@sema-agent/server 7.78.2 → 7.79.1

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 (44) hide show
  1. package/dist/auto-mode-face.d.ts +23 -11
  2. package/dist/auto-mode-face.js +5 -2
  3. package/dist/boot/deferred-sandbox-path-env.d.ts +47 -0
  4. package/dist/boot/deferred-sandbox-path-env.js +1 -5
  5. package/dist/boot/resolve-spec.js +21 -2
  6. package/dist/capabilities/code-review-council.js +2 -1
  7. package/dist/config.js +4 -3
  8. package/dist/env-facts.d.ts +61 -0
  9. package/dist/env-facts.js +32 -6
  10. package/dist/execution-lane-caps.d.ts +12 -0
  11. package/dist/execution-lane-caps.js +8 -0
  12. package/dist/file-history-disclosure.d.ts +53 -0
  13. package/dist/file-history-disclosure.js +89 -0
  14. package/dist/http/admission.d.ts +56 -0
  15. package/dist/http/admission.js +560 -0
  16. package/dist/http/approval-sweeps.d.ts +21 -0
  17. package/dist/http/approval-sweeps.js +138 -0
  18. package/dist/http/dispatch.d.ts +96 -0
  19. package/dist/http/dispatch.js +499 -0
  20. package/dist/http/request-guards.d.ts +51 -0
  21. package/dist/http/request-guards.js +205 -0
  22. package/dist/http/resume-legs.d.ts +80 -0
  23. package/dist/http/resume-legs.js +1230 -0
  24. package/dist/http/server.d.ts +19 -104
  25. package/dist/http/server.js +23 -2589
  26. package/dist/main.js +6 -3
  27. package/dist/model-select.d.ts +4 -0
  28. package/dist/model-select.js +2 -1
  29. package/dist/observability/metrics.js +1 -1
  30. package/dist/plugins/file-history-store-sql.js +12 -11
  31. package/dist/plugins/remote-env-adb.d.ts +11 -0
  32. package/dist/plugins/remote-env-adb.js +1 -1
  33. package/dist/plugins/remote-env-device.d.ts +11 -0
  34. package/dist/plugins/remote-env-device.js +1 -1
  35. package/dist/plugins/remote-env-host.d.ts +11 -0
  36. package/dist/plugins/remote-env-host.js +1 -1
  37. package/dist/plugins/remote-env-k8s.d.ts +11 -0
  38. package/dist/plugins/remote-env-k8s.js +1 -1
  39. package/dist/plugins/remote-env-ssh.d.ts +11 -0
  40. package/dist/plugins/remote-env-ssh.js +1 -1
  41. package/dist/run-local.js +2 -1
  42. package/dist/task-cwd.d.ts +24 -0
  43. package/dist/task-cwd.js +9 -0
  44. package/package.json +2 -2
package/dist/main.js CHANGED
@@ -1,7 +1,8 @@
1
1
  import { readFileSync } from "node:fs";
2
+ import { DEFAULT_MODEL_REF } from "./model-select.js";
2
3
  import { homedir } from "node:os";
3
4
  import { join } from "node:path";
4
- import { Runner, terminalProjection, createPermissionRuleStoreProvider, InMemoryToolResultStore, TtlSessionStore, uuidv7, defaultTaskRegistry, createAllowDenyPolicy, workflowsCapability, probeSearchBackend, describeStaticWiring, expandTiers, DEFAULT_SUBAGENT_TOOL_NAME } from "@sema-agent/core";
5
+ import { Runner, terminalProjection, createPermissionRuleStoreProvider, InMemoryToolResultStore, TtlSessionStore, uuidv7, defaultTaskRegistry, createAllowDenyPolicy, workflowsCapability, probeSearchBackend, describeStaticWiring, DEFAULT_SUBAGENT_TOOL_NAME } from "@sema-agent/core";
5
6
  import { createSessionTitler } from "./session-titler.js";
6
7
  import { posIntEnv } from "./session-watch.js";
7
8
  import { selectEnvironmentTool } from "./capabilities/select-environment-tool.js";
@@ -27,6 +28,7 @@ import { createToolTracer, createPermissionDeniedMeter } from "./observability/t
27
28
  import { createRegistryJwtVerifier } from "./auth-bridge.js";
28
29
  import { createMetrics } from "./observability/metrics.js";
