@inflexa-ai/harness 0.28.1 → 0.30.0

package/dist/agents/conversation-agent.js

@@ -27,7 +27,7 @@ import { composeSystemPrompt } from "./system-prompt.js";
 import { searchGeneTool, lookupAnnotationTool, searchInteractionsTool, chemblTool, pubchemTool, openTargetsTool, searchFaersTool, searchClinicalTrialsTool, searchGeoDatasetsTool, genePreclinicalProfileTool, targetSafetyTool, } from "../tools/bio/index.js";
 import { createNcbiTools, createChemDbTools } from "../tools/bio/keys.js";
 // Dependency-bearing tool factories.
-import { createGeneratePlanTool, createLiteratureReviewerTool, createInspectRunTool, createInspectDataProfileTool, createGenerateAnalogyReportTool, } from "../tools/research/index.js";
+import { createGeneratePlanTool, createInspectRunTool, createInspectDataProfileTool, createGenerateAnalogyReportTool, createSearchSemanticScholarTool, searchArxivTool, } from "../tools/research/index.js";
 import { createFileStatTool, createGrepTool, createListFilesTool, createReadFileTool, createShowPlanTool, createWorkspaceSearchTool, showFileTool, } from "../tools/workspace/index.js";
 import { showUserTool } from "../tools/display/index.js";
 import { createUpdateWorkingMemoryTool } from "../tools/memory/index.js";
@@ -39,7 +39,7 @@ import { createResolveCitationTool } from "../tools/research/resolve-citation.js
 /** Canonical agent id — the single source of truth. */
 export const CONVERSATION_AGENT_ID = "conversation-agent";
 /** Runaway guard — heavy tool-driving turns need generous headroom. */
-const CONVERSATION_MAX_ITERATIONS = 50;
+const CONVERSATION_MAX_ITERATIONS = 200;
 /** Build the conversation `AgentDefinition` with every tool fully dep-bound. */
 export function createConversationAgent(deps) {
     const { provider, pool, embedding, workspaceFs, model, utilityProvider, utilityModel, executeAnalysisWorkflow, runAuthorizer, runLauncher, bioKeys, chrome, resourcePolicy, hostTools, refStorePath, packagesFile, usageRecorder, citationResolver, } = deps;
@@ -57,6 +57,10 @@ export function createConversationAgent(deps) {
         ncbi.pubmed,
         // Verification of one caller-supplied citation (distinct from discovery).
         createResolveCitationTool(citationResolver),
+        // Literature outside PubMed. A question about a method, a model or a
+        // preprint has no MEDLINE record, thus PubMed alone cannot answer it.
+        createSearchSemanticScholarTool({ ...(bioKeys.semanticScholar === undefined ? {} : { apiKey: bioKeys.semanticScholar }) }),
+        searchArxivTool,
         // ChEMBL (compounds / drug / mechanism / bioactivity / targets behind one action).
         chemblTool,
         // PubChem (compound / crossrefs / assays behind one action).
@@ -95,6 +99,7 @@ export function createConversationAgent(deps) {
             pool,
             resourcePolicy,
             usageRecorder,
+            bioKeys,
             ...(refStorePath ? { refStorePath } : {}),
             ...(packagesFile ? { packagesFile } : {}),
             ...(deps.logger ? { logger: deps.logger } : {}),
@@ -123,8 +128,6 @@ export function createConversationAgent(deps) {
         showUserTool,
         createShowPlanTool(pool),
         showFileTool,
-        // Batch literature/biology research (sub-agent as a loop-driving tool).
-        createLiteratureReviewerTool({ provider, model, bioKeys, citationResolver, usageRecorder }),
         // Cross-domain analogy generation (sub-agent as a loop-driving tool).
         createGenerateAnalogyReportTool({ provider, model, bioKeys, usageRecorder }),
         // Workspace semantic search + raw read/grep over the read seam.

