@opengeni/runtime 0.6.1 → 0.7.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (45) hide show
  1. package/dist/{chunk-DYBF7ZEB.js → chunk-RYDCVTMW.js} +1958 -207
  2. package/dist/chunk-RYDCVTMW.js.map +1 -0
  3. package/dist/{index-BTo_bZvN.d.ts → index-DIiiwQY2.d.ts} +689 -52
  4. package/dist/index.d.ts +293 -123
  5. package/dist/index.js +1819 -790
  6. package/dist/index.js.map +1 -1
  7. package/dist/sandbox/index.d.ts +1 -1
  8. package/dist/sandbox/index.js +79 -3
  9. package/package.json +21 -20
  10. package/src/codex-tool-search.ts +185 -31
  11. package/src/context-compaction.ts +78 -243
  12. package/src/history-sanitizer.ts +172 -20
  13. package/src/index.ts +2573 -898
  14. package/src/metrics.ts +28 -2
  15. package/src/sandbox/capabilities.ts +12 -3
  16. package/src/sandbox/channel-a.ts +420 -72
  17. package/src/sandbox/display-stack.ts +18 -4
  18. package/src/sandbox/errors.ts +13 -0
  19. package/src/sandbox/index.ts +398 -112
  20. package/src/sandbox/op-correlation.ts +70 -0
  21. package/src/sandbox/providers/index.ts +9 -3
  22. package/src/sandbox/providers/modal.ts +84 -14
  23. package/src/sandbox/providers/selfhosted.ts +1 -4
  24. package/src/sandbox/recording.ts +146 -90
  25. package/src/sandbox/routing/backend-resolver.ts +144 -10
  26. package/src/sandbox/routing/routing-session.ts +129 -25
  27. package/src/sandbox/select.ts +60 -8
  28. package/src/sandbox/selfhosted/capabilities.ts +59 -25
  29. package/src/sandbox/selfhosted/control-rpc.ts +160 -28
  30. package/src/sandbox/selfhosted/fault-rendering.ts +226 -0
  31. package/src/sandbox/selfhosted/op-observer.ts +54 -0
  32. package/src/sandbox/selfhosted/op-stream.ts +801 -0
  33. package/src/sandbox/selfhosted/op-testing.ts +468 -0
  34. package/src/sandbox/selfhosted/op-transport.ts +151 -0
  35. package/src/sandbox/selfhosted/retry-policy.ts +177 -0
  36. package/src/sandbox/selfhosted/session.ts +498 -55
  37. package/src/sandbox/selfhosted/testing.ts +60 -19
  38. package/src/sandbox/stream-port.ts +17 -5
  39. package/src/sandbox/stream-token.ts +8 -2
  40. package/src/sandbox/terminal-server.ts +18 -4
  41. package/src/sandbox-computer.ts +332 -70
  42. package/src/usage-telemetry.ts +73 -0
  43. package/dist/chunk-DYBF7ZEB.js.map +0 -1
  44. package/src/image-history.ts +0 -149
  45. package/src/screenshot-omitted-card.ts +0 -15
@@ -19,6 +19,7 @@
19
19
  // API/worker pass a resume-by-id closure for the sibling box's lease).
20
20
 
