@alis-build/harness-eval 0.1.0 → 0.1.2

package/README.md

@@ -33,6 +33,8 @@ npx @alis-build/harness-eval run examples/basic.yaml --output report.json
 
 The npm package name is `@alis-build/harness-eval`; the CLI binary is `harness-eval`. With a single bin entry, `npx @alis-build/harness-eval <command>` invokes it directly.
 
+In a git checkout of this repo, npm resolves `npx @alis-build/harness-eval` to the local package (not the registry). Run `pnpm run build` first so `dist/cli/bin.js` exists; the build links `harness-eval` into `node_modules/.bin` for local use.
+
 ### Development (clone & build)
 
 Contributors working from a git checkout:
@@ -40,7 +42,8 @@ Contributors working from a git checkout:
 ```bash
 pnpm install
 pnpm run build
-node dist/cli/bin.js --help
+pnpm exec harness-eval --help
+# or: node dist/cli/bin.js --help
 ```
 
 ---
@@ -93,6 +96,8 @@ cases:
 
 Generic fields (`model`, `cwd`, `timeoutMs`, `env`) sit at the top level. Claude-specific options go under `claudeCode`.
 
+**Full suite & grading YAML reference:** [docs/suite-config.md](docs/suite-config.md) — all case/matrix fields, `reference_trajectory`, `human_ratings`, multi-file layout, and `grading.yaml` options.
+
 ### 2. Run behavioral eval
 
 ```bash
@@ -419,6 +424,8 @@ npx @alis-build/harness-eval --help
 
 Uses a standalone **`grading.yaml`** for judge model, timeout, env, and `claudeCode` flags (Option B — separate from the suite file).
 
+**Field reference:** [docs/suite-config.md — Grading config](docs/suite-config.md#grading-config-gradingyaml)
+
 ```yaml
 # examples/grading.yaml
 judge:
@@ -464,7 +471,7 @@ npx @alis-build/harness-eval envelope report.json --suite examples/basic.yaml --
 # Interchange projections
 npx @alis-build/harness-eval envelope report.json --projection trajectory --output trajectory.jsonl
 npx @alis-build/harness-eval envelope report.json --projection instances --output instances.json
-npx @alis-build/harness-eval envelope report.json --projection agent-trace --output agent-traces.json
+npx @alis-build/harness-eval envelope report.json --projection instances --output instances.jsonl
 ```
 
 | Option                                                      | Description                                               |
@@ -472,7 +479,7 @@ npx @alis-build/harness-eval envelope report.json --projection agent-trace --out
 | `--output <path>`                                           | Write output (stdout if omitted)                          |
 | `--grading <path>`                                          | Merge `grading.json` outcome scores into the envelope     |
 | `--suite <path>`                                            | Suite YAML for provenance (`uri`, `contentHash`)          |
-| `--projection envelope\|trajectory\|instances\|agent-trace` | Output shape (default: `envelope`)                        |
+| `--projection envelope\|trajectory\|instances` | Output shape (default: `envelope`)                        |
 | `--include-raw-stream-events`                               | Include adapter raw stream events in repetition artifacts |
 | `--no-transcript`                                           | Omit judge transcript artifacts                           |
 
@@ -506,6 +513,8 @@ See [Data contracts & schemas](#data-contracts--schemas) for type details.
 
 ## Suite concepts
 
+**Authoring reference:** [docs/suite-config.md](docs/suite-config.md) — complete field list for suite YAML, matrix cells, test cases, reference trajectories, and grading config.
+
 ### Test case
 
 One prompt + assertions + optional expectations, run N times per matrix cell.
@@ -530,6 +539,10 @@ assertions:
 
 Default threshold is `1.0` (every evaluated rep must pass). Reps where the harness crashes are excluded from the denominator and counted as `adapterErrors`.
 
+### Reference trajectory (optional)
+
+Define expected tool calls for Vertex trajectory metrics on the eval envelope. Use `tool_name_mode: bare` when reference steps use short tool names but the harness records MCP-prefixed names. See [docs/suite-config.md — Reference trajectory](docs/suite-config.md#reference-trajectory).
+
 **Full reference:** [docs/assertions.md](docs/assertions.md) — all assertion kinds, predicates, statistical model, and how to add new assertion types or harness adapters.
 
 ---
@@ -687,7 +700,7 @@ pnpm run typecheck
 pnpm run generate-schemas   # Zod → schemas/*.schema.json only
 ```
 
-**Docs:** [Assertion DSL & adapter extension](docs/assertions.md) · [Eval record contract (DB / CI)](docs/eval-record.md)
+**Docs:** [Suite & grading YAML](docs/suite-config.md) · [Assertion DSL & adapter extension](docs/assertions.md) · [Eval record contract (DB / CI)](docs/eval-record.md)
 