29
30
  import { installFailOpenRecorder } from "./observability/fail-open.js";
31
+ import { installFileHistoryFailureRecorder } from "./file-history-disclosure.js";
30
32
  import { installSecretEnvScrubRecorder } from "./observability/secret-env-scrub.js";
31
33
  import { setRunTerminalObserver } from "./observability/run-terminal-log.js";
32
34
  import { setRedactionObserver, redactSecrets } from "./trace/redact.js";
@@ -95,6 +97,7 @@ async function main() {
95
97
  const metrics = createMetrics();
96
98
  installFailOpenRecorder({ logger, metrics });
97
99
  installSecretEnvScrubRecorder({ logger, metrics });
100
+ installFileHistoryFailureRecorder({ logger, metrics });
98
101
  logConfigDiagnostics(logger);
99
102
  for (const w of drainNumEnvWarnings())
100
103
  logger.warn("config_env_invalid_using_default", { env: w.env, raw: w.raw });
@@ -469,7 +472,7 @@ async function main() {
469
472
  const ensureChildSessionDurable = checkpointStore && config.sessionBackend === "tidb" ? (sessionId) => ensureChildSessionDurableWithPromotion(sessionStore, subRunnerSessions, sessionId) : undefined;
470
473
  const webFetchSummarize = createWebFetchSummarizeLane({ brain, config, getKeyResolver: () => configCenter.getKeyResolver(), logger });
471
474
  const scenarioDeps = {
472
- runner, subRunner, handslessSubRunner, model: "default", skills, repoClient, requirePrincipal: config.requirePrincipal,
475
+ runner, subRunner, handslessSubRunner, model: DEFAULT_MODEL_REF, skills, repoClient, requirePrincipal: config.requirePrincipal,
473
476
  webSearch: webSearch ? webSearch : undefined, metrics, logger, webFetchSummarize,
474
477
  backgroundAgentStore: backgroundAgentStore ? backgroundAgentStore : undefined,
475
478
  checkpointStore: checkpointStore ? checkpointStore : undefined,
@@ -692,7 +695,7 @@ async function main() {
692
695
  engineArmsOnIntent: coreArmsAutoOnIntent(),
693
696
  resolveCaps: runtimeCapsPeekResolver,
694
697
  classifierModel: () => {
695
- const route = classifierRouteFor({ models: expandTiers(config.models, config.tiers ?? {}) ?? config.models, roles: config.roles });
698
+ const route = classifierRouteFor(runnerDeps);
696
699
  if (route.fault !== undefined)
697
700
  logger.debug("auto_mode_classifier_route_unresolved", { fault: route.fault });
698
701
  return route.id;
@@ -23,6 +23,10 @@
23
23
  * 🔴 Was a real defect: resolveSpec previously read ONLY the inline @mention, so `body.model` (the picker + the
24
24
  * `--model` flag) was SILENTLY DROPPED — the picker UX was a no-op server-side. This restores it.
25
25
  */
26
+ /** 调用方**什么都没选**时这条 task 落到的目录 ref —— 即 `config.models.default` 那一格(引擎侧
27
+ * `resolveModel("default", catalog)` 解它)。**单点**:读面要说「这台部署的主模型是谁」时读同一个词
28
+ * (`auto-mode-face.ts` 的席解析器入参),两处各写一个 `"default"` 字面串正是漂移的成因。 */
29
+ export declare const DEFAULT_MODEL_REF = "default";
26
30
  export declare function resolveTaskModel(bodyModel: unknown, objective: string, catalog: Record<string, unknown>, allowlist?: readonly string[]): {
27
31
  model: string;
28
32
  cleanedObjective: string;
@@ -1,11 +1,12 @@
1
1
  import { parseModelMention } from "@sema-agent/core";
2
+ export const DEFAULT_MODEL_REF = "default";
2
3
  export function resolveTaskModel(bodyModel, objective, catalog, allowlist) {
3
4
  const mention = parseModelMention(objective, Object.keys(catalog));
4
5
  const explicit = typeof bodyModel === "string" ? matchCatalogModel(bodyModel, catalog) : undefined;