21
21
  import { buildSelfhostedBackendSession, type SelfhostedRelayConfig } from "../selfhosted/session";
22
+ import type { SelfhostedOpObserver } from "../selfhosted/op-observer";
22
23
  import type { ControlRpc } from "../selfhosted/control-rpc";
23
24
  import type {
24
25
  ActivePointer,
@@ -60,8 +61,16 @@ export interface ActiveBackendResolverDeps {
60
61
  * absent, a modal swap target surfaces as unsupported (the caller validated
61
62
  * liveness, so this is the "modal swap not wired in this context" guard). */
62
63
  establishModalTarget?: (sandbox: RoutableSandbox) => Promise<RoutableBackendSession>;
63
- /** Override the selfhosted control-op timeout (tests). */
64
+ /** The selfhosted CONTROL-op timeout (ping/fs/desktop/pty) for a swap/pin target.
65
+ * Absent ⇒ the session's 30s default. */
64
66
  selfhostedTimeoutMs?: number;
67
+ /** The selfhosted EXEC process deadline for a swap/pin target, distinct from the
68
+ * control timeout (a swapped-to machine runs real commands too). Absent ⇒ falls
69
+ * back to the control timeout, as in the session leaf. */
70
+ selfhostedExecTimeoutMs?: number;
71
+ /** The per-op observer threaded into a selfhosted swap/pin target (out-of-band
72
+ * telemetry — metrics + machine.* events). Absent ⇒ no-op. */
73
+ selfhostedOnOp?: SelfhostedOpObserver;
65
74
  /**
66
75
  * The run's declared sandbox environment — the SAME `Record<string,string>` the
67
76
  * worker turn threads into the agent's TARGET manifest (and into the group box at
@@ -86,15 +95,104 @@ export interface ActiveBackendResolverDeps {
86
95
  * path (which always builds fresh — it has no pre-established turn session).
87
96
  */
88
97
  pinnedSelfhosted?: { sandboxId: string; epoch: number; session: RoutableBackendSession };
98
+ /**
99
+ * Whether `defaultBackend` IS the session's home box — i.e. whether the null pointer
100
+ * (== home) may resolve to it. Defaults to TRUE (omitted): a genuine machine-HOME turn
101
+ * (home IS the machine) and a Modal-home turn established on its group box both pass/omit
102
+ * true, so null resolves to `defaultBackend` as before. Set explicitly FALSE only for a
103
+ * machine-primary turn of a Modal-HOME session (pinned to a machine, group box never
104
+ * established this turn): the null pointer then has nothing to resolve to THIS turn, so
105
+ * rather than silently keep serving the pinned machine as if the detach never happened,
106
+ * the null branch throws a typed `home_unavailable_this_turn` error — the detach's pointer
107
+ * commit stands and takes effect on the NEXT turn, which starts null and establishes the
108
+ * home box normally. Lazily establishing the home box mid-turn on such a clear is a
109
+ * deferred follow-up (issue #341); until then this makes the gap fail typed-and-specific,
110
+ * never silent.
111
+ */
112
+ defaultIsHome?: boolean;
89
113
  }
90
114
 
115
+ /** Why a persisted pointer / swap target cannot be turned into a live backend. A
116
+ * typed discriminant so callers (turn-start reconcile, the swap tool, future
117
+ * reports) can distinguish a stale pointer from an unaddressable enrollment from a
118
+ * backend the turn context simply cannot establish — instead of string-matching
119
+ * one opaque message (issue #341: "typed diagnostics that distinguish stale
120
+ * pointer, offline enrollment, unsupported backend context, transient failure").
121
+ * - `stale_pointer` — the pointed-at sandbox row is gone (deleted/orphaned).
122
+ * - `offline_enrollment` — a selfhosted target carries no enrollment/agent id to address.
123
+ * - `unsupported_backend_context` — a target no turn routing context can establish
124
+ * (a non-group Modal sibling; an unknown kind).
125
+ * - `transient_establishment` — a momentary establish failure worth a retry (reserved;
126
+ * control-plane timeouts surface as their own typed error).
127
+ * - `home_unavailable_this_turn` — the pointer was cleared to the session default (home)
128
+ * mid-turn, but this turn started pinned to a machine and
129
+ * never established the home box, so null has nothing to
130
+ * resolve to THIS turn (the detach takes effect next turn). */
131
+ export type BackendUnresolvableCode =
132
+ | "stale_pointer"
133
+ | "offline_enrollment"
134
+ | "unsupported_backend_context"
135
+ | "transient_establishment"
136
+ | "home_unavailable_this_turn";
137
+
91
138
  /** Thrown when a swap target cannot be resolved (unknown sandbox, or a modal
92
- * target with no establisher in this context). The caller maps it to a 409. */
139
+ * target with no establisher in this context). The caller maps it to a 409.
140
+ * Carries a typed {@link BackendUnresolvableCode} in addition to the message. */
93
141
  export class ActiveBackendUnresolvableError extends Error {
94
142
  readonly name = "ActiveBackendUnresolvableError";
95
- constructor(message: string) {
143
+ readonly code: BackendUnresolvableCode;
144
+ constructor(code: BackendUnresolvableCode, message: string) {
96
145
  super(message);
146
+ this.code = code;
147
+ }
148
+ }
149
+
150
+ /** The outcome of {@link swapTargetEstablishability}. */
151
+ export type SwapTargetEstablishability =
152
+ | { ok: true }
153
+ | { ok: false; code: BackendUnresolvableCode; reason: string };
154
+
155
+ /**
156
+ * THE SINGLE SOURCE OF TRUTH for "can a turn's routing context ESTABLISH this swap
157
+ * target?" — shared by swap/seed ADMISSION (`fleet.resolveTarget`, before the CAS
158
+ * pointer commit) and turn-time ESTABLISHMENT (this resolver + `wrapTurnBoxWithRouting`).
159
+ * Admission MUST reject any target this predicate calls unestablishable, so the
160
+ * pointer is never committed to a backend no turn can resume (issue #341 Shape 1:
161
+ * a Modal-sibling swap was admitted + epoch-bumped, then every following op
162
+ * stranded because no context wires an establisher for it).
163
+ *
164
+ * - the session's OWN group box (`isSessionGroup`) is the null/home target and is
165
+ * always establishable (it IS the default backend);
166
+ * - a `selfhosted` machine target is establishable (`buildSelfhostedBackendSession`);
167
+ * admission's separate liveness probe — not this predicate — gates online-ness;
168
+ * - a NON-group `modal` target is NOT establishable: no turn context wires
169
+ * `establishModalTarget` (cross-group Modal resume-by-id is a future, billing-
170
+ * touching feature — sandbox-routing.ts). Wiring that establisher is the ONE
171
+ * toggle that flips this branch; keep admission and the resolver in lockstep.
172
+ */
173
+ export function swapTargetEstablishability(target: {
174
+ kind: string;
175
+ isSessionGroup: boolean;
176
+ }): SwapTargetEstablishability {
177
+ if (target.isSessionGroup) {
178
+ return { ok: true };
179
+ }
180
+ if (target.kind === "selfhosted") {
181
+ return { ok: true };
182
+ }
183
+ if (target.kind === "modal") {
184
+ return {
185
+ ok: false,
186
+ code: "unsupported_backend_context",
187
+ reason:
188
+ "a Modal sandbox other than this session's own box cannot be established by a turn yet; swap back to the session default or attach a Connected Machine",
189
+ };
97
190
  }
191
+ return {
192
+ ok: false,
193
+ code: "unsupported_backend_context",
194
+ reason: `a sandbox of kind "${target.kind}" cannot be established by a turn`,
195
+ };
98
196
  }
99
197
 
