@sanity/ailf 6.0.0 → 6.1.0

package/dist/_vendor/ailf-core/services/diagnosis/cards/weakest-area.js

@@ -89,7 +89,9 @@ export const generateWeakestArea = async (report, ctx) => {
         path: ["confidence", "level"],
     });
     const prompt = buildWeakestAreaPrompt(report);
-    const { value, usage } = await ctx.llm.completeStructured({
+    // Destructure `cost` and `model` from the LLMClient return —
+    // already provided per llm-client.ts:139-144, previously discarded.
+    const { value, usage, cost, model } = await ctx.llm.completeStructured({
         model: CARD_MODEL,
         prompt: `${prompt.system}\n\n${prompt.user}`,
         schema: PerCallSchema,
@@ -109,6 +111,8 @@ export const generateWeakestArea = async (report, ctx) => {
             cardVersion: "weakest-area@0.1.0",
             tokenUsage: { input: usage.promptTokens, output: usage.completionTokens },
             generatedAt: new Date().toISOString(),
+            cost,
+            model,
         },
     };
 };

package/dist/_vendor/ailf-core/services/diagnosis/prompt-builders.js

@@ -131,10 +131,23 @@ export function buildWeakestAreaPrompt(report) {
 export function buildLowConfidenceAttributionPrompt(report, judgmentAttributions) {
     // Filter to low-confidence entries (by any attribution in the set)
     const lowConf = judgmentAttributions.filter((ja) => ja.attributions.some((a) => a.confidence.level === "low"));
-    // If no low-confidence entries, use all sorted by score ascending (most uncertain first)
+    // If no low-confidence entries, use all sorted by score ascending (most
+    // uncertain first). Guard against entries with empty `attributions` arrays:
+    // `Math.min(...[])` is Infinity and produces an unstable sort that ranks
+    // empty-attribution entries identically at the top of the prompt. The card
+    // schema requires `judgmentRefs.min(1)`, so emitting empty-attribution rows
+    // here forces a degraded card downstream. Caller should short-circuit to
+    // missing before reaching here, but defend against the seam regressing.
+    const validAttrs = judgmentAttributions.filter((ja) => ja.attributions.length > 0);
+    if (validAttrs.length === 0) {
+        return {
+            system: LOW_CONFIDENCE_ATTRIBUTION_SYSTEM_PROMPT,
+            user: "(no attribution data with non-empty attributions)",
+        };
+    }
     const source = lowConf.length > 0
         ? lowConf
-        : [...judgmentAttributions].sort((a, b) => {
+        : [...validAttrs].sort((a, b) => {
             const aMin = Math.min(...a.attributions.map((x) => x.score));
             const bMin = Math.min(...b.attributions.map((x) => x.score));
             return aMin - bMin;

package/dist/_vendor/ailf-core/services/diagnosis/prompts/weakest-area.system.d.ts

