@oscharko-dev/keiko 0.1.6 → 0.1.7

package/dist/audit/types.d.ts

@@ -1,188 +1,2 @@
-import type { CostClass } from "../gateway/types.js";
-import type { HarnessCode, HarnessStateName, RunManifest, RunOutcome, RunResult, TaskType } from "../harness/types.js";
-import type { AuditSummary } from "../workspace/types.js";
-import type { VerificationAuditSummary } from "../verification/summary.js";
-import type { EvidenceStore } from "./store.js";
-export declare const EVIDENCE_SCHEMA_VERSION: "1";
-export interface EvidenceRunIdentity {
-    readonly runId: string;
-    readonly fingerprint: string;
-    readonly harnessVersion: string;
-    readonly taskType: EvidenceTaskType;
-    readonly outcome: RunOutcome;
-    readonly startedAt: number;
-    readonly finishedAt: number;
-    readonly durationMs: number;
-}
-export interface EvidenceModel {
-    readonly modelId: string;
-    readonly costClass: CostClass | "unknown";
-}
-export interface EvidenceUsageTotals {
-    readonly promptTokens: number;
-    readonly completionTokens: number;
-    readonly requestCount: number;
-    readonly totalLatencyMs: number;
-}
-export interface EvidenceStateTransition {
-    readonly seq: number;
-    readonly ts: number;
-    readonly from: HarnessStateName;
-    readonly to: HarnessStateName;
-    readonly reason: string;
-}
-export interface EvidenceToolCall {
-    readonly seq: number;
-    readonly ts: number;
-    readonly toolName: string;
-    readonly toolCallId: string;
-    readonly outcome: "completed" | "failed";
-    readonly durationMs?: number | undefined;
-    readonly errorCode?: string | undefined;
-}
-export interface EvidenceCommandExecution {
-    readonly seq: number;
-    readonly ts: number;
-    readonly executable: string;
-    readonly argCount: number;
-    readonly exitCode: number | null;
-    readonly timedOut: boolean;
-    readonly durationMs: number;
-}
-export interface EvidenceSandboxConfiguration {
-    readonly seq: number;
-    readonly ts: number;
-    readonly envAllowlist: readonly string[];
-    readonly network: "inherit" | "none";
-    readonly maxOutputBytes: number;
-    readonly timeoutMs: number;
-    readonly terminationGraceMs: number;
-    readonly cwdRequested: boolean;
-}
-export interface EvidenceVerificationResult {
-    readonly seq: number;
-    readonly ts: number;
-    readonly passed: boolean;
-    readonly detail: string;
-}
-export interface EvidencePatch {
-    readonly proposed: boolean;
-    readonly applied: boolean;
-    readonly targetFileCount: number;
-    readonly patchBytes: number;
-    readonly changedFiles: number;
-    readonly created: number;
-    readonly deleted: number;
-    readonly redactedDiff?: string | undefined;
-}
-export interface EvidenceReasoningEntry {
-    readonly seq: number;
-    readonly ts: number;
-    readonly phase: HarnessStateName;
-    readonly rationale: string;
-    readonly modelResponse?: string | undefined;
-}
-export interface EvidenceFailure {
-    readonly category: HarnessCode;
-    readonly message: string;
-}
-export type EvidenceTaskType = TaskType | "browser-capture" | "terminal-execution";
-export interface EvidenceBrowserViewportPx {
-    readonly width: number;
-    readonly height: number;
-}
-export type EvidenceBrowserEventType = "browser:session-opened" | "browser:navigated" | "browser:screenshot-captured" | "browser:page-content-captured" | "browser:session-closed" | "browser:trust-warning" | "browser:error";
-export interface EvidenceBrowserEvent {
-    readonly schemaVersion: "1";
-    readonly type: EvidenceBrowserEventType;
-    readonly sessionId: string;
-    readonly seq: number;
-    readonly ts: number;
-    readonly originOnly?: string | undefined;
-    readonly httpStatus?: number | null | undefined;
-    readonly captureSeq?: number | undefined;
-    readonly persisted?: boolean | undefined;
-    readonly viewportPx?: EvidenceBrowserViewportPx | undefined;
-    readonly path?: string | undefined;
-    readonly sha256?: string | undefined;
-    readonly bytes?: number | undefined;