100
198
  /**
@@ -115,9 +213,29 @@ export function makeActiveBackendResolver(
115
213
  deps: ActiveBackendResolverDeps,
116
214
  ): (pointer: ActivePointer) => Promise<ResolvedActiveBackend> {
117
215
  return async (pointer: ActivePointer): Promise<ResolvedActiveBackend> => {
118
- // The DEFAULT target: the session's own group sandbox (backward-compat). The
119
- // proxy routes to the already-established box; the lease owns its lifecycle.
216
+ // NULL == the session's HOME backend (issue #341 invariant 1). `defaultBackend`
217
+ // is the already-established home the wiring handed us — the Modal group box for a
218
+ // Modal-home turn, or the machine itself for a machine-home turn. It is NEVER a
219
+ // swap target's SelfhostedSession: a clear-to-null must re-land on the home, never
220
+ // keep serving the machine a prior op swapped to. (The proxy caches by the FULL
221
+ // pointer tuple, so a clear-to-null — epoch-bumped OR a same-epoch FK null —
222
+ // invalidates the cached machine and re-enters this branch.) The lease owns the
223
+ // home box lifecycle; the proxy does not re-establish it.
120
224
  if (pointer.activeSandboxId === null) {
225
+ // A machine-pinned turn (Modal-home session pinned to a machine) never
226
+ // established its home box, so a mid-turn clear-to-null has no home to resolve
227
+ // THIS turn. Fail typed-and-specific — the detach was accepted and its pointer
228
+ // commit STANDS, taking effect on the next turn (which starts null and
229
+ // establishes the home box) — instead of silently serving the pinned machine as
230
+ // if the clear never happened. Only an EXPLICIT `false` throws; omitted (the
231
+ // common path) resolves to the home default as before. (Lazily establishing the
232
+ // home box mid-turn on such a clear is a deferred follow-up; issue #341.)
233
+ if (deps.defaultIsHome === false) {
234
+ throw new ActiveBackendUnresolvableError(
235
+ "home_unavailable_this_turn",
236
+ "the active sandbox was detached to the session default (home), but this turn started pinned to a machine and established no home box; the detach is accepted and takes effect on the next turn — this turn has no active home box",
237
+ );
238
+ }
121
239
  return { session: deps.defaultBackend, sandboxId: null, kind: deps.defaultKind };
122
240
  }
123
241
 
@@ -129,16 +247,21 @@ export function makeActiveBackendResolver(
129
247
  // to a fresh build fenced under the CURRENT epoch (the stale pinned instance is
130
248
  // fenced at the old epoch and must not be reused).
131
249
  if (
132
- deps.pinnedSelfhosted
133
- && pointer.activeSandboxId === deps.pinnedSelfhosted.sandboxId
134
- && pointer.activeEpoch === deps.pinnedSelfhosted.epoch
250
+ deps.pinnedSelfhosted &&
251
+ pointer.activeSandboxId === deps.pinnedSelfhosted.sandboxId &&
252
+ pointer.activeEpoch === deps.pinnedSelfhosted.epoch
135
253
  ) {
136
- return { session: deps.pinnedSelfhosted.session, sandboxId: pointer.activeSandboxId, kind: "selfhosted" };
254
+ return {
255
+ session: deps.pinnedSelfhosted.session,
256
+ sandboxId: pointer.activeSandboxId,
257
+ kind: "selfhosted",
258
+ };
137
259
  }
138
260
 