@@ -8,9 +8,11 @@
  * Mitigations embedded:
  * - failure-mode #3: confidence inflation on small samples — prompt instructs
  *   to hedge when sampleSize < 10; Zod W3 refine enforces at parse time
- * - failure-mode #4: taxonomy drift — full canonical taxonomy enumerated
- *   verbatim in this prompt so the LLM picks from a known list
+ * - failure-mode #4: taxonomy drift — failure-mode lists derived at build
+ *   time from the canonical const arrays in
+ *   `packages/core/src/grader/failure-modes/*.ts`, so the prompt and the
+ *   Zod `.refine(buildFailureModeRefinement())` validator always agree.
  *
  * @see .planning/phases/05-diagnosis-engine-cli-llm-cards/05-AI-SPEC.md §4b
  */
-export declare const SYSTEM_PROMPT = "You are an AILF evaluation analyst identifying the documentation area most in need of improvement.\n\n## Your Output\n\nReturn a JSON object matching this exact shape:\n{\n  \"summary\": \"<1-2 sentence description of the weakest area and why>\",\n  \"area\": \"<feature area name, e.g. 'schema-deploy'>\",\n  \"dimension\": \"<MUST be one of the canonical dimensions listed below>\",\n  \"failureMode\": \"<MUST be from the canonical taxonomy for the chosen dimension>\",\n  \"sampleSize\": <number \u2014 MUST equal the judgmentCount provided for this area>,\n  \"confidence\": {\n    \"level\": \"high\" | \"medium\" | \"low\",\n    \"signalsPresent\": <number of tasks backing this finding>,\n    \"derivation\": \"card-type-specific\"\n  }\n}\n\n## CANONICAL DIMENSIONS AND FAILURE MODES\n\nYou MUST pick dimension and failureMode from this exact taxonomy. Cross-dimension combinations are invalid (e.g., \"security\" dimension with \"missing-docs\" failure mode is rejected).\n\n### Literacy family (dimensions: task-completion, code-correctness, doc-coverage)\nFailure modes:\n- missing-docs \u2014 relevant doc didn't exist\n- outdated-docs \u2014 doc reflects an older API/version\n- incorrect-docs \u2014 doc states something factually wrong\n- poor-structure \u2014 doc exists but is hard to find or follow\nPlus cross-cutting: api-error, model-limitation, false-floor, unclassified\n\n### MCP family (dimensions: mcp-behavior, input-validation, output-correctness, error-handling, security)\nFailure modes:\n- invalid-tool-call \u2014 model called tool with wrong args\n- missing-required-param \u2014 required parameter omitted\n- extra-param \u2014 unexpected extra parameter sent\n- wrong-tool-selected \u2014 chose wrong tool for task\n- tool-call-order \u2014 tools called in wrong sequence\n- no-tool-call \u2014 should have used a tool but didn't\n- schema-mismatch \u2014 response did not match expected schema\n- unsafe-operation \u2014 operation could cause data loss\n- auth-bypass \u2014 security check skipped\nPlus cross-cutting: api-error, model-limitation, false-floor, unclassified\n\n### Knowledge-probe family (dimensions: knowledge-probe, factual-correctness, completeness, currency)\nFailure modes:\n- factual-error \u2014 stated an incorrect fact\n- out-of-date \u2014 used deprecated API or old syntax\n- missing-step \u2014 omitted a required step\n- hallucinated-api \u2014 invented an API that does not exist\n- wrong-version \u2014 used v1 API when v2 was required\n- incomplete-coverage \u2014 missed important edge case\nPlus cross-cutting: api-error, model-limitation, false-floor, unclassified\n\n### Agent-harness family (dimensions: agent-harness, process-quality, agent-output, tool-usage)\nFailure modes:\n- excessive-loops \u2014 agent looped unnecessarily\n- premature-stop \u2014 stopped before completing the task\n- incorrect-output \u2014 output was wrong or incomplete\n- inefficient-path \u2014 completed task but via unnecessary steps\n- assertion-failure \u2014 failed a structural assertion check\nPlus cross-cutting: api-error, model-limitation, false-floor, unclassified\n\n## Confidence Calibration Rules\n\n**CRITICAL:** When sampleSize < 10, you MUST set confidence.level = \"low\".\n\n- sampleSize >= 30 \u2192 \"high\" is appropriate\n- sampleSize >= 10 \u2192 \"medium\" is appropriate\n- sampleSize < 10 \u2192 MUST use \"low\" (small-sample hedge required)\n\nIn your summary, reflect the confidence level: if \"low\", include language like \"small sample (N=X) \u2014 re-run with broader dataset before acting\".";
+export declare const SYSTEM_PROMPT: string;

