@stigmer/runner 3.8.0 → 3.10.0

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 (64) hide show
  1. package/README.md +2 -2
  2. package/dist/.build-fingerprint +1 -1
  3. package/dist/activities/execute-cursor/attachment-resolver.d.ts +16 -0
  4. package/dist/activities/execute-cursor/attachment-resolver.js +63 -4
  5. package/dist/activities/execute-cursor/attachment-resolver.js.map +1 -1
  6. package/dist/activities/execute-cursor/index.d.ts +15 -0
  7. package/dist/activities/execute-cursor/index.js +73 -11
  8. package/dist/activities/execute-cursor/index.js.map +1 -1
  9. package/dist/activities/execute-cursor/prompt-builder.d.ts +14 -1
  10. package/dist/activities/execute-cursor/prompt-builder.js +11 -2
  11. package/dist/activities/execute-cursor/prompt-builder.js.map +1 -1
  12. package/dist/activities/execute-deep-agent/attachment-injector.d.ts +17 -0
  13. package/dist/activities/execute-deep-agent/attachment-injector.js +34 -4
  14. package/dist/activities/execute-deep-agent/attachment-injector.js.map +1 -1
  15. package/dist/activities/execute-deep-agent/hitl.d.ts +15 -7
  16. package/dist/activities/execute-deep-agent/hitl.js +6 -15
  17. package/dist/activities/execute-deep-agent/hitl.js.map +1 -1
  18. package/dist/activities/execute-deep-agent/index.js +4 -1
  19. package/dist/activities/execute-deep-agent/index.js.map +1 -1
  20. package/dist/activities/execute-deep-agent/prompt-builder.d.ts +17 -0
  21. package/dist/activities/execute-deep-agent/prompt-builder.js +13 -2
  22. package/dist/activities/execute-deep-agent/prompt-builder.js.map +1 -1
  23. package/dist/activities/execute-deep-agent/setup.js +49 -3
  24. package/dist/activities/execute-deep-agent/setup.js.map +1 -1
  25. package/dist/runner-manager.js +14 -0
  26. package/dist/runner-manager.js.map +1 -1
  27. package/dist/runner.js +14 -0
  28. package/dist/runner.js.map +1 -1
  29. package/dist/shared/artifact-storage.d.ts +10 -0
  30. package/dist/shared/artifact-storage.js +49 -8
  31. package/dist/shared/artifact-storage.js.map +1 -1
  32. package/dist/shared/attachment-vision.d.ts +244 -0
  33. package/dist/shared/attachment-vision.js +330 -0
  34. package/dist/shared/attachment-vision.js.map +1 -0
  35. package/dist/shared/mcp-manager.d.ts +14 -1
  36. package/dist/shared/mcp-manager.js +20 -21
  37. package/dist/shared/mcp-manager.js.map +1 -1
  38. package/dist/shared/model-registry.d.ts +20 -2
  39. package/dist/shared/model-registry.js +37 -2
  40. package/dist/shared/model-registry.js.map +1 -1
  41. package/package.json +3 -3
  42. package/src/activities/execute-cursor/__tests__/attachment-resolver.test.ts +218 -0
  43. package/src/activities/execute-cursor/__tests__/build-prompt.test.ts +105 -3
  44. package/src/activities/execute-cursor/attachment-resolver.ts +97 -4
  45. package/src/activities/execute-cursor/index.ts +94 -13
  46. package/src/activities/execute-cursor/prompt-builder.ts +27 -2
  47. package/src/activities/execute-deep-agent/__tests__/attachment-injector.test.ts +207 -0
  48. package/src/activities/execute-deep-agent/__tests__/hitl.test.ts +13 -13
  49. package/src/activities/execute-deep-agent/__tests__/vision-input.test.ts +152 -0
  50. package/src/activities/execute-deep-agent/attachment-injector.ts +65 -4
  51. package/src/activities/execute-deep-agent/hitl.ts +14 -19
  52. package/src/activities/execute-deep-agent/index.ts +4 -5
  53. package/src/activities/execute-deep-agent/prompt-builder.ts +37 -2
  54. package/src/activities/execute-deep-agent/setup.ts +58 -3
  55. package/src/runner-manager.ts +19 -0
  56. package/src/runner.ts +19 -0
  57. package/src/shared/__tests__/artifact-storage.test.ts +76 -1
  58. package/src/shared/__tests__/attachment-vision.test.ts +420 -0
  59. package/src/shared/__tests__/mcp-manager.test.ts +87 -1
  60. package/src/shared/__tests__/model-registry.test.ts +71 -0
  61. package/src/shared/artifact-storage.ts +55 -8
  62. package/src/shared/attachment-vision.ts +456 -0
  63. package/src/shared/mcp-manager.ts +22 -22
  64. package/src/shared/model-registry.ts +50 -2
@@ -35,7 +35,7 @@ import type { SubAgentExecution } from "@stigmer/protos/ai/stigmer/agentic/agent
35
35
  import type { PendingApproval } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/approval_pb";
36
36
  import type { AgentExecution, AgentExecutionStatus } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/api_pb";
37
37
  import { ExecutionControlSignal, ExecutionPhase, FileChangeSetStatus, InteractionMode, MessageType, ApprovalAction } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/enum_pb";
38
- import type { Run, ConversationTurn } from "@cursor/sdk";
38
+ import type { Run, ConversationTurn, SDKUserMessage } from "@cursor/sdk";
39
39
 
