@inflexa-ai/harness 0.2.0 → 0.3.0

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 (80) hide show
  1. package/README.md +48 -0
  2. package/dist/agents/conversation-agent.d.ts +3 -0
  3. package/dist/agents/conversation-agent.js +2 -2
  4. package/dist/agents/sandbox/shared.d.ts +3 -0
  5. package/dist/agents/sandbox/shared.js +2 -2
  6. package/dist/app/chat-turn.d.ts +3 -0
  7. package/dist/app/chat-turn.js +3 -1
  8. package/dist/app/synthesize-run.d.ts +3 -0
  9. package/dist/app/synthesize-run.js +3 -1
  10. package/dist/execution/artifact-metadata.d.ts +3 -0
  11. package/dist/execution/artifact-metadata.js +8 -2
  12. package/dist/execution/artifact-registration.d.ts +2 -1
  13. package/dist/execution/artifact-registration.js +8 -2
  14. package/dist/execution/ephemeral-runner.d.ts +3 -0
  15. package/dist/execution/ephemeral-runner.js +8 -4
  16. package/dist/execution/post-step-pipeline.d.ts +3 -0
  17. package/dist/execution/post-step-pipeline.js +24 -4
  18. package/dist/execution/reconcile-manifest.d.ts +7 -0
  19. package/dist/execution/reconcile-manifest.js +19 -5
  20. package/dist/execution/report-runner.d.ts +3 -0
  21. package/dist/execution/report-runner.js +4 -4
  22. package/dist/execution/step-summary.d.ts +3 -0
  23. package/dist/execution/step-summary.js +4 -2
  24. package/dist/index.d.ts +3 -0
  25. package/dist/index.js +5 -0
  26. package/dist/lib/chrome.d.ts +4 -1
  27. package/dist/lib/chrome.js +10 -7
  28. package/dist/lib/console-logger.d.ts +19 -0
  29. package/dist/lib/console-logger.js +62 -0
  30. package/dist/lib/logger.d.ts +88 -0
  31. package/dist/lib/logger.js +37 -0
  32. package/dist/lib/otel.d.ts +2 -1
  33. package/dist/lib/otel.js +8 -3
  34. package/dist/runtime/boot.d.ts +2 -2
  35. package/dist/runtime/boot.js +1 -1
  36. package/dist/runtime/connection-budget.d.ts +2 -2
  37. package/dist/runtime/connection-budget.js +5 -3
  38. package/dist/runtime/dbos.d.ts +7 -7
  39. package/dist/runtime/dbos.js +14 -11
  40. package/dist/runtime/shutdown.d.ts +2 -2
  41. package/dist/runtime/shutdown.js +4 -3
  42. package/dist/sandbox/create-sandbox.d.ts +2 -2
  43. package/dist/sandbox/docker-client.d.ts +2 -2
  44. package/dist/sandbox/docker-client.js +4 -1
  45. package/dist/sandbox/k8s-client.d.ts +3 -0
  46. package/dist/sandbox/k8s-client.js +5 -4
  47. package/dist/sandbox/notification-sweep.d.ts +3 -3
  48. package/dist/sandbox/notification-sweep.js +4 -2
  49. package/dist/sandbox/reaper.d.ts +3 -3
  50. package/dist/sandbox/reaper.js +7 -4
  51. package/dist/sandbox/watchdog.d.ts +3 -3
  52. package/dist/sandbox/watchdog.js +8 -2
  53. package/dist/state/init.d.ts +2 -1
  54. package/dist/state/init.js +5 -6
  55. package/dist/tasks/data-profile.d.ts +5 -0
  56. package/dist/tasks/data-profile.js +24 -17
  57. package/dist/tools/bio/search-dgidb.d.ts +11 -1
  58. package/dist/tools/bio/search-dgidb.js +92 -78
  59. package/dist/tools/iterate-report.d.ts +3 -0
  60. package/dist/tools/iterate-report.js +3 -1
  61. package/dist/tools/research/literature-reviewer.d.ts +3 -0
  62. package/dist/tools/research/literature-reviewer.js +2 -2
  63. package/dist/tools/workspace/execute-command.d.ts +3 -0
  64. package/dist/tools/workspace/execute-command.js +3 -1
  65. package/dist/workflows/execute-analysis.d.ts +3 -0
  66. package/dist/workflows/execute-analysis.js +11 -7
  67. package/dist/workflows/execute-target-assessment.d.ts +3 -0
  68. package/dist/workflows/execute-target-assessment.js +19 -15
  69. package/dist/workflows/sandbox-step.d.ts +9 -0
  70. package/dist/workflows/sandbox-step.js +39 -17
  71. package/dist/workflows/target-assessment/lib/llm-step.d.ts +3 -0
  72. package/dist/workflows/target-assessment/lib/llm-step.js +3 -1
  73. package/dist/workflows/target-assessment/lib/structured-llm.d.ts +3 -0
  74. package/dist/workflows/target-assessment/lib/structured-llm.js +1 -0
  75. package/dist/workflows/target-assessment/phase5-persist.d.ts +3 -0
  76. package/dist/workflows/target-assessment/phase5-persist.js +8 -3
  77. package/dist/workflows/target-assessment/progress.d.ts +2 -1
  78. package/dist/workflows/target-assessment/progress.js +4 -3
  79. package/dist/workspace/filesystem.js +163 -74
  80. package/package.json +95 -96
@@ -28,6 +28,7 @@
28
28
  */
29
29
  import { DBOS, Error as DBOSErrors } from "@dbos-inc/dbos-sdk";
30
30
  import { insertStepExecution, updateStepExecution } from "../state/index.js";
31
+ import { createNoopLogger } from "../lib/console-logger.js";
31
32
  import { unwrapOrThrow } from "../lib/result.js";
32
33
  import { isBudgetExceeded } from "../loop/budget-exceeded.js";
33
34
  import { runAgent } from "../loop/run-agent.js";
@@ -101,6 +102,14 @@ export function registerSandboxStep(deps) {
101
102
  * (the DBOS calls inside still rely on a workflow context being present).