package/dist/_vendor/ailf-core/services/diagnosis/prompts/weakest-area.system.js

@@ -8,11 +8,19 @@
  * Mitigations embedded:
  * - failure-mode #3: confidence inflation on small samples — prompt instructs
  *   to hedge when sampleSize < 10; Zod W3 refine enforces at parse time
- * - failure-mode #4: taxonomy drift — full canonical taxonomy enumerated
- *   verbatim in this prompt so the LLM picks from a known list
+ * - failure-mode #4: taxonomy drift — failure-mode lists derived at build
+ *   time from the canonical const arrays in
+ *   `packages/core/src/grader/failure-modes/*.ts`, so the prompt and the
+ *   Zod `.refine(buildFailureModeRefinement())` validator always agree.
  *
  * @see .planning/phases/05-diagnosis-engine-cli-llm-cards/05-AI-SPEC.md §4b
  */
+import { AGENT_FAILURE_MODES, COMMON_FAILURE_MODES, KP_FAILURE_MODES, LITERACY_FAILURE_MODES, MCP_FAILURE_MODES, } from "../../../grader/failure-modes/index.js";
+const literacyList = LITERACY_FAILURE_MODES.map((m) => `- ${m}`).join("\n");
+const mcpList = MCP_FAILURE_MODES.map((m) => `- ${m}`).join("\n");
+const kpList = KP_FAILURE_MODES.map((m) => `- ${m}`).join("\n");
+const agentList = AGENT_FAILURE_MODES.map((m) => `- ${m}`).join("\n");
+const commonList = COMMON_FAILURE_MODES.join(", ");
 export const SYSTEM_PROMPT = `You are an AILF evaluation analyst identifying the documentation area most in need of improvement.
 
 ## Your Output
@@ -33,47 +41,27 @@ Return a JSON object matching this exact shape:
 
 ## CANONICAL DIMENSIONS AND FAILURE MODES
 
-You MUST pick dimension and failureMode from this exact taxonomy. Cross-dimension combinations are invalid (e.g., "security" dimension with "missing-docs" failure mode is rejected).
+You MUST pick dimension and failureMode from this exact taxonomy. Cross-dimension combinations are invalid (e.g., "security" dimension with a literacy-only failure mode is rejected). The lists below are derived at build time from \`packages/core/src/grader/failure-modes/*.ts\` — the Zod validator on the card schema enforces the same taxonomy.
 
 ### Literacy family (dimensions: task-completion, code-correctness, doc-coverage)
 Failure modes:
-- missing-docs — relevant doc didn't exist
-- outdated-docs — doc reflects an older API/version
-- incorrect-docs — doc states something factually wrong
-- poor-structure — doc exists but is hard to find or follow
-Plus cross-cutting: api-error, model-limitation, false-floor, unclassified
+${literacyList}
+Plus cross-cutting: ${commonList}
 
 ### MCP family (dimensions: mcp-behavior, input-validation, output-correctness, error-handling, security)
 Failure modes:
-- invalid-tool-call — model called tool with wrong args
-- missing-required-param — required parameter omitted
-- extra-param — unexpected extra parameter sent
-- wrong-tool-selected — chose wrong tool for task
-- tool-call-order — tools called in wrong sequence
-- no-tool-call — should have used a tool but didn't
-- schema-mismatch — response did not match expected schema
-- unsafe-operation — operation could cause data loss
-- auth-bypass — security check skipped
-Plus cross-cutting: api-error, model-limitation, false-floor, unclassified
+${mcpList}
+Plus cross-cutting: ${commonList}
 
 ### Knowledge-probe family (dimensions: knowledge-probe, factual-correctness, completeness, currency)
 Failure modes:
-- factual-error — stated an incorrect fact
-- out-of-date — used deprecated API or old syntax
-- missing-step — omitted a required step
-- hallucinated-api — invented an API that does not exist
-- wrong-version — used v1 API when v2 was required
-- incomplete-coverage — missed important edge case
-Plus cross-cutting: api-error, model-limitation, false-floor, unclassified
+${kpList}
+Plus cross-cutting: ${commonList}
 
 ### Agent-harness family (dimensions: agent-harness, process-quality, agent-output, tool-usage)
 Failure modes:
-- excessive-loops — agent looped unnecessarily
-- premature-stop — stopped before completing the task
-- incorrect-output — output was wrong or incomplete
-- inefficient-path — completed task but via unnecessary steps
-- assertion-failure — failed a structural assertion check
-Plus cross-cutting: api-error, model-limitation, false-floor, unclassified
+${agentList}
+Plus cross-cutting: ${commonList}
 