package/dist/agents/report-session-agent.js

@@ -35,7 +35,7 @@ export const REPORT_SESSION_AGENT_ID = "report-session";
  * look, record — drives more small tool calls than a conversation turn, thus it
  * gets the headroom of the report runner (REPORT_AGENT_MAX_STEPS).
  */
-const REPORT_SESSION_MAX_ITERATIONS = 75;
+const REPORT_SESSION_MAX_ITERATIONS = 200;
 /** Build the report `AgentDefinition` with every tool bound to its deps. */
 export function createReportSessionAgent(deps) {
     const { model, pool, embedding, workspaceFs, gateway, resolveWorkspaceRoot, store, threads, chrome, eyes, makeResolver, resolvePageAsset, logger } = deps;

package/dist/agents/sandbox/data-profiler.js

@@ -15,7 +15,6 @@ export const meta = {
     // differently. Declared rather than bolted on, so the resolved roster stays
     // `meta.tools` plus the always-on set.
     tools: [...BASE_SANDBOX_TOOLS, "scanInputs"],
-    defaultMaxSteps: 85,
     plannable: false,
 };
 export function createDataProfilerAgent(deps) {

package/dist/agents/sandbox/immune-profiling-agent.js

@@ -26,7 +26,9 @@ export const meta = {
     // `single-cell` carries the scirpy reference, which is the route from 10x
     // VDJ / AIRR files to called clonotypes — the input every repertoire metric
     // in `immune-profiling` assumes it already has.
-    skills: ["immune-profiling", "single-cell", "shared/omics-general"],
+    // `statistical-modeling` carries the survival-cutpoint correction and the
+    // multiplicity rules that the immune-clinical correlation work demands.
+    skills: ["immune-profiling", "single-cell", "statistical-modeling", "shared/omics-general"],
     tools: [...BASE_SANDBOX_TOOLS, "pubmed", "searchGene", "lookupAnnotation", "searchInteractions", "searchGeoDatasets", "opentargets"],
 };
 export function createImmuneProfilingAgent(deps) {

package/dist/agents/sandbox/network-agent.js

@@ -9,7 +9,6 @@ export const meta = {
     suitableFor: ["expression-matrices", "correlation-matrices", "general-omics"],
     skills: ["network-regulatory", "shared/omics-general"],
     tools: [...BASE_SANDBOX_TOOLS, "pubmed", "searchGene", "lookupAnnotation", "searchInteractions"],
-    defaultMaxSteps: 35,
 };
 export function createNetworkAgent(deps) {
     return createSandboxAgent(deps, meta, networkAgentPrompt);

package/dist/agents/sandbox/types.d.ts

@@ -31,4 +31,4 @@ export interface AgentMeta {
     readonly plannable?: boolean;
 }
 /** Default runaway-guard for sandbox agents. */
-export declare const SANDBOX_AGENT_DEFAULT_MAX_ITERATIONS = 50;
+export declare const SANDBOX_AGENT_DEFAULT_MAX_ITERATIONS = 200;

package/dist/agents/sandbox/types.js

@@ -12,4 +12,4 @@
  * in the tool's `perSource` report; the rest of the call still answers.
  */
 /** Default runaway-guard for sandbox agents. */
-export const SANDBOX_AGENT_DEFAULT_MAX_ITERATIONS = 50;
+export const SANDBOX_AGENT_DEFAULT_MAX_ITERATIONS = 200;