40
40
  import type { Config } from "../../config.js";
41
41
  import { StigmerClient } from "../../client/stigmer-client.js";
@@ -58,6 +58,13 @@ import { readSessionContext } from "../../shared/session-context.js";
58
58
  import { withholdSecretContentFromMessages } from "../../shared/tool-row.js";
59
59
  import { StallTimeoutError, formatStallFailure } from "../../shared/stall-watchdog.js";
60
60
  import { resolveUsableArtifactStorage, loadArtifactStorageConfig, type ArtifactStorage } from "../../shared/artifact-storage.js";
61
+ import {
62
+ CURSOR_VISION_PROFILE,
63
+ VisionBudget,
64
+ toCursorImages,
65
+ type NotViewableEntry,
66
+ } from "../../shared/attachment-vision.js";
67
+ import { getModelVisionCapability } from "../../shared/model-registry.js";
61
68
  import { publishPlanArtifact } from "../../shared/plan-artifact.js";
62
69
  import { DeltaEnricher } from "./delta-enricher.js";
63
70
  import { TodoTracker } from "./todo-tracker.js";
@@ -84,7 +91,7 @@ import { buildCursorSubAgentDefinitions } from "./subagent-config.js";
84
91
  import { resolveSkills } from "./skill-resolver.js";
85
92
  import { removeStigmerSymlink } from "../../shared/workspace/stigmer-link.js";
86
93
  import { resolveAttachments } from "./attachment-resolver.js";
87
- import { buildEnhancedPrompt, buildReinvocationPrompt, formatConversationCatchupSection, formatInteractionModePrefix, formatImplementPlanSection } from "./prompt-builder.js";
94
+ import { buildEnhancedPrompt, buildReinvocationPrompt, formatConversationCatchupSection, formatInputFiles, formatInteractionModePrefix, formatImplementPlanSection } from "./prompt-builder.js";
88
95
  import { installHitlGate, removeHitlGate } from "./workspace-setup.js";
89
96
  import { ensureHitlDir } from "../../shared/workspace/platform-dir.js";
90
97
  import {
@@ -761,14 +768,44 @@ async function executeCursorInner(
761
768
 
762
769
  // Phase 5b: Resolve attachments (fail-hard — explicit user inputs; see
763
770
  // attachment-resolver.ts). Downloads by storage key through the same
764
- // artifactStorage resolved for status offload above.
771
+ // artifactStorage resolved for status offload above. The vision budget
772
+ // rides along so image attachments are selected for inline delivery while
773
+ // their bytes are already in hand (attachment-vision.ts owns all policy).
774
+ // The budget also carries the requested model's registry vision
775
+ // capability, looked up from the raw executionConfig name — full model
776
+ // validation (Phase 6) isn't needed for this, and ""/"default" (the Auto
777
+ // pool) resolves to unknown, which the policy treats as sighted.
778
+ const visionBudget = new VisionBudget(CURSOR_VISION_PROFILE, {
779
+ modelVision: await getModelVisionCapability(spec.executionConfig?.modelName ?? ""),
780
+ });
765
781
  const attachmentResults = await resolveAttachments(spec.attachments, {
766
782
  sessionId,
767
783
  primaryWorkspaceDir,
768
784
  mode: config.mode,
769
785
  storage: artifactStorage,
786
+ visionBudget,
770
787
  });
771
788
  const attachmentPaths = attachmentResults.map((a) => a.relativePath);
789
+ // Vision facts, derived once from the single resolution result: the
790
+ // images the model will see inline (in attachment order) and the ones
791
+ // that degraded to path-only, disclosed in the prompt.
792
+ const visionImages = attachmentResults.flatMap((a) => (a.vision ? [a.vision] : []));
793
+ const visionNotViewable: NotViewableEntry[] = attachmentResults.flatMap((a) =>
794
+ a.visionDegraded ? [{ path: a.relativePath, reason: a.visionDegraded }] : [],
795
+ );
796
+ const visionPromptInfo = visionImages.length > 0 || visionNotViewable.length > 0
797
+ ? {
798
+ inlineFilenames: visionImages.map((v) => v.filename),
799
+ notViewable: visionNotViewable,
800
+ }
801
+ : undefined;
802
+ if (visionPromptInfo) {
803
+ console.log(
804
+ `[attachment-vision] execution=${executionId} inline=${visionImages.length} ` +
805
+ `(${visionImages.reduce((n, v) => n + v.byteSize, 0)} bytes) ` +
806
+ `degraded=${JSON.stringify(visionNotViewable.map((d) => `${d.path}:${d.reason}`))}`,
807
+ );
808
+ }
772
809
  setupTiming.mark("resolve_attachments");
773
810
 