-    readonly byteLength?: number | undefined;
-    readonly reason?: string | undefined;
-    readonly warning?: string | undefined;
-    readonly code?: string | undefined;
-    readonly message?: string | undefined;
-}
-export interface EvidenceBrowserScreenshot {
-    readonly seq: number;
-    readonly path: string;
-    readonly sha256: string;
-    readonly bytes: number;
-    readonly capturedAt: number;
-    readonly viewportPx: EvidenceBrowserViewportPx;
-}
-export interface EvidenceBrowserContentCapture {
-    readonly seq: number;
-    readonly byteLength: number;
-    readonly capturedAt: number;
-    readonly redactedHtml: string;
-}
-export interface EvidenceBrowserCapture {
-    readonly sessionId: string;
-    readonly cdpPort: number;
-    readonly targetId: string;
-    readonly status: "open" | "closed";
-    readonly startedAt: number;
-    readonly closedAt?: number | undefined;
-    readonly closeReason?: string | undefined;
-    readonly lastOriginOnly?: string | undefined;
-    readonly events: readonly EvidenceBrowserEvent[];
-    readonly screenshots?: readonly EvidenceBrowserScreenshot[] | undefined;
-    readonly contentCaptures?: readonly EvidenceBrowserContentCapture[] | undefined;
-}
-export interface EvidenceManifest {
-    readonly evidenceSchemaVersion: "1";
-    readonly run: EvidenceRunIdentity;
-    readonly model: EvidenceModel;
-    readonly usageTotals: EvidenceUsageTotals;
-    readonly context?: AuditSummary | undefined;
-    readonly stateTransitions: readonly EvidenceStateTransition[];
-    readonly toolCalls: readonly EvidenceToolCall[];
-    readonly commandExecutions: readonly EvidenceCommandExecution[];
-    readonly sandboxConfigurations?: readonly EvidenceSandboxConfiguration[] | undefined;
-    readonly verificationResults?: readonly EvidenceVerificationResult[] | undefined;
-    readonly patch?: EvidencePatch | undefined;
-    readonly verification?: VerificationAuditSummary | undefined;
-    readonly failure?: EvidenceFailure | undefined;
-    readonly reasoning?: readonly EvidenceReasoningEntry[] | undefined;
-    readonly browser?: EvidenceBrowserCapture | undefined;
-}
-export interface AuditRedactionConfig {
-    readonly additionalSecrets?: readonly string[] | undefined;
-    readonly redactEnvValues?: readonly string[] | undefined;
-    readonly sensitiveLiterals?: readonly string[] | undefined;
-}
-export interface RetentionPolicy {
-    readonly maxRuns?: number | undefined;
-    readonly maxAgeMs?: number | undefined;
-    readonly maxTotalBytes?: number | undefined;
-    readonly disabled?: boolean | undefined;
-}
-export declare const DEFAULT_RETENTION: RetentionPolicy;
-export interface BuildOptions {
-    readonly includeReasoning?: boolean | undefined;
-    readonly includeDiff?: boolean | undefined;
-}
-export interface EvidenceBuildInput {
-    readonly result: RunResult;
-    readonly manifest: RunManifest;
-    readonly context?: AuditSummary | undefined;
-    readonly verification?: VerificationAuditSummary | undefined;
-    readonly redaction?: AuditRedactionConfig | undefined;
-    readonly options?: BuildOptions | undefined;
-}
-export interface EvidenceDeps {
-    readonly env?: Readonly<Record<string, string | undefined>> | undefined;
-    readonly store?: EvidenceStore | undefined;
-    readonly randomSuffix?: (() => string) | undefined;
-    readonly now?: (() => number) | undefined;
-}
+export type { EvidenceRunIdentity, EvidenceModel, EvidenceUsageTotals, EvidenceStateTransition, EvidenceToolCall, EvidenceCommandExecution, EvidenceVerificationResult, EvidencePatch, EvidenceReasoningEntry, EvidenceFailure, EvidenceTaskType, EvidenceBrowserViewportPx, EvidenceBrowserEventType, EvidenceBrowserEvent, EvidenceBrowserScreenshot, EvidenceBrowserContentCapture, EvidenceBrowserCapture, EvidenceManifest, AuditRedactionConfig, RetentionPolicy, BuildOptions, EvidenceBuildInput, EvidenceDeps, EvidenceStore, } from "@oscharko-dev/keiko-evidence";
+export { EVIDENCE_SCHEMA_VERSION, DEFAULT_RETENTION } from "@oscharko-dev/keiko-evidence";

