@sanity/ailf 5.0.0 → 6.1.0

package/dist/_vendor/ailf-core/services/diagnosis-runner.js

@@ -1,12 +1,22 @@
 /**
  * Diagnosis runner — engine entry point (D0048).
  *
- * Phase 1 lands the version constant only; the runner factory + cache
- * lookup land in Phase 5.
+ * Phase 5 implements the factory body; Phase 1 shipped `diagnosisVersion` only.
+ * `GeneratorContext.judgmentAttributions` is sourced once per `.run({...})` via
+ * `deps.loadAttributions(runId)` reading Phase 4's
+ * `runs/{runId}/attribution/{entryKey}.json` per-entry artifacts (RESEARCH
+ * Landmine 11).
  *
  * @see docs/decisions/D0048-engine-homes-for-cli-api-parity.md
  * @see .planning/phases/01-foundation-contracts-cross-cutting-schemas/01-CONTEXT.md (D-02)
+ * @see .planning/phases/05-diagnosis-engine-cli-llm-cards/05-RESEARCH.md (Landmine 11)
+ * @see .planning/phases/05-diagnosis-engine-cli-llm-cards/05-CONTEXT.md (D-02, D-06, D-10)
  */
+import { z } from "zod";
+import { ARTIFACT_REGISTRY, encodeDiagnosisPathVersion, } from "../artifact-registry.js";
+// ---------------------------------------------------------------------------
+// Version constant (Phase 1 / VER-01 / D-02)
+// ---------------------------------------------------------------------------
 /**
  * Bumped when the runner's selection logic, prompt orchestration, or
  * card-set composition changes in a way that should invalidate cached
@@ -17,3 +27,127 @@
  * across vitest workers (cross-cutting hazard #2).
  */
 export const diagnosisVersion = "0.1.0";
+// ---------------------------------------------------------------------------
+// Private helpers
+// ---------------------------------------------------------------------------
+/**
+ * Build the deterministic cache path that incorporates all four version
+ * segments AND the model id (AI-SPEC §3 lines 463-473 + D-02).
+ *
+ * The artifact path from `ARTIFACT_REGISTRY.diagnosis.objectPath(...)` is
+ * already version-scoped; we append `::${model}` to include model identity
+ * in the key without changing the artifact path shape.
+ */
+function buildCacheKey(report, versions, model) {
+    const artifactPath = ARTIFACT_REGISTRY.diagnosis.objectPath(report.provenance.runId, report.id, encodeDiagnosisPathVersion(versions.diagnosisVersion, versions.cardVersion));
+    // Embed the remaining two version axes + model in the key string. The
+    // artifact path already carries diagnosisVersion + cardVersion; the other
+    // two axes are appended here so any single-segment bump produces a
+    // distinct key.
+    return `${artifactPath}::grader=${versions.graderJudgmentsVersion}::ensemble=${versions.ensembleVersion}::model=${model}`;
+}
+/**
+ * Per-card invocation — never panics. ZodError or any other thrown value
+ * both translate to a degraded card (AI-SPEC §3 lines 530-552).
+ */
+async function runOne(generator, report, ctx, cardType) {
+    try {
+        return await generator(report, ctx);
+    }
+    catch (err) {
+        const meta = {
+            cardVersion: `${cardType}@unknown`,
+            generatedAt: new Date().toISOString(),
+        };
+        const isZodErr = err instanceof z.ZodError;
+        return {
+            status: "degraded",
+            cardType,
+            reason: err instanceof Error ? err.message : String(err),
+            parseFailed: isZodErr,
+            meta,
+        };
+    }
+}
+// ---------------------------------------------------------------------------
+// Factory (AI-SPEC §3 lines 458-523 + D-02 / Landmine-11 deltas)
+// ---------------------------------------------------------------------------
+/**
+ * Build a `DiagnosisRunner` whose `.run({report, versions, baseline?, refresh?})`
+ * produces a `Diagnosis` with cards in registry-order.
+ *
+ * No module-scope `let` — all state lives in the `deps` closure and per-run
+ * local variables (AI-SPEC §3 Pitfall 1).
+ */
+export function createDiagnosisRunner(deps) {
+    return {
+        async run({ report, versions, baseline, refresh }) {
+            const cachePath = buildCacheKey(report, versions, deps.model);
+            // Cache lookup (bypassed when --refresh).
+            if (!refresh) {
+                const cached = await deps.diagnosisReader(cachePath);
+                if (cached !== null)
+                    return cached;
+            }
+            // One-shot attribution load (Landmine 11 — Phase 4 per-entry artifacts).
+            let judgmentAttributions;
+            try {
+                judgmentAttributions = await deps.loadAttributions(report.provenance.runId);
+            }
+            catch (err) {
+                deps.logger.warn("diagnosis-runner: loadAttributions failed", {
+                    runId: report.provenance.runId,
+                    error: err instanceof Error ? err.message : String(err),
+                });
+                judgmentAttributions = undefined;
+            }
+            const ctx = {
+                llm: deps.llm,
+                model: deps.model,
+                logger: deps.logger,
+                progress: deps.progress,
+                versions,
+                runId: report.provenance.runId, // D-10: provenance.runId, NOT report.runId
+                reportId: report.id,
+                judgmentAttributions, // Landmine 11
+                ...(baseline ? { baseline } : {}),
+            };
+            const cardTypes = Object.keys(deps.registry);
+            const cards = [];
+            let parseFailures = 0;
+            for (const cardType of cardTypes) {
+                const generator = deps.registry[cardType];
+                // Budget enforcement: once ≤1 budget is breached, downgrade
+                // subsequent parse-failing cards to "missing" before even running
+                // the generator (AI-SPEC §3 lines 496-510 + must-have #4).
+                // We still RUN the generator here to match the behavior spec —
+                // the budget check happens AFTER the card result is obtained.
+                const card = await runOne(generator, report, ctx, cardType);
+                if (card.status === "degraded" && card.parseFailed) {
+                    if (parseFailures >= 1) {
+                        // Budget exceeded — demote to missing.
+                        deps.logger.warn(`diagnosis-runner: parse-failure budget exceeded for card "${cardType}"; demoting to missing`, { reportId: report.id });
+                        cards.push({
+                            status: "missing",
+                            cardType,
+                            reason: "degraded-budget-exceeded",
+                        });
+                        continue;
+                    }
+                    parseFailures++;
+                }
+                cards.push(card);
+            }
+            const diagnosis = {
+                runId: report.provenance.runId, // D-10: provenance.runId
+                reportId: report.id,
+                inputs: versions,
+                cards,
+                generatedAt: new Date().toISOString(),
+            };
+            // Unconditional write — a refreshed call replaces the cached Diagnosis.
+            await deps.diagnosisWriter(cachePath, diagnosis);
+            return diagnosis;
+        },
+    };
+}