 ---
 

package/dist/adapters/claude-code/index.d.ts

@@ -1,3 +1,3 @@
 import { n as AdapterError, o as ParseErrorRecord, r as AdapterResult, t as AdapterDiagnostics } from "../../types-B9H4IZtA.js";
-import { a as ClaudeCodeAdapterResult, i as ClaudeCodeAdapterConfig, o as ClaudeCodeOptions, r as runClaudeCode, s as PermissionMode, t as claudeCodeAdapter } from "../../index-6Z17eKZx.js";
+import { a as ClaudeCodeAdapterResult, i as ClaudeCodeAdapterConfig, o as ClaudeCodeOptions, r as runClaudeCode, s as PermissionMode, t as claudeCodeAdapter } from "../../index-V22PrR0p.js";
 export { type AdapterDiagnostics, AdapterError, type AdapterResult, type ClaudeCodeAdapterConfig, type ClaudeCodeAdapterResult, type ClaudeCodeOptions, type ParseErrorRecord, type PermissionMode, claudeCodeAdapter, runClaudeCode };

package/dist/adapters/claude-code/index.js

@@ -1,2 +1,2 @@
-import { a as AdapterError, r as runClaudeCode, t as claudeCodeAdapter } from "../../claude-code-ycT0JQZF.js";
+import { a as AdapterError, r as runClaudeCode, t as claudeCodeAdapter } from "../../claude-code-DZ4Vkgp6.js";
 export { AdapterError, claudeCodeAdapter, runClaudeCode };

package/dist/{claude-code-ycT0JQZF.js → claude-code-DZ4Vkgp6.js}

@@ -298,10 +298,15 @@ var AdapterError = class extends Error {
 };
 //#endregion
 //#region src/adapters/claude-code/flags.ts
+/** Append repeated `--flag value` pairs for array config fields. */
 function pushRepeatableFlag(args, flag, values) {
 	if (!values) return;
 	for (const value of values) args.push(flag, value);
 }
+/**
+* Append an optional CLI flag. Boolean `true` emits the flag alone; other
+* scalars emit `--flag value`.
+*/
 function pushOptionalFlag(args, flag, value) {
 	if (value === void 0) return;
 	if (typeof value === "boolean") {
@@ -360,7 +365,12 @@ function buildArgs(config) {
 	appendClaudeCodeFlags(args, config);
 	return args;
 }
-/** Build args for an LLM judge subprocess (`--output-format json`). */
+/**
+* Build args for an LLM judge subprocess (`--output-format json`).
+*
+* Defaults permission mode to `bypassPermissions` so the judge does not
+* block on tool permission prompts during single-shot JSON grading.
+*/
 function buildJudgeArgs(prompt, config = {}) {
 	const args = [
 		"-p",
@@ -402,6 +412,14 @@ const KILL_GRACE_MS = 5e3;
 /**
 * Spawn `claude` in headless mode with isolated config and a process-group
 * lifecycle. See {@link SpawnedClaude} for how to consume the result.
+*
+* **Kill sequence:** timeout and abort both follow the same two-step path:
+* `SIGTERM` to the process group, then `SIGKILL` after {@link KILL_GRACE_MS}
+* if the group is still alive. This avoids leaving MCP/tool subprocesses
+* running while still giving claude a chance to flush stream-json output.
+*
+* @param config - Adapter options; `timeoutMs`, `signal`, and `isolateConfig`
+*   control lifecycle and config isolation.
 */
 async function spawnClaude(config) {
 	const binary = config.binary ?? "claude";
@@ -425,6 +443,10 @@ async function spawnClaude(config) {
 	let timedOut = false;
 	let killEscalation = null;
 	const timeoutMs = config.timeoutMs ?? DEFAULT_TIMEOUT_MS;
+	/**
+	* Arm (or re-arm) the SIGKILL fallback. Each SIGTERM attempt gets its own
+	* grace window so a slow shutdown doesn't leave orphaned MCP servers.
+	*/
 	const scheduleKillEscalation = () => {
 		if (killEscalation) clearTimeout(killEscalation);
 		killEscalation = setTimeout(() => killTree(child, "SIGKILL"), KILL_GRACE_MS);
@@ -487,10 +509,16 @@ async function spawnClaude(config) {
 * group is already gone. This catches MCP server subprocesses and tool
 * processes spawned by claude.
 *
-* Why both? On some platforms the process group dies before we get here
-* (the child itself already cleaned up), in which case `kill(-pid)` throws
-* ESRCH. The fallback handles that edge case without leaking zombies in
-* the common case.
+* **Signal escalation:** callers typically invoke this first with `SIGTERM`,
+* then again with `SIGKILL` after {@link KILL_GRACE_MS}. The group kill is
+* essential — a bare `child.kill()` would leave MCP servers running.
+*
+* **Platform edge case:** when the group leader exits first, `kill(-pid)`
+* throws `ESRCH`. The single-PID fallback covers that without failing the
+* adapter run.
+*
+* @param child - Spawned process handle from {@link spawn}.
+* @param signal - POSIX signal to deliver (`SIGTERM` or `SIGKILL` in practice).
 */
 function killTree(child, signal) {
 	if (child.pid === void 0) return;
@@ -553,6 +581,7 @@ async function runClaudeCode(config) {
 		await spawned.cleanup();
 	}
 }
+/** Registered {@link HarnessAdapter} for Claude Code headless runs. */
 const claudeCodeAdapter = {
 	id: "claude-code",
 	run: runClaudeCode
@@ -560,4 +589,4 @@ const claudeCodeAdapter = {
 //#endregion
 export { isUserMessage as _, AdapterError as a, buildTrajectory as c, isResult as d, isSystemInit as f, isToolUseBlock as g, isToolResultBlock as h, buildJudgeArgs as i, namespaceOf as l, isTextBlock as m, claude_code_exports as n, parseStreamJson as o, isSystemRetry as p, runClaudeCode as r, TrajectoryBuilder as s, claudeCodeAdapter as t, isAssistantMessage as u };
 