package/dist/app/data-profile-orientation.d.ts

@@ -1,42 +1,67 @@
 /**
  * Data-profile orientation — the bounded projection of a persisted
- * `DataProfileResult` into the few hundred characters an agent needs to know
- * WHAT DATASET IT IS HOLDING before it does anything else.
+ * `DataProfileResult` into what an agent needs to know WHAT DATASET IT IS HOLDING
+ * before it does anything else.
  *
- * The persisted profile is deliberately rich (per-file metrics, warnings, tags,
- * the full narrative summary) because the DB row is its only durable home — the
- * profiler's scratch tree is deleted on completion. That richness is right for a
- * pull-on-demand tool (`inspect_data_profile`) and wrong for anything injected
- * unconditionally into a context window, which is what a step seed does. This is
- * the lossy half of that pair: a hard character budget, a fixed field order, and
- * visible truncation counts, so an agent that needs more knows to go pull it.
+ * The persisted profile is deliberately rich because the DB row is its only durable
+ * home — the profiler's scratch tree is deleted on completion. That richness is right
+ * for a pull-on-demand tool (`inspect_data_profile`) and wrong for anything injected
+ * unconditionally into a context window, which is what a step seed does. This is the
+ * lossy half of that pair: a hard character budget, a fixed section order, and visible
+ * truncation counts, so an agent that needs more knows to go pull it.
+ *
+ * What is lossy is the PROSE, not the structure. The submission schema bounds every array
+ * an agent authors — operations, dimensions, annotations — so the resolved record is small
+ * by construction, and a rendering that dropped half the groups to make room for a design
+ * note would be discarding the addressable facts to keep the summary of them. Groups
+ * therefore render one line each, all of them, each followed by the slots that vary across
+ * its members; dimensions and notable files follow; and the design note and the caveats
+ * take whatever budget is left. Groups are elided only when one line apiece already
+ * overruns the whole budget, and then the count that did not fit is stated outright.
+ *
+ * The census — kept files, groups, unclassified, quarantined, and whether the scan reached
+ * the whole tree — rides in the header on a reserved budget, so no clamp can remove the one
+ * line that says how much of the dataset the profile speaks for.
+ *
+ * Every elision is marked. A section fitted short, a list tail dropped, a prose value
+ * clamped: each ends the rendering with {@link ELISION_MARKER}, whose room is reserved
+ * before anything else is fitted. A rendering that silently lost a section reads as a
+ * profile that never had one.
  *
  * A pure function over the record — no I/O, no LLM, no pipeline. It is the whole
  * mechanism.
  *
- * Every path it renders is projected into the frame-independent
- * `/{analysisId}/…` form. The record stores a path relative to the analysis
- * root, which is the correct form to store; but a step agent resolves a relative
- * path against its own working directory, thus the stored form names a file that
- * does not exist there. The absolute form resolves to the same file in every
- * frame, including the conversation agent's, so the projection is unconditional.
+ * Every path it renders is projected into the frame-independent `/{analysisId}/…`
+ * form. The record stores a path relative to the analysis root, which is the correct
+ * form to store; but a step agent resolves a relative path against its own working
+ * directory, thus the stored form names a file that does not exist there. The absolute
+ * form resolves to the same file in every frame, including the conversation agent's, so
+ * the projection is unconditional.
  */
 import type { DataProfileResult } from "../state/data-profile.js";
-/** The character budget the projection guarantees it will not exceed. */
-export declare const DATA_PROFILE_ORIENTATION_MAX_CHARS = 1200;
 /**
- * Project a persisted profile into an orientation blurb of at most `maxChars`
- * characters. The bound is a guarantee, not a target: the assembled text is
- * hard-clamped, so no profile — however verbose — can blow a caller's budget.
+ * The character budget the projection guarantees it will not exceed.
  *
- * Field order is fixed and reflects what an agent must know first: what the
- * dataset IS, how it was designed, what is wrong with it, then what files carry
- * it. Both list sections state their true total, so an elided tail is visible
- * rather than silent.
+ * Sized from the record's own bounds rather than from a feel for how long a blurb should
+ * be: the census header, a group line plus a slots line for each of a submission's worth of
+ * groups, the dimension list, the notable files, and a prose tail. A profile past that
+ * elides its group tail with a stated count — which is the honest outcome, and a far
+ * cheaper one than a briefing that omits the groups an execution agent must address files
+ * by.
+ */
+export declare const DATA_PROFILE_ORIENTATION_MAX_CHARS = 6000;
+/**
+ * Project a persisted profile into an orientation blurb of at most `maxChars`
+ * characters. The bound is a guarantee, not a target: every section is fitted against the
+ * remaining budget before it is rendered, so no profile — however verbose — can blow a
+ * caller's budget, and nothing the budget removed leaves a half-line behind.
  *
- * A snapshot written before the record was widened carries none of the structured
- * fields. Rather than emit a bare file list, that case falls back to the narrative
- * summary — the one orientation a legacy row does have.
+ * Section order is fixed: identity, census, groups with their slots, dimensions, the
+ * members the agent wrote about individually, the experimental design, then the caveats.
+ * The header holds its own reserved budget, so the census survives whatever the body cannot
+ * hold, and the prose sits last so that it — and not the structure — is what a clamp
+ * reaches first. Each list section states its true total, and the trailing marker states
+ * that the budget, rather than a per-section cap, is why something is missing.
  *
  * `analysisId` roots each rendered file path (see the module header). It is the
  * id of the analysis the profile belongs to, and it is what the sandbox mounts