package/dist/_vendor/ailf-core/services/index.d.ts

@@ -13,5 +13,9 @@ export { aggregateAreas, aggregateDimensions, computeEnsembleScore, computeTaskS
 export { extractModelName, extractProvider, mergeConfig, modelMatchesMode, resolveModelVariants, } from "./config-helpers.js";
 export { buildSlimReportSummary } from "./slim-report-summary.js";
 export { reportToMarkdown, type RenderableReport, } from "./report-to-markdown.js";
-export { diagnosisVersion } from "./diagnosis-runner.js";
+export { createDiagnosisRunner, diagnosisVersion, type CardGenerator, type CardRegistry, type DiagnosisRunner, type DiagnosisRunnerDeps, type DiagnosisRunnerRunArgs, type GeneratorContext, } from "./diagnosis-runner.js";
 export { cardRegistry, type CardDefinition } from "./diagnosis/registry.js";
+export { createLLMClient, type LLMClientAdapters, type LLMClientFactoryConfig, type LLMClientKeys, } from "./llm-client-factory.js";
+export { buildFailureModeRefinement, isFailureModeInDimensionTaxonomy, } from "./diagnosis/card-validators.js";
+export { CARD_REGISTRY_VERSION, DIAGNOSIS_CARD_GENERATORS, generateAreaSummary, generateFailureModeSummary, generateNoIssues, generateTopRecommendations, generateWeakestArea, generateLowConfidenceAttribution, generateDocAttributionSpotlight, generateRegressionVsBaseline, } from "./diagnosis/cards/index.js";
+export { buildTopRecommendationsPrompt, buildWeakestAreaPrompt, buildLowConfidenceAttributionPrompt, buildDocAttributionSpotlightPrompt, buildRegressionVsBaselinePrompt, buildDocSlugAllowList, } from "./diagnosis/prompt-builders.js";

