maka-agent 0.2.0-dev.35.20260915 → 0.2.0-dev.36.20260915

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.
Files changed (41) hide show
  1. package/native/runtime-host-windows-task-launcher/prebuilds/win32-x64/maka-runtime-host-task-launcher.exe +0 -0
  2. package/node_modules/@maka/core/dist/model-metadata.generated.js +175 -116
  3. package/node_modules/@maka/core/dist/models-dev-projection.js +1 -0
  4. package/node_modules/@maka/core/dist/provider-registry.js +17 -0
  5. package/node_modules/@maka/eval/dist/harbor-maka-subject.js +13 -3
  6. package/node_modules/@maka/eval/dist/maka-artifacts.js +28 -14
  7. package/node_modules/@maka/eval/dist/maka-subject.js +28 -3
  8. package/node_modules/@maka/eval/harbor/relay_agent.py +42 -2
  9. package/node_modules/@maka/runtime/dist/ai-sdk-backend.js +0 -1
  10. package/node_modules/@maka/runtime/dist/ai-sdk-compaction.js +40 -53
  11. package/node_modules/@maka/runtime/dist/ai-sdk-turn.js +26 -29
  12. package/node_modules/@maka/runtime/dist/builtin-tools.js +20 -12
  13. package/node_modules/@maka/runtime/dist/runtime-kernel.js +1 -0
  14. package/node_modules/@maka/runtime/dist/telemetry/model-pricing.generated.js +144 -77
  15. package/node_modules/@maka/runtime/node_modules/@ai-sdk/code-mode/dist/run-code-mode.js +6 -19
  16. package/node_modules/@maka/runtime-host/dist/client/connect-or-spawn.js +24 -3
  17. package/node_modules/@maka/runtime-host/dist/client/connection.js +20 -2
  18. package/node_modules/@maka/runtime-host/dist/client/hosted-execution-target.js +27 -6
  19. package/node_modules/@maka/runtime-host/dist/client/hosted-execution.js +10 -28
  20. package/node_modules/@maka/runtime-host/dist/client/startup-error.js +1 -1
  21. package/node_modules/@maka/runtime-host/dist/execution-candidate-main.js +8 -1
  22. package/node_modules/@maka/runtime-host/dist/protocol/agent-graph.js +1 -11
  23. package/node_modules/@maka/runtime-host/dist/protocol/client-capability.js +2 -2
  24. package/node_modules/@maka/runtime-host/dist/protocol/codec.js +11 -0
  25. package/node_modules/@maka/runtime-host/dist/protocol/index.js +1 -1
  26. package/node_modules/@maka/runtime-host/dist/protocol/workhub-coordination.js +10 -2
  27. package/node_modules/@maka/runtime-host/dist/server/bootstrap-runtime-policy.js +65 -0
  28. package/node_modules/@maka/runtime-host/dist/server/execution-candidate.js +4 -1
  29. package/node_modules/@maka/runtime-host/dist/server/execution-composition-factory.js +9 -3
  30. package/node_modules/@maka/runtime-host/dist/server/execution-composition.js +11 -0
  31. package/node_modules/@maka/runtime-host/dist/server/hosted-execution-tool-profile.js +3 -15
  32. package/node_modules/@maka/runtime-host/dist/server/interactive-run-composer.js +9 -4
  33. package/node_modules/@maka/runtime-host/dist/server/message-coordinator.js +1 -2
  34. package/node_modules/@maka/runtime-host/dist/server/session-catalog-coordinator.js +4 -1
  35. package/node_modules/@maka/storage/dist/agent-run-store-contract.js +11 -8
  36. package/node_modules/@maka/storage/dist/agent-run-store.js +1 -1
  37. package/node_modules/@maka/storage/dist/execution-stores.js +1 -1
  38. package/node_modules/@maka/storage/dist/message-admission-store.js +5 -3
  39. package/node_modules/@maka/storage/dist/sqlite-session-metadata-store.js +14 -6
  40. package/package.json +1 -1
  41. package/node_modules/@maka/eval/dist/maka-runtime-policy.js +0 -42
@@ -42,6 +42,7 @@ export const MODELS_DEV_PROVIDERS = {
42
42
  'minimax-coding-plan': 'minimax-coding-plan',
43
43
  mistral: 'mistral',
44
44
  moonshot: 'moonshotai-cn',
45
+ 'moonshot-global': 'moonshotai',
45
46
  nvidia: 'nvidia',
46
47
  'ollama-cloud': 'ollama-cloud',
47
48
  openai: 'openai',
@@ -400,6 +400,10 @@ const moonshotModelIds = toolCallingModelIds('Moonshot', GENERATED_MODELS_DEV_ME
400
400
  'kimi-k2.6',
401
401
  'kimi-k2.7-code',
402
402
  ]).filter((id) => GENERATED_MODELS_DEV_METADATA.moonshot[id]?.lifecycle !== 'deprecated');
403
+ const moonshotGlobal = GENERATED_MODELS_DEV_PROVIDER_FACTS['moonshot-global'];
404
+ if (!moonshotGlobal.api)
405
+ throw new Error('models.dev Moonshot Global provider is missing its API');
406
+ const moonshotGlobalModelIds = toolCallingModelIds('Moonshot Global', GENERATED_MODELS_DEV_METADATA['moonshot-global'], ['kimi-k3']).filter((id) => GENERATED_MODELS_DEV_METADATA['moonshot-global'][id]?.lifecycle !== 'deprecated');