102
103
  */
103
104
  async function runSandboxStepBody(input, deps) {
105
+ // Step coordinates bind once — every record below carries them without each
106
+ // call site restating them. REPLAY: a recovered workflow re-executes this body
107
+ // and re-emits its records, so a reader dedups by (runId, stepId), not by count.
108
+ const logger = (deps.logger ?? createNoopLogger()).named("sandbox-step").with({
109
+ runId: input.runId,
110
+ stepId: input.stepId,
111
+ agentId: input.agentId,
112
+ });
104
113
  const childWorkflowId = DBOS.workflowID ?? `${input.runId}-${input.stepId}`;
105
114
  // Checkpointed so the value (and every `durationMs` derived from it) replays
106
115
  // identically — a raw `Date.now()` in the workflow body drifts on recovery.
@@ -206,7 +215,7 @@ async function runSandboxStepBody(input, deps) {
206
215
  await emitToParentStream(part);
207
216
  }
208
217
  catch (err) {
209
- console.warn(`[sandbox-step] emitToParentStream failed (non-fatal):`, err instanceof Error ? err.message : err);
218
+ logger.warn("emitToParentStream failed (non-fatal)", logger.errorFields(err));
210
219
  }
211
220
  };
212
221
  // Schema-conformant step-activity emitter — stable id per (runId, stepId)
@@ -230,10 +239,15 @@ async function runSandboxStepBody(input, deps) {
230
239
  const failStep = async (errorClass, err) => {
231
240
  if (err instanceof DBOSErrors.DBOSWorkflowCancelledError)
232
241
  throw err;
233
- await tryTeardown(deps, sandbox);
242
+ await tryTeardown(logger, deps, sandbox);
234
243
  const durationMs = (await DBOS.now()) - startedAt;
235
- const detail = err instanceof Error ? (err.stack ?? err.message) : String(err);
236
- console.error(`[sandbox-step] ${errorClass} failure for ${input.runId}/${input.stepId}: ${detail}`);
244
+ // The scrub below is deliberate and load-bearing, which makes this record
245
+ // the ONLY account of why the step died — the thrown error carries just the
246
+ // generic phrase, so nothing downstream can reconstruct the cause.
247
+ logger.error("step failure", {
248
+ errorClass,
249
+ ...logger.errorFields(err),
250
+ });
237
251
  const safe = userFacingStepFailure(errorClass);
238
252
  await DBOS.runStep(async () => {
239
253
  unwrapOrThrow(await updateStepExecution(deps.pool, input.runId, input.stepId, {
@@ -334,7 +348,9 @@ async function runSandboxStepBody(input, deps) {
334
348
  await DBOS.runStep(() => DBOS.send(input.parentWorkflowId, notification, BUDGET_EXCEEDED_TOPIC), { name: "notify-parent-budget-exceeded" });
335
349
  }
336
350
  catch (sendErr) {
337
- console.warn(`[sandbox-step] notify-parent-budget-exceeded failed (non-fatal): ${sendErr instanceof Error ? sendErr.message : sendErr}`);
351
+ logger.warn("notify-parent-budget-exceeded failed (non-fatal)", {
352
+ ...logger.errorFields(sendErr),
353
+ });
338
354
  }
339
355
  // Self-cancel to CANCELLED so the parent's `getResult` observes the
340
356
  // cancellation and triggers the 402 pause cascade. ERROR is NOT
@@ -369,7 +385,7 @@ async function runSandboxStepBody(input, deps) {
369
385
  // terminal-failure status so the parent's fail-fast cascade fires.
370
386
  const blockerOutcome = blockerHolder.outcome;
371
387
  if (blockerOutcome) {
372
- await tryTeardown(deps, sandbox);
388
+ await tryTeardown(logger, deps, sandbox);
373
389
  const durationMs = (await DBOS.now()) - startedAt;
374
390
  await DBOS.runStep(async () => {
375
391
  unwrapOrThrow(await updateStepExecution(deps.pool, input.runId, input.stepId, {
@@ -413,7 +429,7 @@ async function runSandboxStepBody(input, deps) {
413
429
  lineageCollector,
414
430
  };
415
431
  await emitActivity("generating-metadata", "Describing output files");
416
- const manifest = await safeRunValue(() => walkStepArtifacts({
432
+ const manifest = await safeRunValue(logger, () => walkStepArtifacts({
417
433
  writePrefix,
418
434
  stepId: input.stepId,
419
435
  runId: input.runId,
@@ -423,11 +439,11 @@ async function runSandboxStepBody(input, deps) {
423
439
  // gate (`data-step-summary`, file-tree) stay replay-stable and the billed
424
440
  // LLM calls are not re-issued on recovery. The remaining stages (walk /
425
441
  // reconcile / sync / index) stay inline — a separate follow-up.