package/dist/_vendor/ailf-core/services/index.js

@@ -14,7 +14,20 @@ export { extractModelName, extractProvider, mergeConfig, modelMatchesMode, resol
 export { buildSlimReportSummary } from "./slim-report-summary.js";
 export { reportToMarkdown, } from "./report-to-markdown.js";
 // ---------------------------------------------------------------------------
-// Actionability ladder Phase 1 — diagnosis runner + card registry
+// Actionability ladder Phase 1 + Phase 5 — diagnosis runner + card registry
 // ---------------------------------------------------------------------------
-export { diagnosisVersion } from "./diagnosis-runner.js";
+export { createDiagnosisRunner, diagnosisVersion, } from "./diagnosis-runner.js";
 export { cardRegistry } from "./diagnosis/registry.js";
+// ---------------------------------------------------------------------------
+// Phase 5 — LLM client factory (D-01 hoist)
+// ---------------------------------------------------------------------------
+export { createLLMClient, } from "./llm-client-factory.js";
+// ---------------------------------------------------------------------------
+// Phase 5 — card validators (D-05 refine helpers)
+// ---------------------------------------------------------------------------
+export { buildFailureModeRefinement, isFailureModeInDimensionTaxonomy, } from "./diagnosis/card-validators.js";
+// ---------------------------------------------------------------------------
+// Phase 5 Plan 05 — card generators barrel + prompt builders
+// ---------------------------------------------------------------------------
+export { CARD_REGISTRY_VERSION, DIAGNOSIS_CARD_GENERATORS, generateAreaSummary, generateFailureModeSummary, generateNoIssues, generateTopRecommendations, generateWeakestArea, generateLowConfidenceAttribution, generateDocAttributionSpotlight, generateRegressionVsBaseline, } from "./diagnosis/cards/index.js";
+export { buildTopRecommendationsPrompt, buildWeakestAreaPrompt, buildLowConfidenceAttributionPrompt, buildDocAttributionSpotlightPrompt, buildRegressionVsBaselinePrompt, buildDocSlugAllowList, } from "./diagnosis/prompt-builders.js";

package/dist/_vendor/ailf-core/services/llm-client-factory.d.ts

@@ -0,0 +1,64 @@
+/**
+ * LLM client factory — hoisted from packages/eval/src/composition-root.ts
+ * so packages/api can build a DiagnosisRunner without importing eval (D-01).
+ *
+ * Adapter CLASSES stay in packages/eval/src/adapters/llm/. Only the factory
+ * function lives here. Adapter constructors are injected via `LLMClientAdapters`
+ * so core never static-imports vendor SDK code (D0051 invariant / T-05-01-01).
+ *
+ * @see docs/decisions/D0051-llm-client-port.md
+ * @see packages/eval/src/composition-root.ts — call site (updated to use this)
+ */
+import type { LLMClient } from "../ports/llm-client.js";
+import type { Logger } from "../ports/logger.js";
+/**
+ * Narrow config slice consumed by the LLM client factory.
+ * Does NOT depend on `ResolvedConfig` from packages/eval — only the
+ * llmProvider field is needed here.
+ */
+export interface LLMClientFactoryConfig {
+    readonly llmProvider?: "anthropic" | "openai";
+}
+/**
+ * Typed key bag passed to `createLLMClient`. The composition root reads
+ * env once and supplies values here; the factory stays pure so tests don't
+ * have to mutate `process.env`.
+ */
+export interface LLMClientKeys {
+    readonly anthropicApiKey?: string;
+    readonly openaiApiKey?: string;
+}
+/**
+ * Constructor callbacks for adapter classes that live in packages/eval.
+ * The eval composition root passes real constructors; tests pass spies.
+ *
+ * This pattern satisfies T-05-01-01: core never static-imports
+ * openai / @anthropic-ai/sdk. The vendor code stays in eval.
+ */
+export interface LLMClientAdapters {
+    readonly newAnthropicClient: (opts: {
+        apiKey: string;
+        logger: Logger;
+    }) => LLMClient;
+    readonly newOpenAIClient: (opts: {
+        apiKey: string;
+        logger: Logger;
+    }) => LLMClient;
+}
+/**
+ * Select the LLMClient adapter based on `config.llmProvider` and the
+ * supplied API keys. Returns `undefined` when no usable credential is
+ * present — `AppContext.llmClient` stays unset and consumers handle that
+ * explicitly.
+ *
+ * Adapters never read `process.env` themselves (per
+ * `.claude/rules/typescript.md`); env mapping happens at the call site
+ * (typically `createAppContext`).
+ *
+ * Adapter classes stay in packages/eval; they are passed in via `deps.adapters`
+ * so this factory has zero eval imports (D-01 / T-05-01-01).
+ */
+export declare function createLLMClient(config: LLMClientFactoryConfig, keys: LLMClientKeys, deps: {
+    logger: Logger;
+    adapters: LLMClientAdapters;
+}): LLMClient | undefined;

