@sanity/ailf 2.7.1 → 2.9.0

package/dist/pipeline/emit-eval-results.js

@@ -0,0 +1,100 @@
+/**
+ * emit-eval-results.ts — decompose the promptfoo results file into the
+ * per-entry descriptors that W0049's registry expects.
+ *
+ * Replaces the Phase-B-stopgap "route the aggregated JSON through the
+ * deprecated `evalResults` bulk descriptor" path. For each test in the
+ * promptfoo output we emit:
+ *
+ *   - `rawResults`      per (run, mode, task, model)           — the full result
+ *   - `renderedPrompts` per (run, mode, task, model)           — prompt the model saw
+ *   - `graderPrompts`   per (run, mode, task, model, grader)   — rubric text
+ *   - `graderJudgments` per (run, mode, task, model, grader)   — {score, reason, pass}
+ *
+ * `testOutputs` is still emitted separately by `calculate-scores-step`
+ * via `uploadTestOutputs()` (carried forward from W0048 for byte-
+ * equivalence with the original rollout).
+ *
+ * `traces` is NOT produced here — agentic trace data flows through the
+ * agent-observer, not through the promptfoo result shape. Traces
+ * emission is out of scope for this helper and lands when the observer
+ * integration migrates (follow-up; not in W0050).
+ *
+ * The "grader" axis value is the rubric dimension string produced by
+ * `classifyRubric` (e.g. "task-completion", "code-correctness"). Non-
+ * LLM-rubric component assertions (javascript, contains, etc.) don't
+ * have a natural grader identifier and are skipped — their outcomes
+ * still live inside the full `rawResults` object.
+ */
+import { readFileSync } from "node:fs";
+import { classifyRubric, parseRubricScore, } from "../_vendor/ailf-core/index.js";
+// ---------------------------------------------------------------------------
+// Public entry point
+// ---------------------------------------------------------------------------
+/**
+ * Parse a promptfoo results file and emit the per-entry artifacts.
+ *
+ * Non-blocking: any individual emit failure warns but does not halt.
+ * File read/parse errors are caught and logged; the caller keeps going.
+ */
+export async function emitPerEntryEvalResults(writer, ctx, mode, resultsPath) {
+    let raw;
+    try {
+        raw = JSON.parse(readFileSync(resultsPath, "utf-8"));
+    }
+    catch (err) {
+        const message = err instanceof Error ? err.message : String(err);
+        console.warn(`  ⚠️  emitPerEntryEvalResults: failed to read ${resultsPath} — ${message}`);
+        return;
+    }
+    // Promptfoo wraps results in either `{ results: { results: [...] } }`
+    // (older shape) or directly as `{ results: [...] }` (some adapters).
+    const wrapper = raw.results && "results" in raw.results
+        ? raw.results
+        : raw;
+    const rows = wrapper?.results ?? [];
+    if (rows.length === 0) {
+        console.warn(`  ⚠️  emitPerEntryEvalResults: ${resultsPath} has no results[]`);
+        return;
+    }
+    for (const result of rows) {
+        const taskId = result.testCase?.description ?? "unknown-task";
+        const modelId = result.provider?.id ?? result.provider?.label ?? "unknown-model";
+        const baseAssoc = {
+            run: ctx.runId,
+            mode,
+            task: taskId,
+            model: modelId,
+        };
+        // rawResults — full raw entry (bounded by descriptor capBytes: 1 MB)
+        await writer.emit("rawResults", baseAssoc, result);
+        // renderedPrompts — what the model saw + which provider it went to
+        if (result.prompt !== undefined) {
+            await writer.emit("renderedPrompts", baseAssoc, {
+                prompt: result.prompt,
+                provider: result.provider,
+            });
+        }
+        // Per-grader decomposition — only LLM-rubric assertions have a
+        // natural grader identity. Code assertions (javascript/contains/…)
+        // show up in rawResults but not as standalone graderJudgments.
+        const components = result.gradingResult?.componentResults ?? [];
+        for (const comp of components) {
+            if (comp.assertion?.type !== "llm-rubric")
+                continue;
+            const dimension = classifyRubric(comp);
+            if (!dimension)
+                continue;
+            const graderAssoc = { ...baseAssoc, grader: dimension };
+            await writer.emit("graderPrompts", graderAssoc, {
+                dimension,
+                assertion: comp.assertion,
+            });
+            await writer.emit("graderJudgments", graderAssoc, {
+                score: parseRubricScore(comp) ?? 0,
+                reason: comp.reason ?? "",
+                pass: comp.pass,
+            });
+        }
+    }
+}