426
- const metadataEntries = await safeRunValue(() => DBOS.runStep(() => generateStepFileMetadata(deps, postCtx, manifest), {
442
+ const metadataEntries = await safeRunValue(logger, () => DBOS.runStep(() => generateStepFileMetadata(deps, postCtx, manifest), {
427
443
  name: "post-step.generate-file-metadata",
428
444
  }), "post-step.metadata", []);
429
445
  await emitActivity("generating-summary", "Summarizing step results");
430
- const summary = await safeRunValue(() => DBOS.runStep(() => generateStepSummaryAndWrite(deps, postCtx, manifest), {
446
+ const summary = await safeRunValue(logger, () => DBOS.runStep(() => generateStepSummaryAndWrite(deps, postCtx, manifest), {
431
447
  name: "post-step.generate-step-summary",
432
448
  }), "post-step.summary", undefined);
433
449
  // Lineage attestation (registration + sync) is fail-fast (see the artifact-manifest spec): unlike
@@ -459,10 +475,10 @@ async function runSandboxStepBody(input, deps) {
459
475
  stepOutputs = collectStepOutputs(postCtx, postArtifacts);
460
476
  }
461
477
  catch (err) {
462
- console.warn(`[sandbox-step] post-step.collect failed (non-fatal):`, err instanceof Error ? err.message : err);
478
+ logger.warn("post-step.collect failed (non-fatal)", logger.errorFields(err));
463
479
  }
464
480
  await emitActivity("indexing", "Indexing outputs for search");
465
- await safeRun(() => vectorIndexStepOutputs(deps, postCtx, postArtifacts), "post-step.vector-index");
481
+ await safeRun(logger, () => vectorIndexStepOutputs(deps, postCtx, postArtifacts), "post-step.vector-index");
466
482
  // (9) teardown — destroy sandbox + clear sandbox_ref.
467
483
  await DBOS.runStep(() => deps.sandboxClient.teardown(sandbox), { name: "sandbox.teardown" });
468
484
  // (10) mark-terminal + emit.
@@ -537,14 +553,17 @@ async function runSandboxStepBody(input, deps) {
537
553
  * swallows non-fatal failures — these are reflected in the step summary
538
554
  * but do not fail the step itself (synthesis/sync/index can be retried).
539
555
  */
540
- async function safeRun(fn, label) {
556
+ async function safeRun(logger, fn, label) {
541
557
  try {
542
558
  await fn();
543
559
  }
544
560
  catch (err) {
545
561
  if (err instanceof DBOSErrors.DBOSWorkflowCancelledError)
546
562
  throw err;
547
- console.warn(`[sandbox-step] ${label} failed (non-fatal):`, err instanceof Error ? err.message : err);
563
+ // The stage rides as a namespace, not interpolated prose: `label` is what
564
+ // says WHICH post-step stage degraded, and a reader must be able to filter
565
+ // on it without parsing the message.
566
+ logger.named(label).warn("failed (non-fatal)", logger.errorFields(err));
548
567
  }
549
568
  }
550
569
  /**
@@ -552,22 +571,25 @@ async function safeRun(fn, label) {
552
571
  * logs and yields `fallback` so the body can thread a safe default downstream
553
572
  * instead of aborting the post-step pipeline.
554
573
  */
555
- async function safeRunValue(fn, label, fallback) {
574
+ async function safeRunValue(logger, fn, label, fallback) {
556
575
  try {
557
576
  return await fn();
558
577
  }
559
578
  catch (err) {
560
579
  if (err instanceof DBOSErrors.DBOSWorkflowCancelledError)
561
580
  throw err;
562
- console.warn(`[sandbox-step] ${label} failed (non-fatal):`, err instanceof Error ? err.message : err);
581
+ // The stage rides as a namespace, not interpolated prose: `label` is what
582
+ // says WHICH post-step stage degraded, and a reader must be able to filter
583
+ // on it without parsing the message.
584
+ logger.named(label).warn("failed (non-fatal)", logger.errorFields(err));
563
585
  return fallback;
564
586
  }
565
587
  }
566
- async function tryTeardown(deps, sandbox) {
588
+ async function tryTeardown(logger, deps, sandbox) {
567
589
  try {
568
590
  await DBOS.runStep(() => deps.sandboxClient.teardown(sandbox), { name: "sandbox.teardown" });
569
591
  }
570
592
  catch (err) {
571
- console.warn(`[sandbox-step] teardown failed (non-fatal):`, err instanceof Error ? err.message : err);
593
+ logger.warn("teardown failed (non-fatal)", logger.errorFields(err));
572
594
  }
573
595
  }
@@ -23,6 +23,7 @@
23
23
  * the caller bumps the attempt counter so the resumed call lands a fresh
24
24
  * DBOS cache slot rather than replaying the cancelled prior attempt.