package/dist/_vendor/ailf-core/services/llm-client-factory.js

@@ -0,0 +1,54 @@
+/**
+ * LLM client factory — hoisted from packages/eval/src/composition-root.ts
+ * so packages/api can build a DiagnosisRunner without importing eval (D-01).
+ *
+ * Adapter CLASSES stay in packages/eval/src/adapters/llm/. Only the factory
+ * function lives here. Adapter constructors are injected via `LLMClientAdapters`
+ * so core never static-imports vendor SDK code (D0051 invariant / T-05-01-01).
+ *
+ * @see docs/decisions/D0051-llm-client-port.md
+ * @see packages/eval/src/composition-root.ts — call site (updated to use this)
+ */
+// ---------------------------------------------------------------------------
+// Factory function
+// ---------------------------------------------------------------------------
+/**
+ * Select the LLMClient adapter based on `config.llmProvider` and the
+ * supplied API keys. Returns `undefined` when no usable credential is
+ * present — `AppContext.llmClient` stays unset and consumers handle that
+ * explicitly.
+ *
+ * Adapters never read `process.env` themselves (per
+ * `.claude/rules/typescript.md`); env mapping happens at the call site
+ * (typically `createAppContext`).
+ *
+ * Adapter classes stay in packages/eval; they are passed in via `deps.adapters`
+ * so this factory has zero eval imports (D-01 / T-05-01-01).
+ */
+export function createLLMClient(config, keys, deps) {
+    const { logger, adapters } = deps;
+    const explicit = config.llmProvider;
+    const anthropicKey = keys.anthropicApiKey;
+    const openaiKey = keys.openaiApiKey;
+    // Auto-select: prefer Anthropic when both are present (matches the
+    // current grader's default model in config/models.ts).
+    const provider = explicit ?? (anthropicKey ? "anthropic" : openaiKey ? "openai" : undefined);
+    if (!provider) {
+        logger.debug("LLM client: not wired — no Anthropic or OpenAI API key supplied");
+        return undefined;
+    }
+    if (provider === "anthropic") {
+        if (!anthropicKey) {
+            logger.warn('llmProvider="anthropic" but no Anthropic API key supplied — LLMClient not wired');
+            return undefined;
+        }
+        logger.debug("LLM client: AnthropicLLMClient");
+        return adapters.newAnthropicClient({ apiKey: anthropicKey, logger });
+    }
+    if (!openaiKey) {
+        logger.warn('llmProvider="openai" but no OpenAI API key supplied — LLMClient not wired');
+        return undefined;
+    }
+    logger.debug("LLM client: OpenAILLMClient");
+    return adapters.newOpenAIClient({ apiKey: openaiKey, logger });
+}