package/dist/audit/types.js

@@ -1,8 +1,4 @@
-// All Evidence* interfaces, the retention/redaction config, the injectable deps, and the frozen
-// EVIDENCE_SCHEMA_VERSION / DEFAULT_RETENTION tables (ADR-0010 D2/D6/D10). No runtime logic beyond
-// the two frozen tables. Everything here is plain-JSON, deeply readonly, and JSON-serializable:
-// timestamps are epoch-ms numbers sourced from events/RunResult, never Date objects.
-// The schema discriminant — distinct from the harness event `schemaVersion`. A breaking change
-// produces "2" as a NEW union member rather than mutating "1" (ADR-0010 D2).
-export const EVIDENCE_SCHEMA_VERSION = "1";
-export const DEFAULT_RETENTION = { maxRuns: 50 };
+// Re-export shim: all Evidence* interfaces, the retention/redaction config tables, and the frozen
+// EVIDENCE_SCHEMA_VERSION / DEFAULT_RETENTION constants live in @oscharko-dev/keiko-contracts
+// (#158) and reach @oscharko-dev/keiko-evidence's public surface in #163 (ADR-0019).
+export { EVIDENCE_SCHEMA_VERSION, DEFAULT_RETENTION } from "@oscharko-dev/keiko-evidence";

package/dist/audit/workflow-evidence.d.ts

@@ -1,27 +1 @@
-import { type EvidenceReport } from "./report.js";
-import { type EvidenceManifest } from "./types.js";
-import type { EvidenceStore } from "./store.js";
-import type { EnvSource } from "../gateway/index.js";
-export type WorkflowRunKind = "unit-tests" | "bug-investigation";
-export type WorkflowTerminalStatus = "completed" | "cancelled" | "failed";
-export interface EvidencePersistContext {
-    readonly store: EvidenceStore;
-    readonly env: EnvSource;
-    readonly additionalSecrets?: readonly string[] | undefined;
-}
-export interface WorkflowRunIdentity {
-    readonly runId: string;
-    readonly fingerprint: string;
-    readonly modelId: string;
-    readonly kind: WorkflowRunKind;
-    readonly status: WorkflowTerminalStatus;
-    readonly startedAt: number;
-    readonly finishedAt: number;
-    readonly workspaceRoot?: string | undefined;
-}
-export interface WorkflowEventLike {
-    readonly type: string;
-}
-export declare function foldWorkflowUsage(events: readonly WorkflowEventLike[]): EvidenceManifest["usageTotals"];
-export declare function buildWorkflowManifest(identity: WorkflowRunIdentity, events: readonly WorkflowEventLike[], report: unknown): EvidenceManifest;
-export declare function persistWorkflowEvidence(identity: WorkflowRunIdentity, report: unknown, events: readonly WorkflowEventLike[], ctx: EvidencePersistContext): EvidenceReport;
+export { buildWorkflowManifest, foldWorkflowUsage, persistWorkflowEvidence, type EvidencePersistContext, type WorkflowEventLike, type WorkflowRunIdentity, type WorkflowRunKind, type WorkflowTerminalStatus, } from "@oscharko-dev/keiko-evidence";

