@inflexa-ai/harness 0.19.2 → 0.20.0

package/dist/literature/sources/pubmed.js

@@ -11,7 +11,8 @@ const IdConvResponseSchema = z.object({
 });
 export function ncbiUrl(apiKey, base, params) {
     const url = new URL(base);
-    if (apiKey !== undefined)
+    // The URL must not include an empty key. NCBI rejects `api_key=` with HTTP 400.
+    if (apiKey)
         url.searchParams.set("api_key", apiKey);
     for (const [key, value] of Object.entries(params))
         if (value !== undefined)

package/dist/sandbox/k8s-client.js

@@ -32,10 +32,6 @@ const MANAGED_BY_VALUE = "cortex";
 const OWNER_WORKFLOW_LABEL = "cortex/owner-workflow-id";
 const RUN_ID_LABEL = "cortex/run-id";
 const STEP_ID_LABEL = "cortex/step-id";
-// OpenCost sanitizes `/` and `-` to `_`, so these arrive at the metering
-// reconciler as cortex_billing_context / cortex_user_id / cortex_analysis_id.
-const BILLING_CONTEXT_LABEL = "cortex/billing-context";
-const USER_ID_LABEL = "cortex/user-id";
 const ANALYSIS_ID_LABEL = "cortex/analysis-id";
 /**
  * Coerce an arbitrary identifier into a valid K8s label value:
@@ -219,17 +215,15 @@ function buildJobSpec(meta, config, identity) {
         podSpec.tolerations = config.tolerations;
     if (config.runtimeClassName)
         podSpec.runtimeClassName = config.runtimeClassName;
-    // Pod-template placement is load-bearing: the OpenCost reconciler allocates
-    // by pod labels and filters on a non-empty cortex_billing_context — a
-    // Job-only label is invisible to compute metering.
-    const billingLabels = meta.billing
-        ? {
-            [BILLING_CONTEXT_LABEL]: sanitizeLabelValue(meta.billing.billingContextId),
-            [USER_ID_LABEL]: sanitizeLabelValue(meta.billing.userId),
-            [ANALYSIS_ID_LABEL]: sanitizeLabelValue(meta.analysisId),
-            [RUN_ID_LABEL]: sanitizeLabelValue(meta.runId),
-        }
-        : {};
+    // Pod-template placement is load-bearing: a cost reconciler allocates by pod
+    // labels, thus a Job-only label is invisible to it. The host labels stay
+    // opaque — every value passes through `sanitizeLabelValue`, because one
+    // malformed value makes the API server reject the whole Job at admission.
+    const attributionLabels = {
+        [ANALYSIS_ID_LABEL]: sanitizeLabelValue(meta.analysisId),
+        [RUN_ID_LABEL]: sanitizeLabelValue(meta.runId),
+        ...Object.fromEntries(Object.entries(meta.podLabels ?? {}).map(([key, value]) => [key, sanitizeLabelValue(value)])),
+    };
     return {
         apiVersion: "batch/v1",
         kind: "Job",
@@ -241,9 +235,8 @@ function buildJobSpec(meta, config, identity) {
                 role: "sandbox",
                 "cortex/sandbox-id": sandboxId,
                 [OWNER_WORKFLOW_LABEL]: sanitizeLabelValue(meta.childWorkflowId),
-                [RUN_ID_LABEL]: sanitizeLabelValue(meta.runId),
                 [STEP_ID_LABEL]: sanitizeLabelValue(meta.stepId),
-                ...billingLabels,
+                ...attributionLabels,
             },
         },
         spec: {
@@ -254,7 +247,7 @@ function buildJobSpec(meta, config, identity) {
                     labels: {
                         role: "sandbox",
                         "cortex/sandbox-id": sandboxId,
-                        ...billingLabels,
+                        ...attributionLabels,
                     },
                 },
                 spec: podSpec,

package/dist/sandbox/types.d.ts

@@ -243,12 +243,10 @@ export interface CreateSandboxMeta {
     /** Enforced read-only: provision with no read-write step mount, only the
      *  read-only analysis tree for generic read-only agents. */
     readOnly?: boolean;
-    /** Billing attribution stamped as pod labels for OpenCost compute metering.
-     *  Absent ⇒ the pod is invisible to the metering reconciler (filter-level). */
-    billing?: {
-        billingContextId: string;
-        userId: string;
-    };
+    /** Host-supplied labels stamped onto the sandbox pod, verbatim. Opaque here:
+     *  the harness reads no key and no value, it only sanitizes each value into a
+     *  valid label. Absent ⇒ the pod carries the harness's own labels only. */
+    podLabels?: Record<string, string>;
 }
 /**
  * The identity minted for a sandbox machine *before* it is spawned — the