5
6
  const present = bodyModel !== undefined && bodyModel !== null;
6
7
  const unknown = present && explicit === undefined;
7
8
  const repr = typeof bodyModel !== "string" ? `<non-string:${typeof bodyModel}>` : bodyModel === "" ? "<empty-string>" : bodyModel;
8
- const model = explicit ?? mention.model ?? "default";
9
+ const model = explicit ?? mention.model ?? DEFAULT_MODEL_REF;
9
10
  const picked = explicit !== undefined ? { ref: explicit, source: "explicit", requested: repr } : mention.model !== undefined ? { ref: mention.model, source: "mention", requested: mention.model } : undefined;
10
11
  const notAllowed = picked !== undefined && !isModelAllowlisted(picked.ref, catalog, allowlist)
11
12
  ? { requested: picked.requested, resolved: terminalCatalogName(picked.ref, catalog) ?? picked.ref, source: picked.source }
@@ -221,7 +221,7 @@ export function createMetrics() {
221
221
  m.counter("degraded_total", "Tasks degraded to a cheaper model (1.40), by reason (budget/rate_limit/breaker_open)");
222
222
  m.counter("task_model_not_allowlisted_resume_total", "Resumed legs whose model/compactionModel is outside the @-mention allowlist (allowed to keep running by design — the allowlist gates fresh picks only)");
223
223
  m.counter("prompt_constitution_total", "Prompt assembly constitution mode per task (core 1.243 提示词主权): steady state = core; provider-assembled/replaced/legacy warrant eyes");
224
- m.counter("rewind_snapshot_failed_total", "Per-turn rewind file snapshots that failed, by code (too_large = big working tree, expected/benign; other codes warrant eyes)");
224
+ m.counter("rewind_snapshot_failed_total", "Per-turn rewind file snapshots that failed, by the engine's FileHistoryResult.error.code (snapshot_failed/aborted/…; `unknown` = the disclosure text carried no readable code). COARSE on purpose — whether a failure is permanent (the same file will fail every turn until the deployment changes) is counted by file_history_permanent_failure_total instead, from the typed error object rather than from prose");
225
225
  m.counter("verifications_total", "Tasks run through the verification gate (1.44), by final verdict");
226
226
  m.counter("cascade_total", "Tasks run through the model cascade (1.45), by outcome (passed/exhausted)");
227
227
  m.counter("degenerate_total", "Degenerate-output failures (1.59), by salvage class (a/b/unknown)");
@@ -4,6 +4,7 @@ import { SqlBlobBackend } from "./blob-backend.js";
4
4
  import { mysqlDriver, pgDriver } from "./sql-driver.js";
5
5
  import { isMissingColumnError } from "./sql-errors.js";
6
6
  import { recordFailOpen } from "../observability/fail-open.js";
7
+ import { discloseCaughtFailure } from "../file-history-disclosure.js";
7
8
  export const FILE_HISTORY_BLOB_TABLE = "file_history_blob";
8
9
  export const FILE_HISTORY_BLOB_KEY_PREFIX = "file-history/";
9
10
  const sha256hex = (s) => createHash("sha256").update(s).digest("hex");
@@ -116,7 +117,7 @@ export class SqlFileHistoryStore {
116
117
  return { ok: false, error: { code: "track_failed", message: `first-touch record for ${absPath} kept racing a concurrent retraction — retry` } };
117
118
  }
118
119
  catch (err) {
119
- return { ok: false, error: { code: "track_failed", message: `first-touch record for ${absPath} could not be persisted: ${err.message}` } };
120
+ return { ok: false, error: { code: "track_failed", message: `first-touch record for ${absPath} could not be persisted: ${discloseCaughtFailure(err)}` } };