package/dist/audit/workflow-evidence.js

@@ -1,145 +1,3 @@
-// Shared workflow→EvidenceManifest mapping (ADR-0010 + ADR-0011 AC5 + ADR-0012 D9/C2). This is the
-// PURE, surface-agnostic core that folds a terminated workflow run (its typed report + buffered
-// events) into a redacted, versioned EvidenceManifest and writes it through the #10 EvidenceStore.
-//
-// It was extracted from src/ui/evidence.ts so BOTH the UI BFF and the evaluation harness build the
-// manifest from one implementation. The dependency direction is preserved: this module lives in the
-// audit layer and imports only audit + harness + gateway primitives. It defines its own narrow
-// `WorkflowRunKind` / `WorkflowTerminalStatus` so it never depends on src/ui types. The UI
-// re-exports it (behaviour-preserving); the evaluation runner imports it directly.
-//
-import { buildEvidenceReport } from "./report.js";
-import { resolveCostClass } from "./aggregate.js";
-import { createAuditRedactor, deepRedactStrings } from "./redaction.js";
-import { EVIDENCE_SCHEMA_VERSION } from "./types.js";
-import { HARNESS_VERSION } from "../harness/index.js";
-const KIND_TO_TASK_TYPE = {
-    "unit-tests": "generate-unit-tests",
-    "bug-investigation": "investigate-bug",
-};
-// Folds the buffered `workflow:model:call:completed` events into usage totals. The workflow event
-// carries token/latency fields at the TOP level (not under `usage` like the harness event), so this
-// sums the four dimensions directly rather than reusing the audit `aggregateUsage` fold.
-export function foldWorkflowUsage(events) {
-    let promptTokens = 0;
-    let completionTokens = 0;
-    let requestCount = 0;
-    let totalLatencyMs = 0;
-    for (const event of events) {
-        if (event.type !== "workflow:model:call:completed" &&
-            event.type !== "bug:model:call:completed") {
-            continue;
-        }
-        const record = event;
-        promptTokens += numberOf(record.promptTokens);
-        completionTokens += numberOf(record.completionTokens);
-        totalLatencyMs += numberOf(record.latencyMs);
-        requestCount += 1;
-    }
-    return { promptTokens, completionTokens, requestCount, totalLatencyMs };
-}
-function numberOf(value) {
-    return typeof value === "number" && Number.isFinite(value) ? value : 0;
-}
-export function buildWorkflowManifest(identity, events, report) {
-    return {
-        evidenceSchemaVersion: EVIDENCE_SCHEMA_VERSION,
-        run: {
-            runId: identity.runId,
-            fingerprint: identity.fingerprint,
-            harnessVersion: HARNESS_VERSION,
-            taskType: KIND_TO_TASK_TYPE[identity.kind],
-            outcome: identity.status,
-            startedAt: identity.startedAt,
-            finishedAt: identity.finishedAt,
-            durationMs: Math.max(0, identity.finishedAt - identity.startedAt),
-        },
-        model: { modelId: identity.modelId, costClass: resolveCostClass(identity.modelId) },
-        usageTotals: foldWorkflowUsage(events),
-        ...(contextOf(identity.workspaceRoot) === undefined
-            ? {}
-            : { context: contextOf(identity.workspaceRoot) }),
-        stateTransitions: [],
-        toolCalls: [],
-        commandExecutions: [],
-        verification: verificationOf(report),
-        patch: patchOf(report),
-        failure: undefined,
-    };
-}
-function contextOf(workspaceRoot) {
-    if (workspaceRoot === undefined) {
-        return undefined;
-    }
-    return {
-        workspaceRoot,
-        totalCandidates: 0,
-        usedBytes: 0,
-        budgetBytes: 0,
-        droppedForBudget: 0,
-        entries: [],
-    };
-}
-// Builds, redacts, writes the workflow manifest through the store, and returns the structured
-// EvidenceReport. Errors intentionally surface to the caller so UI/evaluation paths cannot silently
-// claim a terminal run without a durable evidence artifact.
-export function persistWorkflowEvidence(identity, report, events, ctx) {
-    const manifest = buildWorkflowManifest(identity, events, report);
-    const redactor = createAuditRedactor({ additionalSecrets: ctx.additionalSecrets ?? [] }, ctx.env);
-    const redacted = deepRedactStrings(manifest, redactor);
-    const location = ctx.store.put(redacted.run.runId, JSON.stringify(redacted));
-    return buildEvidenceReport(redacted, location);
-}
-function isRecord(value) {
-    return typeof value === "object" && value !== null && !Array.isArray(value);
-}
-// Extracts the verification audit summary from a workflow report when present. The summary is already
-// the audit shape; the deep redact in persist re-scrubs every string leaf for defense in depth.
-function verificationOf(report) {
-    if (!isRecord(report)) {
-        return undefined;
-    }
-    const summary = report.verificationSummary ?? verifiedVerificationOf(report);
-    return isRecord(summary) ? summary : undefined;
-}
-function verifiedVerificationOf(report) {
-    const verified = report.verified;
-    return isRecord(verified) ? verified.verification : undefined;
-}
-// Builds patch metadata (counts/bytes only, never the raw diff) from a workflow report. unit-tests
-// reports carry `addedTestFiles`; bug-investigation reports carry `changedFiles`.
-function patchOf(report) {
-    if (!isRecord(report)) {
-        return undefined;
-    }
-    const proposedDiff = report.proposedDiff;
-    const proposed = typeof proposedDiff === "string" && proposedDiff.length > 0;
-    const changedFiles = changedFileCount(report);
-    if (!proposed && changedFiles === 0) {
-        return undefined;
-    }
-    return {
-        proposed,
-        applied: patchApplied(report),
-        targetFileCount: changedFiles,
-        patchBytes: typeof proposedDiff === "string" ? Buffer.byteLength(proposedDiff, "utf8") : 0,
-        changedFiles,
-        created: 0,
-        deleted: 0,
-    };
-}
-function patchApplied(report) {
-    if (report.status === "completed" || report.status === "fix-applied") {
-        return true;
-    }
-    const verified = report.verified;
-    return isRecord(verified) && verified.patchApplied === true;
-}
-function changedFileCount(report) {
-    const added = report.addedTestFiles;
-    if (Array.isArray(added)) {
-        return added.length;
-    }
-    const changed = report.changedFiles;
-    return Array.isArray(changed) ? changed.length : 0;
-}
+// Re-export shim: the workflow→EvidenceManifest mapping lives in @oscharko-dev/keiko-evidence
+// (issue #163, ADR-0019).
+export { buildWorkflowManifest, foldWorkflowUsage, persistWorkflowEvidence, } from "@oscharko-dev/keiko-evidence";