 ## Confidence Calibration Rules
 

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

@@ -17,5 +17,5 @@ export { createDiagnosisRunner, diagnosisVersion, type CardGenerator, type CardR
 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 { DIAGNOSIS_CARD_GENERATORS, generateAreaSummary, generateFailureModeSummary, generateNoIssues, generateTopRecommendations, generateWeakestArea, generateLowConfidenceAttribution, generateDocAttributionSpotlight, generateRegressionVsBaseline, } from "./diagnosis/cards/index.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

@@ -29,5 +29,5 @@ export { buildFailureModeRefinement, isFailureModeInDimensionTaxonomy, } from ".
 // ---------------------------------------------------------------------------
 // Phase 5 Plan 05 — card generators barrel + prompt builders
 // ---------------------------------------------------------------------------
-export { DIAGNOSIS_CARD_GENERATORS, generateAreaSummary, generateFailureModeSummary, generateNoIssues, generateTopRecommendations, generateWeakestArea, generateLowConfidenceAttribution, generateDocAttributionSpotlight, generateRegressionVsBaseline, } from "./diagnosis/cards/index.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/types/diagnosis.d.ts

@@ -19,6 +19,7 @@
 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
@@ -54,6 +55,8 @@ export interface CardMeta {
     latencyMs?: number;
     /** ISO 8601 UTC timestamp. */
     generatedAt: string;
+    cost?: number;
+    model?: ModelId;
 }
 /**
  * A single actionable suggestion surfaced by a recommendations card.

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

@@ -37,6 +37,7 @@ export type { ArtifactId, AssociationAxis, AssociationValues, Brand, EntryKey, E
 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, 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[];
+}

package/dist/_vendor/ailf-core/types/synthesis-telemetry.js

@@ -0,0 +1,18 @@
+/**
+ * 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
+ */
+export {};

package/dist/adapters/config-sources/file-config-adapter.js

@@ -115,12 +115,10 @@ function mapEvalConfigToResolvedConfig(config, rootDir) {
         compareBaseline: config.compareBaseline,
         gapAnalysisEnabled: config.execution?.gapAnalysis ?? true,
         // W0077 Phase 4 — `publish` is now a policy object. Map the auto value
-        // directly to a boolean for the file-config path; the runtime
-        // smart-default logic in pipeline-action.ts isn't relevant here because
-        // the user has explicitly handed us a config file.
-        publishEnabled: config.publish?.auto === "never"
-            ? false
-            : config.publish?.auto !== undefined,
+        // to a boolean for the file-config path. Absence of publish.auto mirrors
+        // the CLI's "full-runs" default (enable publish; composition root gates on
+        // token availability). Only "never" explicitly disables auto-publish.
+        publishEnabled: config.publish?.auto !== "never",
         publishTag: config.publish?.tag,
         noCache: config.noCache ?? false,
         noRemoteCache: config.noRemoteCache ?? false,
@@ -150,5 +148,9 @@ function mapEvalConfigToResolvedConfig(config, rootDir) {
             ? resolve(rootDir, config.taskSource.repoTasksPath)
             : undefined,
         presets: config.presets,
+        // Phase 6 / DIAG-06 — thread summary.onRun into ResolvedConfig so the
+        // file-config exit branch in executePipeline can pass it to
+        // runPostPipelineHooks.
+        summaryOnRun: config.summary?.onRun,
     };
 }

package/dist/adapters/llm/index.d.ts

@@ -5,5 +5,5 @@ export type { FakeCallRecord, FakeCompletionResponse, FakeStructuredResponse, }
 export { OpenAILLMClient } from "./openai-llm-client.js";
 export type { OpenAILLMClientOptions } from "./openai-llm-client.js";
 export type { ModelPricing } from "./pricing.js";
-export { DEFAULT_RETRY_POLICY, LLMHttpError, isRetryableStatus, parseRetryAfterSeconds, runWithRetry, } from "./retry.js";
+export { DEFAULT_RETRY_POLICY, LLMHttpError, LLMParseError, isRetryableStatus, parseRetryAfterSeconds, runWithRetry, } from "./retry.js";
 export type { RetryPolicy } from "./retry.js";