120
121
  }
121
122
  }
122
123
  async annulIneligibleReason(scope, keyHash, absPath) {
@@ -189,7 +190,7 @@ export class SqlFileHistoryStore {
189
190
  return { ok: false, error: { code: "conflict", message: `first-touch record for ${absPath} was replaced while this retraction was being proven` } };
190
191
  }
191
192
  catch (err) {
192
- return { ok: false, error: { code: "track_failed", message: `annulling the first-touch record for ${absPath} failed: ${err.message}` } };
193
+ return { ok: false, error: { code: "track_failed", message: `annulling the first-touch record for ${absPath} failed: ${discloseCaughtFailure(err)}` } };
193
194
  }
194
195
  }
195
196
  async loadTracked(exec, scope, forUpdate = false) {
@@ -296,7 +297,7 @@ export class SqlFileHistoryStore {
296
297
  }
297
298
  }
298
299
  catch (err) {
299
- return { ok: false, error: { code: "snapshot_failed", message: err.message } };
300
+ return { ok: false, error: { code: "snapshot_failed", message: discloseCaughtFailure(err) } };
300
301
  }
301
302
  this.activeAttempts.add(lease);
302
303
  try {
@@ -398,7 +399,7 @@ export class SqlFileHistoryStore {
398
399
  catch (err) {
399
400
  await conn.rollback().catch(() => undefined);
400
401
  this.spentAttempts.add(lease);
401
- return { ok: false, error: { code: "snapshot_failed", message: `boundary for entry "${entryId}" could not be committed: ${err.message}` } };
402
+ return { ok: false, error: { code: "snapshot_failed", message: `boundary for entry "${entryId}" could not be committed: ${discloseCaughtFailure(err)}` } };
402
403
  }
403
404
  finally {
404
405
  conn.release();
@@ -408,7 +409,7 @@ export class SqlFileHistoryStore {
408
409
  }
409
410
  catch (err) {
410
411
  this.spentAttempts.add(lease);
411
- return { ok: false, error: { code: "snapshot_failed", message: err.message } };
412
+ return { ok: false, error: { code: "snapshot_failed", message: discloseCaughtFailure(err) } };
412
413
  }
413
414
  finally {
414
415
  this.activeAttempts.delete(lease);
@@ -475,7 +476,7 @@ export class SqlFileHistoryStore {
475
476
  });
476
477
  }
477
478
  catch (err) {
478
- applied = { outcome: "failed", reason: err.message };
479
+ applied = { outcome: "failed", reason: discloseCaughtFailure(err) };
479
480
  }
480
481
  if (applied.outcome === "applied")
481
482
  ledger.filesChanged.push(t.trackKey);
@@ -492,7 +493,7 @@ export class SqlFileHistoryStore {
492
493
  }
493
494
  catch (err) {
494
495
  const touched = empty.filesChanged.length > 0 || empty.identicalSkipped.length > 0 || empty.refused.length > 0 || empty.failed.length > 0;
495
- return { outcome: touched ? "partial" : "failed", ...empty, error: { code: "restore_failed", message: err.message } };
496
+ return { outcome: touched ? "partial" : "failed", ...empty, error: { code: "restore_failed", message: discloseCaughtFailure(err) } };
496
497
  }
497
498
  }
498
499
  async restorePlan(scope, entryId) {
@@ -593,7 +594,7 @@ export class SqlFileHistoryStore {
593
594
  return { ok: true };
594
595
  }
595
596
  catch (err) {
596
- return { ok: false, error: { code: "snapshot_failed", message: `putBlob ${hash} write/upload failed: ${err.message}` } };
597
+ return { ok: false, error: { code: "snapshot_failed", message: `putBlob ${hash} write/upload failed: ${discloseCaughtFailure(err)}` } };
597
598
  }
598
599
  }
599
600