package/dist/cli/index.js

@@ -1,7 +1,26 @@
 #!/usr/bin/env node
-import { runCli } from "./runner.js";
-// runCli returns a number for synchronous commands and a Promise<number> for the async
-// `run` command. Promise.resolve normalises both before exiting with the resulting code.
+import { resolve, dirname } from "node:path";
+import { fileURLToPath } from "node:url";
+import { runCli } from "@oscharko-dev/keiko-cli";
+// Re-export shim: the CLI command modules now live in @oscharko-dev/keiko-cli
+// (issue #168, ADR-0019). This file remains the entry the published bin points
+// at (`bin: dist/cli/index.js` in the root package.json) and is the only
+// `src/cli/` file with a shebang.
+//
+// The bin entry is the authoritative reference point for two installation-
+// dependent paths the cli package needs:
+//   - the packaged UI static export at `dist/ui/static`, used by `keiko ui`;
+//   - the bin executable path itself, used by `keiko start` when it re-execs
+//     the cli as a detached child to serve the UI.
+// We surface both via env vars before dispatch so the cli package does not
+// have to deep-import the bin or know its own installation layout. Tests can
+// override either variable to point at fixtures.
+const HERE = dirname(fileURLToPath(import.meta.url));
+const ROOT_DIST = resolve(HERE, "..");
+process.env.KEIKO_CLI_BIN_PATH ??= resolve(HERE, "index.js");
+process.env.KEIKO_UI_STATIC_ROOT ??= resolve(ROOT_DIST, "ui", "static");
+// runCli returns a number for synchronous commands and a Promise<number> for
+// the async `run` command; Promise.resolve normalises both before exiting.
 void Promise.resolve(runCli(process.argv.slice(2), {
     out: (text) => {
         process.stdout.write(text);