403
407
  const cloudflareWorkersAi = GENERATED_MODELS_DEV_PROVIDER_FACTS['cloudflare-workers-ai'];
404
408
  if (cloudflareWorkersAi.id !== 'cloudflare-workers-ai') {
405
409
  throw new Error('models.dev Cloudflare Workers AI provider facts are missing stable id cloudflare-workers-ai');
@@ -656,6 +660,19 @@ const providerRegistry = {
656
660
  signupUrl: 'https://platform.kimi.com/console/api-keys',
657
661
  catalogOrder: 4,
658
662
  },
663
+ 'moonshot-global': {
664
+ label: 'Moonshot Global',
665
+ baseUrl: moonshotGlobal.api,
666
+ authKind: 'api_key',
667
+ fallbackModels: moonshotGlobalModelIds,
668
+ status: 'ready',
669
+ runtimeAdapter: { kind: 'openai', apiProtocol: 'openai-responses' },
670
+ modelDiscovery: { kind: 'protocol' },
671
+ category: 'overseas',
672
+ catalogGroup: 'api',
673
+ signupUrl: 'https://platform.kimi.ai/console/api-keys',
674
+ catalogOrder: 4.1,
675
+ },
659
676
  'zai-coding-plan': {
660
677
  label: 'Z.AI Coding Plan',
661
678
  menuLabel: 'Z.AI',
@@ -16,11 +16,10 @@
16
16
  * specific language governing permissions and limitations
17
17
  * under the License.
18
18
  */
19
- import { mkdir, writeFile } from 'node:fs/promises';
19
+ import { mkdir } from 'node:fs/promises';
20
20
  import { dirname, join } from 'node:path';
21
21
  import { runHostedExecution } from '@maka/runtime-host/client';
22
22
  import { captureMakaRuntimeArtifacts, writeMakaArtifactCollectionError } from './maka-artifacts.js';
23
- import { makaEvalRuntimePolicyDocument } from './maka-runtime-policy.js';
24
23
  import { takeRelayResultToken, writeRelayResult } from './relay-result-frame.js';
25
24
  const resultToken = takeRelayResultToken();
26
25
  const payload = JSON.parse(Buffer.from(process.argv[2] ?? '', 'base64url').toString());
@@ -50,12 +49,23 @@ process.once('SIGTERM', stop);
50
49
  const runtimeHome = join(dirname(payload.rootPath), `${process.pid}-home`);
51
50
  await mkdir(payload.rootPath, { recursive: true });
52
51
  await mkdir(runtimeHome, { recursive: true, mode: 0o700 });
53
- await writeFile(join(payload.rootPath, 'runtime-policy.json'), `${JSON.stringify(makaEvalRuntimePolicyDocument(process.env.HTTPS_PROXY))}\n`, { flag: 'wx', mode: 0o600 });
54
52
  process.env.HOME = runtimeHome;
55
53
  process.env.DEEPSEEK_BASE_URL = payload.baseUrl;
56
54
  let result;
57
55
  try {
58
56
  result = await runHostedExecution({
57
+ initialization: {
58
+ incognito: true,
59
+ ...(process.env.HTTPS_PROXY ? { proxyUrl: process.env.HTTPS_PROXY } : {}),
60
+ },
61
+ ...(payload.connection
62
+ ? {
63
+ connection: {
64
+ providerType: payload.connection.providerType,
65
+ apiKey: process.env[payload.connection.apiKeyEnvironment] ?? '',
66
+ },
67
+ }
68
+ : {}),
59
69
  rootPath: payload.rootPath,
60
70
  baseUrl: payload.baseUrl,
61
71
  execution: payload.execution,
@@ -33,24 +33,33 @@ export async function captureMakaRuntimeArtifacts(input) {
33
33
  await mkdir(stagingRoot, { recursive: true, mode: 0o700 });
34
34
  const sourcePath = join(stateRoot, 'runtime.sqlite');
35
35
  const destinationPath = join(stagingRoot, 'runtime.sqlite');
36
- const source = new DatabaseSync(sourcePath, { readOnly: true });
37
36
  try {
38
- await backup(source, destinationPath);
39
- }
40
- finally {
41
- source.close();
42
- }
43
- const snapshot = new DatabaseSync(destinationPath);
44
- try {
45
- snapshot.exec('PRAGMA wal_checkpoint(TRUNCATE)');
46
- snapshot.exec('PRAGMA journal_mode=DELETE');
37
+ await stat(sourcePath);
38
+ const source = new DatabaseSync(sourcePath, { readOnly: true });
39
+ try {
40
+ await backup(source, destinationPath);
41
+ }
42
+ finally {
43
+ source.close();
44
+ }
45
+ const snapshot = new DatabaseSync(destinationPath);
46
+ try {
47
+ snapshot.exec('PRAGMA wal_checkpoint(TRUNCATE)');
48
+ snapshot.exec('PRAGMA journal_mode=DELETE');
49
+ }
50
+ finally {
51
+ snapshot.close();
52
+ }
53
+ await chmod(destinationPath, 0o600);
47
54
  }
48
- finally {
49
- snapshot.close();
55
+ catch (error) {
56
+ await rm(destinationPath, { force: true });
57
+ if (error.code !== 'ENOENT') {
58
+ await writeMakaArtifactCollectionError(stagingRoot, error);
59
+ }
50
60
  }
51
61
  await rm(`${destinationPath}-wal`, { force: true });
52
62