package/dist/pipeline/provenance.d.ts

@@ -1,65 +1,45 @@
 /**
  * pipeline/provenance.ts
  *
- * Builds ReportProvenance from data available during a pipeline run.
+ * Builds `ReportProvenance` from data available during a pipeline run.
  *
- * Provenance captures what produced an evaluation report: which models,
- * which source, which mode, what triggered it, git metadata, etc.
- * Most of this data already flows through the pipeline — this module
- * just captures what would otherwise be ephemeral.
+ * `ReportProvenance extends RunContext` (D0032). This module derives
+ * RunContext via `buildRunContext()` and attaches report-specific extras
+ * (lineage, autoScope, promptfoo URLs, targetDocuments, runId). A single
+ * derivation path for RunContext foreclosures drift between the run
+ * manifest (GCS) and the report provenance (Content Lake).
  *
- * @see docs/design-docs/report-store/domain-model.md
- * @see docs/design-docs/report-store/architecture.md — Provenance collection
+ * @see packages/eval/src/pipeline/run-context.ts — the shared derivation path
+ * @see docs/decisions/D0032-run-anchored-artifact-store.md (§ Drift Prevention)
  */
-import type { Logger } from "../_vendor/ailf-core/index.d.ts";
-import type { ResolvedSourceConfig } from "../sources.js";
-import type { EvalMode, PromptfooUrlEntry, ReportAutoScope, ReportProvenance } from "./types.js";
-export interface ProvenanceInput {
-    /** Feature areas that were evaluated */
-    areas: string[];
-    /** Logger instance (defaults to ConsoleLogger) */
-    logger?: Logger;
+import type { PromptfooUrlEntry, ReportAutoScope, ReportProvenance, RunId } from "./types.js";
+import { type RunContextInput } from "./run-context.js";
+/**
+ * Inputs needed to build a ReportProvenance. Extends `RunContextInput` so
+ * the RunContext derivation path is shared.
+ */
+export interface ProvenanceInput extends RunContextInput {
     /** Release auto-scope metadata (when perspective evaluation was scoped) */
     autoScope?: ReportAutoScope;
-    /**
-     * Git metadata from the *calling* repository (cross-repo evaluations).
-     * When provided, overrides CI env var detection so provenance attributes
-     * to the caller — not the AILF core repo where the workflow executes.
-     */
-    callerGit?: {
-        branch?: string;
-        prNumber?: number;
-        repo: string;
-        sha?: string;
-    };
     /** SHA-256 hash of the doc context files (from cache system) */
     contextHash?: string;
-    /** Evaluation fingerprint for cross-environment cache lookup */
-    evalFingerprint?: string;
-    /** Evaluation mode */
-    mode: EvalMode;
     /** @deprecated Use `promptfooUrls` — kept for backward compatibility */
     promptfooUrl?: string;
     /** Per-mode Promptfoo share URLs */
     promptfooUrls?: PromptfooUrlEntry[];
-    /** Path to the package root (for reading config/models) */
-    rootDir: string;
-    /** Report ID that triggered this re-run (becomes lineage.rerunOf) */
-    sourceReportId?: string;
+    /** Identity of the pipeline run that produced this report (D0032) */
+    runId: RunId;
     /** Sanity document IDs targeted */
     sanityDocumentIds?: string[];
-    /** Resolved documentation source */
-    source: ResolvedSourceConfig;
-    /** Specific task IDs evaluated (if scoped) */
-    taskIds?: string[];
+    /** Report ID that triggered this re-run (becomes lineage.rerunOf) */
+    sourceReportId?: string;
 }
 /**
- * Build a ReportProvenance object from pipeline context.
+ * Build a ReportProvenance from pipeline context.
  *
- * Assembles provenance from:
- * - Pipeline options (mode, source, areas, tasks)
- * - config/models.ts (model list, grader)
- * - Environment variables (CI metadata, trigger detection)
- * - Optional metadata (context hash, Promptfoo URL)
+ * RunContext fields (mode, areas, taskIds, models, graderModel, source,
+ * evalFingerprint, trigger, git) come from `buildRunContext`. Report-
+ * specific fields (autoScope, contextHash, lineage, promptfoo*, runId,
+ * targetDocuments) are attached here.
  */
 export declare function buildProvenance(input: ProvenanceInput): ReportProvenance;