139
261
  const sandbox = await deps.getSandbox(pointer.activeSandboxId);
140
262
  if (!sandbox) {
141
263
  throw new ActiveBackendUnresolvableError(
264
+ "stale_pointer",
142
265
  `active sandbox ${pointer.activeSandboxId} not found in workspace ${deps.workspaceId}`,
143
266
  );
144
267
  }
@@ -146,6 +269,7 @@ export function makeActiveBackendResolver(
146
269
  if (sandbox.kind === "selfhosted") {
147
270
  if (!sandbox.enrollmentId) {
148
271
  throw new ActiveBackendUnresolvableError(
272
+ "offline_enrollment",
149
273
  `selfhosted sandbox ${sandbox.id} has no enrollment (agent id) to address`,
150
274
  );
151
275
  }
@@ -161,6 +285,10 @@ export function makeActiveBackendResolver(
161
285
  agentId: sandbox.enrollmentId,
162
286
  epoch: pointer.activeEpoch,
163
287
  ...(deps.selfhostedTimeoutMs !== undefined ? { timeoutMs: deps.selfhostedTimeoutMs } : {}),
288
+ ...(deps.selfhostedExecTimeoutMs !== undefined
289
+ ? { execTimeoutMs: deps.selfhostedExecTimeoutMs }
290
+ : {}),
291
+ ...(deps.selfhostedOnOp !== undefined ? { onOp: deps.selfhostedOnOp } : {}),
164
292
  // The turn's declared environment → the session's manifest.environment, so
165
293
  // the SDK's per-turn manifest-env delta is empty (no "cannot change manifest
166
294
  // environment variables" throw on a pin-to-vm turn).
@@ -169,12 +297,17 @@ export function makeActiveBackendResolver(
169
297
  // for this selfhosted backend. Absent/empty ⇒ the default workspace_root.
170
298
  ...(pointer.workingDir ? { workingDir: pointer.workingDir } : {}),
171
299
  });
172
- return { session: session as RoutableBackendSession, sandboxId: sandbox.id, kind: "selfhosted" };
300
+ return {
301
+ session: session as RoutableBackendSession,
302
+ sandboxId: sandbox.id,
303
+ kind: "selfhosted",
304
+ };
173
305
  }
174
306
 
175
307
  if (sandbox.kind === "modal") {
176
308
  if (!deps.establishModalTarget) {
177
309
  throw new ActiveBackendUnresolvableError(
310
+ "unsupported_backend_context",
178
311
  `modal swap target ${sandbox.id} cannot be established in this context (no establisher wired)`,
179
312
  );
180
313
  }
@@ -183,6 +316,7 @@ export function makeActiveBackendResolver(
183
316
  }
184
317
 
185
318
  throw new ActiveBackendUnresolvableError(
319
+ "unsupported_backend_context",
186
320
  `unsupported swap target kind "${sandbox.kind}" for sandbox ${sandbox.id}`,
187
321
  );
188
322
  };
@@ -28,6 +28,8 @@
28
28
  // `@opengeni/db`.
29
29
 
30
30
  import type { ExposedPortEndpoint } from "../stream-port";
31
+ import { SelfhostedControlError } from "../selfhosted/control-rpc";
32
+ import { renderSelfhostedFault } from "../selfhosted/fault-rendering";
31
33
 
32
34
  /** The per-session active-sandbox pointer the proxy re-reads on every op. Mirror
33
35
  * of `@opengeni/db`'s `ActiveSandboxPointer` (structural, so the leaf does not
@@ -74,7 +76,13 @@ export interface RoutableBackendSession {
74
76
  // `event` is kept `unknown` (mirroring the interface's structural style + avoiding
75
77
  // a proto import into the leaf); the SelfhostedSession takes `DesktopInputRequest["event"]`.
76
78
  desktopInput?(event: unknown): Promise<void>;
77
- screenshot?(): Promise<{ png: Uint8Array; width: number; height: number; nativeWidth: number; nativeHeight: number }>;
79
+ screenshot?(): Promise<{
80
+ png: Uint8Array;
81
+ width: number;
82
+ height: number;
83
+ nativeWidth: number;
84
+ nativeHeight: number;
85
+ }>;
78
86
  }
79
87
 