package/dist/_vendor/ailf-core/types/diagnosis.d.ts

@@ -8,14 +8,18 @@
  * discriminator inside the `ready` variant.
  *
  * Phase 1 lands placeholder body shapes; Phase 5 enriches each per
- * Doc 05 specs.
+ * AI-SPEC §3 and CONTEXT D-05/D-07. The `DiagnosisCard` discriminated
+ * union surface (arms + `cardType` literals) is stable — only the
+ * `body: <BodyInterface>` references resolve to richer shapes.
  *
  * @see docs/decisions/D0049-shared-confidence-contract.md
  * @see docs/decisions/D0052-judgment-ref-granularity.md
  * @see docs/decisions/D0050-artifact-registry-post-hoc-versioned-extension.md
  */
+import type { Confidence } from "./confidence.js";
 import type { RunId } from "./branded-ids.js";
 import type { ReportId } from "./index.js";
+import type { ModelId } from "../ports/llm-client.js";
 /**
  * The four-version cache envelope. Every cached `Diagnosis` carries the
  * versions of the inputs that produced it; any bump in any segment
@@ -51,47 +55,144 @@ export interface CardMeta {
     latencyMs?: number;
     /** ISO 8601 UTC timestamp. */
     generatedAt: string;
+    cost?: number;
+    model?: ModelId;
 }
 /**
  * A single actionable suggestion surfaced by a recommendations card.
- * The full Phase 5 shape may add fields (per Doc 05 specs); Phase 1
- * locks the minimum required surface.
+ *
+ * Phase 5 adds `docSlug` (the canonical doc page to rewrite) and
+ * `sectionHeading` (null when the suggestion targets the whole page)
+ * per AI-SPEC actionability-specificity rubric + failure-mode #2
+ * mitigation.
  */
 export interface ActionSuggestion {
     title: string;
     body: string;
     priority: "high" | "medium" | "low";
+    /** Canonical slug of the documentation page this suggestion targets. */
+    docSlug: string;
+    /**
+     * Heading within `docSlug` that should be revised, or `null` when the
+     * suggestion targets the page as a whole.
+     */
+    sectionHeading: string | null;
+}
+/**
+ * Minimal judgment reference per D0052 (taskId × modelId × dimension).
+ * Used by `LowConfidenceAttributionBody.judgmentRefs` to cite the
+ * specific judgments that drove a low-confidence finding.
+ */
+export interface JudgmentRef {
+    taskId: string;
+    modelId: string;
+    dimension: string;
 }
 /**
- * Phase 1 body placeholders. Each shape is intentionally minimal; Phase 5
- * card files enrich them per Doc 05 specs and assert
- * `satisfies z.ZodType<Extract<DiagnosisCard, { status: "ready"; cardType: "X" }>["body"]>`
- * against these declarations.
+ * Phase 5 enriched body shapes. Each keeps `summary: string` (load-bearing
+ * for CLI default render per AI-SPEC §6) and adds fields the corresponding
+ * Zod schema needs (asserting `satisfies z.ZodType<T>` in the card file).
  */
+/** area-summary: deterministic — keep only summary (no behavioral claims). */
 export interface AreaSummaryBody {
     summary: string;
 }
+/**
+ * failure-mode-summary: deterministic + D-05 dimension/failureMode gate.
+ * `count` = frequency in the report; `sampleSize` = judgment count for the
+ * dimension (per AI-SPEC failure-mode #3 mitigation).
+ */
 export interface FailureModeSummaryBody {
     summary: string;
+    /** Rubric dimension this summary targets (e.g. "task-completion"). */
+    dimension: string;
+    /** Canonical failure mode within this dimension. */
+    failureMode: string;
+    /** Number of judgments in this report with this failure mode. */
+    count: number;
+    /** Total judgments for this dimension — calibration denominator. */
+    sampleSize: number;
 }
+/**
+ * no-issues: deterministic + AI-SPEC failure-mode #7 sycophancy guard.