package/dist/evaluations/runner-support.d.ts

@@ -1,6 +1,6 @@
 import type { SpawnFn, WorkspaceWriter } from "../tools/index.js";
-import type { UnitTestWorkflowInput } from "../workflows/unit-tests/types.js";
-import type { BugInvestigationInput } from "../workflows/bug-investigation/types.js";
+import type { UnitTestWorkflowInput } from "@oscharko-dev/keiko-workflows";
+import type { BugInvestigationInput } from "@oscharko-dev/keiko-workflows";
 import type { ScoringInput } from "./scorer.js";
 import type { EvaluationFixture } from "./types.js";
 export interface MaterializedWorkspace {

package/dist/evaluations/runner.js

@@ -6,10 +6,11 @@
 // the #10 store, scores every dimension, aggregates the suite, and cleans up the temp dir. No
 // network or live-model call is made in offline mode; no Date.now / Math.random touches a scored path.
 import { createHash, randomUUID } from "node:crypto";
-import { generateUnitTests } from "../workflows/unit-tests/workflow.js";
-import { investigateBug } from "../workflows/bug-investigation/workflow.js";
+import { generateUnitTests } from "@oscharko-dev/keiko-workflows";
+import { investigateBug } from "@oscharko-dev/keiko-workflows";
 import { createNodeEvidenceStore, persistWorkflowEvidence, resolveEvidenceDir, } from "../audit/index.js";
 import { canonicalise, HARNESS_VERSION } from "../harness/index.js";
+import { resolveCostClass } from "../gateway/index.js";
 import { createEvaluationModelProvider } from "./model-provider.js";
 import { aggregateScorecard, scoreFixture, summarizeScorecard } from "./scorer.js";
 import { checkSurfaceParity } from "./surface-parity.js";
@@ -68,7 +69,7 @@ function persistAndCheck(fixture, report, store, env, runId, workspaceRoot, mode
         startedAt,
         finishedAt,
         workspaceRoot,
-    }, report, events, { store, env });
+    }, report, events, { store, env, costClassResolver: resolveCostClass });
     const raw = store.get(runId);
     return {
         manifestValid: raw !== undefined && isManifestValid(raw),

package/dist/evaluations/surface-parity.js

@@ -3,8 +3,7 @@
 // contracts. It is NOT a scored dimension: it is a fixed structural invariant of the codebase, so it
 // has its own scorecard section and its own test file. A parity failure is a hard blocker that causes
 // `keiko evaluate` to exit 1 regardless of dimension scores.
-import { runGenTestsCli } from "../cli/gen-tests.js";
-import { runInvestigateCli } from "../cli/investigate.js";
+import { runGenTestsCli, runInvestigateCli } from "@oscharko-dev/keiko-cli";
 import { BUG_INVESTIGATION_WORKFLOW_DESCRIPTOR, UNIT_TEST_WORKFLOW_DESCRIPTOR, } from "../workflows/index.js";
 const DESCRIPTOR_EXPECTATIONS = [
     {
@@ -140,7 +139,7 @@ async function checkSdkExports() {
 // time guarantee is enforced by the TypeScript check; this is the runtime shape assertion (D7 d).
 // Composer-launched workflow runs must also carry the selected local project context.
 async function checkRunRequestShapes() {
-    const { parseRunRequest } = await import("../ui/run-request.js");
+    const { parseRunRequest } = await import("../ui/index.js");
     return RUN_REQUEST_EXPECTATIONS.map((expectation) => {
         const parsed = parseRunRequest(JSON.stringify({
             workflowId: expectation.workflowId,