  async getBlob(hash) {
@@ -636,7 +637,7 @@ export class SqlFileHistoryStore {
636
637
  }
637
638
  }
638
639
  catch (err) {
639
- return { ok: false, error: { code: "conflict", message: `adoptScope failed: ${err.message}` } };
640
+ return { ok: false, error: { code: "conflict", message: `adoptScope failed: ${discloseCaughtFailure(err)}` } };
640
641
  }
641
642
  }
642
643
  async scopeEmpty(exec, scope, forUpdate = false) {
@@ -735,7 +736,7 @@ export class SqlFileHistoryStore {
735
736
  bytes = await srcGetBlob(hash);
736
737
  }
737
738
  catch (err) {
738
- return { ok: false, error: { code: "unsupported_format", message: `source blob ${hash} fetch failed: ${err.message}` } };
739
+ return { ok: false, error: { code: "unsupported_format", message: `source blob ${hash} fetch failed: ${discloseCaughtFailure(err)}` } };
739
740
  }
740
741
  if (bytes === undefined)
741
742
  return { ok: false, error: { code: "not_found", message: `source blob ${hash} is missing — the envelope's graph cannot be completed, so nothing is published` } };
@@ -786,7 +787,7 @@ export class SqlFileHistoryStore {
786
787
  }
787
788
  }
788
789
  catch (err) {
789
- return { ok: false, error: { code: "unsupported_format", message: `importHistory failed: ${err.message}` } };
790
+ return { ok: false, error: { code: "unsupported_format", message: `importHistory failed: ${discloseCaughtFailure(err)}` } };
790
791
  }
791
792
  }
792
793
  }
@@ -44,9 +44,20 @@ export interface AdbEnvDeps {
44
44
  * typed retryable (reconnect + retry per idempotency), never a fake command result. Exported for tests. */
45
45
  export declare function isAdbTransportLost(stderr: string): boolean;
46
46
  export declare class RemoteAdbExecutionEnv implements RemoteExecutionEnv {
47
+ /**
48
+ * 🔴 **S-354 / core 7.20.0([ref])`canonicalPathAuthoritative: true`** —— 本适配器的 `canonicalPath`
49
+ * 就是**命令真正跑的那块文件系统**的权威答案(`posix-shell-fs` 的 `canonicalPath` 原语,在设备 shell 里解到落点)。声明它 = 告诉引擎:整链答不出来是**解析器出了事**
50
+ * (RPC 断、传输掉),不是「一个它看不见的命名空间」⇒ 读边界的执行期复核**拒**,而不是让词法判决站着。
51
+ * 缺席是今天每个适配器的读法(整链沉默 ⇒ 词法判决成立、命令照跑),所以这一声明是**收紧**方向。
52
+ * 严格读法 `=== true`(core `canonicalPathIsAuthoritative`);`isSuspendable` 一族同样严格,对本仓的
53
+ * 布尔字面量零影响(三键都写死字面 `true`/`false`,不存在「非布尔值被读成缺席」那一形 —— 那一形由
54
+ * core 的 `config.execution_env_capability_invalid` 普查响亮播报,本仓的受众表已登记)。
55
+ * 谁该声明由闭集门看着(`test/execution-lane-caps.test.ts` 的逐 lane 对账:新 lane 不表态 = 编译红)。
56
+ */
47
57
  readonly capabilities: {
48
58
  isolation: boolean;
49
59
  suspendable: boolean;
60
+ canonicalPathAuthoritative: boolean;
50
61
  };
51
62
  cwd: string;
52
63
  private connected;
@@ -18,7 +18,7 @@ export function isAdbTransportLost(stderr) {
18
18
  return /device offline|device '.*' not found|no devices\/emulators found|connection reset|protocol fault|device disconnected/i.test(stderr);
19
19
  }
20
20
  export class RemoteAdbExecutionEnv {
21
- capabilities = { isolation: false, suspendable: false };
21
+ capabilities = { isolation: false, suspendable: false, canonicalPathAuthoritative: true };
22
22
  cwd;
23
23
  connected = false;
24
24
  destroyed = false;
@@ -94,9 +94,20 @@ export declare class RemoteDeviceExecutionEnv implements RemoteExecutionEnv {
94
94
  * 谁要把 executor 的工作区改成临时目录,先来改这一行:那一刻这个 `false` 就变成谎话,park 赎回会
95
95
  * 落到一个空工作区上而链路上没有任何一层会报错。
96
96
  */
97
+ /**
98
+ * 🔴 **S-354 / core 7.20.0([ref])`canonicalPathAuthoritative: true`** —— 本适配器的 `canonicalPath`
99