25
25
  */
26
+ import type { Logger } from "../../../lib/logger.js";
26
27
  import type { AgentChat, ChatRequest, ChatResponse } from "../../../providers/types.js";
27
28
  import type { AgentSession } from "../../../auth/types.js";
28
29
  /**
@@ -38,6 +39,8 @@ export interface BudgetExceededMarker {
38
39
  readonly error: string;
39
40
  }
40
41
  export interface RunLlmStepOptions {
42
+ /** Operational logging seam; omitted falls back to no-op. */
43
+ readonly logger?: Logger;
41
44
  /** Durable-step name — already attempt-numbered by the caller. */
42
45
  readonly stepName: string;
43
46
  /** Agent id stamped in the marker for telemetry. */
@@ -24,6 +24,7 @@
24
24
  * DBOS cache slot rather than replaying the cancelled prior attempt.
25
25
  */
26
26
  import { DBOS } from "@dbos-inc/dbos-sdk";
27
+ import { createNoopLogger } from "../../../lib/console-logger.js";
27
28
  import { unwrapOrThrow } from "../../../lib/result.js";
28
29
  import { isBudgetExceeded } from "../../../loop/budget-exceeded.js";
29
30
  /**
@@ -82,7 +83,8 @@ export async function runLlmStep(opts) {
82
83
  await DBOS.runStep(() => DBOS.send(workflowId, marker, BUDGET_EXCEEDED_TOPIC), { name: `${stepName}:notify-budget-exceeded` });
83
84
  }
84
85
  catch (sendErr) {
85
- console.warn(`[ta-llm-step] notify-budget-exceeded send failed (non-fatal): ${sendErr instanceof Error ? sendErr.message : sendErr}`);
86
+ const logger = (opts.logger ?? createNoopLogger()).named("ta-llm-step");
87
+ logger.warn("notify-budget-exceeded send failed (non-fatal)", { stepName, ...logger.errorFields(sendErr) });
86
88
  }
87
89
  return { kind: "budget-exceeded", sentinel: BUDGET_EXCEEDED_SENTINEL };
88
90
  }
@@ -18,9 +18,12 @@
18
18
  */
19
19
  import { z } from "zod";
20
20
  import type { AgentSession } from "../../../auth/types.js";
21
+ import type { Logger } from "../../../lib/logger.js";
21
22
  import type { AgentChat } from "../../../providers/types.js";
22
23
  import { BUDGET_EXCEEDED_SENTINEL } from "./llm-step.js";
23
24
  export interface StructuredLlmCallOptions<TSchema extends z.ZodType> {
25
+ /** Operational logging seam, forwarded to `runLlmStep`; omitted falls back to no-op. */
26
+ readonly logger?: Logger;
24
27
  /** Attempt-numbered durable-step name (e.g. `"ta-decision:modulator-triage:0"`). */
25
28
  readonly stepName: string;
26
29
  /** Agent id stamped on telemetry and billing-gateway provenance. */
@@ -78,6 +78,7 @@ export async function structuredLlmCall(opts) {
78
78
  toolChoice: { type: "tool", toolName: SUBMIT_TOOL_NAME },
79
79
  };
80
80
  const result = await runLlmStep({
81
+ ...(opts.logger ? { logger: opts.logger } : {}),
81
82
  stepName: opts.stepName,
82
83
  agentId: opts.agentId,
83
84
  provider: opts.provider,
@@ -15,6 +15,7 @@
15
15
  * in §17 if a harness-native audit pipeline is wired.
16
16
  */
17
17
  import { z } from "zod";
18
+ import type { Logger } from "../../lib/logger.js";
18
19
  import { type DossierV4Body } from "@inflexa-ai/harness/contracts/target-dossier.js";
19
20
  import type { Phase2Bundle } from "./steps/phase2-aggregate.js";
20
21
  import type { DossierRecommendationStepOutput, LiabilityBulletsStepOutput, SafetyFlagsTrailStepOutput, TranslationalCommentaryStepOutput } from "./synthesis/index.js";
@@ -27,6 +28,8 @@ export declare class DossierDerivedInvariantError extends Error {
27
28
  constructor(message: string, cause?: unknown | undefined);
28
29
  }
29
30
  export interface Phase5PersistInput {
31
+ /** Operational logging seam; omitted falls back to no-op. */
32
+ readonly logger?: Logger;
30
33
  readonly assessmentId: string;
31
34
  readonly phase4Dossier: DossierV4Body;
32
35
  readonly phase2: Phase2Bundle;
@@ -15,6 +15,7 @@
15
15
  * in §17 if a harness-native audit pipeline is wired.