package/dist/evaluations/types.d.ts

@@ -1,74 +1,2 @@
-import type { NormalizedResponse } from "../gateway/types.js";
-export type EvaluationDimension = "task-completion" | "patch-correctness" | "test-pass-rate" | "verification-completeness" | "patch-size" | "audit-completeness" | "unsafe-action-rejection";
-export declare const EVALUATION_DIMENSIONS: readonly EvaluationDimension[];
-export interface FixtureOracle {
-    readonly expectedStatuses: readonly string[];
-    readonly expectPatch: boolean;
-    readonly expectVerificationSkip: boolean;
-    readonly maxExpectedChangedFiles: number;
-    readonly maxExpectedPatchBytes: number;
-}
-export type WorkflowKind = "unit-tests" | "bug-investigation";
-export interface EvaluationFixture {
-    readonly name: string;
-    readonly workflowKind: WorkflowKind;
-    readonly workspaceFiles: Record<string, string>;
-    readonly workflowInput: Record<string, unknown>;
-    readonly apply?: boolean | undefined;
-    readonly mockTranscript: readonly (NormalizedResponse | Error)[];
-    readonly dimensions: ReadonlySet<EvaluationDimension>;
-    readonly oracle: FixtureOracle;
-}
-export type DimensionOutcome = "pass" | "fail" | "not-applicable";
-export interface DimensionResult {
-    readonly dimension: EvaluationDimension;
-    readonly outcome: DimensionOutcome;
-    readonly reason?: string | undefined;
-}
-export interface FixtureRunResult {
-    readonly fixtureName: string;
-    readonly workflowKind: WorkflowKind;
-    readonly durationMs: number;
-    readonly dimensionResults: readonly DimensionResult[];
-    readonly report: Record<string, unknown>;
-}
-export interface ScorecardEntry {
-    readonly dimension: EvaluationDimension;
-    readonly passCount: number;
-    readonly failCount: number;
-    readonly notApplicableCount: number;
-    readonly passRate: number | null;
-}
-export interface SurfaceParityCheckResult {
-    readonly check: string;
-    readonly workflowKind: WorkflowKind;
-    readonly passed: boolean;
-    readonly reason?: string | undefined;
-}
-export interface SurfaceParityResult {
-    readonly allPassed: boolean;
-    readonly checks: readonly SurfaceParityCheckResult[];
-}
-export interface LiveRunContext {
-    readonly modelId: string;
-    readonly configDescriptor: string;
-    readonly evidenceRefs: readonly string[];
-}
-export declare const EVAL_SCORECARD_SCHEMA_VERSION: "1";
-export interface ScorecardSummary {
-    readonly totalFixtures: number;
-    readonly fullyPassedFixtures: number;
-    readonly safetyGatePassed: boolean;
-    readonly pilotReadyIndicator: boolean;
-}
-export interface EvalScorecard {
-    readonly schemaVersion: typeof EVAL_SCORECARD_SCHEMA_VERSION;
-    readonly evaluatedAt: string;
-    readonly mode: EvaluationMode;
-    readonly liveRunContext?: LiveRunContext | undefined;
-    readonly dimensions: readonly ScorecardEntry[];
-    readonly surfaceParity: SurfaceParityResult;
-    readonly fixtureResults: readonly FixtureRunResult[];
-    readonly summary: ScorecardSummary;
-}
-export type EvaluationMode = "offline" | "live";
+export type { EvaluationDimension, FixtureOracle, WorkflowKind, EvaluationFixture, DimensionOutcome, DimensionResult, FixtureRunResult, ScorecardEntry, SurfaceParityCheckResult, SurfaceParityResult, LiveRunContext, ScorecardSummary, EvalScorecard, EvaluationMode, } from "@oscharko-dev/keiko-contracts";
+export { EVALUATION_DIMENSIONS, EVAL_SCORECARD_SCHEMA_VERSION, } from "@oscharko-dev/keiko-contracts";