80
88
  /** The resolved active backend for an epoch: the live session + the sandbox id it
@@ -148,10 +156,12 @@ function isFenceError(error: unknown): boolean {
148
156
  if ((error as { fenced?: unknown }).fenced === true) {
149
157
  return true;
150
158
  }
151
- const name = typeof (error as { name?: unknown }).name === "string" ? (error as { name: string }).name : "";
152
- const message = error instanceof Error ? error.message : String((error as { message?: unknown }).message ?? "");
159
+ const name =
160
+ typeof (error as { name?: unknown }).name === "string" ? (error as { name: string }).name : "";
161
+ const message =
162
+ error instanceof Error ? error.message : String((error as { message?: unknown }).message ?? "");
153
163
  const haystack = `${name} ${message}`.toLowerCase();
154
- return haystack.includes("fenced") || haystack.includes("epoch") && haystack.includes("super");
164
+ return haystack.includes("fenced") || (haystack.includes("epoch") && haystack.includes("super"));
155
165
  }
156
166
 
157
167
  /**
@@ -172,9 +182,17 @@ function isFenceError(error: unknown): boolean {
172
182
  export class RoutingSandboxSession implements RoutableBackendSession {
173
183
  private readonly deps: RoutingSandboxSessionDeps;
174
184
  private readonly maxFenceRetries: number;
175
- // The per-epoch resolved-backend cache. Keyed by activeEpoch: a swap bumps the
176
- // epoch, invalidating the cache so the NEXT op re-resolves the new backend.
185
+ // The resolved-backend cache. Keyed by the FULL pointer tuple
186
+ // `(activeEpoch, activeSandboxId)` — NOT the epoch alone. A swap bumps the epoch,
187
+ // but a pointer can also change its target id WITHOUT an epoch bump: the
188
+ // `sessions.active_sandbox_id` FK is `ON DELETE SET NULL`, so a cascade that
189
+ // deletes the pointed-at sandbox row nulls the id at the SAME epoch. Keying on the
190
+ // epoch alone would then keep serving the deleted/stale backend for that epoch
191
+ // (issue #341 §5.2 — a swap-free route to the Shape-3 symptom). Keying on the tuple
192
+ // makes any target change — epoch-bumped OR not — invalidate the cache so the next
193
+ // op re-resolves (and, for a null id, re-resolves the session HOME).
177
194
  private cachedEpoch: number | undefined;
195
+ private cachedSandboxId: string | null | undefined;
178
196
  private cached: ResolvedActiveBackend | undefined;
179
197
  // The last-resolved backend, exposed via the `state` getter (a method-free read
180
198
  // of the active backend's `state`). Updated on every resolve.
@@ -189,7 +207,13 @@ export class RoutingSandboxSession implements RoutableBackendSession {
189
207
  // that cannot serve them. So the constructor assigns them ONLY when the
190
208
  // construction-time default backend actually implements the native surface (below).
191
209
  desktopInput?: (event: unknown) => Promise<void>;
192
- screenshot?: () => Promise<{ png: Uint8Array; width: number; height: number; nativeWidth: number; nativeHeight: number }>;
210
+ screenshot?: () => Promise<{
211
+ png: Uint8Array;
212
+ width: number;
213
+ height: number;
214
+ nativeWidth: number;
215
+ nativeHeight: number;
216
+ }>;
193
217
 