package/dist/tasks/data-profile.d.ts

@@ -23,7 +23,6 @@ import type { RunAuthorization, RunAuthorizer } from "../execution/run-authorize
 import type { StagedInput } from "../execution/staged-input.js";
 import type { BioToolKeys } from "../tools/bio/keys.js";
 import type { Logger } from "../lib/logger.js";
-import type { ResolveBilling } from "../billing/resolver.js";
 import type { UsageRecorder } from "../billing/usage-recorder.js";
 import type { ChatProvider, EmbeddingProvider } from "../providers/types.js";
 import type { SandboxClient } from "../sandbox/client.js";
@@ -57,9 +56,13 @@ export interface DataProfileDeps extends EnvironmentStorePaths {
     readonly skillsDir: string;
     /** LLM usage-accounting seam for the profiler agent loop; omitted falls back to the no-op recorder. */
     readonly usageRecorder?: UsageRecorder;
-    /** Resolves the billing context stamped as the sandbox pod's OpenCost
-     *  labels. Absent in upstream-less (dev/OSS) wiring — pods spawn unlabeled. */
-    readonly resolveBilling?: ResolveBilling;
+    /**
+     * Host-supplied labels for the profiler's sandbox pod, resolved under the
+     * profiling session. The map is opaque to the harness: it stamps each entry
+     * and interprets none of it. Absent in a wiring that attributes nothing —
+     * the pod then carries the harness's own labels only.
+     */
+    readonly resolvePodLabels?: (session: RunSession) => Promise<Record<string, string>>;
 }
 /**
  * Workflow input. JSON-serialisable — DBOS persists it as the workflow's

package/dist/tasks/data-profile.js

@@ -42,8 +42,6 @@ import { ensureSearchIndex, searchIndexName } from "../workspace/search-config.j
 import { DATA_PROFILE_RUN_LITERAL } from "../contracts/data-profile.js";
 const DATA_PROFILE_STEP_LITERAL = "profile";
 const DATA_PROFILE_AGENT_ID = "data-profiler";
-/** Resolved-header key carrying the billing-context id (see `ResolveBilling`). */
-const BILLING_CONTEXT_HEADER = "X-Inflexa-Billing-Context";
 /** Sandbox-server exec budget for the profile run. */
 const DEFAULT_DEADLINE_MS = 300_000;
 /** Build the canonical artifact path for a staged input file. */
@@ -228,16 +226,18 @@ export async function runDataProfileBody(input, deps) {
         // from the agent loop would leave the longest wait of all unreported — the same defect the
         // run panel's activity readout was built to remove.
         await activity.sandboxInit();
-        let billingContextId;
-        if (deps.resolveBilling) {
+        let podLabels;
+        if (deps.resolvePodLabels) {
             try {
-                billingContextId = (await deps.resolveBilling(childSession))[BILLING_CONTEXT_HEADER];
+                podLabels = await deps.resolvePodLabels(childSession);
             }
             catch (err) {
-                logger.error("[billing] data-profile billing resolution failed", logger.errorFields(err));
+                logger.warn("pod-label resolution failed", logger.errorFields(err));
             }
-            if (!billingContextId)
-                logger.error("[billing] sandbox spawned without billing labels");
+            // A wired resolver that yields nothing is the one loud case: the host
+            // asked for attribution and the pod spawns without it.
+            if (!podLabels || Object.keys(podLabels).length === 0)
+                logger.warn("sandbox spawned with no pod labels");
         }
         const sandbox = await deps.sandboxClient.createSandbox({
             runId: DATA_PROFILE_RUN_LITERAL,
@@ -245,7 +245,7 @@ export async function runDataProfileBody(input, deps) {
             analysisId,
             childWorkflowId: workflowId,
             resources: estimateDataProfileResources(stagedInputs),
-            billing: billingContextId ? { billingContextId, userId: childSession.identity.user } : undefined,
+            podLabels,
         }, mintSandboxIdentity(DATA_PROFILE_RUN_LITERAL));
         try {
             // Checkpointed clock, not `Date.now()`: `awaitExec` gates on this absolute