  await rm(`${destinationPath}-shm`, { force: true });
53
- await chmod(destinationPath, 0o600);
54
63
  for (const name of ['runtime-host-candidate.log', 'runtime-policy.json']) {
55
64
  const source = join(stateRoot, name);
56
65
  try {
@@ -66,7 +75,12 @@ export async function captureMakaRuntimeArtifacts(input) {
66
75
  throw error;
67
76
  }
68
77
  }
69
- const files = await Promise.all(['runtime.sqlite', 'runtime-host-candidate.log', 'runtime-policy.json'].map((name) => describeFile(join(stagingRoot, name), name)));
78
+ const files = await Promise.all([
79
+ 'runtime.sqlite',
80
+ 'runtime-host-candidate.log',
81
+ 'runtime-policy.json',
82
+ 'collection-error.json',
83
+ ].map((name) => describeFile(join(stagingRoot, name), name)));
70
84
  const manifest = {
71
85
  schemaVersion: 'maka.eval.runtime_artifacts.v1',
72
86
  reason: input.reason,
@@ -17,6 +17,8 @@
17
17
  * under the License.
18
18
  */
19
19
  import { randomUUID } from 'node:crypto';
20
+ import { PROVIDER_REGISTRY } from '@maka/core/llm-connections';
21
+ import { isThinkingLevel } from '@maka/core/model-thinking';
20
22
  import { isSessionToolProfile } from '@maka/core/session';
21
23
  import { decodeHostedExecutionProjection } from '@maka/runtime-host/protocol';
22
24
  import { MAKA_RUNTIME_ARTIFACT_PATH, MAKA_SUBJECT_STDERR_PATH, MAKA_SUBJECT_STDOUT_PATH, } from './maka-artifacts.js';
@@ -24,7 +26,14 @@ import { deepSeekCostUsd } from './provider-metering.js';
24
26
  export function createMakaSubjectAdapter() {
25
27
  return {
26
28
  kind: 'maka',
27
- validate: (cell) => decodeConfig(cell.subject.config),
29
+ validate: (cell) => {
30
+ const config = decodeConfig(cell.subject.config);
31
+ if (config.apiKeyEnvironment &&
32
+ !cell.subject.credentials.includes(config.apiKeyEnvironment)) {
33
+ throw new Error('Maka apiKeyEnvironment must name a declared subject credential');
34
+ }
35
+ return config;
36
+ },
28
37
  async execute({ cell, context }) {
29
38
  const config = decodeConfig(cell.subject.config);
30
39
  const executionId = randomUUID();
@@ -37,7 +46,7 @@ export function createMakaSubjectAdapter() {
37
46
  connectionSlug: config.connectionSlug,
38
47
  model: config.model,
39
48
  },
40
- thinkingLevel: config.thinkingLevel,
49
+ ...(config.thinkingLevel === undefined ? {} : { thinkingLevel: config.thinkingLevel }),
41
50
  permissionMode: config.permissionMode,
42
51
  collaborationMode: config.collaborationMode,
43
52
  orchestrationMode: config.orchestrationMode,
@@ -50,6 +59,14 @@ export function createMakaSubjectAdapter() {
50
59
  rootPath: `${config.runtimeHostsPath}/${executionId}`,
51
60
  artifactRoot: MAKA_RUNTIME_ARTIFACT_PATH,
52
61
  baseUrl: config.baseUrl,
62
+ ...(config.providerType
63
+ ? {
64
+ connection: {
65
+ providerType: config.providerType,
66
+ apiKeyEnvironment: config.apiKeyEnvironment,
67
+ },
68
+ }
69
+ : {}),
53
70
  hostSettlementTimeoutMs: config.hostSettlementTimeoutMs,
54
71
  execution: input,
55
72
  })).toString('base64url');
@@ -219,7 +236,8 @@ function decodeConfig(value) {
219
236
  'baseUrl',
220
237
  'connectionSlug',
221
238
  'model',
222
- 'thinkingLevel',
239
+ ...(Object.hasOwn(value, 'providerType') ? ['providerType', 'apiKeyEnvironment'] : []),
240
+ ...(Object.hasOwn(value, 'thinkingLevel') ? ['thinkingLevel'] : []),
223
241
  'permissionMode',
224
242
  'collaborationMode',
225
243
  'orchestrationMode',
@@ -227,6 +245,13 @@ function decodeConfig(value) {
227
245
  'toolProfile',
228
246
  ];
229
247
  const config = exact(value, fields);
248
+ if (config.providerType !== undefined &&
249
+ !Object.hasOwn(PROVIDER_REGISTRY, String(config.providerType))) {
250
+ throw new Error('Maka config.providerType is invalid');
251
+ }
252
+ if (config.thinkingLevel !== undefined && !isThinkingLevel(config.thinkingLevel)) {
253
+ throw new Error('Maka config.thinkingLevel is invalid');
254
+ }
230
255
  if (!URL.canParse(String(config.baseUrl)))
231
256
  throw new Error('Maka baseUrl is invalid');
232
257
  const hostSettlementTimeoutMs = positiveInteger(config.hostSettlementTimeoutMs, 'Maka config.hostSettlementTimeoutMs');
@@ -100,6 +100,7 @@ class RelayAgent(BaseAgent):
100
100
  if not isinstance(teardown_timeout_ms, int) or teardown_timeout_ms <= 0:
101
101
  raise RuntimeError("Maka Eval teardown timeout is invalid")
102
102