-//# sourceMappingURL=claude-code-ycT0JQZF.js.map
+//# sourceMappingURL=claude-code-DZ4Vkgp6.js.map

package/dist/{claude-code-ycT0JQZF.js.map → claude-code-DZ4Vkgp6.js.map}

@@ -1 +1 @@
-{"version":3,"file":"claude-code-ycT0JQZF.js","names":[],"sources":["../src/types/stream.ts","../src/types/trajectory.ts","../src/trajectory/builder.ts","../src/parsers/stream-json.ts","../src/adapters/types.ts","../src/adapters/claude-code/flags.ts","../src/adapters/claude-code/process.ts","../src/adapters/claude-code/index.ts"],"sourcesContent":["/**\n * Discriminated union of events emitted by Claude Code's\n * `--output-format stream-json` mode.\n *\n * The format is NDJSON (one JSON object per line on stdout). Each line has\n * a required `type` field and often a `subtype` for further disambiguation.\n *\n * Source notes: the stream-json schema is not formally documented as of mid-2026.\n * These types are derived from:\n *   - https://code.claude.com/docs/en/headless\n *   - https://github.com/anthropics/claude-code/issues/24612 (event-types tracking issue)\n *   - https://takopi.dev/reference/runners/claude/stream-json-cheatsheet/\n *   - The `@anthropic-ai/claude-agent-sdk` TypeScript declaration files,\n *     which are the de-facto source of truth.\n *\n * When adding new event types, prefer extending the union here rather than\n * branching on `any` in callers. Unknown events should be tolerated silently\n * by the builder (the schema evolves and we don't want CI to break on a new\n * event type we haven't modelled yet).\n */\n\n/** Top-level discriminated union of stream-json events. */\nexport type StreamEvent =\n  | SystemInitEvent\n  | SystemRetryEvent\n  | SystemPluginInstallEvent\n  | SystemCompactBoundaryEvent\n  | SystemUnknownEvent\n  | AssistantMessageEvent\n  | UserMessageEvent\n  | ResultEvent;\n\n// system events\n\n/** Emitted once at session start. Carries the session-level metadata. */\nexport interface SystemInitEvent {\n  type: \"system\";\n  subtype: \"init\";\n  session_id: string;\n  cwd: string;\n  model: string;\n  permissionMode?: string;\n  apiKeySource?: string;\n  /** Names of tools available in the session (built-in + MCP). */\n  tools: string[];\n  /** MCP servers configured for this session, with connection status. */\n  mcp_servers: McpServerStatus[];\n}\n\nexport interface McpServerStatus {\n  name: string;\n  status: \"connected\" | \"disconnected\" | \"error\" | string;\n}\n\n/** Emitted when the API rate-limits us or otherwise asks for a retry. */\nexport interface SystemRetryEvent {\n  type: \"system\";\n  subtype: \"api_retry\";\n  session_id: string;\n  /** Implementation-defined retry payload (delay, reason, etc). */\n  [key: string]: unknown;\n}\n\n/** Emitted while marketplace plugins are installing pre-session. */\nexport interface SystemPluginInstallEvent {\n  type: \"system\";\n  subtype: \"plugin_install\";\n  session_id: string;\n  [key: string]: unknown;\n}\n\n/** Emitted when Claude Code compacts the context window mid-session. */\nexport interface SystemCompactBoundaryEvent {\n  type: \"system\";\n  subtype: \"compact_boundary\";\n  session_id: string;\n  [key: string]: unknown;\n}\n\n/**\n * Catch-all for `type: \"system\"` events we haven't modelled.\n *\n * Keeps the union exhaustive while tolerating schema evolution. Callers should\n * either explicitly handle a known subtype or fall through to ignore.\n */\nexport interface SystemUnknownEvent {\n  type: \"system\";\n  subtype: string;\n  session_id?: string;\n  [key: string]: unknown;\n}\n\n// conversational events\n\n/** One assistant turn. The `message` field mirrors the Anthropic Messages API shape. */\nexport interface AssistantMessageEvent {\n  type: \"assistant\";\n  session_id: string;\n  message: AssistantMessage;\n}\n\nexport interface AssistantMessage {\n  id: string;\n  type: \"message\";\n  role: \"assistant\";\n  content: ContentBlock[];\n  model?: string;\n  stop_reason?: StopReason | null;\n  usage?: Usage;\n}\n\n/**\n * A user-role message in the stream.\n *\n * In stream-json these are usually *synthetic* — the harness injects them to\n * feed tool results back into the conversation after dispatching a tool. The\n * very first user message (the prompt) is also emitted here for completeness.\n */\nexport interface UserMessageEvent {\n  type: \"user\";\n  session_id: string;\n  message: UserMessage;\n}\n\nexport interface UserMessage {\n  role: \"user\";\n  /** String for the initial prompt, array of blocks when carrying tool results. */\n  content: ContentBlock[] | string;\n}\n\n// content blocks\n\nexport type ContentBlock = TextBlock | ToolUseBlock | ToolResultBlock;\n\nexport interface TextBlock {\n  type: \"text\";\n  text: string;\n}\n\nexport interface ToolUseBlock {\n  type: \"tool_use\";\n  /** Unique id assigned by the model; used to match tool_result back to this call. */\n  id: string;\n  /** Tool name. MCP tools follow the convention `mcp__<server>__<tool>`. */\n  name: string;\n  /** Arguments the model passed. Schema is per-tool. */\n  input: unknown;\n}\n\nexport interface ToolResultBlock {\n  type: \"tool_result\";\n  /** The id of the corresponding tool_use block. */\n  tool_use_id: string;\n  /** Tool output. May be plain text or further content blocks for richer tools. */\n  content: string | ContentBlock[];\n  is_error?: boolean;\n}\n\n// result envelope\n\n/** Emitted once at session end. Carries aggregate usage and cost. */\nexport interface ResultEvent {\n  type: \"result\";\n  subtype: \"success\" | \"error\";\n  session_id: string;\n  total_cost_usd: number;\n  is_error: boolean;\n  duration_ms: number;\n  duration_api_ms?: number;\n  num_turns: number;\n  /** The final text the harness returned, if any. */\n  result?: string;\n  usage?: Usage;\n}\n\n// shared scalars\n\n/**\n * Reasons the model can stop a turn. Open-ended string union because new\n * stop reasons appear over time.\n */\nexport type StopReason =\n  | \"end_turn\"\n  | \"tool_use\"\n  | \"max_tokens\"\n  | \"stop_sequence\"\n  | (string & {});\n\nexport interface Usage {\n  input_tokens: number;\n  output_tokens: number;\n  cache_creation_input_tokens?: number;\n  cache_read_input_tokens?: number;\n}\n\n// type guards\n\n/** Type guards. Prefer these over manual `e.type === \"...\"` checks at call sites. */\n\nexport function isSystemInit(e: StreamEvent): e is SystemInitEvent {\n  return e.type === \"system\" && (e as SystemInitEvent).subtype === \"init\";\n}\n\nexport function isSystemRetry(e: StreamEvent): e is SystemRetryEvent {\n  return e.type === \"system\" && (e as SystemRetryEvent).subtype === \"api_retry\";\n}\n\nexport function isAssistantMessage(e: StreamEvent): e is AssistantMessageEvent {\n  return e.type === \"assistant\";\n}\n\nexport function isUserMessage(e: StreamEvent): e is UserMessageEvent {\n  return e.type === \"user\";\n}\n\nexport function isResult(e: StreamEvent): e is ResultEvent {\n  return e.type === \"result\";\n}\n\nexport function isTextBlock(b: ContentBlock): b is TextBlock {\n  return b.type === \"text\";\n}\n\nexport function isToolUseBlock(b: ContentBlock): b is ToolUseBlock {\n  return b.type === \"tool_use\";\n}\n\nexport function isToolResultBlock(b: ContentBlock): b is ToolResultBlock {\n  return b.type === \"tool_result\";\n}\n","/**\n * TrajectoryView — the assertion-friendly projection of a Claude Code session.\n *\n * The view is derived from the stream of {@link StreamEvent} values produced by\n * the harness, but is optimized for the queries that the assertion DSL needs to\n * express:\n *\n *   - did tool X get called?  (look at `toolCalls`)\n *   - did tool A come before tool B?  (compare `turnIndex` / `callIndex`)\n *   - was a tool called with arguments matching predicate P?  (`toolCalls[i].args`)\n *   - did the agent answer without using any tool?  (`toolCalls.length === 0`)\n *\n * The view is reconstructable from the raw events (lossless w.r.t. assertions),\n * but operating on it directly is dramatically simpler than walking event\n * streams or OTel span trees.\n *\n * Design notes:\n *  - `turnIndex` and `callIndex` are the right primitives for ordering.\n *    Wall-clock timestamps from the stream are unreliable for sub-second\n *    ordering and parallel tool dispatch.\n *  - Parallel tool calls (multiple `tool_use` blocks in one assistant message)\n *    share a `turnIndex` but have distinct `callIndex` values in emission order.\n *  - `namespace` is precomputed so assertions like `called(pattern: \"mcp__api__*\")`\n *    can do a cheap string check.\n */\n\nimport type { StopReason } from \"./stream\";\n\nexport interface TrajectoryView {\n  /** Session metadata, captured from the `system/init` event. */\n  meta: SessionMeta;\n\n  /** Every tool call, in global emission order. */\n  toolCalls: ToolCall[];\n\n  /** Each assistant turn: text content + any tool calls emitted in that turn. */\n  turns: AssistantTurn[];\n\n  /** All assistant text concatenated across turns. Useful for `response_contains`. */\n  finalResponse: string;\n\n  /** Stop reason of the *last* assistant turn. */\n  finalStopReason: StopReason | null;\n\n  /** Aggregate usage and cost from the result event. */\n  usage: UsageSummary;\n\n  /** Retry events observed during the run (rate limits, transient errors). */\n  retries: RetryRecord[];\n\n  /** Whether the result envelope indicated success. */\n  success: boolean;\n}\n\nexport interface SessionMeta {\n  sessionId: string;\n  model: string;\n  cwd: string;\n  permissionMode?: string;\n  /** Tool names the harness reported as available at session start. */\n  availableTools: string[];\n  /** MCP servers configured for the session, with connection status. */\n  mcpServers: { name: string; status: string }[];\n}\n\nexport interface ToolCall {\n  /** Fully-qualified tool name, e.g. `\"mcp__api__search_skills\"` or `\"Bash\"`. */\n  name: string;\n\n  /**\n   * Namespace prefix for MCP-style names (`\"mcp__api\"`), or null for built-ins.\n   * Precomputed via {@link namespaceOf} for cheap pattern matching.