package/dist/pipeline/provenance.js

@@ -1,188 +1,40 @@
 /**
  * pipeline/provenance.ts
  *
- * Builds ReportProvenance from data available during a pipeline run.
+ * Builds `ReportProvenance` from data available during a pipeline run.
  *
- * Provenance captures what produced an evaluation report: which models,
- * which source, which mode, what triggered it, git metadata, etc.
- * Most of this data already flows through the pipeline — this module
- * just captures what would otherwise be ephemeral.
+ * `ReportProvenance extends RunContext` (D0032). This module derives
+ * RunContext via `buildRunContext()` and attaches report-specific extras
+ * (lineage, autoScope, promptfoo URLs, targetDocuments, runId). A single
+ * derivation path for RunContext foreclosures drift between the run
+ * manifest (GCS) and the report provenance (Content Lake).
  *
- * @see docs/design-docs/report-store/domain-model.md
- * @see docs/design-docs/report-store/architecture.md — Provenance collection
+ * @see packages/eval/src/pipeline/run-context.ts — the shared derivation path
+ * @see docs/decisions/D0032-run-anchored-artifact-store.md (§ Drift Prevention)
  */
-import { ConsoleLogger } from "../adapters/loggers/index.js";
-import { tryLoadConfigFile } from "./compiler/config-loader.js";
+import { buildRunContext } from "./run-context.js";
 /**
- * Build a ReportProvenance object from pipeline context.
+ * Build a ReportProvenance from pipeline context.
  *
- * Assembles provenance from:
- * - Pipeline options (mode, source, areas, tasks)
- * - config/models.ts (model list, grader)
- * - Environment variables (CI metadata, trigger detection)
- * - Optional metadata (context hash, Promptfoo URL)
+ * RunContext fields (mode, areas, taskIds, models, graderModel, source,
+ * evalFingerprint, trigger, git) come from `buildRunContext`. Report-
+ * specific fields (autoScope, contextHash, lineage, promptfoo*, runId,
+ * targetDocuments) are attached here.
  */
 export function buildProvenance(input) {
-    const log = input.logger ?? new ConsoleLogger();
-    const models = loadModelsConfig(input.rootDir, log);
-    log.debug("Assembling provenance input", {
-        mode: input.mode,
-        sourceName: input.source.name,
-        sourceBaseUrl: input.source.baseUrl,
-        areas: input.areas,
-        taskIds: input.taskIds,
-        hasContextHash: Boolean(input.contextHash),
-        hasEvalFingerprint: Boolean(input.evalFingerprint),
-        hasCallerGit: Boolean(input.callerGit),
-        hasSourceReportId: Boolean(input.sourceReportId),
-        modelCount: models.models.length,
-    });
-    // Cross-repo evaluations: prefer explicit caller git metadata over
-    // CI env vars (which always reflect the AILF core repo).
-    const git = input.callerGit
-        ? {
-            branch: input.callerGit.branch ?? "unknown",
-            prNumber: input.callerGit.prNumber,
-            repo: input.callerGit.repo,
-            sha: input.callerGit.sha ?? "unknown",
-        }
-        : detectGitMetadata();
+    const runContext = buildRunContext(input);
     // Build lineage from explicit relationships
     const lineage = input.sourceReportId
         ? { rerunOf: input.sourceReportId }
         : undefined;
-    const trigger = detectTrigger();
-    log.debug("Provenance computed", {
-        triggerType: trigger.type,
-        gitRepo: git?.repo,
-        gitBranch: git?.branch,
-        evalFingerprint: input.evalFingerprint,
-        hasLineage: Boolean(lineage),
-    });
-    // Non-literacy modes (agent-harness, mcp-server, etc.) don't use the
-    // config/models.ts model matrix — listing those models would be misleading.
-    // Only include them for literacy mode where they're the actual eval targets.
-    const evaluatedModels = input.mode === "literacy"
-        ? models.models.map((m) => ({ id: m.id, label: m.label }))
-        : [];
     return {
-        areas: input.areas,
+        ...runContext,
         autoScope: input.autoScope,
         contextHash: input.contextHash,
-        evalFingerprint: input.evalFingerprint,
-        git,
-        graderModel: models.grader.id,
         lineage,
-        mode: input.mode,
-        models: evaluatedModels,