16
16
  */
17
17
  import { z } from "zod";
18
+ import { createNoopLogger } from "../../lib/console-logger.js";
18
19
  import { DossierV5Schema, } from "@inflexa-ai/harness/contracts/target-dossier.js";
19
20
  import { getDrugPrimaryTargetUniprots } from "../../tools/lib/chembl-client.js";
20
21
  import { deterministicTranslationalCommentary } from "./assemblers/index.js";
@@ -84,7 +85,7 @@ function pickCommentarySeverity(row) {
84
85
  }
85
86
  return "ok";
86
87
  }
87
- function stampSynthesis(phase4Dossier, syn, phase2) {
88
+ function stampSynthesis(logger, phase4Dossier, syn, phase2) {
88
89
  const next = structuredClone(phase4Dossier);
89
90
  if (syn.bullets.coverage === "available") {
90
91
  next.liability_summary.liability_bullets = syn.bullets.data.bullets.map((b) => ({ text: b.text, rationale: b.rationale, category: b.category }));
@@ -147,7 +148,10 @@ function stampSynthesis(phase4Dossier, syn, phase2) {
147
148
  try {
148
149
  const audit = validateRecommendationCitations(syn.recommendation.data, next);
149
150
  if (audit.citations_unresolved.length > 0) {
150
- console.warn(`[phase5-persist] recommendation_audit: ${audit.citations_unresolved.length} unresolved citation(s)`);
151
+ logger.warn("recommendation_audit: unresolved citation(s)", {
152
+ unresolvedCount: audit.citations_unresolved.length,
153
+ unresolved: audit.citations_unresolved,
154
+ });
151
155
  }
152
156
  next.analytics.recommendation_audit = {
153
157
  coverage: "available",
@@ -388,12 +392,13 @@ async function upgradeDossierToV5(body, phase2) {
388
392
  * rejects the final dossier.
389
393
  */
390
394
  export async function phase5Persist(input) {
395
+ const logger = (input.logger ?? createNoopLogger()).named("phase5-persist").with({ assessmentId: input.assessmentId });
391
396
  // The dossier rides through this function as an untyped working blob; the
392
397
  // `as unknown as DossierV{4,5}Body` casts below are version-shim views that
393
398
  // let the typed helpers operate on it. None of these casts is load-bearing
394
399
  // for soundness: `DossierV5Schema.safeParse(fullDossier)` at the end is the
395
400
  // single validation gate, and it throws before anything is returned.
396
- let dossier = stampSynthesis(input.phase4Dossier, input.synthesis, input.phase2);
401
+ let dossier = stampSynthesis(logger, input.phase4Dossier, input.synthesis, input.phase2);
397
402
  // Demote organ-claim flagged bullets into coverage_qualifier.unverified_bullets.
398
403
  const exec = dossier.executive_recommendation;
399
404
  if (exec?.coverage === "available") {
@@ -12,6 +12,7 @@
12
12
  */
13
13
  import type { Pool } from "pg";
14
14
  import type { TargetAssessmentPhase, TargetAssessmentProgressEvent } from "@inflexa-ai/harness/contracts/target-dossier.js";
15
+ import type { Logger } from "../../lib/logger.js";
15
16
  /** DBOS stream key the SSE route reads via `DBOS.readStream(wfId, "progress")`. */
16
17
  export declare const TA_PROGRESS_STREAM_KEY = "progress";
17
18
  export type ProgressPhase = TargetAssessmentPhase;
@@ -35,4 +36,4 @@ export declare const PROGRESS_MESSAGES: Record<ProgressPhase, string>;
35
36
  * Best-effort: a DB or stream failure logs and continues — progress
36
37
  * emission MUST NOT block the workflow.
37
38
  */
38
- export declare function emitProgress(pool: Pool, assessmentId: string, phase: ProgressPhase): Promise<void>;
39
+ export declare function emitProgress(pool: Pool, logger: Logger, assessmentId: string, phase: ProgressPhase): Promise<void>;
@@ -46,7 +46,8 @@ export const PROGRESS_MESSAGES = {
46
46
  * Best-effort: a DB or stream failure logs and continues — progress
47
47
  * emission MUST NOT block the workflow.
48
48
  */
49
- export async function emitProgress(pool, assessmentId, phase) {
49
+ export async function emitProgress(pool, logger, assessmentId, phase) {
50
+ const log = logger.named("ta-progress").with({ assessmentId, phase });
50
51
  const message = PROGRESS_MESSAGES[phase];
51
52
  const percent = PROGRESS_PERCENTS[phase];
52
53
  await DBOS.runStep(async () => {
@@ -58,7 +59,7 @@ export async function emitProgress(pool, assessmentId, phase) {
58
59
  await DBOS.writeStream(TA_PROGRESS_STREAM_KEY, part);
59
60
  }
60
61
  catch (err) {
61
- console.warn(`[ta-progress] writeStream failed for ${assessmentId} (${phase}): ${err instanceof Error ? err.message : err}`);
62
+ log.warn("writeStream failed", log.errorFields(err));
62
63
  }
63
64
  // The "suspended" phase is owned by the terminal handler
64
65
  // (`markAssessmentSuspended` already set status + progress text). A
@@ -69,7 +70,7 @@ export async function emitProgress(pool, assessmentId, phase) {
69
70
  // emission MUST NOT block the workflow. `updateProgress` returns
70
71
  // its DB failure as an `Err`, so handle both branches here.
71
72
  (await updateProgress(pool, assessmentId, message, rowStatus)).match(() => { }, (error) => {
72
- console.warn(`[ta-progress] updateProgress failed for ${assessmentId} (${phase}): ${describeDbError(error)}`);
73
+ log.warn("updateProgress failed", { err: describeDbError(error) });
73
74
  });
74
75
  }
75
76
  }, { name: `ta-progress:${phase}` });
@@ -28,8 +28,8 @@
28
28
  * OUTSIDE `tryFs`, so only the discrete `fs.open` / `fs.read` / `fs.readFile`
29
29
  * calls become `err`.