\n   */\n  namespace: string | null;\n\n  /** The `tool_use` block's `id`; matches a later `tool_result.tool_use_id`. */\n  callId: string;\n\n  /** Args the model emitted on this call. Tool-specific schema. */\n  args: unknown;\n\n  /** Tool result, or null if no result was observed (e.g. process killed). */\n  result: unknown | null;\n\n  /** Whether the tool reported an error in its result. */\n  isError: boolean;\n\n  /**\n   * Which assistant turn produced this call. Parallel calls within a single\n   * assistant message share a `turnIndex`.\n   */\n  turnIndex: number;\n\n  /** Index in the global ordered tool-call sequence. */\n  callIndex: number;\n}\n\nexport interface AssistantTurn {\n  turnIndex: number;\n  /** Text emitted in this turn (may be empty if turn was tool-only). */\n  text: string;\n  /** Tool calls emitted in this turn, in their block order. */\n  toolCalls: ToolCall[];\n  /** Stop reason reported by the model for this turn. */\n  stopReason: StopReason | null;\n}\n\nexport interface UsageSummary {\n  inputTokens: number;\n  outputTokens: number;\n  totalCostUsd: number;\n  durationMs: number;\n  numTurns: number;\n}\n\nexport interface RetryRecord {\n  /** ms since session start (approximate; the stream doesn't include precise ts). */\n  offsetMs: number;\n  /** Raw payload from the `system/api_retry` event for diagnostics. */\n  raw: unknown;\n}\n\n// helpers\n\n/**\n * Extract the MCP namespace prefix from a tool name.\n *\n * Claude Code formats MCP tool names as `mcp__<server>__<tool>`. The namespace\n * is the first two segments joined: `mcp__<server>`. Returns null for non-MCP\n * tool names (built-ins like `Bash`, `Read`, `Edit`).\n *\n * @example\n *   namespaceOf(\"mcp__api__search_skills\") // \"mcp__api\"\n *   namespaceOf(\"Bash\")                     // null\n */\nexport function namespaceOf(toolName: string): string | null {\n  if (!toolName.startsWith(\"mcp__\")) return null;\n  const parts = toolName.split(\"__\");\n  if (parts.length < 3) return null;\n  return `${parts[0]}__${parts[1]}`;\n}\n","/**\n * TrajectoryBuilder — consumes a stream of {@link StreamEvent} values and\n * produces a {@link TrajectoryView}.\n *\n * State machine: the builder is a small, tolerant state machine. Invariants:\n *\n *   - Exactly one `system/init` event opens the session. The builder requires\n *     it to be present before `build()`.\n *   - Each `assistant` event begins a new turn. Text blocks accumulate into\n *     the turn's text; `tool_use` blocks become `ToolCall` records.\n *   - `user` events with `tool_result` blocks deliver tool results back. We\n *     match them to pending calls by `tool_use_id`.\n *   - One `result` event closes the session and carries aggregate usage.\n *\n * The builder is *tolerant of partial streams*: a process killed mid-run\n * produces a coherent (but flagged) view. Tool calls without matching results\n * keep `result: null`. The `success` flag reflects whether a successful result\n * event was actually observed.\n *\n * Why a class (not a reducer)?\n *   The internal `pendingCalls` map is mutable by design — we modify ToolCall\n *   objects in place when results arrive, so other parts of the view (which\n *   hold references to the same objects) see the update for free. A reducer\n *   would force a deep copy per result event, which is wasteful and would\n *   complicate identity-based queries.\n */\n\nimport {\n  isAssistantMessage,\n  isResult,\n  isSystemInit,\n  isTextBlock,\n  isToolResultBlock,\n  isToolUseBlock,\n  isUserMessage,\n  type StreamEvent,\n  type Usage,\n} from \"../types/stream\";\nimport {\n  namespaceOf,\n  type AssistantTurn,\n  type RetryRecord,\n  type SessionMeta,\n  type ToolCall,\n  type TrajectoryView,\n} from \"../types/trajectory\";\n\nexport class TrajectoryBuilder {\n  private meta: SessionMeta | null = null;\n  private sessionStartTs: number | null = null;\n\n  private turns: AssistantTurn[] = [];\n  private allToolCalls: ToolCall[] = [];\n\n  /**\n   * tool_use_id → ToolCall, for matching results back to calls.\n   * Entries are removed once a result is observed.\n   */\n  private pendingCalls: Map<string, ToolCall> = new Map();\n\n  private retries: RetryRecord[] = [];\n\n  private finalUsage: Usage | null = null;\n  private finalCostUsd = 0;\n  private finalDurationMs = 0;\n  private finalNumTurns = 0;\n  private finalResultText = \"\";\n  private sawResultEvent = false;\n  private resultIsError = false;\n\n  /**\n   * Consume one event. Safe to call with events in stream order.\n   *\n   * Unknown event types are silently ignored — the schema evolves and we\n   * don't want CI to break on a new event type we haven't modelled.