         promptfooUrl: input.promptfooUrl,
         promptfooUrls: input.promptfooUrls,
-        source: {
-            baseUrl: input.source.baseUrl,
-            dataset: input.source.dataset,
-            name: input.source.name,
-            perspective: input.source.perspective,
-            projectId: input.source.projectId,
-        },
+        runId: input.runId,
         targetDocuments: input.sanityDocumentIds,
-        taskIds: input.taskIds,
-        trigger: detectTrigger(),
-    };
-}
-// ---------------------------------------------------------------------------
-// Trigger detection
-// ---------------------------------------------------------------------------
-/**
- * Extract git metadata from GitHub Actions environment variables.
- * Returns undefined when not running in CI.
- */
-function detectGitMetadata() {
-    const repo = process.env.GITHUB_REPOSITORY;
-    if (!repo)
-        return undefined;
-    const sha = process.env.GITHUB_SHA ?? "unknown";
-    const ref = process.env.GITHUB_REF ?? "";
-    // Extract branch name from ref (refs/heads/main → main)
-    const branch = ref.startsWith("refs/heads/")
-        ? ref.slice("refs/heads/".length)
-        : ref.startsWith("refs/pull/")
-            ? `pr-${ref.split("/")[2]}`
-            : ref;
-    // Extract PR number from GITHUB_REF (refs/pull/123/merge)
-    const prMatch = ref.match(/^refs\/pull\/(\d+)\//);
-    const prNumber = prMatch ? parseInt(prMatch[1], 10) : undefined;
-    return { branch, prNumber, repo, sha };
-}
-// ---------------------------------------------------------------------------
-// Git metadata
-// ---------------------------------------------------------------------------
-/**
- * Infer what triggered this evaluation from environment variables.
- *
- * Detection order:
- * 1. AILF_TRIGGER_TYPE — explicit override (for custom integrations)
- * 2. GITHUB_EVENT_NAME === "schedule" — cron-triggered
- * 3. GITHUB_EVENT_NAME === "repository_dispatch" — cross-repo trigger
- * 4. GITHUB_ACTIONS === "true" — CI-triggered
- * 5. Default: manual
- */
-function detectTrigger() {
-    const explicit = process.env.AILF_TRIGGER_TYPE;
-    if (explicit === "scheduled") {
-        return {
-            schedule: process.env.AILF_SCHEDULE ?? "unknown",
-            type: "scheduled",
-        };
-    }
-    if (explicit === "webhook") {
-        return {
-            documentId: process.env.AILF_WEBHOOK_DOCUMENT_ID,
-            source: process.env.AILF_WEBHOOK_SOURCE ?? "unknown",
-            type: "webhook",
-        };
-    }
-    // GitHub Actions context
-    const eventName = process.env.GITHUB_EVENT_NAME;
-    if (eventName === "schedule") {
-        return {
-            schedule: process.env.GITHUB_SCHEDULE ?? "unknown",
-            type: "scheduled",
-        };
-    }
-    if (eventName === "repository_dispatch") {
-        return {
-            callerRef: process.env.GITHUB_REF,
-            // Note: callerRepo here is a fallback. The accurate caller repo
-            // comes from callerGit (injected into the PipelineRequest payload).
-            // GITHUB_REPOSITORY_OWNER_ID is just the org ID, not owner/repo.
-            callerRepo: process.env.GITHUB_REPOSITORY_OWNER_ID ?? "unknown",
-            type: "cross-repo",
-        };
-    }
-    if (process.env.GITHUB_ACTIONS === "true") {
-        return {
-            runId: process.env.GITHUB_RUN_ID ?? "unknown",
-            type: "ci",
-            workflow: process.env.GITHUB_WORKFLOW ?? "unknown",
-        };
-    }
-    return { type: "manual" };
-}
-// ---------------------------------------------------------------------------
-// Model config loading
-// ---------------------------------------------------------------------------
-/**
- * Load config/models to extract model list and grader info.
- * Falls back to a minimal config if the file can't be read.
- */
-function loadModelsConfig(rootDir, log) {
-    const result = tryLoadConfigFile("models", rootDir);
-    if (result)
-        return result.data;
-    log.warn("Could not read config/models for provenance");
-    return {
-        defaults: {},
-        grader: { id: "unknown" },
-        models: [],
     };
 }