+ * `thresholdScore` surfaces the threshold used to qualify as "no issues"
+ * so readers can see the criterion behind the positive assessment.
+ */
 export interface NoIssuesBody {
     summary: string;
+    /** Minimum composite score that qualified this area as "no issues". */
+    thresholdScore: number;
 }
+/**
+ * top-recommendations: LLM-driven. `suggestions` reuses the enriched
+ * `ActionSuggestion` shape (docSlug + sectionHeading per AI-SPEC
+ * actionability-specificity rubric + failure-mode #2 mitigation).
+ */
 export interface TopRecommendationsBody {
     summary: string;
     suggestions: ActionSuggestion[];
 }
+/**
+ * weakest-area: LLM-driven. Adds area identification, dimension/failureMode
+ * context, and a small-sample calibration guard (AI-SPEC failure-mode #3).
+ */
 export interface WeakestAreaBody {
     summary: string;
+    /** Documentation area with the lowest composite score. */
+    area: string;
+    /** Primary dimension driving the low score. */
+    dimension: string;
+    /** Dominant failure mode in this area. */
+    failureMode: string;
+    /** Number of judgments sampled for this area — calibration denominator. */
+    sampleSize: number;
+    /** Calibrated confidence per D0049 (ensemble-stdev derivation). */
+    confidence: Confidence;
 }
-export interface LowConfidenceAttributionBody {
+/**
+ * regression-vs-baseline: LLM-driven. `deltas` is the per-area diff
+ * (JS-computed pre-call, max 10 entries); `drivers` is LLM prose;
+ * `overallTrend` is a 4-bucket summary per AI-SPEC §3 lines 605-613.
+ */
+export interface RegressionVsBaselineBody {
     summary: string;
+    /**
+     * Per-area score deltas (max 10). `drivers` carries the LLM's prose
+     * reasoning about what caused the change.
+     */
+    deltas: {
+        area: string;
+        direction: "improved" | "regressed" | "unchanged";
+        pointsDelta: number;
+        drivers: string[];
+    }[];
+    /** 4-bucket aggregate trend across all deltas. */
+    overallTrend: "net-improved" | "net-regressed" | "mixed" | "stable";
 }
-export interface DocAttributionSpotlightBody {
+/**
+ * low-confidence-attribution: LLM-driven. `judgmentRefs` cites the
+ * specific judgments (D0052 triple) that drove the low-confidence finding.
+ */
+export interface LowConfidenceAttributionBody {
     summary: string;
+    /** Judgment references (D0052) driving this low-confidence finding. */
+    judgmentRefs: JudgmentRef[];
 }
-export interface RegressionVsBaselineBody {
+/**
+ * doc-attribution-spotlight: LLM-driven. `docCitations` carries per-doc
+ * attribution roles and confidence calibration (AI-SPEC failure-mode #5).
+ */
+export interface DocAttributionSpotlightBody {
     summary: string;
+    /**
+     * Per-doc attribution records. `role` classifies how the doc contributed;
+     * `confidence` calibrates the attribution certainty (D0049).
+     */
+    docCitations: {
+        docSlug: string;
+        confidence: Confidence;
+        role: "supports" | "contradicts" | "missing" | "irrelevant";
+    }[];
 }
 /**
  * Outer-`status` discriminated union: 8 ready variants (one per
@@ -102,6 +203,10 @@ export interface RegressionVsBaselineBody {
  * No `not-yet-generated` variant — old-report fallback is a Phase 7
  * concern at the slim-shape boundary, handled at fetch-time, not in
  * `DiagnosisCard` itself.
+ *
+ * D-07: only the `body: <BodyInterface>` references resolve to richer
+ * shapes. The union arms, status literals, and cardType literals are
+ * identical to Phase 1.
  */
 export type DiagnosisCard = {
     status: "ready";