package/dist/evaluations/types.js

@@ -1,16 +1,4 @@
-// All evaluation interfaces and type aliases for the Wave 1 evaluation harness (ADR-0012 D6/D8/D11).
-// No runtime logic lives here beyond the frozen schema-version constant the type layer exposes as a
-// value. `readonly` everywhere; optional props are `| undefined` because exactOptionalPropertyTypes
-// is on. Every scorecard shape is plain JSON-serializable so the harness can emit it to stdout/file.
-// The seven scored dimensions in their canonical scorecard order. A scorecard always carries one
-// ScorecardEntry per name here, even when no fixture exercises the dimension (all not-applicable).
-export const EVALUATION_DIMENSIONS = [
-    "task-completion",
-    "patch-correctness",
-    "test-pass-rate",
-    "verification-completeness",
-    "patch-size",
-    "audit-completeness",
-    "unsafe-action-rejection",
-];
-export const EVAL_SCORECARD_SCHEMA_VERSION = "1";
+// Re-export shim: evaluation contract types and the EVAL_SCORECARD_SCHEMA_VERSION constant live in
+// @oscharko-dev/keiko-contracts (issue #158). `verbatimModuleSyntax` is on, so type-only names use
+// `export type` and value-emitting frozen constants use `export`.
+export { EVALUATION_DIMENSIONS, EVAL_SCORECARD_SCHEMA_VERSION, } from "@oscharko-dev/keiko-contracts";

package/dist/gateway/capabilities.d.ts

@@ -1,12 +1,2 @@
-import type { ModelCapability, ModelKind } from "./types.js";
-export declare const CAPABILITY_REGISTRY: readonly ModelCapability[];
-export interface CapabilityQuery {
-    readonly kind?: ModelKind | undefined;
-    readonly toolCalling?: boolean | undefined;
-    readonly structuredOutput?: boolean | undefined;
-    readonly minContextWindow?: number | undefined;
-}
-export declare function findCapability(modelId: string): ModelCapability | undefined;
-export declare function listCapabilities(): readonly ModelCapability[];
-export declare function createDefaultChatCapability(modelId: string): ModelCapability;
-export declare function selectCheapest(query: CapabilityQuery): ModelCapability | undefined;
+export { CAPABILITY_REGISTRY, createDefaultChatCapability, findCapability, listCapabilities, selectCheapest, } from "@oscharko-dev/keiko-model-gateway";
+export type { CapabilityQuery } from "@oscharko-dev/keiko-model-gateway";

package/dist/gateway/capabilities.data.d.ts

@@ -1,2 +1 @@
-import type { ModelCapability } from "./types.js";
-export declare const CAPABILITY_DATA: readonly ModelCapability[];
+export { CAPABILITY_DATA } from "@oscharko-dev/keiko-model-gateway";

package/dist/gateway/capabilities.data.js

@@ -1,5 +1,3 @@
-// Raw built-in capability registry data.
-//
-// Keiko intentionally ships no customer or deployment-specific model ids. Private model
-// capabilities are supplied by local config or discovered at runtime during UI onboarding.
-export const CAPABILITY_DATA = [];
+// Re-export shim: capabilities.data now lives in @oscharko-dev/keiko-model-gateway (issue #160, ADR-0019).
+// All existing import sites (`from "../gateway/capabilities.data.js"`) keep resolving unchanged via this barrel.
+export { CAPABILITY_DATA } from "@oscharko-dev/keiko-model-gateway";