package/dist/adapters/llm/index.js

@@ -1,4 +1,4 @@
 export { AnthropicLLMClient } from "./anthropic-llm-client.js";
 export { FakeLLMClient } from "./fake-llm-client.js";
 export { OpenAILLMClient } from "./openai-llm-client.js";
-export { DEFAULT_RETRY_POLICY, LLMHttpError, isRetryableStatus, parseRetryAfterSeconds, runWithRetry, } from "./retry.js";
+export { DEFAULT_RETRY_POLICY, LLMHttpError, LLMParseError, isRetryableStatus, parseRetryAfterSeconds, runWithRetry, } from "./retry.js";

package/dist/adapters/llm/openai-llm-client.js

@@ -12,7 +12,7 @@
  */
 import { z } from "zod";
 import { OpenAIChatResponseSchema, splitModelId, } from "../../_vendor/ailf-core/index.js";
-import { DEFAULT_RETRY_POLICY, parseRetryAfterSeconds, runWithRetry, } from "./retry.js";
+import { DEFAULT_RETRY_POLICY, LLMParseError, parseRetryAfterSeconds, runWithRetry, } from "./retry.js";
 const DEFAULT_BASE_URL = "https://api.openai.com/v1/chat/completions";
 /**
  * Conservative defaults for the models in `packages/eval/config/models.ts`.
@@ -98,7 +98,12 @@ export class OpenAILLMClient {
             parsed = JSON.parse(raw);
         }
         catch (err) {
-            throw new Error(`OpenAI structured completion returned invalid JSON for model ${args.model}: ${err instanceof Error ? err.message : String(err)}`, { cause: err });
+            // Sanitize: SyntaxError.message embeds a snippet at the failure offset,
+            // which can leak prompt text or user content echoed back by the model.
+            // Keep the raw body on the instance for callers that opt in via .raw,
+            // mirroring the LLMHttpError pattern (verified by the "does not leak
+            // the response body" test in openai-llm-client.test.ts).
+            throw new LLMParseError(`OpenAI structured completion returned invalid JSON for model ${args.model}`, raw, { cause: err });
         }
         // strict:true guarantees a valid-against-the-schema JSON document, but
         // the Zod parse is still load-bearing — it brands the result as T and is

package/dist/adapters/llm/retry.d.ts

@@ -33,6 +33,24 @@ export declare class LLMHttpError extends Error {
     readonly body: string;
     constructor(status: number, body: string, attempts: number);
 }
+/**
+ * Sanitized error raised when an LLM adapter receives an HTTP-200 response
+ * whose body is not valid JSON. The raw response body (which may echo back
+ * user prompt content or even API-key fragments from prompts) is kept on the
+ * instance for callers that opt in via `.raw`, NOT in the message string.
+ *
+ * Mirrors the LLMHttpError pattern verified by the
+ * "does not leak the response body" test in openai-llm-client.test.ts.
+ */
+export declare class LLMParseError extends Error {
+    /** Full raw response body (kept on the instance, NOT in `message`). */
+    readonly raw: string;
+    /** Byte length of `raw` — safe to include in the message. */
+    readonly rawLength: number;
+    constructor(message: string, raw: string, options?: {
+        cause?: unknown;
+    });
+}
 export declare function isRetryableStatus(status: number): boolean;
 export interface RunWithRetryArgs<T> {
     policy: RetryPolicy;

package/dist/adapters/llm/retry.js

@@ -29,6 +29,27 @@ export class LLMHttpError extends Error {
         this.body = body;
     }
 }
+/**
+ * Sanitized error raised when an LLM adapter receives an HTTP-200 response
+ * whose body is not valid JSON. The raw response body (which may echo back
+ * user prompt content or even API-key fragments from prompts) is kept on the
+ * instance for callers that opt in via `.raw`, NOT in the message string.
+ *
+ * Mirrors the LLMHttpError pattern verified by the
+ * "does not leak the response body" test in openai-llm-client.test.ts.
+ */
+export class LLMParseError extends Error {
+    /** Full raw response body (kept on the instance, NOT in `message`). */
+    raw;
+    /** Byte length of `raw` — safe to include in the message. */
+    rawLength;
+    constructor(message, raw, options) {
+        super(`${message} (raw=${raw.length}B)`, options);
+        this.name = "LLMParseError";
+        this.raw = raw;
+        this.rawLength = raw.length;
+    }
+}
 export function isRetryableStatus(status) {
     return status === 429 || (status >= 500 && status < 600);
 }