package/dist/_vendor/ailf-core/types/diagnosis.js

@@ -8,7 +8,9 @@
  * discriminator inside the `ready` variant.
  *
  * Phase 1 lands placeholder body shapes; Phase 5 enriches each per
- * Doc 05 specs.
+ * AI-SPEC §3 and CONTEXT D-05/D-07. The `DiagnosisCard` discriminated
+ * union surface (arms + `cardType` literals) is stable — only the
+ * `body: <BodyInterface>` references resolve to richer shapes.
  *
  * @see docs/decisions/D0049-shared-confidence-contract.md
  * @see docs/decisions/D0052-judgment-ref-granularity.md

package/dist/_vendor/ailf-core/types/index.d.ts

@@ -36,7 +36,8 @@ export { CONVENTIONAL_DERIVATIONS, isConfidence } from "./confidence.js";
 export type { ArtifactId, AssociationAxis, AssociationValues, Brand, EntryKey, Err, FixtureId, IdValidationError, JudgmentId, NewReportId, Ok, ProviderId, PromptId, Result, ResultId, RubricId, RunFingerprint, RunId, SuiteId, TaskId, TaskSlug, TraceId, } from "./branded-ids.js";
 export { err, fixtureId, generateJudgmentId, generateRunId, judgmentId, ok, providerId, resultId, runId, suiteId, taskId, traceId, } from "./branded-ids.js";
 export type { AgentHarnessTaskDefinition, ContentLakeAuthorableMode, ContentLakeAuthorableTask, CriterionRef, CustomTaskDefinition, GeneralizedAssertionDefinition, GeneralizedDocRef, GeneralizedTaskDefinition, GeneralizedTemplatedAssertion, GeneralizedValueAssertion, IdDocRef, KnowledgeProbeTaskDefinition, LiteracyTaskDefinition, MCPServerTaskDefinition, PromptVars, PathDocRef, PerspectiveDocRef, ReservedPromptVarKey, RubricRef, SlugDocRef, TaskCommonFields, TaskDifficulty, TaskOptions, TaskProviderConfig, TaskStatus, } from "./generalized-task.js";
-export type { ActionSuggestion, AreaSummaryBody, CardMeta, CardType, Diagnosis, DiagnosisCard, DocAttributionSpotlightBody, FailureModeSummaryBody, LowConfidenceAttributionBody, NoIssuesBody, RegressionVsBaselineBody, TopRecommendationsBody, VersionedInputs, WeakestAreaBody, } from "./diagnosis.js";
+export type { ActionSuggestion, AreaSummaryBody, CardMeta, CardType, Diagnosis, DiagnosisCard, DocAttributionSpotlightBody, FailureModeSummaryBody, JudgmentRef, LowConfidenceAttributionBody, NoIssuesBody, RegressionVsBaselineBody, TopRecommendationsBody, VersionedInputs, WeakestAreaBody, } from "./diagnosis.js";
+export type { SynthesisCostTelemetry, SynthesisPerCardTelemetry, } from "./synthesis-telemetry.js";
 export type { AttributionMeta, DocAttribution, JudgmentAttribution, } from "./attribution.js";
 export type { CriterionSubJudgment, DocCitation, DocCitationRole, GraderJudgment, } from "./grader-judgment.js";
 export type { LegacyGraderJudgment } from "./legacy-grader-judgment.js";
@@ -754,6 +755,12 @@ export interface PipelineResult {
     promptfooUrls?: PromptfooUrlEntry[];
     /** Results per step */
     steps: Record<string, StepResult>;
+    /** Report ID produced by PublishReportStep (when publish was enabled). Used by
+     * post-run hooks (e.g. runPostPipelineHooks) to target diagnosis and telemetry
+     * writeback at the correct Content Lake document. Absent when publish was
+     * skipped or the publish step did not produce a report. (Phase 6 / DIAG-06)
+     */
+    reportId?: string;
     /** Overall success (all non-skipped steps succeeded) */
     success: boolean;
     /** Summary of test execution outcomes. */

package/dist/_vendor/ailf-core/types/repo-config.d.ts

@@ -106,6 +106,21 @@ export interface RepoTriggersConfig {
     "pr-task-change"?: TriggerConfig;
     schedule?: ScheduleTriggerConfig;
 }
+/**
+ * Post-run diagnosis summary policy (Phase 6 / DIAG-06).
+ * Controls whether `ailf run` auto-fires the in-process diagnosis runner
+ * at the end of a published pipeline. Precedence is resolved at the CLI
+ * layer — see `shouldRunPostSummary()` in `pipeline-action.ts`.
+ */
+export interface RepoSummaryConfig {
+    /**
+     * - `"auto"`   — fire only when `process.stdout.isTTY === true` AND
+     *                `CI !== "true"`.
+     * - `"always"` — fire unconditionally (bypasses TTY check).
+     * - `"never"`  — never fire.
+     */
+    onRun?: "auto" | "always" | "never";
+}
 /**
  * Parsed shape of `.ailf/config.yaml`.
  *
@@ -124,6 +139,7 @@ export interface RepoConfig {
     publish?: RepoPublishConfig;
     reportStore?: RepoReportStoreConfig;
     source?: RepoSourceConfig;
+    summary?: RepoSummaryConfig;
     taskSource?: RepoTaskSourceConfig;
     triggers?: RepoTriggersConfig;
 }