774
811
  // Phase 5b3: Exact-apply approved whole-file writes (HITL "what you approve
@@ -1094,6 +1131,7 @@ async function executeCursorInner(
1094
1131
  workspaceDirs: blueprint.workspaceDirs,
1095
1132
  workspaceFileRefs: spec.workspaceFileRefs ?? [],
1096
1133
  attachmentPaths,
1134
+ vision: visionPromptInfo,
1097
1135
  pendingApprovals: adjudicatedApprovals,
1098
1136
  appliedToolCallIds,
1099
1137
  interactionMode,
@@ -1111,6 +1149,20 @@ async function executeCursorInner(
1111
1149
  effectivePrompt += `\n\n---\nCRITICAL OUTPUT REQUIREMENT:\nYour final response MUST be a single valid JSON object (no markdown, no commentary, no code fences) that matches this schema:\n${schemaStr}\n\nRespond with ONLY the JSON object. Nothing else.`;
1112
1150
  }
1113
1151
 
1152
+ // Phase 10a1: The turn's vision payload. The invariant is "images
1153
+ // accompany the user's turn message — where the message goes, they go":
1154
+ // every send that delivers this turn's message carries them (the primary
1155
+ // send and both fresh-agent recovery retries, whose empty conversations
1156
+ // genuinely need the re-send), while a HITL re-invocation — whose prompt
1157
+ // carries no user message and whose resumed agent already holds the
1158
+ // images in its native conversation — carries none. Computed ONCE here so
1159
+ // all send sites agree by construction.
1160
+ const turnImages = isHitlReinvocation(approvalDecisions)
1161
+ ? []
1162
+ : toCursorImages(visionImages);
1163
+ const toSendMessage = (sendPrompt: string): string | SDKUserMessage =>
1164
+ turnImages.length > 0 ? { text: sendPrompt, images: turnImages } : sendPrompt;
1165
+
1114
1166
  // Phase 10a2: Log Stigmer preamble size for context trimming diagnostics
1115
1167
  const promptChars = effectivePrompt.length;
1116
1168
  const promptEstimatedTokens = Math.ceil(promptChars / 4);
@@ -1225,7 +1277,7 @@ async function executeCursorInner(
1225
1277
  // The stall watchdog is armed inside consumeCursorTurnStream (it needs the
1226
1278
  // run to cancel), stored on turnState.stallWatchdog so this shared onDelta can
1227
1279
  // reset it and the activity's finally can stop it as a backstop.
1228
- const run = await resolution.agent.send(effectivePrompt, {
1280
+ const run = await resolution.agent.send(toSendMessage(effectivePrompt), {
1229
1281
  onDelta: (event) => {
1230
1282
  if (!turnFirstEventEmitted) {
1231
1283
  turnFirstEventEmitted = true;
@@ -1596,7 +1648,10 @@ async function executeCursorInner(
1596
1648
  // poisoned-handle path leaked the fresh agent — it closed the stale one.)
1597
1649
  resolution = { ...resolution, agent: freshAgent, agentId: freshAgent.agentId, isNew: true };
1598
1650
  turnState.streamErrorMessage = undefined;
1599
- const retryRun = await freshAgent.send(retryPrompt, {
1651
+ // The retry carries the turn's images too (same toSendMessage): the
1652
+ // fresh agent's conversation is empty, so skipping them here would
1653
+ // silently lose the user's photo on a recovered turn.
1654
+ const retryRun = await freshAgent.send(toSendMessage(retryPrompt), {
1600
1655
  onDelta: makeCursorTurnOnDelta(onDeltaDeps),
1601
1656
  });
1602
1657
  await consumeCursorTurnStream(retryRun, streamDeps);
@@ -1735,6 +1790,7 @@ async function executeCursorInner(
1735
1790
  workspaceDirs: blueprint.workspaceDirs,
1736
1791
  workspaceFileRefs: spec.workspaceFileRefs ?? [],
1737
1792
  attachmentPaths,
1793
+ vision: visionPromptInfo,
1738
1794
  pendingApprovals: adjudicatedApprovals,
1739
1795
  interactionMode,
1740
1796
  // buildFromPlan was silently dropped here until T03 Sitting 3 —
@@ -2310,6 +2366,12 @@ export interface BuildPromptInput {
2310
2366
  workspaceDirs: string[];
2311
2367
  workspaceFileRefs: string[];
2312
2368
  attachmentPaths: string[];
2369
+ /**
2370
+ * Vision facts for the input-files section (T04): which attachments the
2371
+ * model sees inline and which degraded to path-only. PER-TURN like the
2372
+ * catchup — it rides both the enhanced prompt and a resumed turn's prefix.
2373
+ */
2374
+ vision?: import("./prompt-builder.js").VisionPromptInfo;
2313
2375
  pendingApprovals: import("@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/approval_pb").PendingApproval[];
2314
2376
  /**
2315
2377
  * Approved whole-file writes the runner already applied itself (exact-apply).
@@ -2372,6 +2434,20 @@ export interface BuildPromptInput {
2372
2434
  * 3. first execution / fresh -> buildEnhancedPrompt (full instructions +
2373
2435
  * agent after resume failure skills; no prior conversation to inherit)
2374
2436
  */
2437
+ /**
2438
+ * Whether this activity invocation is a HITL re-invocation — the turn resumes
2439
+ * an agent purely to convey approval decisions, carrying NO user message.
2440
+ * The single discriminator for everything that must ride with the user's
2441
+ * message and nothing else: the reinvocation prompt shape (below) and the
2442
+ * vision payload (images accompany the message; a resumed agent already holds
2443
+ * them in its native conversation).
2444
+ */
2445
+ export function isHitlReinvocation(
2446
+ approvalDecisions: Map<string, ApprovalAction> | undefined,
2447
+ ): approvalDecisions is Map<string, ApprovalAction> {
2448
+ return approvalDecisions !== undefined && approvalDecisions.size > 0;
2449
+ }
2450
+
2375
2451
  export function buildPrompt(input: BuildPromptInput): string {
2376
2452
  const {
2377
2453
  resolution,
@@ -2388,12 +2464,10 @@ export function buildPrompt(input: BuildPromptInput): string {
2388
2464
  conversationCatchup,
2389
2465
  } = input;
2390
2466
 