\n   */\n  consume(event: StreamEvent): void {\n    if (isSystemInit(event)) {\n      this.meta = {\n        sessionId: event.session_id,\n        model: event.model,\n        cwd: event.cwd,\n        permissionMode: event.permissionMode,\n        availableTools: event.tools ?? [],\n        mcpServers: (event.mcp_servers ?? []).map((s) => ({\n          name: s.name,\n          status: s.status,\n        })),\n      };\n      this.sessionStartTs = Date.now();\n      return;\n    }\n\n    if (event.type === \"system\" && event.subtype === \"api_retry\") {\n      this.retries.push({\n        offsetMs: this.sessionStartTs ? Date.now() - this.sessionStartTs : 0,\n        raw: event,\n      });\n      return;\n    }\n\n    if (isAssistantMessage(event)) {\n      this.handleAssistantMessage(event);\n      return;\n    }\n\n    if (isUserMessage(event)) {\n      this.handleUserMessage(event);\n      return;\n    }\n\n    if (isResult(event)) {\n      this.sawResultEvent = true;\n      this.resultIsError = event.is_error;\n      this.finalUsage = event.usage ?? null;\n      this.finalCostUsd = event.total_cost_usd ?? 0;\n      this.finalDurationMs = event.duration_ms ?? 0;\n      this.finalNumTurns = event.num_turns ?? 0;\n      this.finalResultText = event.result ?? \"\";\n      return;\n    }\n\n    // Unknown event: ignored. See class doc.\n  }\n\n  /**\n   * Finalize the view. Call after consuming the last event from the stream.\n   *\n   * Throws if no `system/init` was observed — at that point we have no model,\n   * no session id, and no available-tools list, which means assertions like\n   * \"called any mcp__api__* tool\" can't even be evaluated meaningfully.\n   */\n  build(): TrajectoryView {\n    if (this.meta === null) {\n      throw new Error(\n        \"TrajectoryBuilder.build() called before any system/init event was observed. \" +\n          \"The harness may have failed to start, or the stream was truncated before init.\",\n      );\n    }\n\n    const lastTurn = this.turns[this.turns.length - 1];\n\n    // Prefer the assistant text we accumulated turn-by-turn over the\n    // `result.result` field, because the latter is sometimes a summary\n    // and the former is exactly what the model said.\n    const accumulatedText = this.turns\n      .map((t) => t.text)\n      .filter((t) => t.length > 0)\n      .join(\"\\n\\n\")\n      .trim();\n\n    return {\n      meta: this.meta,\n      toolCalls: this.allToolCalls,\n      turns: this.turns,\n      finalResponse: accumulatedText || this.finalResultText,\n      finalStopReason: lastTurn?.stopReason ?? null,\n      usage: {\n        inputTokens: this.finalUsage?.input_tokens ?? 0,\n        outputTokens: this.finalUsage?.output_tokens ?? 0,\n        totalCostUsd: this.finalCostUsd,\n        durationMs: this.finalDurationMs,\n        // Fall back to observed turn count if the result event was missing.\n        numTurns: this.finalNumTurns || this.turns.length,\n      },\n      retries: this.retries,\n      // Successful = saw a non-error result envelope. Streams that ended without\n      // a result event are reported as unsuccessful regardless of tool outcomes.\n      success: this.sawResultEvent && !this.resultIsError,\n    };\n  }\n\n  // private handlers\n\n  private handleAssistantMessage(\n    event: Extract<StreamEvent, { type: \"assistant\" }>,\n  ): void {\n    const turnIndex = this.turns.length;\n    const textChunks: string[] = [];\n    const toolCallsThisTurn: ToolCall[] = [];\n\n    for (const block of event.message.content) {\n      if (isTextBlock(block)) {\n        textChunks.push(block.text);\n        continue;\n      }\n      if (isToolUseBlock(block)) {\n        const call: ToolCall = {\n          name: block.name,\n          namespace: namespaceOf(block.name),\n          callId: block.id,\n          args: block.input,\n          result: null,\n          isError: false,\n          turnIndex,\n          callIndex: this.allToolCalls.length,\n        };\n        this.allToolCalls.push(call);\n        this.pendingCalls.set(block.id, call);\n        toolCallsThisTurn.push(call);\n        continue;\n      }\n      // tool_result blocks don't appear in assistant messages — those arrive\n      // via user messages. If one does appear, ignore it; we'd rather drop\n      // an unexpected block than crash the eval.\n    }\n\n    this.turns.push({\n      turnIndex,\n      text: textChunks.join(\"\").trim(),\n      toolCalls: toolCallsThisTurn,\n      stopReason: event.message.stop_reason ?? null,\n    });\n  }\n\n  private handleUserMessage(\n    event: Extract<StreamEvent, { type: \"user\" }>,\n  ): void {\n    const content = event.message.content;\n\n    // The very first user message carries the prompt as a plain string. We\n    // already know the prompt (the caller passed it to the adapter), so we\n    // ignore this case — there's nothing assertion-relevant in it.\n    if (typeof content === \"string\") return;\n\n    for (const block of content) {\n      if (!isToolResultBlock(block)) continue;\n\n      const call = this.pendingCalls.get(block.tool_use_id);\n      if (!call) {\n        // Unmatched result: ignore. Can happen if events arrive out of order\n        // or the corresponding tool_use was emitted in an earlier run that\n        // we're resuming. Either way, dropping is safer than throwing.\n        continue;\n      }\n\n      call.result = block.content;\n      call.isError = block.is_error ?? false;\n      this.pendingCalls.delete(block.tool_use_id);\n    }\n  }\n}\n\n/**\n * Convenience: drain an async iterable of events through a fresh builder.