package/dist/_vendor/ailf-core/types/synthesis-telemetry.d.ts

@@ -0,0 +1,101 @@
+/**
+ * Synthesis cost telemetry types — canonical TS-first shapes for
+ * Phase 6 DIAG-06 cost and parse-failure observability.
+ *
+ * These interfaces are authored independently of their Zod adapter schema
+ * (Plan 06-02) per D0045: the Zod schema declares
+ * `satisfies z.ZodType<SynthesisCostTelemetry>` against this independent
+ * type so drift is a build error, not a runtime bug.
+ *
+ * The 14 attribute paths on `SynthesisCostTelemetry` + `SynthesisPerCardTelemetry`
+ * land on the `ailf.report` Sanity doc under `summary.synthesis.diagnosis.*`
+ * (D6-09). No new sibling doc type (D0033 / D6-09).
+ *
+ * @see docs/decisions/D0045-type-architecture-and-contract-enforcement.md
+ * @see .planning/phases/06-post-run-integration-cost-telemetry/06-CONTEXT.md §D6-09
+ * @see .planning/phases/06-post-run-integration-cost-telemetry/06-CONTEXT.md §D6-12
+ */
+import type { CardType } from "./diagnosis.js";
+/**
+ * Per-card telemetry row for the `synthesis_per_card` Airbyte stream
+ * (D6-11) and the `summary.synthesis.diagnosis.perCard[]` Sanity doc path
+ * (D6-09).
+ *
+ * Fields map directly to the 8 per-card attribute paths in D6-09:
+ * `…perCard[].cardType`, `…perCard[].cost`, `…perCard[].parseFailed`,
+ * `…perCard[].latencyMs`, `…perCard[].tokenInput`, `…perCard[].tokenOutput`,
+ * `…perCard[].cardVersion`, `…perCard[].generatedAt`.
+ *
+ * `cost` is undefined when the card did not make an LLM call (deterministic
+ * cards) and contributes 0 to the roll-up.
+ */
+export interface SynthesisPerCardTelemetry {
+    /** Card archetype — reuses `CardType` from diagnosis.ts:55-63; not redeclared. */
+    cardType: CardType;
+    /**
+     * Per-call USD cost captured from `LLMStructuredCompletion.cost`.
+     * `undefined` for deterministic cards (area-summary, failure-mode-summary,
+     * no-issues) which make no LLM call.
+     */
+    cost?: number;
+    /**
+     * Whether the card's Zod schema parse failed (produces a degraded card).
+     * Used for the 7-day rolling parse-failure rate in BigQuery (D6-15).
+     */
+    parseFailed: boolean;
+    /**
+     * End-to-end latency for the LLM call in milliseconds.
+     * `undefined` for deterministic cards.
+     */
+    latencyMs?: number;
+    /**
+     * Prompt tokens consumed by the LLM call.
+     * `undefined` for deterministic cards.
+     */
+    tokenInput?: number;
+    /**
+     * Completion tokens produced by the LLM call.
+     * `undefined` for deterministic cards.
+     */
+    tokenOutput?: number;
+    /** Per-card version string (e.g. `"top-recommendations@0.1.0"`). */
+    cardVersion: string;
+    /** ISO 8601 UTC timestamp when this card was generated. */
+    generatedAt: string;
+}
+/**
+ * Aggregate synthesis cost telemetry for a single Diagnosis run.
+ * Lands on the `ailf.report` Sanity doc under `summary.synthesis.diagnosis.*`
+ * (D6-09 / D6-10: parallel to `summary.overall.cost` — not additive).
+ *
+ * Written by the post-run hook (D6-08); not written by standalone
+ * `ailf interpret`.
+ *
+ * Field set matches the 4 top-level D6-09 attribute paths:
+ * `summary.synthesis.diagnosis.cost`,
+ * `summary.synthesis.diagnosis.parseFailureCount`,
+ * `summary.synthesis.diagnosis.parseFailureRate`,
+ * `summary.synthesis.diagnosis.perCard`.
+ */
+export interface SynthesisCostTelemetry {
+    /**
+     * Total USD cost across all LLM cards in this Diagnosis run.
+     * Roll-up: `sum(perCard[].cost ?? 0)` for ready + degraded cards.
+     * Missing cards contribute 0.
+     */
+    cost: number;
+    /**
+     * Number of cards whose Zod parse failed in this Diagnosis run.
+     * Counted across all 8 card types (including deterministic cards;
+     * a deterministic-card parse failure indicates a code bug).
+     */
+    parseFailureCount: number;
+    /**
+     * Parse-failure rate: `parseFailureCount / 8` (8 = fixed card registry size).
+     * Range: 0–1. Used as the denominator for the D6-15 BigQuery 7-day
+     * rolling rate view (`synthesis_parse_failure_rate_7d.sql`).
+     */
+    parseFailureRate: number;
+    /** Per-card telemetry rows — one entry per card in registry-order. */
+    perCard: SynthesisPerCardTelemetry[];
+}