2391
- const isHitlReinvocation = approvalDecisions !== undefined && approvalDecisions.size > 0;
2392
-
2393
2467
  // HITL reinvocation: the agent is resumed, so its native context carries the
2394
2468
  // prior conversation; the reinvocation prompt conveys the approval decisions
2395
2469
  // (and which approved writes the runner already exact-applied).
2396
- if (isHitlReinvocation) {
2470
+ if (isHitlReinvocation(approvalDecisions)) {
2397
2471
  return buildReinvocationPrompt(
2398
2472
  input.pendingApprovals,
2399
2473
  approvalDecisions,
@@ -2407,15 +2481,21 @@ export function buildPrompt(input: BuildPromptInput): string {
2407
2481
  // session's first turn: the interaction-mode prefix (a follow-up can switch
2408
2482
  // Agent→Plan mid-session, and for Cursor the prompt is the only plan-mode
2409
2483
  // enforcement), the implement-plan directive (the build turn is usually a
2410
- // follow-up on a resumed agent), and the conversation catchup (handback
2411
- // ALWAYS lands mid-session on a resumed agent this prefix is the property
2412
- // the metadata lane structurally cannot deliver, cloud DD-006). Catchup
2413
- // last: it is context, and context sits closest to the task (the enhanced
2414
- // prompt's own ordering doctrine).
2484
+ // follow-up on a resumed agent), THIS turn's attachments (spec.attachments
2485
+ // is per-execution a file sent on a follow-up turn materializes for this
2486
+ // turn and would otherwise never be announced at all), and the conversation
2487
+ // catchup (handback ALWAYS lands mid-session on a resumed agent this
2488
+ // prefix is the property the metadata lane structurally cannot deliver,
2489
+ // cloud DD-006). Catchup last: it is context, and context sits closest to
2490
+ // the task (the enhanced prompt's own ordering doctrine); the input files
2491
+ // precede it because they are this turn's payload, not background.
2415
2492
  if (resolution.reason === "resumed_successfully") {
2416
2493
  const prefixes = [
2417
2494
  formatInteractionModePrefix(interactionMode),
2418
2495
  formatImplementPlanSection(buildFromPlan, attachmentPaths),
2496
+ attachmentPaths.length > 0
2497
+ ? formatInputFiles(attachmentPaths, input.vision)
2498
+ : undefined,
2419
2499
  conversationCatchup !== undefined
2420
2500
  ? formatConversationCatchupSection(conversationCatchup)
2421
2501
  : undefined,
@@ -2438,6 +2518,7 @@ export function buildPrompt(input: BuildPromptInput): string {
2438
2518
  workspaceDirs,
2439
2519
  workspaceFileRefs,
2440
2520
  attachmentPaths,
2521
+ vision: input.vision,
2441
2522
  interactionMode,
2442
2523
  buildFromPlan,
2443
2524
  contextBridge: input.contextBridge,
@@ -32,6 +32,10 @@ import {
32
32
  type SenderIdentity,
33
33
  } from "../../shared/sender-identity.js";
34
34
  import { formatSessionContextText } from "../../shared/session-context.js";
35
+ import {
36
+ visionDisclosureLines,
37
+ type NotViewableEntry,
38
+ } from "../../shared/attachment-vision.js";
35
39
  import { PLAN_MODE_DIRECTIVE } from "../../shared/plan-mode-prompt.js";
36
40
  import {
37
41
  buildImplementPlanDirective,
@@ -56,6 +60,17 @@ export interface SkillMetadata {
56
60
  path: string;
57
61
  }
58
62
 