+ * 就是**命令真正跑的那块文件系统**的权威答案(设备 WS 协议的**必答** `canonicalPath` op)。声明它 = 告诉引擎:整链答不出来是**解析器出了事**
100
+ * (RPC 断、传输掉),不是「一个它看不见的命名空间」⇒ 读边界的执行期复核**拒**,而不是让词法判决站着。
101
+ * 缺席是今天每个适配器的读法(整链沉默 ⇒ 词法判决成立、命令照跑),所以这一声明是**收紧**方向。
102
+ * 严格读法 `=== true`(core `canonicalPathIsAuthoritative`);`isSuspendable` 一族同样严格,对本仓的
103
+ * 布尔字面量零影响(三键都写死字面 `true`/`false`,不存在「非布尔值被读成缺席」那一形 —— 那一形由
104
+ * core 的 `config.execution_env_capability_invalid` 普查响亮播报,本仓的受众表已登记)。
105
+ * 谁该声明由闭集门看着(`test/execution-lane-caps.test.ts` 的逐 lane 对账:新 lane 不表态 = 编译红)。
106
+ */
97
107
  readonly capabilities: {
98
108
  isolation: boolean;
99
109
  suspendable: boolean;
110
+ canonicalPathAuthoritative: boolean;
100
111
  };
101
112
  readonly externalContentTarget = true;
102
113
  /** [ref]:设备路径不在 worker 宿主文件系统上 —— host-path 回执消费方据此拒绝把它当宿主路径呈现。 */
@@ -110,7 +110,7 @@ function assertPosixDevicePlatform(os, deviceId) {
110
110
  export class RemoteDeviceExecutionEnv {
111
111
  cfg;
112
112
  homeDir;
113
- capabilities = { isolation: false, suspendable: false };
113
+ capabilities = { isolation: false, suspendable: false, canonicalPathAuthoritative: true };
114
114
  externalContentTarget = true;
115
115
  hostLocalPaths = false;
116
116
  cwd;
@@ -58,9 +58,20 @@ export interface HostEnvConfig {
58
58
  }
59
59
  export declare class RemoteHostExecutionEnv implements RemoteExecutionEnv, SchedulerCapability, BackgroundShellCapability {
60
60
  /** [ref] §9 / DUAL-MODE-DESIGN §5: a `host` env is the operator's real machine — no isolation, no snapshot. */
61
+ /**
62
+ * 🔴 **S-354 / core 7.20.0([ref])`canonicalPathAuthoritative: true`** —— 本适配器的 `canonicalPath`
63
+ * 就是**命令真正跑的那块文件系统**的权威答案(worker 本机的 `fs.realpath`)。声明它 = 告诉引擎:整链答不出来是**解析器出了事**
64
+ * (RPC 断、传输掉),不是「一个它看不见的命名空间」⇒ 读边界的执行期复核**拒**,而不是让词法判决站着。
65
+ * 缺席是今天每个适配器的读法(整链沉默 ⇒ 词法判决成立、命令照跑),所以这一声明是**收紧**方向。
66
+ * 严格读法 `=== true`(core `canonicalPathIsAuthoritative`);`isSuspendable` 一族同样严格,对本仓的
67
+ * 布尔字面量零影响(三键都写死字面 `true`/`false`,不存在「非布尔值被读成缺席」那一形 —— 那一形由
68
+ * core 的 `config.execution_env_capability_invalid` 普查响亮播报,本仓的受众表已登记)。
69
+ * 谁该声明由闭集门看着(`test/execution-lane-caps.test.ts` 的逐 lane 对账:新 lane 不表态 = 编译红)。
70
+ */
61
71
  readonly capabilities: {
62
72
  isolation: boolean;
63
73
  suspendable: boolean;
74
+ canonicalPathAuthoritative: boolean;
64
75
  };
65
76
  /** R7 自唤醒: structural delegation to the injected scheduler backend. `hasScheduler(env)` checks BOTH the method
66
77
  * presence (always, since we implement the interface) AND `schedulerCapabilities.supported===true` — which is only
@@ -34,7 +34,7 @@ const UNSUPPORTED_SCHEDULER_CAPS = {
34
34
  maxScheduledHorizonSec: 0,
35
35
  };
36
36
  export class RemoteHostExecutionEnv {
37
- capabilities = { isolation: false, suspendable: false };
37
+ capabilities = { isolation: false, suspendable: false, canonicalPathAuthoritative: true };
38
38
  get schedulerCapabilities() {
39
39
  return this.cfg.scheduler?.schedulerCapabilities ?? UNSUPPORTED_SCHEDULER_CAPS;
40
40
  }
@@ -119,9 +119,20 @@ export { execSocketTlsOptions } from "./k8s-exec-protocol.js";
119
119
  export declare function errorMessageOf(e: unknown): string;
120
120
  export { exitCodeFromStatus } from "./k8s-exec-protocol.js";
121
121
  export declare class RemoteK8sExecutionEnv implements RemoteExecutionEnv, BackgroundShellCapability {
122
+ /**
123
+ * 🔴 **S-354 / core 7.20.0([ref])`canonicalPathAuthoritative: true`** —— 本适配器的 `canonicalPath`
124