package/dist/workflows/sandbox-step.d.ts

@@ -29,7 +29,6 @@
 import type { TokenUsageRollup } from "@inflexa-ai/harness/contracts/usage.js";
 import type { Pool } from "pg";
 import type { Logger } from "../lib/logger.js";
-import type { ResolveBilling } from "../billing/resolver.js";
 import type { UsageRecorder } from "../billing/usage-recorder.js";
 import type { CitationResolver } from "../citations/types.js";
 import type { AgentDefinition, LoopMessage } from "../loop/types.js";
@@ -233,9 +232,13 @@ export interface SandboxStepDeps {
     /** Write-side embedder for the post-step vector index. */
     readonly embedding: EmbeddingProvider;
     readonly sandboxClient: SandboxClient;
-    /** Resolves the billing context stamped as the sandbox pod's OpenCost
-     *  labels. Absent in upstream-less (dev/OSS) wiring — pods spawn unlabeled. */
-    readonly resolveBilling?: ResolveBilling;
+    /**
+     * Host-supplied labels for this step's sandbox pod, resolved under the
+     * step's session. The map is opaque to the harness: it stamps each entry
+     * and interprets none of it. Absent in a wiring that attributes nothing —
+     * the pod then carries the harness's own labels only.
+     */
+    readonly resolvePodLabels?: (session: RunSession) => Promise<Record<string, string>>;
     /**
      * External artifact registration + sync seam. The harness's post-step pipeline
      * registers each step's outputs through it (filesystem index in the

package/dist/workflows/sandbox-step.js

@@ -46,8 +46,6 @@ import { createBlockerHolder } from "../tools/sandbox/report-blocker.js";
  * declare it explicitly.
  */
 export const DEFAULT_STEP_TIMEOUT_SECONDS = 3600;
-/** Resolved-header key carrying the billing-context id (see `ResolveBilling`). */
-const BILLING_CONTEXT_HEADER = "X-Inflexa-Billing-Context";
 /**
  * DBOS message topic the child uses to notify the parent that it is
  * about to self-cancel because of a 402 budget-exceeded error. The
@@ -153,7 +151,7 @@ export async function runSandboxStepBody(input, deps) {
     const identity = await DBOS.runStep(() => Promise.resolve(mintSandboxIdentity(input.runId)), { name: "sandbox.mint" });
     // (2b) sandbox.create — spawn (or adopt) the machine under the minted
     // identity. The handle (secret included) is cached so recovery picks the
-    // same machine back up without re-provisioning. Billing resolution lives
+    // same machine back up without re-provisioning. Pod-label resolution lives
     // INSIDE this step (not as its own step) so the child's step sequence is
     // unchanged — in-flight workflows resumed across a deploy replay cleanly.
     const sandboxMeta = {
@@ -166,20 +164,20 @@ export async function runSandboxStepBody(input, deps) {
         resources: input.resources,
     };
     const sandbox = await DBOS.runStep(async () => {
-        let billing;
-        if (deps.resolveBilling) {
+        let podLabels;
+        if (deps.resolvePodLabels) {
             try {
-                const bcId = (await deps.resolveBilling(session))[BILLING_CONTEXT_HEADER];
-                if (bcId)
-                    billing = { billingContextId: bcId, userId: input.runSession.identity.user };
+                podLabels = await deps.resolvePodLabels(session);
             }
             catch (err) {
-                logger.error("[billing] step billing resolution failed", { analysisId: input.analysisId, ...logger.errorFields(err) });
+                logger.warn("pod-label resolution failed", { analysisId: input.analysisId, ...logger.errorFields(err) });
             }
-            if (!billing)
-                logger.error("[billing] sandbox spawned without billing labels", { analysisId: input.analysisId });
+            // A wired resolver that yields nothing is the one loud case: the
+            // host asked for attribution and the pod spawns without it.
+            if (!podLabels || Object.keys(podLabels).length === 0)
+                logger.warn("sandbox spawned with no pod labels", { analysisId: input.analysisId });
         }
-        return deps.sandboxClient.createSandbox({ ...sandboxMeta, billing }, identity);
+        return deps.sandboxClient.createSandbox({ ...sandboxMeta, podLabels }, identity);
     }, { name: "sandbox.create" });
     // Recovery path: re-check `isAlive` on the persisted ref before continuing.
     // A classified-dead sandbox triggers a fresh `createSandbox` (task 4.6).

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@inflexa-ai/harness",
-    "version": "0.19.2",
+    "version": "0.20.0",
     "description": "Host-agnostic bioinformatics agent harness \u2014 hand-rolled agent loop, DBOS-durable workflows, sandboxed compute.",
     "license": "Apache-2.0",
     "type": "module",