63
+ /**
64
+ * Vision facts about this turn's attachments, rendered inside the
65
+ * input-files section: which images the model can see inline (in send
66
+ * order) and which degraded to the file-pointer story. Derived per turn
67
+ * from the resolved attachments — never carried across turns.
68
+ */
69
+ export interface VisionPromptInfo {
70
+ inlineFilenames: string[];
71
+ notViewable: NotViewableEntry[];
72
+ }
73
+
59
74
  export interface EnhancedPromptOptions {
60
75
  instructions: string;
61
76
  userMessage: string;
@@ -76,6 +91,8 @@ export interface EnhancedPromptOptions {
76
91
  workspaceDirs: string[];
77
92
  workspaceFileRefs: string[];
78
93
  attachmentPaths: string[];
94
+ /** Inline/degraded image facts for the input-files section (T04 vision). */
95
+ vision?: VisionPromptInfo;
79
96
  interactionMode?: InteractionMode;
80
97
  /**
81
98
  * The execution is a Build-from-plan turn (spec.execution_config
@@ -179,7 +196,7 @@ export function buildEnhancedPrompt(options: EnhancedPromptOptions): string {
179
196
  }
180
197
 
181
198
  if (options.attachmentPaths.length > 0) {
182
- sections.push(formatInputFiles(options.attachmentPaths));
199
+ sections.push(formatInputFiles(options.attachmentPaths, options.vision));
183
200
  }
184
201
 
185
202
  if (options.workspaceFileRefs.length > 0) {
@@ -462,12 +479,20 @@ export function formatWorkspaceContext(dirs: string[]): string {
462
479
  ].join("\n");
463
480
  }
464
481
 
465
- export function formatInputFiles(paths: string[]): string {
482
+ export function formatInputFiles(paths: string[], vision?: VisionPromptInfo): string {
466
483
  const entries = paths.map((p) => `- \`${p}\``);
484
+ // The vision lines (shared wording, attachment-vision.ts) tell the model
485
+ // which of these files it can already SEE inline versus which degraded to
486
+ // path-only — without them an agent silently ignores a photo the user
487
+ // believes it can see.
488
+ const disclosure = vision
489
+ ? visionDisclosureLines(vision.inlineFilenames, vision.notViewable)
490
+ : [];
467
491
  return [
468
492
  "<input_files>",
469
493
  "The following files have been provided as inputs. Read them when relevant to the task:",
470
494
  ...entries,
495
+ ...disclosure,
471
496
  "</input_files>",
472
497
  ].join("\n");
473
498
  }
@@ -15,6 +15,10 @@ import {
15
15
  } from "../attachment-injector.js";
16
16
  import { mockWorkspaceBackend } from "../../../__test-utils__/mock-workspace.js";
17
17
  import { makeInMemoryArtifactStorage } from "../../../__test-utils__/fake-artifact-storage.js";
18
+ import {
19
+ DEEP_AGENT_VISION_PROFILE,
20
+ VisionBudget,
21
+ } from "../../../shared/attachment-vision.js";
18
22
 
