifecycleCallbacks.onAgentEvent,
deferTerminalLifecycle: true,
});
},
});
result = fallbackResult.result;
terminal = fallbackResult.terminal;
if (isAgentRunRestartAbortReason(params.opts.abortSignal?.reason)) {
throw params.opts.abortSignal?.reason;
}
fallbackProvider = fallbackResult.provider;
fallbackModel = fallbackResult.model;
fallbackExhausted = fallbackResult.outcome === "exhausted";
await fallbackResult.settleSessionOverride();
if (fallbackResult.attempts.length > 0 && result.meta.agentMeta) {
result = {
...result,
meta: {
...result.meta,
agentMeta: {
...result.meta.agentMeta,
fallbackAttempts: fallbackResult.attempts,
},
},
};
}
if (!fallbackExhausted) {
lifecycle.emitFinishing(terminal);
}
break;
} catch (err) {
if (err instanceof LiveSessionModelSwitchError) {
if (isModelSelectionLocked(sessionEntry)) {
if (!attemptLifecycleState.lifecycleEnded) {
emitAgentEvent({
runId,
lifecycleGeneration,
stream: "lifecycle",
data: {
phase: "error",
startedAt,
endedAt: Date.now(),
error: MODEL_SELECTION_LOCKED_MESSAGE,
},
});
}
await fallbackTrajectoryRecorder?.flush();
throw new ModelSelectionLockedError();
}
if (
sessionKey &&
hasNewGeneratedMediaTaskForSessionKey(sessionKey, liveSwitchMediaTaskIds)
) {
throw err;
}
liveSwitchRetries += 1;
if (liveSwitchRetries > MAX_LIVE_SWITCH_RETRIES) {
log.error(
`Live session model switch in subagent run ${runId}: exceeded maximum retries (${MAX_LIVE_SWITCH_RETRIES})`,
);
if (!attemptLifecycleState.lifecycleEnded) {
emitAgentEvent({
runId,
lifecycleGeneration,
stream: "lifecycle",
data: {
phase: "error",
startedAt,
endedAt: Date.now(),
error: "Agent run failed",
},
});
}
await fallbackTrajectoryRecorder?.flush();
throw new Error(
`Exceeded maximum live model switch retries (${MAX_LIVE_SWITCH_RETRIES})`,
{
cause: err,
},
);
}
const switchRef = normalizeAgentCommandModelRef(
cfg,
err.provider,
err.model,
modelManifestContext,
);
if (!visibilityPolicy.allowsKey(modelKey(switchRef.provider, switchRef.model))) {
log.info(
`Live session model switch in subagent run ${runId}: ` +
`rejected ${sanitizeForLog(err.provider)}/${sanitizeForLog(err.model)} (not in allowlist)`,
);
if (!attemptLifecycleState.lifecycleEnded) {
emitAgentEvent({
runId,
lifecycleGeneration,
stream: "lifecycle",
data: {
phase: "error",
startedAt,
endedAt: Date.now(),
error: "Agent run failed",
},
});
}
await fallbackTrajectoryRecorder?.flush();
throw new Error(
`Live model switch rejected: ${sanitizeForLog(err.provider)}/${sanitizeForLog(err.model)} is not in the agent allowlist`,
{ cause: err },
);
}
const previousProvider = provider;
const previousModel = model;
if (autoFallbackPrimaryProbe) {
autoFallbackPrimaryProbeInterruptedByLiveSwitch = true;
}
provider = err.provider;
model = err.model;
fallbackProvider = err.provider;
fallbackModel = err.model;
providerForAuthProfileValidation = err.provider;
if (sessionEntry) {
sessionEntry = { ...sessionEntry };
if (err.agentRuntimeOverride) {
sessionEntry.agentRuntimeOverride = err.agentRuntimeOverride;
} else {
delete sessionEntry.agentRuntimeOverride;
}
sessionEntry.authProfileOverride = err.authProfileId;
sessionEntry.authProfileOverrideSource = err.authProfileId
? err.authProfileIdSource
: undefined;
sessionEntry.authProfileOverrideCompactionCount = undefined;
sessionEntryForAttempt = sessionEntry;
}
if (
storedModelOverride ||
err.model !== previousModel ||
err.provider !== previousProvider
) {
storedModelOverride = err.model;
storedModelOverrideSource = "user";
}
attemptLifecycleState.lifecycleEnded = false;
log.info(
`Live session model switch in subagent run ${runId}: switching to ${sanitizeForLog(err.provider)}/${sanitizeForLog(err.model)}`,
);
continue;
}
if (!attemptLifecycleState.lifecycleEnded) {
const errorLifecycleFields = isAgentRunDirectAbortReason(err)
? { aborted: true as const, stopReason: "aborted" as const }
: resolveAgentRunErrorLifecycleFields(err, params.opts.abortSignal);
emitAgentEvent({
runId,
lifecycleGeneration,
stream: "lifecycle",
data: {
phase: "error",
startedAt,
endedAt: Date.now(),
error: err instanceof Error ? err.message : "Agent run failed",
...errorLifecycleFields,
},
});
}
await fallbackTrajectoryRecorder?.flush();
throw err;
}
}
return {
result,
fallbackProvider,
fallbackModel,
fallbackExhausted,
provider,
model,
sessionEntry,
lifecycleGeneration,
effectiveTurnThinkLevel,
internalSessionTarget,
attemptExecutionRuntime,
messageChannel,
suppressUserTurnPersistence,
userTurnTranscriptRecorder,
fallbackTrajectoryRecorder,
lifecycle,
terminal,
};
}
export type EmbeddedAgentAttempt = Awaited<ReturnType<typeof runEmbeddedAgentAttempt>>;