30
30
  */
31
- import { open as fsOpen, readFile as fsReadFile, readdir as fsReaddir, stat as fsStat } from "node:fs/promises";
32
- import { createReadStream } from "node:fs";
31
+ import { lstat as fsLstat, open as fsOpen, readdir as fsReaddir, stat as fsStat } from "node:fs/promises";
32
+ import { constants as fsConstants } from "node:fs";
33
33
  import { createInterface } from "node:readline";
34
34
  import { join } from "node:path";
35
35
  import { ResultAsync, err, errAsync, ok } from "neverthrow";
@@ -123,10 +123,20 @@ export function createWorkspaceFilesystem(deps) {
123
123
  return okAsync({ kind: "out_of_scope" });
124
124
  }
125
125
  const absolute = resolved.absolute;
126
- return tryFs("workspace.list", () => fsReaddir(absolute, { withFileTypes: true }), { path: absolute, onAbsent: () => ({ notFound: true }) }).andThen((dirents) => {
127
- if ("notFound" in dirents)
128
- return okAsync({ kind: "not_found" });
129
- return collectEntries(absolute, dirents).map((entries) => ({ kind: "ok", entries }));
126
+ // `lstat` first so a symlinked directory leaf is refused, not
127
+ // followed: a plain `readdir` would list the target's entries,
128
+ // and a leaf swapped to an escaping symlink after the realpath
129
+ // check would leak an off-tree directory's names. Node has no
130
+ // fd-based readdir, so this narrows (does not fully close) the
131
+ // readdir race — the atomic guarantee is on the content reads.
132
+ return safeLstat(absolute).andThen((s) => {
133
+ if (s?.isSymbolicLink())
134
+ return okAsync({ kind: "out_of_scope" });
135
+ return tryFs("workspace.list", () => fsReaddir(absolute, { withFileTypes: true }), { path: absolute, onAbsent: () => ({ notFound: true }) }).andThen((dirents) => {
136
+ if ("notFound" in dirents)
137
+ return okAsync({ kind: "not_found" });
138
+ return collectEntries(absolute, dirents).map((entries) => ({ kind: "ok", entries }));
139
+ });
130
140
  });
131
141
  });
132
142
  },
@@ -135,9 +145,16 @@ export function createWorkspaceFilesystem(deps) {
135
145
  if (resolved.kind === "out_of_scope") {
136
146
  return okAsync({ kind: "out_of_scope" });
137
147
  }
138
- return safeStat(resolved.absolute).map((s) => {
148
+ // `lstat`, not `stat`: a leaf symlink is refused rather than
149
+ // followed — same policy as the read open's `O_NOFOLLOW`.
150
+ // Reporting a followed target's type/size would re-open the
151
+ // metadata side of the confused-deputy leak, and it is atomic
152
+ // (a single non-following syscall, no check/use gap).
153
+ return safeLstat(resolved.absolute).map((s) => {
139
154
  if (!s)
140
155
  return { kind: "not_found" };
156
+ if (s.isSymbolicLink())
157
+ return { kind: "out_of_scope" };
141
158
  return {
142
159
  kind: "ok",
143
160
  type: s.isDirectory() ? "directory" : "file",
@@ -160,29 +177,89 @@ function safeStat(absolute) {
160
177
  onAbsent: () => null,
161
178
  });
162
179
  }
163
- /** Stat each directory entry to attach a size; absence/I/O failures degrade to no size. */
180
+ /** Like {@link safeStat} but NEVER follows a leaf symlink (`lstat`) — the seam's
181
+ * read-side no-follow policy for the `stat`/`list` metadata surface. */
182
+ function safeLstat(absolute) {
183
+ return tryFs("workspace.lstat", () => fsLstat(absolute), {
184
+ path: absolute,
185
+ onAbsent: () => null,
186
+ });
187
+ }
188
+ /** Lstat each directory entry to attach a size; absence/I/O failures degrade to
189
+ * no size. `lstat` (not `stat`) so a symlink entry reports its own size, never
190
+ * its target's — a followed target could be off-tree. `d.isDirectory()` comes
191
+ * from the dirent, which does not follow, so a symlink is a plain `file` entry. */
164
192
  function collectEntries(dir, dirents) {
165
193
  return ResultAsync.combine(dirents.map((d) => {
166
194
  if (d.isDirectory())
167
195
  return okAsync({ name: d.name, type: "directory" });
168
- return safeStat(join(dir, d.name)).map((s) => ({ name: d.name, type: "file", size: s?.size ?? undefined }));
196
+ return safeLstat(join(dir, d.name)).map((s) => ({ name: d.name, type: "file", size: s?.size ?? undefined }));
169
197
  }));
170
198
  }
199
+ /**
200
+ * Read-only open flags that REFUSE to follow a final-component symlink.