package/dist/pipeline/report-title.d.ts

@@ -15,7 +15,7 @@
  * @see docs/design-docs/report-store/domain-model.md
  * @see packages/eval/src/pipeline/provenance.ts — builds the provenance input
  */
-import type { EvalMode, ReportTrigger } from "./types.js";
+import type { EvalMode, RunTrigger } from "./types.js";
 /** Input required to generate a human-readable report title. */
 export interface ReportTitleInput {
     provenance: {
@@ -31,7 +31,7 @@ export interface ReportTitleInput {
         /** Sanity document IDs targeted (when scoped to specific documents) */
         targetDocuments?: string[];
         /** What triggered the evaluation */
-        trigger: ReportTrigger;
+        trigger: RunTrigger;
     };
     /**
      * Total number of known feature areas in the system.

package/dist/pipeline/run-context.d.ts

@@ -0,0 +1,57 @@
+/**
+ * buildRunContext — the single code path that derives `RunContext` from
+ * pipeline inputs.
+ *
+ * `RunContext` is the 9-field shape shared between `RunManifest.context`
+ * (in GCS) and `ReportProvenance` (in Content Lake, which `extends
+ * RunContext`). Routing every consumer through this function makes it
+ * structurally impossible for the two to disagree: there is no second
+ * code path to drift against.
+ *
+ * Contract test: `packages/eval/src/__tests__/run-context-parity.test.ts`
+ *
+ * @see docs/decisions/D0032-run-anchored-artifact-store.md (§ Move 5 — Drift Prevention)
+ */
+import type { Logger, RunContext } from "../_vendor/ailf-core/index.d.ts";
+import type { ResolvedSourceConfig } from "../sources.js";
+import type { EvalMode } from "./types.js";
+/**
+ * Inputs required to derive a `RunContext`. `ProvenanceInput` extends this
+ * so every caller of `buildProvenance` is automatically a valid input to
+ * `buildRunContext`.
+ */
+export interface RunContextInput {
+    /** Feature areas that were evaluated */
+    areas: string[];
+    /**
+     * Git metadata from the *calling* repository (cross-repo evaluations).
+     * When provided, overrides CI env var detection so context attributes
+     * to the caller — not the AILF core repo where the workflow executes.
+     */
+    callerGit?: {
+        branch?: string;
+        prNumber?: number;
+        repo: string;
+        sha?: string;
+    };
+    /** Evaluation fingerprint for cross-environment cache lookup */
+    evalFingerprint?: string;
+    /** Logger instance (defaults to ConsoleLogger) */
+    logger?: Logger;
+    /** Evaluation mode */
+    mode: EvalMode;
+    /** Path to the package root (for reading config/models) */
+    rootDir: string;
+    /** Resolved documentation source */
+    source: ResolvedSourceConfig;
+    /** Specific task IDs evaluated (if scoped) */
+    taskIds?: string[];
+}
+/**
+ * Derive `RunContext` from pipeline inputs. The only construction path.
+ *
+ * Both `FinalizeRunStep` (via `RunManifest.context`) and
+ * `PublishReportStep` (via `ReportProvenance`) call this function — the
+ * former directly, the latter transitively through `buildProvenance`.
+ */
+export declare function buildRunContext(input: RunContextInput): RunContext;