package/dist/adapters/synthesis/synthesis-telemetry-schema.d.ts

@@ -0,0 +1,49 @@
+/**
+ * Zod adapter schema for SynthesisCostTelemetry at the trust boundary.
+ *
+ * This schema sits at `packages/eval/src/adapters/**` and is therefore
+ * scanned by `pnpm check-trust-boundary-satisfies` (D0045). The
+ * `satisfies z.ZodType<SynthesisCostTelemetry>` clause makes schema/type
+ * drift a build error, not a runtime bug.
+ *
+ * Used by:
+ * - Plan 06-04 `ReportStore.patchSynthesis` — validates telemetry before
+ *   writing to Sanity (process memory → Sanity write boundary, T-06-04).
+ * - Any future Sanity-side reader of `summary.synthesis.diagnosis.*`
+ *   (Sanity Content Lake → eval process boundary, T-06-04).
+ *
+ * Security constraints:
+ * - No `.passthrough()` — schema is closed to prevent PII leakage from
+ *   card body text into the telemetry shape (T-06-05).
+ * - Satisfies clause is load-bearing (T-06-06); no exemption marker.
+ *
+ * @see packages/core/src/types/synthesis-telemetry.ts — independently authored domain types
+ * @see docs/decisions/D0045-type-architecture-and-contract-enforcement.md
+ * @see .planning/phases/06-post-run-integration-cost-telemetry/06-CONTEXT.md §D6-09
+ */
+import { z } from "zod";
+export declare const SynthesisCostTelemetrySchema: z.ZodObject<{
+    cost: z.ZodNumber;
+    parseFailureCount: z.ZodNumber;
+    parseFailureRate: z.ZodNumber;
+    perCard: z.ZodArray<z.ZodObject<{
+        cardType: z.ZodEnum<{
+            "area-summary": "area-summary";
+            "failure-mode-summary": "failure-mode-summary";
+            "no-issues": "no-issues";
+            "top-recommendations": "top-recommendations";
+            "weakest-area": "weakest-area";
+            "low-confidence-attribution": "low-confidence-attribution";
+            "doc-attribution-spotlight": "doc-attribution-spotlight";
+            "regression-vs-baseline": "regression-vs-baseline";
+        }>;
+        cost: z.ZodOptional<z.ZodNumber>;
+        parseFailed: z.ZodBoolean;
+        latencyMs: z.ZodOptional<z.ZodNumber>;
+        tokenInput: z.ZodOptional<z.ZodNumber>;
+        tokenOutput: z.ZodOptional<z.ZodNumber>;
+        cardVersion: z.ZodString;
+        generatedAt: z.ZodString;
+    }, z.core.$strip>>;
+}, z.core.$strip>;
+export type { SynthesisCostTelemetry, SynthesisPerCardTelemetry, } from "../../_vendor/ailf-core/index.d.ts";