  self._teardown_timeout = teardown_timeout_ms / 1000
103
+ self._subject_owner: str | None = None
103
104
 
104
105
  @staticmethod
105
106
  def name() -> str:
@@ -109,7 +110,35 @@ class RelayAgent(BaseAgent):
109
110
  return "1"
110
111
 
111
112
  async def setup(self, environment: Any) -> None:
112
- return None
113
+ identity = await environment.exec('printf "%s:%s" "$(id -u)" "$(id -g)"')
114
+ owner = str(identity.stdout or "").strip()
115
+ if identity.return_code != 0 or re.fullmatch(r"[0-9]+:[0-9]+", owner) is None:
116
+ raise RuntimeError("Maka Eval could not resolve the task user")
117
+ prepared = await environment.exec(
118
+ "mkdir -p /logs/agent /logs/artifacts && "
119
+ f"chown {owner} /logs/agent /logs/artifacts && "
120
+ "chmod 700 /logs/agent /logs/artifacts",
121
+ user="root",
122
+ )
123
+ if prepared.return_code != 0:
124
+ raise RuntimeError("Maka Eval could not prepare task-owned artifact directories")
125
+ # Storage coordinates by the system account home, not the subject's
126
+ # temporary HOME. Provision an absent home without changing task identity
127
+ # or taking ownership of an existing directory.
128
+ uid = owner.split(":", 1)[0]
129
+ home_script = (
130
+ f"set -eu; account=$(getent passwd {uid}); "
131
+ "task_home=$(printf '%s' \"$account\" | cut -d: -f6) && "
132
+ 'case "$task_home" in /*) ;; *) exit 1 ;; esac; '
133
+ 'if [ ! -e "$task_home" ] && [ ! -L "$task_home" ]; then '
134
+ f'mkdir -p -- "$task_home" && chown {owner} "$task_home" && '
135
+ 'chmod 700 "$task_home" || exit 1; fi; '
136
+ 'test -d "$task_home"'
137
+ )
138
+ prepared_home = await environment.exec(home_script, user="root")
139
+ if prepared_home.return_code != 0:
140
+ raise RuntimeError("Maka Eval could not prepare the task account home")
141
+ self._subject_owner = owner
113
142
 
114
143
  async def run(self, instruction: str, environment: Any, context: Any) -> None:
115
144
  reader, writer = await asyncio.open_connection(self._host, self._port)
@@ -145,7 +174,9 @@ class RelayAgent(BaseAgent):
145
174
  raise RelayTransportClosed("Maka Eval relay transport closed before ready")
146
175
  request = await _receive(reader)
147
176
  _require_message(request, self._token, "execute")
148
- command = await _prepare_command(environment, request, self._token, scope_path)
177
+ command = await _prepare_command(
178
+ environment, request, self._token, scope_path, self._subject_owner
179
+ )
149
180
  execution = asyncio.create_task(environment.exec(command, cwd=cwd))
150
181
  decision = asyncio.create_task(_receive(reader))
151
182
  done, _ = await asyncio.wait({execution, decision}, return_when=asyncio.FIRST_COMPLETED)
@@ -278,6 +309,7 @@ async def _prepare_command(
278
309
  request: dict[str, Any],
279
310
  token: str,
280
311
  scope_path: str,
312
+ subject_owner: str | None = None,
281
313
  ) -> str:
282
314
  credentials = request.get("credentials")
283
315
  public_environment = request.get("environment", {})
@@ -317,6 +349,14 @@ async def _prepare_command(
317
349
  }.items():
318
350
  secret.write(f"export {key}={shlex.quote(value)}\n")
319
351
  await environment.upload_file(secret_path, container_path)
352
+ if subject_owner is not None:
353
+ prepared = await environment.exec(
354
+ f"chown {subject_owner} {shlex.quote(container_path)} && "
355
+ f"chmod 600 {shlex.quote(container_path)}",
356
+ user="root",
357
+ )
358
+ if prepared.return_code != 0:
359
+ raise RuntimeError("Maka Eval could not prepare task-owned credentials")
320
360
  finally:
321
361
  if secret_path is not None:
322
362
  secret_path.unlink(missing_ok=True)
@@ -162,7 +162,6 @@ export class AiSdkBackend {
162
162
  targetConnectionId: input.header.llmConnectionId,
163
163
  targetProviderStateIdentity: input.providerStateIdentity,
164
164
  now: this.now,
165
- modelAdapter: this.modelAdapter,
166
165
  createProviderRequestTracker: (trackerInput) => this.providerTelemetry.createTracker(trackerInput),
167
166
  materializeRuntimeReplayPlan: (plan, imageBudget, checkpoint, providerReasoningReplayEventIds) => this.messageProjection.materializeRuntimeReplayPlan(plan, imageBudget, checkpoint, providerReasoningReplayEventIds),
168
167
  canReplayProviderNative: (plan) => this.messageProjection.canReplayProviderNative(plan),
@@ -35,7 +35,6 @@ export class AiSdkCompaction {
35
35
  targetConnectionId;