+ * 就是**命令真正跑的那块文件系统**的权威答案(`posix-shell-fs` 的 `canonicalPath` 原语,经 pod exec 解到落点)。声明它 = 告诉引擎:整链答不出来是**解析器出了事**
125
+ * (RPC 断、传输掉),不是「一个它看不见的命名空间」⇒ 读边界的执行期复核**拒**,而不是让词法判决站着。
126
+ * 缺席是今天每个适配器的读法(整链沉默 ⇒ 词法判决成立、命令照跑),所以这一声明是**收紧**方向。
127
+ * 严格读法 `=== true`(core `canonicalPathIsAuthoritative`);`isSuspendable` 一族同样严格,对本仓的
128
+ * 布尔字面量零影响(三键都写死字面 `true`/`false`,不存在「非布尔值被读成缺席」那一形 —— 那一形由
129
+ * core 的 `config.execution_env_capability_invalid` 普查响亮播报,本仓的受众表已登记)。
130
+ * 谁该声明由闭集门看着(`test/execution-lane-caps.test.ts` 的逐 lane 对账:新 lane 不表态 = 编译红)。
131
+ */
122
132
  readonly capabilities: {
123
133
  isolation: boolean;
124
134
  suspendable: boolean;
135
+ canonicalPathAuthoritative: true;
125
136
  };
126
137
  cwd: string;
127
138
  private podName?;
@@ -78,7 +78,7 @@ export class RemoteK8sExecutionEnv {
78
78
  this.cwd = this.cfg.mountPath;
79
79
  this.homeDir = config.homeDir;
80
80
  this.deps = deps ?? {};
81
- this.capabilities = { isolation: true, suspendable: !!this.cfg.s3Snapshot };
81
+ this.capabilities = { isolation: true, suspendable: !!this.cfg.s3Snapshot, canonicalPathAuthoritative: true };
82
82
  const bgMax = Math.max(1, Math.min(K8S_BG_MAX_TIMEOUT_SEC, Math.floor(this.cfg.timeoutMs / 1000)));
83
83
  this.backgroundCapabilities = {
84
84
  supported: true,
@@ -61,9 +61,20 @@ export interface SshEnvDeps {
61
61
  connectOnce?: (signal?: AbortSignal) => Promise<ConnectResult>;
62
62
  }
63
63
  export declare class RemoteSshExecutionEnv implements RemoteExecutionEnv {
64
+ /**
65
+ * 🔴 **S-354 / core 7.20.0([ref])`canonicalPathAuthoritative: true`** —— 本适配器的 `canonicalPath`
66
+ * 就是**命令真正跑的那块文件系统**的权威答案(`posix-shell-fs` 的 `canonicalPath` 原语,在真机上解到落点)。声明它 = 告诉引擎:整链答不出来是**解析器出了事**
67
+ * (RPC 断、传输掉),不是「一个它看不见的命名空间」⇒ 读边界的执行期复核**拒**,而不是让词法判决站着。
68
+ * 缺席是今天每个适配器的读法(整链沉默 ⇒ 词法判决成立、命令照跑),所以这一声明是**收紧**方向。
69
+ * 严格读法 `=== true`(core `canonicalPathIsAuthoritative`);`isSuspendable` 一族同样严格,对本仓的
70
+ * 布尔字面量零影响(三键都写死字面 `true`/`false`,不存在「非布尔值被读成缺席」那一形 —— 那一形由
71
+ * core 的 `config.execution_env_capability_invalid` 普查响亮播报,本仓的受众表已登记)。
72
+ * 谁该声明由闭集门看着(`test/execution-lane-caps.test.ts` 的逐 lane 对账:新 lane 不表态 = 编译红)。
73
+ */
64
74
  readonly capabilities: {
65
75
  isolation: boolean;
66
76
  suspendable: boolean;
77
+ canonicalPathAuthoritative: boolean;
67
78
  };
68
79
  /** Working directory; relative paths resolve against it (ExecutionEnv contract). Starts at mountPath. */
69
80
  cwd: string;
@@ -16,7 +16,7 @@ function exitCodeForSignalKill(signal) {
16
16
  return 128 + (signalNumber(`SIG${signal}`) ?? 9);
17
17
  }
18
18
  export class RemoteSshExecutionEnv {
19
- capabilities = { isolation: false, suspendable: false };
19
+ capabilities = { isolation: false, suspendable: false, canonicalPathAuthoritative: true };
20
20
  cwd;
21
21
  conn;
22
22
  sftpPromise;
package/dist/run-local.js CHANGED
@@ -1,4 +1,5 @@
1
1
  #!/usr/bin/env node
2
+ import { DEFAULT_MODEL_REF } from "./model-select.js";
2
3
  import { terminalProjection } from "@sema-agent/core";
3
4
  import { homedir } from "node:os";
4
5
  import { join } from "node:path";
@@ -429,7 +430,7 @@ export async function runLocal(argv, deps = {}) {
429
430
  const webSearch = webSearchCfg ? createWebSearchBackend(webSearchCfg) : undefined;
430
431
  const taskListLane = createTaskListLane({ logger });
431
432
  const scenarioDeps = {
432
- runner, subRunner, handslessSubRunner, model: "default", skills, repoClient, metrics, logger,
433