\n *\n * Suitable when you have the full event stream and just want the view.\n * For interactive/incremental scenarios (e.g. surfacing partial state in a UI)\n * instantiate {@link TrajectoryBuilder} directly and call `consume()` /\n * `build()` yourself.\n */\nexport async function buildTrajectory(\n  events: AsyncIterable<StreamEvent>,\n): Promise<TrajectoryView> {\n  const builder = new TrajectoryBuilder();\n  for await (const event of events) {\n    builder.consume(event);\n  }\n  return builder.build();\n}\n","/**\n * Line-buffered NDJSON parser for Claude Code's `--output-format stream-json`.\n *\n * Claude Code emits one JSON object per line on stdout. The parser:\n *   - buffers across chunk boundaries (a single JSON line may arrive in two reads)\n *   - skips empty lines (defensive — shouldn't occur, but harmless if it does)\n *   - emits a discriminated `ParseResult` per line so callers can decide whether\n *     a malformed line should abort the run or just be logged.\n *\n * Why a generator (and not a Transform stream)?\n *   The eval adapter consumes events sequentially and synchronously updates a\n *   builder. Async iteration is the simplest interface for that pattern and\n *   composes cleanly with `for await` in the adapter. A Transform would force\n *   the builder into event-handler style.\n */\n\nimport type { Readable } from \"node:stream\";\nimport type { StreamEvent } from \"../types/stream\";\n\n/**\n * Result of attempting to parse a single line.\n *\n * Successful parses yield `{ ok: true }` with the typed event and the raw line\n * (kept for diagnostics and OTel `events.attributes.raw`). Failed parses yield\n * `{ ok: false }` with the parse error and the raw line — callers can log,\n * skip, or fail the run as they see fit.\n */\nexport type ParseResult =\n  | { ok: true; event: StreamEvent; rawLine: string }\n  | { ok: false; error: Error; rawLine: string };\n\n/**\n * Parse a readable stream of NDJSON into a sequence of typed stream-json events.\n *\n * @example\n *   const child = spawn(\"claude\", [\"-p\", prompt, \"--output-format\", \"stream-json\", \"--verbose\"]);\n *   for await (const result of parseStreamJson(child.stdout)) {\n *     if (result.ok) builder.consume(result.event);\n *     else console.warn(\"malformed stream line:\", result.rawLine, result.error);\n *   }\n */\nexport async function* parseStreamJson(\n  stream: Readable,\n): AsyncGenerator<ParseResult, void, void> {\n  let buffer = \"\";\n  // The Node child_process stdout is a binary stream by default. Setting the\n  // encoding here means `for await (const chunk of stream)` yields strings.\n  stream.setEncoding(\"utf8\");\n\n  for await (const chunk of stream) {\n    buffer += chunk as string;\n\n    // Drain every complete line currently in the buffer before reading more.\n    // Multiple JSON objects can arrive in one chunk (e.g. when the harness\n    // emits a burst of events at session start).\n    let newlineIdx: number;\n    while ((newlineIdx = buffer.indexOf(\"\\n\")) !== -1) {\n      const line = buffer.slice(0, newlineIdx).trim();\n      buffer = buffer.slice(newlineIdx + 1);\n      if (line.length === 0) continue;\n      yield tryParseLine(line);\n    }\n  }\n\n  // Flush any trailing content that arrived without a final newline. Stream-json\n  // typically ends with a newline-terminated `result` event, but a killed\n  // process may not flush, so we still try to emit what we have.\n  const trailing = buffer.trim();\n  if (trailing.length > 0) {\n    yield tryParseLine(trailing);\n  }\n}\n\n/**\n * Parse a single line. Extracted as a helper so the generator stays readable.\n *\n * Note: we do not validate the event structure beyond `JSON.parse`. Runtime\n * validation (e.g. zod) is overkill here — the schema is stable enough at\n * runtime, and the TrajectoryBuilder is tolerant of missing fields. Adding\n * validation would be premature.\n */\nfunction tryParseLine(line: string): ParseResult {\n  try {\n    const event = JSON.parse(line) as StreamEvent;\n    return { ok: true, event, rawLine: line };\n  } catch (err) {\n    return {\n      ok: false,\n      error: err instanceof Error ? err : new Error(String(err)),\n      rawLine: line,\n    };\n  }\n}\n","/**\n * Generic harness adapter contract.\n *\n * Every harness adapter produces a {@link TrajectoryView} plus process\n * diagnostics. The runner and assertion engine depend only on these types —\n * not on any specific harness implementation.