201
+ * `classifyWithinRoot` validated a snapshot of the path; a symlink swapped into
202
+ * the leaf between that check and the open here would otherwise be followed
203
+ * off-tree, re-opening the read-exfiltration hole for the length of the race
204
+ * (CWE-367). `O_NOFOLLOW` makes the leaf open atomic — a symlinked final
205
+ * component fails `ELOOP` in the same syscall that opens it, so there is no
206
+ * check/use gap. `?? 0` keeps the mask valid on a platform that lacks the flag.
207
+ *
208
+ * This refuses EVERY leaf symlink, including an in-tree one the persistent check
209
+ * would allow: pure Node cannot express "follow only if it stays in-tree" (that
210
+ * needs `openat2(RESOLVE_BENEATH)`), so the atomic, race-free choice is to not
211
+ * follow the leaf at all. Hard-linked staged inputs are unaffected — a hard link
212
+ * is not a symlink. An intermediate-directory symlink is beyond this guard (Node
213
+ * exposes no `openat2`); the upstream realpath check still rejects the
214
+ * persistent case of that.
215
+ */
216
+ const READ_NO_FOLLOW = fsConstants.O_RDONLY | (fsConstants.O_NOFOLLOW ?? 0);
217
+ /**
218
+ * Open a workspace file read-only without following a leaf symlink (see
219
+ * {@link READ_NO_FOLLOW}). A symlinked final component surfaces as `out_of_scope`
220
+ * — the same verdict a persistent escape gets — never as off-tree bytes. Every
221
+ * content read (`readWithCap`, `readHeadLines`, `readTailLines`) opens through
222
+ * here so the bytes it streams come from a non-symlink leaf. Absence /
223
+ * permission / other I/O throws flow the `FsError` channel unchanged.
224
+ */
225
+ function openReadNoFollow(absolute) {
226
+ return tryFs("workspace.openNoFollow", async () => {
227
+ try {
228
+ return { kind: "ok", fh: await fsOpen(absolute, READ_NO_FOLLOW) };
229
+ }
230
+ catch (cause) {
231
+ if (cause.code === "ELOOP")
232
+ return { kind: "out_of_scope" };
233
+ throw cause;
234
+ }
235
+ }, { path: absolute });
236
+ }
171
237
  function readWithCap(absolute, totalSize, maxBytes) {
172
- if (maxBytes !== undefined && totalSize > maxBytes) {
173
- return tryFs("workspace.readCapped", async () => {
174
- const fh = await fsOpen(absolute, "r");
238
+ return openReadNoFollow(absolute).andThen((opened) => {
239
+ if (opened.kind === "out_of_scope")
240
+ return okAsync({ kind: "out_of_scope" });
241
+ const fh = opened.fh;
242
+ if (maxBytes !== undefined && totalSize > maxBytes) {
243
+ return tryFs("workspace.readCapped", async () => {
244
+ try {
245
+ const buf = Buffer.alloc(maxBytes);
246
+ await fh.read(buf, 0, maxBytes, 0);
247
+ return { kind: "truncated", content: buf, totalSize };
248
+ }
249
+ finally {
250
+ await fh.close();
251
+ }
252
+ }, { path: absolute });
253
+ }
254
+ return tryFs("workspace.readFile", async () => {
175
255
  try {
176
- const buf = Buffer.alloc(maxBytes);
177
- await fh.read(buf, 0, maxBytes, 0);
178
- return { kind: "truncated", content: buf, totalSize };
256
+ return { kind: "ok", content: await fh.readFile(), truncated: false };
179
257
  }
180
258
  finally {
181
259
  await fh.close();
182
260
  }
183
261
  }, { path: absolute });
184
- }
185
- return tryFs("workspace.readFile", async () => ({ kind: "ok", content: await fsReadFile(absolute), truncated: false }), { path: absolute });
262
+ });
186
263
  }
187
264
  function capBuffer(content, maxBytes) {
188
265
  if (maxBytes !== undefined && content.length > maxBytes) {
@@ -204,43 +281,51 @@ function capBuffer(content, maxBytes) {
204
281
  * follow-up `safeStat` (a discrete `fs` call) flows the `FsError` channel.
205
282
  */
206
283
  function readHeadLines(absolute, headLines, maxBytes) {
207
- return new ResultAsync((async () => {
208
- const stream = createReadStream(absolute, { highWaterMark: 16 * 1024 });
209
- const rl = createInterface({ input: stream, crlfDelay: Infinity });
210
- const collected = [];
211
- let bytes = 0;
212
- let truncated = false;
213
- try {
214
- for await (const line of rl) {
215
- const lineBytes = Buffer.byteLength(line, "utf8") + 1;
216
- if (maxBytes !== undefined && bytes + lineBytes > maxBytes) {
217
- truncated = true;
218
- break;
284
+ return openReadNoFollow(absolute).andThen((opened) => {
285
+ if (opened.kind === "out_of_scope")
286
+ return okAsync({ kind: "out_of_scope" });
287
+ const fh = opened.fh;
288
+ return new ResultAsync((async () => {
289
+ // `autoClose: false` keeps the FileHandle ours to close — destroying
290
+ // the stream must not race a close against our own `fh.close()`.