package/dist/pipeline/run-context.js

@@ -0,0 +1,156 @@
+/**
+ * buildRunContext — the single code path that derives `RunContext` from
+ * pipeline inputs.
+ *
+ * `RunContext` is the 9-field shape shared between `RunManifest.context`
+ * (in GCS) and `ReportProvenance` (in Content Lake, which `extends
+ * RunContext`). Routing every consumer through this function makes it
+ * structurally impossible for the two to disagree: there is no second
+ * code path to drift against.
+ *
+ * Contract test: `packages/eval/src/__tests__/run-context-parity.test.ts`
+ *
+ * @see docs/decisions/D0032-run-anchored-artifact-store.md (§ Move 5 — Drift Prevention)
+ */
+import { ConsoleLogger } from "../adapters/loggers/index.js";
+import { tryLoadConfigFile } from "./compiler/config-loader.js";
+/**
+ * Derive `RunContext` from pipeline inputs. The only construction path.
+ *
+ * Both `FinalizeRunStep` (via `RunManifest.context`) and
+ * `PublishReportStep` (via `ReportProvenance`) call this function — the
+ * former directly, the latter transitively through `buildProvenance`.
+ */
+export function buildRunContext(input) {
+    const log = input.logger ?? new ConsoleLogger();
+    const models = loadModelsConfig(input.rootDir, log);
+    // Cross-repo evaluations: prefer explicit caller git metadata over
+    // CI env vars (which always reflect the AILF core repo).
+    const git = input.callerGit
+        ? {
+            branch: input.callerGit.branch ?? "unknown",
+            prNumber: input.callerGit.prNumber,
+            repo: input.callerGit.repo,
+            sha: input.callerGit.sha ?? "unknown",
+        }
+        : detectGitMetadata();
+    const trigger = detectTrigger();
+    // Non-literacy modes (agent-harness, mcp-server, etc.) don't use the
+    // config/models.ts model matrix — listing those models would be
+    // misleading. Only include them for literacy mode where they're the
+    // actual eval targets.
+    const evaluatedModels = input.mode === "literacy"
+        ? models.models.map((m) => ({ id: m.id, label: m.label }))
+        : [];
+    return {
+        areas: input.areas,
+        evalFingerprint: input.evalFingerprint,
+        git,
+        graderModel: models.grader.id,
+        mode: input.mode,
+        models: evaluatedModels,
+        source: {
+            baseUrl: input.source.baseUrl,
+            dataset: input.source.dataset,
+            name: input.source.name,
+            perspective: input.source.perspective,
+            projectId: input.source.projectId,
+        },
+        taskIds: input.taskIds,
+        trigger,
+    };
+}
+// ---------------------------------------------------------------------------
+// Environment-derived context
+// ---------------------------------------------------------------------------
+/**
+ * Extract git metadata from GitHub Actions environment variables.
+ * Returns undefined when not running in CI.
+ */
+function detectGitMetadata() {
+    const repo = process.env.GITHUB_REPOSITORY;
+    if (!repo)
+        return undefined;
+    const sha = process.env.GITHUB_SHA ?? "unknown";
+    const ref = process.env.GITHUB_REF ?? "";
+    // Extract branch name from ref (refs/heads/main → main)
+    const branch = ref.startsWith("refs/heads/")
+        ? ref.slice("refs/heads/".length)
+        : ref.startsWith("refs/pull/")
+            ? `pr-${ref.split("/")[2]}`
+            : ref;
+    // Extract PR number from GITHUB_REF (refs/pull/123/merge)
+    const prMatch = ref.match(/^refs\/pull\/(\d+)\//);
+    const prNumber = prMatch ? parseInt(prMatch[1], 10) : undefined;
+    return { branch, prNumber, repo, sha };
+}
+/**
+ * Infer what triggered this evaluation from environment variables.
+ *
+ * Detection order:
+ * 1. AILF_TRIGGER_TYPE — explicit override (for custom integrations)
+ * 2. GITHUB_EVENT_NAME === "schedule" — cron-triggered
+ * 3. GITHUB_EVENT_NAME === "repository_dispatch" — cross-repo trigger
+ * 4. GITHUB_ACTIONS === "true" — CI-triggered
+ * 5. Default: manual
+ */
+function detectTrigger() {
+    const explicit = process.env.AILF_TRIGGER_TYPE;
+    if (explicit === "scheduled") {
+        return {
+            schedule: process.env.AILF_SCHEDULE ?? "unknown",
+            type: "scheduled",
+        };
+    }
+    if (explicit === "webhook") {
+        return {
+            documentId: process.env.AILF_WEBHOOK_DOCUMENT_ID,
+            source: process.env.AILF_WEBHOOK_SOURCE ?? "unknown",
+            type: "webhook",
+        };
+    }
+    // GitHub Actions context
+    const eventName = process.env.GITHUB_EVENT_NAME;
+    if (eventName === "schedule") {
+        return {
+            schedule: process.env.GITHUB_SCHEDULE ?? "unknown",
+            type: "scheduled",
+        };
+    }
+    if (eventName === "repository_dispatch") {
+        return {
+            callerRef: process.env.GITHUB_REF,
+            // Note: callerRepo here is a fallback. The accurate caller repo
+            // comes from callerGit (injected into the PipelineRequest payload).
+            // GITHUB_REPOSITORY_OWNER_ID is just the org ID, not owner/repo.
+            callerRepo: process.env.GITHUB_REPOSITORY_OWNER_ID ?? "unknown",
+            type: "cross-repo",
+        };
+    }
+    if (process.env.GITHUB_ACTIONS === "true") {
+        return {
+            runId: process.env.GITHUB_RUN_ID ?? "unknown",
+            type: "ci",
+            workflow: process.env.GITHUB_WORKFLOW ?? "unknown",
+        };
+    }
+    return { type: "manual" };
+}
+// ---------------------------------------------------------------------------
+// Model config loading
+// ---------------------------------------------------------------------------
+/**
+ * Load config/models to extract model list and grader info.
+ * Falls back to a minimal config if the file can't be read.
+ */
+function loadModelsConfig(rootDir, log) {
+    const result = tryLoadConfigFile("models", rootDir);
+    if (result)
+        return result.data;
+    log.warn("Could not read config/models for run context");
+    return {
+        defaults: {},
+        grader: { id: "unknown" },
+        models: [],
+    };
+}