19
23
  // ── ZIP Construction Helpers ─────────────────────────────────────────
20
24
 
@@ -552,6 +556,28 @@ describe("injectAttachments", () => {
552
556
  })).rejects.toThrow(/missing storageKey/);
553
557
  });
554
558
 
559
+ it("rejects a caller-supplied mountPath that escapes the workspace root", async () => {
560
+ // `resolveMountPath` only stripped leading slashes, so `..` segments on a
561
+ // non-`.stigmer/` mount path reached the unchecked join(rootDir, path).
562
+ const storage = makeMockStorage();
563
+ storage.download.mockResolvedValue(Buffer.from("owned"));
564
+ const backend = mockWorkspaceBackend();
565
+
566
+ const escapes = ["../../escape.txt", "../etc/evil", "foo/../../bar"];
567
+ for (const mountPath of escapes) {
568
+ await expect(injectAttachments({
569
+ backend,
570
+ attachments: [makeAttachment({
571
+ filename: "data.txt",
572
+ storageKey: "attachments/xyz/data.txt",
573
+ mountPath,
574
+ })],
575
+ storage,
576
+ isLocalMode: false,
577
+ })).rejects.toThrow(/mount path .* escapes the workspace root|traversal/i);
578
+ }
579
+ });
580
+
555
581
  it("preserves binary content via writeFileBuffer", async () => {
556
582
  const binaryContent = Buffer.from([0x00, 0x01, 0xFF, 0xFE, 0x89, 0x50, 0x4E, 0x47]);
557
583
  const localFile = join(tempDir, "image.png");
@@ -705,3 +731,184 @@ describe("injectAttachments", () => {
705
731
  })).rejects.toThrow(AttachmentValidationError);
706
732
  });
707
733
  });
734
+
735
+ // ═══════════════════════════════════════════════════════════════════════
736
+ // Vision selection during injection (T04)
737
+ // ═══════════════════════════════════════════════════════════════════════
738
+ // Vision is strictly additive: every case also asserts the file was written
739
+ // exactly as it would be without a budget.
740
+
741
+ describe("injectAttachments — vision selection", () => {
742
+ const PNG_BYTES = Buffer.concat([
743
+ Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a]),
744
+ Buffer.alloc(56, 0xab),
745
+ ]);
746
+ const WEBP_BYTES = Buffer.concat([
747
+ Buffer.from("RIFF", "ascii"),
748
+ Buffer.from([0x24, 0x00, 0x00, 0x00]),
749
+ Buffer.from("WEBP", "ascii"),
750
+ Buffer.alloc(52, 0xcd),
751
+ ]);
752
+
753
+ it("accepts a PNG into the vision payload (deep-agent profile) and still writes the file", async () => {
754
+ const backend = mockWorkspaceBackend();
755
+ const storage = makeMockStorage();
756
+ await storage.upload("attachments/abc/photo.png", PNG_BYTES);
757
+
758
+ const [injected] = await injectAttachments({
759
+ backend,
760
+ attachments: [makeAttachment({
761
+ filename: "photo.png",
762
+ storageKey: "attachments/abc/photo.png",
763
+ contentType: "image/png",
764
+ })],
765
+ storage,
766
+ isLocalMode: false,
767
+ visionBudget: new VisionBudget(DEEP_AGENT_VISION_PROFILE),
768
+ });
769
+
770
+ expect(injected.vision).toMatchObject({
771
+ filename: "photo.png",
772
+ mimeType: "image/png",
773
+ byteSize: PNG_BYTES.length,
774
+ });
775
+ expect(Buffer.from(injected.vision!.base64, "base64").equals(PNG_BYTES)).toBe(true);
776
+ expect(backend.writeFileBuffer).toHaveBeenCalledWith(".stigmer/inputs/photo.png", PNG_BYTES);
777
+ });
778
+
779
+ it("degrades an image with model_no_vision on a blind model, file written intact", async () => {
780
+ const backend = mockWorkspaceBackend();
781
+ const storage = makeMockStorage();
782
+ await storage.upload("attachments/abc/photo.png", PNG_BYTES);
783
+
784
+ const [injected] = await injectAttachments({
785
+ backend,
786
+ attachments: [makeAttachment({
787
+ filename: "photo.png",
788
+ storageKey: "attachments/abc/photo.png",
789
+ contentType: "image/png",
790
+ })],
791
+ storage,
792
+ isLocalMode: false,
793
+ visionBudget: new VisionBudget(DEEP_AGENT_VISION_PROFILE, { modelVision: false }),
794
+ });
795
+
796
+ expect(injected.vision).toBeUndefined();
797
+ expect(injected.visionDegraded).toBe("model_no_vision");
798
+ expect(backend.writeFileBuffer).toHaveBeenCalledWith(".stigmer/inputs/photo.png", PNG_BYTES);
799
+ });
800
+
801
+ it("accepts WebP on the deep-agent profile (unlike the Cursor harness)", async () => {