291
+ const stream = fh.createReadStream({ highWaterMark: 16 * 1024, autoClose: false });
292
+ const rl = createInterface({ input: stream, crlfDelay: Infinity });
293
+ const collected = [];
294
+ let bytes = 0;
295
+ let truncated = false;
296
+ try {
297
+ for await (const line of rl) {
298
+ const lineBytes = Buffer.byteLength(line, "utf8") + 1;
299
+ if (maxBytes !== undefined && bytes + lineBytes > maxBytes) {
300
+ truncated = true;
301
+ break;
302
+ }
303
+ collected.push(line);
304
+ bytes += lineBytes;
305
+ if (collected.length >= headLines)
306
+ break;
219
307
  }
220
- collected.push(line);
221
- bytes += lineBytes;
222
- if (collected.length >= headLines)
223
- break;
224
308
  }
225
- }
226
- finally {
227
- rl.close();
228
- stream.destroy();
229
- }
230
- const content = Buffer.from(collected.join("\n"));
231
- if (truncated) {
232
- // Total size is unknown without a full stat; report it as a separate read.
233
- return (await safeStat(absolute)).map((s) => ({
234
- kind: "truncated",
235
- content,
236
- totalSize: s?.size ?? content.length,
237
- }));
238
- }
239
- // Whether the loop exhausted the file or hit the line cap, the agent asked
240
- // for exactly this window, so this is `ok` not `truncated`. The tool layer
241
- // reports `mode: "head"` so the agent knows it's a window.
242
- return ok({ kind: "ok", content, truncated: false });
243
- })());
309
+ finally {
310
+ rl.close();
311
+ stream.destroy();
312
+ await fh.close();
313
+ }
314
+ const content = Buffer.from(collected.join("\n"));
315
+ if (truncated) {
316
+ // Total size is unknown without a full stat; report it as a separate read.
317
+ return (await safeStat(absolute)).map((s) => ({
318
+ kind: "truncated",
319
+ content,
320
+ totalSize: s?.size ?? content.length,
321
+ }));
322
+ }
323
+ // Whether the loop exhausted the file or hit the line cap, the agent asked
324
+ // for exactly this window, so this is `ok` not `truncated`. The tool layer
325
+ // reports `mode: "head"` so the agent knows it's a window.
326
+ return ok({ kind: "ok", content, truncated: false });
327
+ })());
328
+ });
244
329
  }
245
330
  /**
246
331
  * Read the last `tailLines` lines via a windowed read from the end. The
@@ -252,30 +337,34 @@ function readHeadLines(absolute, headLines, maxBytes) {
252
337
  * slicing is pure and runs inside the same `fn`.
253
338
  */
254
339
  function readTailLines(absolute, totalSize, tailLines, maxBytes) {
255
- return tryFs("workspace.readTail", async () => {
256
- const windowSize = Math.min(totalSize, maxBytes ?? totalSize);
257
- const offset = totalSize - windowSize;
258
- const fh = await fsOpen(absolute, "r");
259
- try {
260
- const buf = Buffer.alloc(windowSize);
261
- if (windowSize > 0) {
262
- await fh.read(buf, 0, windowSize, offset);
340
+ return openReadNoFollow(absolute).andThen((opened) => {
341
+ if (opened.kind === "out_of_scope")
342
+ return okAsync({ kind: "out_of_scope" });
343
+ const fh = opened.fh;
344
+ return tryFs("workspace.readTail", async () => {
345
+ try {
346
+ const windowSize = Math.min(totalSize, maxBytes ?? totalSize);
347
+ const offset = totalSize - windowSize;
348
+ const buf = Buffer.alloc(windowSize);
349
+ if (windowSize > 0) {
350
+ await fh.read(buf, 0, windowSize, offset);
351
+ }
352
+ const allLines = buf.toString("utf8").split(/\r?\n/);
353
+ // Drop the partial leading line if the window doesn't cover the file start.
354
+ const candidateLines = offset > 0 ? allLines.slice(1) : allLines;
355
+ const sliced = candidateLines.length > tailLines ? candidateLines.slice(-tailLines) : candidateLines;
356
+ const content = Buffer.from(sliced.join("\n"));
357
+ const gotEnoughLines = sliced.length >= tailLines || offset === 0;
358
+ if (!gotEnoughLines) {
359
+ return { kind: "truncated", content, totalSize };
360
+ }
361
+ return { kind: "ok", content, truncated: false };
263
362
  }
264
- const allLines = buf.toString("utf8").split(/\r?\n/);
265
- // Drop the partial leading line if the window doesn't cover the file start.
266
- const candidateLines = offset > 0 ? allLines.slice(1) : allLines;
267
- const sliced = candidateLines.length > tailLines ? candidateLines.slice(-tailLines) : candidateLines;
268
- const content = Buffer.from(sliced.join("\n"));
269
- const gotEnoughLines = sliced.length >= tailLines || offset === 0;
270
- if (!gotEnoughLines) {
271
- return { kind: "truncated", content, totalSize };
363
+ finally {
364
+ await fh.close();
272
365
  }
273
- return { kind: "ok", content, truncated: false };
274
- }
275
- finally {
276
- await fh.close();
277
- }
278
- }, { path: absolute });
366
+ }, { path: absolute });
367
+ });
279
368
  }
280
369
  function sliceHeadLines(content, headLines, maxBytes) {
281
370
  const capped = maxBytes !== undefined && content.length > maxBytes ? content.subarray(0, maxBytes) : content;