36
36
  targetProviderStateIdentity;
37
37
  now;
38
- modelAdapter;
39
38
  createProviderRequestTracker;
40
39
  materializeRuntimeReplayPlan;
41
40
  canReplayProviderNative;
@@ -52,7 +51,6 @@ export class AiSdkCompaction {
52
51
  this.targetConnectionId = deps.targetConnectionId;
53
52
  this.targetProviderStateIdentity = deps.targetProviderStateIdentity;
54
53
  this.now = deps.now;
55
- this.modelAdapter = deps.modelAdapter;
56
54
  this.createProviderRequestTracker = deps.createProviderRequestTracker;
57
55
  this.materializeRuntimeReplayPlan = deps.materializeRuntimeReplayPlan;
58
56
  this.canReplayProviderNative = deps.canReplayProviderNative;
@@ -556,7 +554,7 @@ export class AiSdkCompaction {
556
554
  * the next request neither describes (tool results, user text, images
557
555
  * appended since) is judged by the provider when the request goes out. This
558
556
  * hook never terminates the turn: every failure fails open with a diagnostic
559
- * and the request is sent; a rejection is recovered by one reactive fold
557
+ * and the request is sent; a rejection is recovered by a reactive fold
560
558
  * (#4559).
561
559
  */
562
560
  buildMidTurnCapacityCompactProjection(turnId, state, queue, providerTools, onDiagnosticPatch, origin, memoryCompactionDecision, onMemoryCompaction, abortSignal) {
@@ -616,10 +614,9 @@ export class AiSdkCompaction {
616
614
  const overWindow = state.capacity !== undefined &&
617
615
  state.baselineTokens !== undefined &&
618
616
  state.baselineTokens + state.replyReserveTokens >= state.capacity;
619
- if (!overWindow || state.compactionAttemptedThisSend) {
617
+ if (!overWindow) {
620
618
  return keepProjection();
621
619
  }
622
- state.compactionAttemptedThisSend = true;
623
620
  const activeToolsForStep = options.resolveDispatch(options.activeTools).activeTools;
624
621
  // Fold a safe completed prefix of the durable turn ledger into a
625
622
  // replacement projection (validate → persist), shared with the reactive
@@ -642,13 +639,12 @@ export class AiSdkCompaction {
642
639
  }
643
640
  // The fold replaced the request; the baseline described the old one.
644
641
  // The next accepted request is the first measurement of the new shape.
645
- state.compactionAppliedThisSend = true;
646
642
  state.baselineTokens = undefined;
647
643
  acceptedProjection = {
648
644
  sourceSignatures: incomingMessages.map(modelMessageSignature),
649
645
  projectedMessages: outcome.replacementMessages,
650
646
  };
651
- state.replacedStepNumber = options.stepNumber;
647
+ state.stepShaping.add('fold');
652
648
  onDiagnosticPatch(buildActiveRequestCompactionDiagnosticPatch({
653
649
  checkpoint: outcome.checkpoint,
654
650
  estimatedTokensBefore: outcome.estimatedTokensBefore,
@@ -664,10 +660,10 @@ export class AiSdkCompaction {
664
660
  * messages — the compaction core shared by the proactive projection stage
665
661
  * (issue #882 PR 1) and the reactive overflow recovery (PR 2). It waits for
666
662
  * the seq-ack durability boundary, reads the ledger, plans the fold, then
667
- * validates (materializable ∧ smaller than the reference request ∧
668
- * replay-admissible) and persists BEFORE returning the replacement, so a
669
- * recovery re-projection never re-injects a covered raw span. It only shapes:
670
- * the pass/terminate verdict and the diagnostic emission are the caller's.
663
+ * validates (materializable ∧ replay-admissible) and persists BEFORE
664
+ * returning the replacement, so a recovery re-projection never re-injects a
665
+ * covered raw span. It only shapes: the pass/terminate verdict and the
666
+ * diagnostic emission are the caller's.
671
667
  */
672
668
  async compactActiveRequestHistory(input) {
673
669
  const { turnId, state, queue, activeToolsForStep, abortSignal } = input;
@@ -810,7 +806,8 @@ export class AiSdkCompaction {
810
806
  // next step evaluates the same condition and dispatches the same doomed
811
807
  // summarizer call: a provider that answers slowly and fails (kimi's HTTP
812
808
  // 200 with an error body) produced 15 such calls over 47 minutes before
813
- // one main request (#4634). One attempt per send, then fail open (#4559).
809
+ // one main request (#4634). The latch covers the whole Turn and only
810
+ // fail-opens; a successful fold spends no part of it (#4559).