194
218
  constructor(deps: RoutingSandboxSessionDeps) {
195
219
  this.deps = deps;
@@ -252,16 +276,34 @@ export class RoutingSandboxSession implements RoutableBackendSession {
252
276
  */
253
277
  private async resolve(): Promise<ResolvedActiveBackend> {
254
278
  const pointer = await this.deps.readPointer();
255
- if (this.cachedEpoch === pointer.activeEpoch && this.cached) {
279
+ if (
280
+ this.cachedEpoch === pointer.activeEpoch &&
281
+ this.cachedSandboxId === pointer.activeSandboxId &&
282
+ this.cached
283
+ ) {
256
284
  return this.cached;
257
285
  }
258
286
  const fromEpoch = this.cachedEpoch ?? pointer.activeEpoch;
259
287
  const resolved = await this.deps.resolveActiveBackend(pointer);
288
+ // Re-entrancy guard: a resolver that returns THIS proxy as the active backend
289
+ // makes every op (exec/readFile/…) dispatch back into resolve() -> the same
290
+ // backend -> forever (a silent async infinite recursion that HANGS the turn,
291
+ // not a stack overflow). Fail loud instead — a wiring bug must surface as a
292
+ // legible error, never a hung turn.
293
+ if ((resolved.session as unknown) === this) {
294
+ throw new Error(
295
+ "RoutingSandboxSession.resolveActiveBackend returned the proxy itself as the active backend (re-entrancy) — the resolver must return the underlying box session, not the routing proxy.",
296
+ );
297
+ }
260
298
  this.cachedEpoch = pointer.activeEpoch;
299
+ this.cachedSandboxId = pointer.activeSandboxId;
261
300
  this.cached = resolved;
262
301
  this.lastResolved = resolved;
263
302
  this.deps.onTransition?.({
264
- type: this.cachedEpoch !== undefined && fromEpoch !== pointer.activeEpoch ? "epoch-changed" : "resolved",
303
+ type:
304
+ this.cachedEpoch !== undefined && fromEpoch !== pointer.activeEpoch
305
+ ? "epoch-changed"
306
+ : "resolved",
265
307
  fromEpoch,
266
308
  toEpoch: pointer.activeEpoch,
267
309
  sandboxId: resolved.sandboxId,
@@ -281,7 +323,10 @@ export class RoutingSandboxSession implements RoutableBackendSession {
281
323
  * A non-fence error propagates immediately (it is a real op failure, not a swap
282
324
  * race).
283
325
  */
284
- private async dispatch<T>(op: string, fn: (session: RoutableBackendSession) => Promise<T>): Promise<T> {
326
+ private async dispatch<T>(
327
+ op: string,
328
+ fn: (session: RoutableBackendSession) => Promise<T>,
329
+ ): Promise<T> {
285
330
  let attempt = 0;
286
331
  let lastError: unknown;
287
332
  while (attempt <= this.maxFenceRetries) {
@@ -297,6 +342,7 @@ export class RoutingSandboxSession implements RoutableBackendSession {
297
342
  // active sandbox (the fenced-retry role). Bounded by maxFenceRetries.
298
343
  lastError = error;
299
344
  this.cachedEpoch = undefined;
345
+ this.cachedSandboxId = undefined;
300
346
  this.cached = undefined;
301
347
  this.deps.onTransition?.({
302
348
  type: "fenced-retry",
@@ -313,6 +359,31 @@ export class RoutingSandboxSession implements RoutableBackendSession {
313
359
  throw lastError ?? new Error(`routing op "${op}" exhausted fence retries`);
314
360
  }
315
361
 
362
+ /**
363
+ * The failure-visibility boundary for the `exec_command` SDK capability tool — the
364
+ * dominant fault surface, and the one whose thrown `SelfhostedControlError` reaches
365
+ * the model wrapped by the SDK's generic tool-error function as
366
+ * "…Please try again. Error: …" — actively wrong for a machine that is offline, a
367
+ * consent that is not granted, or an oversized reply. Since the SDK closure-captures
368
+ * its `errorFunction` (there is no seam to attach one to its internally-built tools),
369
+ * we render the fault into the doctrine's four fields HERE and return it as the
370
+ * tool's string result — legible, in-band, and free of the misleading wrapper.
371
+ * (`apply_patch` already surfaces `error.message` via the SDK's own catch; the
372
+ * skills / `view_image` tools consume their session methods internally, so they are
373
+ * rendered elsewhere or left to their own SDK renderers — not this string boundary.)
374
+ *
375
+ * A FENCE error is re-thrown, NEVER rendered: `dispatch` already retries it against
376
+ * a re-resolved backend, and a fence that escapes retries is a routing condition the
377
+ * turn handles, not a model-facing fault. Any non-selfhosted error (a Modal fault, a
378
+ * `RoutingUnsupportedError`) is re-thrown unchanged.
379
+ */
380
+ private renderSelfhostedFaultOrThrow(error: unknown): string {
381
+ if (error instanceof SelfhostedControlError && !error.fenced) {
382
+ return renderSelfhostedFault(error);
383
+ }
384
+ throw error;
385
+ }
386
+
316
387
  // ── The forwarded structural surface ──────────────────────────────────────
317
388
  // Every method is PRESENT on the proxy (the SDK binds presence once) and
318
389
  // dispatches to the active backend at call-time. A missing backend method
@@ -332,16 +403,22 @@ export class RoutingSandboxSession implements RoutableBackendSession {
332
403
  }
333
404
 