802
+ const backend = mockWorkspaceBackend();
803
+ const storage = makeMockStorage();
804
+ await storage.upload("attachments/abc/pic.webp", WEBP_BYTES);
805
+
806
+ const [injected] = await injectAttachments({
807
+ backend,
808
+ attachments: [makeAttachment({
809
+ filename: "pic.webp",
810
+ storageKey: "attachments/abc/pic.webp",
811
+ contentType: "image/webp",
812
+ })],
813
+ storage,
814
+ isLocalMode: false,
815
+ visionBudget: new VisionBudget(DEEP_AGENT_VISION_PROFILE),
816
+ });
817
+
818
+ expect(injected.vision?.mimeType).toBe("image/webp");
819
+ });
820
+
821
+ it("carries no vision fields for a non-image attachment", async () => {
822
+ const backend = mockWorkspaceBackend();
823
+ const storage = makeMockStorage();
824
+ await storage.upload("attachments/abc/doc.pdf", Buffer.from("%PDF-1.7"));
825
+
826
+ const [injected] = await injectAttachments({
827
+ backend,
828
+ attachments: [makeAttachment({
829
+ filename: "doc.pdf",
830
+ storageKey: "attachments/abc/doc.pdf",
831
+ contentType: "application/pdf",
832
+ })],
833
+ storage,
834
+ isLocalMode: false,
835
+ visionBudget: new VisionBudget(DEEP_AGENT_VISION_PROFILE),
836
+ });
837
+
838
+ expect(injected.vision).toBeUndefined();
839
+ expect(injected.visionDegraded).toBeUndefined();
840
+ });
841
+
842
+ it("degrades a declared image whose bytes are not one (type_mismatch)", async () => {
843
+ const backend = mockWorkspaceBackend();
844
+ const storage = makeMockStorage();
845
+ await storage.upload("attachments/abc/photo.jpg", Buffer.from("actually HEIC"));
846
+
847
+ const [injected] = await injectAttachments({
848
+ backend,
849
+ attachments: [makeAttachment({
850
+ filename: "photo.jpg",
851
+ storageKey: "attachments/abc/photo.jpg",
852
+ contentType: "image/jpeg",
853
+ })],
854
+ storage,
855
+ isLocalMode: false,
856
+ visionBudget: new VisionBudget(DEEP_AGENT_VISION_PROFILE),
857
+ });
858
+
859
+ expect(injected.vision).toBeUndefined();
860
+ expect(injected.visionDegraded).toBe("type_mismatch");
861
+ });
862
+
863
+ it("never offers an extract archive to the budget — extracted files carry no vision fields", async () => {
864
+ // A PNG inside a ZIP has no attachment-level bytes; the archive rides the
865
+ // normal extraction story with no vision involvement or disclosure.
866
+ const zip = makeZip({ "inner.png": PNG_BYTES });
867
+ const backend = mockWorkspaceBackend();
868
+ const storage = makeMockStorage();
869
+ await storage.upload("attachments/abc/bundle.zip", zip);
870
+
871
+ const injected = await injectAttachments({
872
+ backend,
873
+ attachments: [makeAttachment({
874
+ filename: "bundle.zip",
875
+ storageKey: "attachments/abc/bundle.zip",
876
+ extract: true,
877
+ })],
878
+ storage,
879
+ isLocalMode: false,
880
+ visionBudget: new VisionBudget(DEEP_AGENT_VISION_PROFILE),
881
+ });
882
+
883
+ expect(injected.length).toBeGreaterThan(0);
884
+ for (const file of injected) {
885
+ expect(file.vision).toBeUndefined();
886
+ expect(file.visionDegraded).toBeUndefined();
887
+ }
888
+ });
889
+
890
+ it("degrades over the total budget in attachment order (budget_exhausted)", async () => {
891
+ const backend = mockWorkspaceBackend();
892
+ const storage = makeMockStorage();
893
+ await storage.upload("attachments/a/a.png", PNG_BYTES);
894
+ await storage.upload("attachments/b/b.png", PNG_BYTES);
895
+
896
+ const injected = await injectAttachments({
897
+ backend,
898
+ attachments: [
899
+ makeAttachment({ filename: "a.png", storageKey: "attachments/a/a.png", contentType: "image/png" }),
900
+ makeAttachment({ filename: "b.png", storageKey: "attachments/b/b.png", contentType: "image/png" }),
901
+ ],
902
+ storage,
903
+ isLocalMode: false,
904
+ visionBudget: new VisionBudget(DEEP_AGENT_VISION_PROFILE, {
905
+ maxImageBytes: PNG_BYTES.length,
906
+ maxTotalBytes: PNG_BYTES.length,
907
+ }),
908
+ });
909
+
910
+ expect(injected[0].vision).toBeDefined();
911
+ expect(injected[1].vision).toBeUndefined();
912
+ expect(injected[1].visionDegraded).toBe("budget_exhausted");
913
+ });
914
+ });
@@ -70,23 +70,23 @@ function makeGraphState(
70
70
  }
71
71
 