814
811
  state.summarizerFailure = diagnosticReason;
815
812
  return {
816
813
  decision: 'fail',
@@ -860,7 +857,6 @@ export class AiSdkCompaction {
860
857
  // Memory extraction is fail-open and must never perturb Compaction.
861
858
  }
862
859
  }
863
- state.previousCheckpoint = plan.checkpoint;
864
860
  state.projectionCheckpoint = plan.checkpoint;
865
861
  return {
866
862
  decision: 'compacted',
@@ -873,36 +869,38 @@ export class AiSdkCompaction {
873
869
  /**
874
870
  * Reactive overflow recovery (issue #882 PR 2): the second line of defense.
875
871
  * When a provider rejects a request with a context-length error, fold the
876
- * durable turn ledger once and resend once — a single compact-and-retry
877
- * latch (pi's `_overflowRecoveryAttempted`). Returns the compacted messages
878
- * to resend, or undefined when recovery is impossible or already spent, in
879
- * which case the caller surfaces the real provider error rather than a
880
- * fabricated success or a locally synthesized verdict (the provider — not the
881
- * runtime — rejected the request). Non-context-length
882
- * errors and turns without the mid-turn seam never reach compaction, so the
883
- * default (no seam) behavior is already better than the old fake end_turn.
872
+ * durable turn ledger and resend — once per step: a step that a fold or an
873
+ * image omission has already reshaped gets no second recovery, so the
874
+ * resend's own rejection is terminal, while a step the provider has since
875
+ * accepted earns a fresh attempt. Returns the compacted messages to resend,
876
+ * or undefined when recovery is impossible or already spent, in which case
877
+ * the caller surfaces the real provider error rather than a fabricated
878
+ * success or a locally synthesized verdict (the provider — not the runtime —
879
+ * rejected the request). Non-context-length errors and turns without the
880
+ * mid-turn seam never reach compaction, so the default (no seam) behavior is
881
+ * already better than the old fake end_turn.
884
882
  */
885
883
  async recoverFromOverflowError(input) {
886
884
  const state = input.midTurnState;
887
- if (!state || state.compactionAttemptedThisSend || state.omittedImageToolResults.size > 0) {
885
+ if (!state || state.stepShaping.has('fold') || state.stepShaping.has('omit_images')) {
888
886
  return undefined;
889
887
  }
890
- if (this.modelAdapter.classifyError(input.error) !== 'context_overflow')
891
- return undefined;
892
888
  const eligibleImages = collectHistoricalImageToolResults(state.priorContentEvents);
893
889
  const imageOmission = omitHistoricalImageToolResults(input.currentMessages, eligibleImages);
894
890
  if (imageOmission.omittedParts > 0) {
895
- state.omittedImageToolResults = new Map([...imageOmission.omittedToolCallIds].flatMap((toolCallId) => {
891
+ state.stepShaping.add('omit_images');
892
+ for (const toolCallId of imageOmission.omittedToolCallIds) {
896
893
  const image = eligibleImages.get(toolCallId);
897
- return image ? [[toolCallId, image]] : [];
898
- }));
894
+ // Earlier steps' omissions stay omitted: every later request is
895
+ // projected through this map, so losing an entry would put its image
896
+ // back into the very history the provider just rejected.
897
+ if (image)
898
+ state.omittedImageToolResults.set(toolCallId, image);
899
+ }
899
900
  state.baselineTokens = undefined;
900
901
  return { messages: imageOmission.messages };
901
902
  }
902
903
  const phase = state.compactionPhase(input.stepNumber);
903
- // Entering the module spends the send's one attempt whether or not a fold
904
- // comes out of it; only a selected projection sets `applied`.
905
- state.compactionAttemptedThisSend = true;
906
904
  const outcome = await this.compactActiveRequestHistory({
907
905
  turnId: input.turnId,
908
906
  phase,
@@ -946,7 +944,7 @@ export class AiSdkCompaction {
946
944
  reason: 'overflow',
947
945
  }));
948
946
  // The fold replaced the request; the baseline described the rejected one.
949
- state.compactionAppliedThisSend = true;
947
+ state.stepShaping.add('fold');
950
948
  state.baselineTokens = undefined;
951
949
  return { messages: outcome.replacementMessages };
952
950
  }
@@ -1026,16 +1024,21 @@ export class MidTurnCapacityCompactState {
1026
1024
  * be accepted but leave the reply no room (#4559).
1027
1025
  */
1028
1026
  replyReserveTokens = 0;
1029
- /** Latest durable checkpoint (loaded or written) for roll-forward summaries. */
1030
- previousCheckpoint;
1027
+ /** Pre-turn checkpoint the turn starts from, before any fold of its own. */
1028
+ seedCheckpoint;
1031
1029
  /** Checkpoint accepted during this send; pins every later durable projection. */
1032
1030
  projectionCheckpoint;
1031
+ /** Latest durable checkpoint (loaded or written) for roll-forward summaries. */
1032
+ get previousCheckpoint() {
1033
+ return this.projectionCheckpoint ?? this.seedCheckpoint;
1034
+ }
1033
1035
  /**
1034
- * Step whose request the capacity hook replaced. Semantic/active-full
1035
- * compaction yields on that exact step so one step never runs two
1036
- * summarizers or double-projects.