+ runner, subRunner, handslessSubRunner, model: DEFAULT_MODEL_REF, skills, repoClient, metrics, logger,
433
434
  ...(webSearch ? { webSearch } : {}),
434
435
  backgroundAgentStore: fileBackend.backgroundAgentStore,
435
436
  taskListStoreFor: taskListLane.storeFor,
@@ -18,6 +18,30 @@ export declare const MAX_CWD_CHARS = 4096;
18
18
  * 先改这里**:把判据换成 core 的 `isAbsoluteForFamily`,并把家族从这条 cwd 自己的拼法取。
19
19
  * 钉见 `test/task-cwd.test.ts` 的「[ref] 路径家族接缝」格(实测读数,不是推断)。 */
20
20
  export declare function isValidCwd(cwd: unknown): cwd is string;
21
+ /**
22
+ * 词法归一:纯字符串运算,折 `.` / `..` / 重复分隔符,**不碰 fs**。
23
+ * `posix.normalize` 把 `..` 在根部截断(`/../x` → `/x`),与「根之上没有东西」的语义一致;尾部分隔符
24
+ * 统一剥掉(根 `/` 除外),让同一目标只有一个 key —— 前缀判域(见 {@link dirCoversPath})靠的就是
25
+ * key 的唯一性。
26
+ *
27
+ * 🔴 **单点**(S-351 起):此前它的一份私有副本住在 `boot/deferred-sandbox-path-env.ts`(沙箱代理的
28
+ * `absolutePath` 词法臂),而写门豁免域与任务根的判域现在读的是同一条律 —— 两份各自归一就是两个 key
29
+ * 空间,而「豁免域盖没盖住任务根」这种判断在两个 key 空间里可以同时答是和否。
30
+ * (core 内部有同形的 `normalizeAbsPathLexically`,但未从包根导出;此处是 Node 标准库的等价运算,
31
+ * 不是它的抄本——若日后 core 导出,这里应改为直接复用。)
32
+ */
33
+ export declare function normalizeAbsolutePathLexically(p: string): string;
34
+ /**
35
+ * **词法**判域:`dir` 是不是 `target` 本身或它的祖先(= `target` ⊆ `dir`,「等于 ∪ 后代」)。
36
+ *
37
+ * 这是 core `assertExemptDirsDoNotCoverRoot` 那条律的本仓同形(亲读安装树
38
+ * `dist/core/fs-write-gate-policy.js:11-28` → `swallowedFenceRoot` → `isWithin(root, p)`:
39
+ * 先各自词法归一,再判 `p === root` 或 `p` 以 `root + "/"` 起头)。**两端都只做词法**,不 realpath ——
40
+ * 写门的构造期本来就没有 I/O(core 同理),承诺 canonical 是承诺一件此刻做不到的事。
41
+ *
42
+ * 段感知由「归一后 + 分隔符前缀」保证:`/a/bc` **不在** `/a/b` 域内。
43
+ */
44
+ export declare function dirCoversPath(dir: string, target: string): boolean;
21
45
  /**
22
46
  * The deployment-level gate: does THIS deployment honor a caller-supplied `cwd`? TRUE only for the single-user
23
47
  * `host` lane (see module header). All cloud/container lanes + any multi-tenant deployment → FALSE (cwd ignored).
package/dist/task-cwd.js CHANGED
@@ -8,6 +8,15 @@ export function isValidCwd(cwd) {
8
8
  path.isAbsolute(cwd) &&
9
9
  !cwd.split(/[/\\]/).includes(".."));
10
10
  }
11
+ export function normalizeAbsolutePathLexically(p) {
12
+ const collapsed = path.posix.normalize(p.replace(/^\/+/, "/"));
13
+ return collapsed.length > 1 ? collapsed.replace(/\/+$/, "") : collapsed;
14
+ }
15
+ export function dirCoversPath(dir, target) {
16
+ const root = normalizeAbsolutePathLexically(dir);
17
+ const p = normalizeAbsolutePathLexically(target);
18
+ return p === root || p.startsWith(root.endsWith("/") ? root : `${root}/`);
19
+ }
11
20
  export function cwdHonored(config) {
12
21
  return config.remoteExec?.provider === "host" && config.requirePrincipal !== true;
13
22
  }
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@sema-agent/server",
3
- "version": "7.78.2",
3
+ "version": "7.79.1",
4
4
  "description": "Sema Server — the server/API implementation layer for Sema, wiring core, registry, model providers, and cloud agent execution. Built on @sema-agent/core.",
5
5
  "type": "module",
6
6
  "license": "BUSL-1.1",
@@ -54,7 +54,7 @@
54
54
  "build:binary:run-local:darwin-arm64": "bun build --compile --target=bun-darwin-arm64 src/run-local.ts --outfile dist/run-local-darwin-arm64"
55
55
  },
56
56
  "dependencies": {
57
- "@sema-agent/core": "7.19.0",
57
+ "@sema-agent/core": "7.20.0",
58
58
  "@sema-agent/settings-schema": "2.1.0",
59
59
  "e2b": "^2.28.0",
60
60
  "libsodium-wrappers": "^0.8.4",