package/dist/pipeline/upload-test-outputs.d.ts

@@ -0,0 +1,26 @@
+/**
+ * upload-test-outputs.ts — shared helper for the testOutputs artifact upload.
+ *
+ * CalculateScoresStep calls this once its score-summary.json is complete.
+ * Each {taskId, modelId} pair becomes one GCS object under
+ * `runs/{runId}/test-outputs/{taskId}--{modelId}.json` carrying the full
+ * response output and truncation flag. The returned ArtifactRef's
+ * `entries[]` catalog lists every uploaded entry so Studio can render
+ * drill-down state without a second listing call.
+ *
+ * PublishReportStep later strips responseOutput from the inline
+ * testResults[] when this upload succeeds, so the Content Lake document
+ * stays slim — the full output lives in GCS and is fetched per-entry
+ * on click.
+ *
+ * @see docs/decisions/D0032-run-anchored-artifact-store.md
+ */
+import type { ArtifactRef, ArtifactWriter, RunId, StoredTestResult } from "../_vendor/ailf-core/index.d.ts";
+/**
+ * Upload testOutputs as per-entry GCS objects under
+ * `runs/{runId}/test-outputs/`, one per `{taskId}::{modelId}` pair.
+ *
+ * Returns the `ArtifactRef` on success, or `null` when upload is skipped or
+ * fails (P5: non-blocking).
+ */
+export declare function uploadTestOutputs(writer: ArtifactWriter, runId: RunId, testResults: StoredTestResult[]): Promise<ArtifactRef | null>;