1036
+ * What has reshaped the request since the last one a provider accepted. The
1037
+ * turn clears it at that acceptance, every shaper records itself here, and
1038
+ * every consumer reads it here — an empty set is the only proof that the next
1039
+ * request is a pure append of its predecessor.
1037
1040
  */
1038
- replacedStepNumber;
1041
+ stepShaping = new Set();
1039
1042
  /**
1040
1043
  * finish-step boundaries the event pump has flushed into the session-event
1041
1044
  * queue. The capacity hook's durability wait needs it: only after the pump
@@ -1047,22 +1050,6 @@ export class MidTurnCapacityCompactState {
1047
1050
  omittedImageToolResults = new Map();
1048
1051
  /** Malformed summaries spend one bounded repair budget for this whole Turn. */
1049
1052
  summarizerFailure;
1050
- /**
1051
- * The compaction module has been entered in this send.
1052
- *
1053
- * One attempt per send, whatever its outcome: the summarizer's own failure
1054
- * circuit already latches for the rest of the send, so a second entry would
1055
- * dispatch nothing new. This is the budget, and only the budget (#4559).
1056
- */
1057
- compactionAttemptedThisSend = false;
1058
- /**
1059
- * A folded projection was actually selected in this send.
1060
- *
1061
- * Distinct from the attempt: a fold that fails open leaves the request
1062
- * carrying its full raw history, so nothing may be concluded from a later
1063
- * rejection about what remains in it.
1064
- */
1065
- compactionAppliedThisSend = false;
1066
1053
  /**
1067
1054
  * Input tokens of the last request a provider accepted before this send.
1068
1055
  *
@@ -726,10 +726,6 @@ export class AiSdkTurn {
726
726
  contextCompactedNoteWritten = await this.recordSystemNote('context_compacted', turnId);
727
727
  }
728
728
  };
729
- // Request index (0-based) at which the active prune last rewrote the
730
- // request. A step Maka pruned is not append-only, so usage may legitimately
731
- // shrink.
732
- let pruneAppliedAtStep;
733
729
  const trace = new RunTrace({
734
730
  sessionId: this.deps.backend.sessionId,
735
731
  turnId,
@@ -888,7 +884,7 @@ export class AiSdkTurn {
888
884
  // Roll-forward seed: the latest durable checkpoint (loaded or written at
889
885
  // turn start) so a mid-turn summary only re-reads the newly folded span.
890
886
  const checkpoint = priorReplay.latestHistoryCompactCheckpoint;
891
- midTurnState.previousCheckpoint =
887
+ midTurnState.seedCheckpoint =
892
888
  checkpoint &&
893
889
  canContinueHistoryCompactCheckpointForModel(checkpoint, this.deps.backend.connection, this.deps.backend.header.llmConnectionId, this.deps.backend.modelId)
894
890
  ? checkpoint
@@ -977,7 +973,7 @@ export class AiSdkTurn {
977
973
  const pruned = await this.deps.compaction.pruneToolResults(turnEvents, turnId);
978
974
  if (pruned.stats) {
979
975
  if (pruned.stats.prunedToolResults > 0)
980
- pruneAppliedAtStep = runtimeSteps;
976
+ midTurnState?.stepShaping.add('prune');
981
977
  contextBudgetForTelemetry = addToolResultPruneStats(contextBudgetForTelemetry ?? minimalContextBudgetDiagnostic(), pruned.stats);
982
978
  }
983
979
  const projectionCheckpoint = midTurnState?.projectionCheckpoint;
@@ -1132,6 +1128,14 @@ export class AiSdkTurn {
1132
1128
  : undefined;
1133
1129
  const projectedMessages = shaped?.messages ?? contextualRequestMessages;
1134
1130
  const activeToolsForRequest = resolveDispatch(shaped?.activeTools).activeTools;
1131
+ // A finalization step resolves an empty tool set, so its request
1132
+ // legitimately drops several thousand schema tokens with no fold,
1133
+ // prune or image omission. Maka shaped that request; the provider did
1134
+ // not drop anything.
1135
+ if (lastStepActiveToolCount !== undefined &&
1136
+ activeToolsForRequest.length < lastStepActiveToolCount) {
1137
+ midTurnState?.stepShaping.add('tools');
1138
+ }
1135
1139
  const requestCompositionId = this.runId && this.deps.backend.recordRequestComposition
1136
1140
  ? await this.deps.backend.recordRequestComposition(this.runId, {
1137
1141
  compositionId: this.deps.newId(),
@@ -1416,12 +1420,6 @@ export class AiSdkTurn {
1416
1420
  // reply's reasoning may not be resent, so input + output is
1417
1421
  // not the floor of the next input on every wire.
1418
1422
  const completedRequestIndex = runtimeSteps - 1;
1419
- // A finalization step resolves an empty tool set, so its
1420
- // request legitimately drops several thousand schema tokens
1421
- // with no fold, prune or image omission. Maka shaped that
1422
- // request; the provider did not drop anything.