334
405
  async execCommand(args: unknown): Promise<string> {
335
- return this.dispatch("execCommand", async (s) => {
336
- if (s.execCommand) {
337
- return s.execCommand(args);
338
- }
339
- if (s.exec) {
340
- const r = (await s.exec(args)) as { stdout?: string; output?: string };
341
- return r.stdout ?? r.output ?? "";
342
- }
343
- throw new RoutingUnsupportedError("execCommand", this.cached?.kind ?? "unknown");
344
- });
406
+ try {
407
+ return await this.dispatch("execCommand", async (s) => {
408
+ if (s.execCommand) {
409
+ return s.execCommand(args);
410
+ }
411
+ if (s.exec) {
412
+ const r = (await s.exec(args)) as { stdout?: string; output?: string };
413
+ return r.stdout ?? r.output ?? "";
414
+ }
415
+ throw new RoutingUnsupportedError("execCommand", this.cached?.kind ?? "unknown");
416
+ });
417
+ } catch (error) {
418
+ // Render a terminal selfhosted fault as the tool's result (four fields, correct
419
+ // verdict) instead of letting the SDK mislabel it "Please try again".
420
+ return this.renderSelfhostedFaultOrThrow(error);
421
+ }
345
422
  }
346
423
 
347
424
  async writeStdin(args: unknown): Promise<string> {
@@ -417,12 +494,39 @@ export class RoutingSandboxSession implements RoutableBackendSession {
417
494
  return Boolean(s?.supportsPty?.());
418
495
  }
419
496
 
420
- /** createEditor is a synchronous factory in the SDK surface; it binds to the
421
- * last-resolved backend's editor (or the default backend before the first op).
422
- * Returns undefined when the active backend has no editor (channel-a falls back
423
- * to its exec-based write path). */
497
+ /** createEditor is a synchronous factory in the SDK surface. The SDK's filesystem
498
+ * capability calls it ONCE at tool-BIND time — `FilesystemCapability.tools()`,
499
+ * every turn, before any tool runs — and throws "Filesystem sandbox sessions must
500
+ * provide createEditor()" if it returns falsy. When a backend is already resolved
501
+ * (eager/selfhosted routing) we bind to its editor directly, byte-for-byte as
502
+ * before. But under LAZY provisioning the backend is not established yet
503
+ * (defaultResolved is the synthetic unprovisioned session with no editor), so a
504
+ * direct delegate returns undefined and every lazy turn would die at bind. Return a
505
+ * LAZY EDITOR PROXY instead: a non-null editor whose async ops resolve the active
506
+ * backend (establishing the box on first use, via `dispatch`) and delegate to its
507
+ * real editor — mirroring how this proxy defers exec/readFile. */
424
508
  createEditor(runAs?: string): unknown {
425
- return (this.lastResolved ?? this.deps.defaultResolved)?.session.createEditor?.(runAs);
509
+ const eager = (this.lastResolved ?? this.deps.defaultResolved)?.session.createEditor?.(runAs);
510
+ if (eager) {
511
+ return eager;
512
+ }
513
+ const op =
514
+ (name: "createFile" | "updateFile" | "deleteFile") =>
515
+ (operation: unknown, context?: unknown): Promise<unknown> =>
516
+ this.dispatch(`editor.${name}`, async (s) => {
517
+ const editor = s.createEditor?.(runAs) as
518
+ | Record<string, (operation: unknown, context?: unknown) => Promise<unknown>>
519
+ | undefined;
520
+ if (!editor?.[name]) {
521
+ throw new RoutingUnsupportedError(`editor.${name}`, this.cached?.kind ?? "unknown");
522
+ }
523
+ return editor[name](operation, context);
524
+ });
525
+ return {
526
+ createFile: op("createFile"),
527
+ updateFile: op("updateFile"),
528
+ deleteFile: op("deleteFile"),
529
+ };
426
530
  }
427
531
 
428
532
  async resolveExposedPort(port: number): Promise<ExposedPortEndpoint> {
@@ -165,7 +165,14 @@ export function negotiateCapabilities(ctx: NegotiationContext): SessionCapabilit
165
165
 