72
72
  describe("resolveResumeInput", () => {
73
- it("returns fresh input when no interrupts exist", () => {
73
+ it("reports not-a-resume when no interrupts exist (caller falls back to setup.langgraphInput)", () => {
74
74
  const execution = makeExecution([]);
75
75
  const state: GraphStateSnapshot = { values: {}, tasks: [] };
76
76
 
77
- const result = resolveResumeInput(execution, state, "hello");
77
+ const result = resolveResumeInput(execution, state);
78
78
 
79
79
  expect(result.isResumeFromApproval).toBe(false);
80
- expect(result.graphInput).toEqual({
81
- messages: [{ role: "user", content: "hello" }],
82
- });
80
+ // No graphInput on this branch — the turn's user message has exactly ONE
81
+ // construction site (setup.langgraphInput), never a second copy here.
82
+ expect("graphInput" in result).toBe(false);
83
83
  });
84
84
 
85
- it("returns fresh input when interrupts exist but no decisions", () => {
85
+ it("reports not-a-resume when interrupts exist but no decisions", () => {
86
86
  const execution = makeExecution([]);
87
87
  const state = makeGraphState([{ taskId: "task-1", toolCallId: "call-1" }]);
88
88
 
89
- const result = resolveResumeInput(execution, state, "hello");
89
+ const result = resolveResumeInput(execution, state);
90
90
 
91
91
  expect(result.isResumeFromApproval).toBe(false);
92
92
  });
@@ -100,9 +100,9 @@ describe("resolveResumeInput", () => {
100
100
 
101
101
  const state = makeGraphState([{ taskId: "task-1", toolCallId: "call-1" }]);
102
102
 
103
- const result = resolveResumeInput(execution, state, "hello");
103
+ const result = resolveResumeInput(execution, state);
104
104
 
105
- expect(result.isResumeFromApproval).toBe(true);
105
+ if (!result.isResumeFromApproval) throw new Error("expected an approval resume");
106
106
  expect(result.graphInput).toBeInstanceOf(Command);
107
107
  });
108
108
 
@@ -115,13 +115,13 @@ describe("resolveResumeInput", () => {
115
115
 
116
116
  const state = makeGraphState([{ taskId: "task-1", toolCallId: "call-1" }]);
117
117
 
118
- const result = resolveResumeInput(execution, state, "hello");
118
+ const result = resolveResumeInput(execution, state);
119
119
 
120
120
  // REJECT resumes the gate like any other decision — the gate returns a
121
121
  // denial ToolMessage and the run continues. There is no execution-level
122
122
  // "rejection" flag any more; the terminal tool status is set by
123
123
  // reconcileNonExecutingDecisions.
124
- expect(result.isResumeFromApproval).toBe(true);
124
+ if (!result.isResumeFromApproval) throw new Error("expected an approval resume");
125
125
  expect(result.graphInput).toBeInstanceOf(Command);
126
126
  });
127
127
 
@@ -144,7 +144,7 @@ describe("resolveResumeInput", () => {
144
144
  { taskId: "task-2", toolCallId: "call-2" },
145
145
  ]);
146
146
 
147
- const result = resolveResumeInput(execution, state, "hello");
147
+ const result = resolveResumeInput(execution, state);
148
148
 
149
149
  expect(result.isResumeFromApproval).toBe(true);
150
150
  });
@@ -160,7 +160,7 @@ describe("resolveResumeInput", () => {
160
160
  { taskId: "task-1", toolCallId: "call-1", hasResume: true },
161
161
  ]);
162
162
 
163
- const result = resolveResumeInput(execution, state, "hello");
163
+ const result = resolveResumeInput(execution, state);
164
164
 
165
165
  expect(result.isResumeFromApproval).toBe(false);
166
166
  });