1423
- const toolSchemaShrank = lastStepActiveToolCount !== undefined &&
1424
- activeToolsForRequest.length < lastStepActiveToolCount;
1425
1423
  // Across the send boundary the comparison is the same one,
1426
1424
  // against the last request a provider accepted before this
1427
1425
  // send. A provider that truncates to a fixed window reports
@@ -1431,21 +1429,15 @@ export class AiSdkTurn {
1431
1429
  // input tokens across eight turns with nothing reported
1432
1430
  // (#4623). The first request of a send therefore compares
1433
1431
  // against the persisted anchor, which is route-validated
1434
- // where it is read; a fold before that request would explain
1435
- // a smaller input by itself, so it disables the comparison.
1432
+ // where it is read.
1436
1433
  const acrossSends = completedRequestIndex === 0;
1437
1434
  const priorInput = acrossSends
1438
- ? midTurnState?.compactionAppliedThisSend === true
1439
- ? undefined
1440
- : midTurnState?.priorAcceptedInputTokens
1435
+ ? midTurnState?.priorAcceptedInputTokens
1441
1436
  : lastStepInputTokens;
1442
1437
  if (!this.deps.session.contextProviderDroppingReported &&
1443
- !toolSchemaShrank &&
1444
1438
  midTurnState &&
1445
1439
  priorInput !== undefined &&
1446
- midTurnState.replacedStepNumber !== completedRequestIndex &&
1447
- pruneAppliedAtStep !== completedRequestIndex &&
1448
- midTurnState.omittedImageToolResults.size === 0 &&
1440
+ midTurnState.stepShaping.size === 0 &&
1449
1441
  stepUsage !== undefined &&
1450
1442
  Number.isFinite(stepUsage.inputTokens) &&
1451
1443
  stepUsage.inputTokens > 0 &&
@@ -1467,6 +1459,11 @@ export class AiSdkTurn {
1467
1459
  priorInputTokens: priorInput,
1468
1460
  });
1469
1461
  }
1462
+ // Clear here, not at the top of the next step: the continuation
1463
+ // lane archives its prune below, after this point, and archiving
1464
+ // is idempotent — the next step's re-projection reports no prune,
1465
+ // so a clear above it would lose the only record of this one.
1466
+ midTurnState?.stepShaping.clear();
1470
1467
  // Fail closed: reset on every step boundary so a missing final
1471
1468
  // step's usage does not leave a stale value from an earlier step.
1472
1469
  // The reply needed more room than the declared window had
@@ -1581,10 +1578,10 @@ export class AiSdkTurn {
1581
1578
  // nothing left to grant it, so the error is terminal.
1582
1579
  const stepBudgetRemains = maxSteps === undefined || runtimeSteps < maxSteps;
1583
1580
  const recovered = stepBudgetRemains &&
1581
+ failure.kind === 'context_overflow' &&
1584
1582
  providerAttempt < MAX_PROVIDER_ATTEMPTS_PER_STEP &&
1585
1583
  attemptHasNoObservableOutput()
1586
1584
  ? await this.deps.compaction.recoverFromOverflowError({
1587
- error: failure,
1588
1585
  midTurnState,
1589
1586
  turnId,
1590
1587
  stepNumber: runtimeSteps,
@@ -1607,8 +1604,8 @@ export class AiSdkTurn {
1607
1604
  continue;
1608
1605
  }
1609
1606
  // Window suggestion (#4559): the provider rejected a request and
1610
- // no recovery is left — the one fold is spent, or there was no
1611
- // seam. The baseline is a proven-fit total (input + output of an
1607
+ // no recovery is left — this step's fold is spent, or there was
1608
+ // no seam. The baseline is a proven-fit total (input + output of an
1612
1609
  // accepted request), so it is a number the user can declare; the
1613
1610
  // trigger is `>=`, so declaring exactly it folds before this
1614
1611
  // point next time. Once per send, and only when the turn is
@@ -1637,7 +1634,7 @@ export class AiSdkTurn {
1637
1634
  // its size can be concluded (#4559).
1638
1635
  if (!contextOverflowAfterCompactionNoteWritten &&
1639
1636
  failure.kind === 'context_overflow' &&
1640
- midTurnState?.compactionAppliedThisSend === true) {
1637
+ midTurnState?.projectionCheckpoint !== undefined) {
1641
1638
  contextOverflowAfterCompactionNoteWritten = true;
1642
1639
  await this.recordSystemNote('context_overflow_after_compaction', turnId);
1643
1640
  }
@@ -1696,10 +1693,10 @@ export class AiSdkTurn {
1696
1693
  });
1697
1694
  continue;
1698
1695
  }
1699
- // Unrecoverable (not context-length, latch spent, no seam, or no
1700
- // safe fold): surface the real provider error via the terminal
1701
- // handler after settling any authoritative usage — never a
1702
- // fabricated success.
1696
+ // Unrecoverable (not context-length, this step's attempt spent,
1697
+ // no seam, or no safe fold): surface the real provider error via
1698
+ // the terminal handler after settling any authoritative usage —
1699
+ // never a fabricated success.
1703
1700
  terminalProviderError = failure;
1704
1701
  terminalRetry = { error: terminalProviderError, retry };
1705
1702
  terminalProviderErrorReason =