package/dist/adapters/synthesis/synthesis-telemetry-schema.js

@@ -0,0 +1,55 @@
+/**
+ * Zod adapter schema for SynthesisCostTelemetry at the trust boundary.
+ *
+ * This schema sits at `packages/eval/src/adapters/**` and is therefore
+ * scanned by `pnpm check-trust-boundary-satisfies` (D0045). The
+ * `satisfies z.ZodType<SynthesisCostTelemetry>` clause makes schema/type
+ * drift a build error, not a runtime bug.
+ *
+ * Used by:
+ * - Plan 06-04 `ReportStore.patchSynthesis` — validates telemetry before
+ *   writing to Sanity (process memory → Sanity write boundary, T-06-04).
+ * - Any future Sanity-side reader of `summary.synthesis.diagnosis.*`
+ *   (Sanity Content Lake → eval process boundary, T-06-04).
+ *
+ * Security constraints:
+ * - No `.passthrough()` — schema is closed to prevent PII leakage from
+ *   card body text into the telemetry shape (T-06-05).
+ * - Satisfies clause is load-bearing (T-06-06); no exemption marker.
+ *
+ * @see packages/core/src/types/synthesis-telemetry.ts — independently authored domain types
+ * @see docs/decisions/D0045-type-architecture-and-contract-enforcement.md
+ * @see .planning/phases/06-post-run-integration-cost-telemetry/06-CONTEXT.md §D6-09
+ */
+import { z } from "zod";
+/**
+ * Enum of all valid card types — mirrors `CardType` from diagnosis.ts.
+ * Using `z.enum()` (not `z.string()`) so the schema satisfies
+ * `z.ZodType<SynthesisPerCardTelemetry>` (which requires `cardType: CardType`).
+ */
+const CardTypeSchema = z.enum([
+    "area-summary",
+    "failure-mode-summary",
+    "no-issues",
+    "top-recommendations",
+    "weakest-area",
+    "low-confidence-attribution",
+    "doc-attribution-spotlight",
+    "regression-vs-baseline",
+]);
+const SynthesisPerCardSchema = z.object({
+    cardType: CardTypeSchema,
+    cost: z.number().nonnegative().optional(),
+    parseFailed: z.boolean(),
+    latencyMs: z.number().int().nonnegative().optional(),
+    tokenInput: z.number().int().nonnegative().optional(),
+    tokenOutput: z.number().int().nonnegative().optional(),
+    cardVersion: z.string(),
+    generatedAt: z.string().datetime({ offset: false }), // ISO 8601 UTC required
+});
+export const SynthesisCostTelemetrySchema = z.object({
+    cost: z.number().nonnegative(),
+    parseFailureCount: z.number().int().nonnegative(),
+    parseFailureRate: z.number().min(0).max(1),
+    perCard: z.array(SynthesisPerCardSchema),
+});

package/dist/adapters/task-sources/content-lake-task-source.js

@@ -286,16 +286,21 @@ function mapAssertions(raw) {
                     .map((c) => ({ id: c.id, text: c.text })),
                 template: a.template,
                 type: "llm-rubric",
-                ...(a.weight !== undefined ? { weight: a.weight } : {}),
+                // Use `!= null` (loose) so we drop both `undefined` AND `null`.
+                // GROQ projects missing scalar fields as `null`, but the domain
+                // schema's `z.number().optional()` accepts `T | undefined`, not
+                // `T | null` — a strict `!== undefined` check would forward
+                // `weight: null` and trigger Zod's "Invalid input" on assertions.
+                ...(a.weight != null ? { weight: a.weight } : {}),
             };
         }
-        // Value-based assertion
+        // Value-based assertion — same null-vs-undefined hazard as above.
         const result = { type: a.type };
-        if (a.value !== undefined)
+        if (a.value != null)
             result.value = a.value;
-        if (a.threshold !== undefined)
+        if (a.threshold != null)
             result.threshold = a.threshold;
-        if (a.weight !== undefined)
+        if (a.weight != null)
             result.weight = a.weight;
         return result;
     });

package/dist/adapters/task-sources/repo-schemas.d.ts

@@ -1561,6 +1561,13 @@ export declare const RepoConfigSchema: z.ZodObject<{
         dir: z.ZodOptional<z.ZodString>;
         exclude: z.ZodOptional<z.ZodArray<z.ZodString>>;
     }, z.core.$strip>>;
+    summary: z.ZodOptional<z.ZodObject<{
+        onRun: z.ZodOptional<z.ZodEnum<{
+            never: "never";
+            always: "always";
+            auto: "auto";
+        }>>;
+    }, z.core.$strip>>;
     taskSource: z.ZodOptional<z.ZodObject<{
         type: z.ZodOptional<z.ZodEnum<{
             "content-lake": "content-lake";