166
166
  const fileSystem = (() => {
167
167
  if (osReason) {
168
- return { available: false, readOnly: true, root: descriptor.workspaceRoot, pathSep: "/" as const, treeMode: "lazy" as const, reason: osReason };
168
+ return {
169
+ available: false,
170
+ readOnly: true,
171
+ root: descriptor.workspaceRoot,
172
+ pathSep: "/" as const,
173
+ treeMode: "lazy" as const,
174
+ reason: osReason,
175
+ };
169
176
  }
170
177
  const cap = descriptor.capabilities.FileSystem;
171
178
  return {
@@ -181,10 +188,26 @@ export function negotiateCapabilities(ctx: NegotiationContext): SessionCapabilit
181
188
  const terminal = (() => {
182
189
  const cap = descriptor.capabilities.Terminal;
183
190
  if (osReason) {
184
- return { transport: null, ptyCapable: false, shell: "/bin/bash", url: null, token: null, expiresAt: null, reason: osReason };
191
+ return {
192
+ transport: null,
193
+ ptyCapable: false,
194
+ shell: "/bin/bash",
195
+ url: null,
196
+ token: null,
197
+ expiresAt: null,
198
+ reason: osReason,
199
+ };
185
200
  }
186
201
  if (!cap.available) {
187
- return { transport: null, ptyCapable: false, shell: "/bin/bash", url: null, token: null, expiresAt: null, reason: "backend_unsupported" as const };
202
+ return {
203
+ transport: null,
204
+ ptyCapable: false,
205
+ shell: "/bin/bash",
206
+ url: null,
207
+ token: null,
208
+ expiresAt: null,
209
+ reason: "backend_unsupported" as const,
210
+ };
188
211
  }
189
212
  // The REAL PTY (ttyd pty-ws) rides the SAME tunnel as the desktop, so it is
190
213
  // gated identically: a real-PTY backend (cap.pty), the terminal policy toggle
@@ -231,7 +254,11 @@ export function negotiateCapabilities(ctx: NegotiationContext): SessionCapabilit
231
254
  if (osReason) {
232
255
  return { available: false, repos: [], reason: osReason };
233
256
  }
234
- return { available: cap.available, repos: [], reason: cap.available ? null : ("backend_unsupported" as const) };
257
+ return {
258
+ available: cap.available,
259
+ repos: [],
260
+ reason: cap.available ? null : ("backend_unsupported" as const),
261
+ };
235
262
  })();
236
263
 
237
264
  const desktop = (() => {
@@ -317,14 +344,32 @@ export function negotiateCapabilities(ctx: NegotiationContext): SessionCapabilit
317
344
  const recording = (() => {
318
345
  const cap = descriptor.capabilities.Recording;
319
346
  if (osReason) {
320
- return { available: false, modes: [] as ("manual" | "on-turn" | "on-verify")[], codecs: [] as ("h264-mp4" | "vp9-webm")[], reason: osReason };
347
+ return {
348
+ available: false,
349
+ modes: [] as ("manual" | "on-turn" | "on-verify")[],
350
+ codecs: [] as ("h264-mp4" | "vp9-webm")[],
351
+ reason: osReason,
352
+ };
321
353
  }
322
354
  if (!cap.available) {
323
- return { available: false, modes: [] as ("manual" | "on-turn" | "on-verify")[], codecs: [] as ("h264-mp4" | "vp9-webm")[], reason: descriptor.tier === "headless" ? ("tier_headless" as const) : ("backend_unsupported" as const) };
355
+ return {
356
+ available: false,
357
+ modes: [] as ("manual" | "on-turn" | "on-verify")[],
358
+ codecs: [] as ("h264-mp4" | "vp9-webm")[],
359
+ reason:
360
+ descriptor.tier === "headless"
361
+ ? ("tier_headless" as const)
362
+ : ("backend_unsupported" as const),
363
+ };
324
364
  }
325
365
  // Recording feasibility tracks desktop; policy-gate it the same way.
326
366
  if (!ctx.desktopEnabled) {
327
- return { available: false, modes: [] as ("manual" | "on-turn" | "on-verify")[], codecs: [] as ("h264-mp4" | "vp9-webm")[], reason: "disabled_by_policy" as const };
367
+ return {
368
+ available: false,
369
+ modes: [] as ("manual" | "on-turn" | "on-verify")[],
370
+ codecs: [] as ("h264-mp4" | "vp9-webm")[],
371
+ reason: "disabled_by_policy" as const,
372
+ };
328
373
  }
329
374
  return {
330
375
  available: true,
@@ -345,7 +390,14 @@ export function negotiateCapabilities(ctx: NegotiationContext): SessionCapabilit
345
390
  return { available: false, readOnly, reason: osReason };
346
391
  }
347
392
  if (!desktopCapable) {
348
- return { available: false, readOnly, reason: descriptor.tier === "headless" ? ("tier_headless" as const) : ("backend_unsupported" as const) };
393
+ return {
394
+ available: false,
395
+ readOnly,
396
+ reason:
397
+ descriptor.tier === "headless"
398
+ ? ("tier_headless" as const)
399
+ : ("backend_unsupported" as const),
400
+ };
349
401
  }
350
402
  if (!ctx.desktopEnabled || ctx.computerUseEnabled === false) {
351
403
  return { available: false, readOnly, reason: "disabled_by_policy" as const };