\n */\n\nimport type { TrajectoryView } from \"../types/trajectory\";\n\n/** Base config every adapter must accept. */\nexport interface BaseAdapterConfig {\n  prompt: string;\n  model?: string;\n  timeoutMs?: number;\n  signal?: AbortSignal;\n  env?: Record<string, string>;\n  cwd?: string;\n}\n\n/** Suite-level config: generic fields plus adapter-specific nested blocks. */\nexport type SuiteConfig = Partial<BaseAdapterConfig> & {\n  /** Claude Code adapter options (when `adapter` is `claude-code`). */\n  claudeCode?: Record<string, unknown>;\n};\n\n/** Generic harness adapter interface. */\nexport interface HarnessAdapter<\n  TConfig extends BaseAdapterConfig = BaseAdapterConfig,\n> {\n  readonly id: string;\n  run(config: TConfig): Promise<AdapterResult>;\n}\n\n/** Successful adapter run. */\nexport interface AdapterResult {\n  view: TrajectoryView;\n  diagnostics: AdapterDiagnostics;\n}\n\n/** Process-level diagnostics from any adapter. */\nexport interface AdapterDiagnostics {\n  exitCode: number | null;\n  signal: NodeJS.Signals | null;\n  stderr: string;\n  parseErrors: ParseErrorRecord[];\n  timedOut: boolean;\n  durationMs: number;\n}\n\nexport interface ParseErrorRecord {\n  line: string;\n  error: string;\n}\n\n/**\n * Thrown when the harness fails to produce a usable trajectory.\n *\n * Most commonly this means the process failed before emitting a usable\n * session init event. Inspect `diagnostics.stderr` for the cause.\n */\nexport class AdapterError extends Error {\n  constructor(\n    message: string,\n    public readonly diagnostics: Partial<AdapterDiagnostics>,\n  ) {\n    super(message);\n    this.name = \"AdapterError\";\n  }\n}\n","/**\n * Build CLI args for Claude Code judge subprocesses (JSON output, not stream-json).\n */\n\nimport type { ClaudeCodeAdapterConfig, ClaudeCodeOptions } from \"./types\";\n\nfunction pushRepeatableFlag(args: string[], flag: string, values?: string[]): void {\n  if (!values) return;\n  for (const value of values) {\n    args.push(flag, value);\n  }\n}\n\nfunction pushOptionalFlag(\n  args: string[],\n  flag: string,\n  value: string | number | boolean | undefined,\n): void {\n  if (value === undefined) return;\n  if (typeof value === \"boolean\") {\n    if (value) args.push(flag);\n    return;\n  }\n  args.push(flag, String(value));\n}\n\n/** Append Claude Code CLI flags shared by harness runs and grading judges. */\nexport function appendClaudeCodeFlags(\n  args: string[],\n  config: ClaudeCodeOptions & { model?: string },\n): void {\n  pushRepeatableFlag(args, \"--plugin-dir\", config.pluginDirs);\n  pushRepeatableFlag(args, \"--plugin-url\", config.pluginUrls);\n  pushRepeatableFlag(args, \"--add-dir\", config.addDirs);\n\n  pushOptionalFlag(args, \"--mcp-config\", config.mcpConfig);\n  pushOptionalFlag(args, \"--model\", config.model);\n  pushOptionalFlag(args, \"--permission-mode\", config.permissionMode);\n  pushOptionalFlag(args, \"--effort\", config.effort);\n  pushOptionalFlag(args, \"--agent\", config.agent);\n  pushOptionalFlag(args, \"--fallback-model\", config.fallbackModel);\n  pushOptionalFlag(args, \"--tools\", config.tools);\n  pushOptionalFlag(args, \"--settings\", config.settings);\n  pushOptionalFlag(args, \"--setting-sources\", config.settingSources);\n  pushOptionalFlag(args, \"--max-turns\", config.maxTurns);\n  pushOptionalFlag(args, \"--max-budget-usd\", config.maxBudgetUsd);\n  pushOptionalFlag(args, \"--system-prompt\", config.systemPrompt);\n  pushOptionalFlag(args, \"--system-prompt-file\", config.systemPromptFile);\n  pushOptionalFlag(args, \"--append-system-prompt\", config.appendSystemPrompt);\n  pushOptionalFlag(\n    args,\n    \"--append-system-prompt-file\",\n    config.appendSystemPromptFile,\n  );\n  pushOptionalFlag(args, \"--debug\", config.debug);\n  pushOptionalFlag(args, \"--debug-file\", config.debugFile);\n\n  if (config.allowedTools && config.allowedTools.length > 0) {\n    args.push(\"--allowedTools\", config.allowedTools.join(\",\"));\n  }\n\n  if (config.disallowedTools && config.disallowedTools.length > 0) {\n    args.push(\"--disallowedTools\", config.disallowedTools.join(\",\"));\n  }\n\n  pushOptionalFlag(args, \"--strict-mcp-config\", config.strictMcpConfig);\n  pushOptionalFlag(args, \"--include-hook-events\", config.includeHookEvents);\n  pushOptionalFlag(args, \"--no-session-persistence\", config.noSessionPersistence);\n  pushOptionalFlag(args, \"--disable-slash-commands\", config.disableSlashCommands);\n  pushOptionalFlag(args, \"--bare\", config.bare);\n  pushOptionalFlag(args, \"--safe-mode\", config.safeMode);\n  pushOptionalFlag(\n    args,\n    \"--allow-dangerously-skip-permissions\",\n    config.allowDangerouslySkipPermissions,\n  );\n  pushOptionalFlag(\n    args,\n    \"--dangerously-skip-permissions\",\n    config.dangerouslySkipPermissions,\n  );\n}\n\n/**\n * Build the argument vector for spawning `claude`.\n *\n * Order matters only for flags that take values — value flags must come\n * after their flag name. Everything else is order-independent.\n */\nexport function buildArgs(config: ClaudeCodeAdapterConfig): string[] {\n  const args: string[] = [\n    \"-p\",\n    config.prompt,\n    \"--output-format\",\n    \"stream-json\",\n    \"--verbose\",\n  ];\n\n  appendClaudeCodeFlags(args, config);\n\n  return args;\n}\n\n/** Build args for an LLM judge subprocess (`--output-format json`). */\nexport function buildJudgeArgs(\n  prompt: string,\n  config: ClaudeCodeOptions & { model?: string } = {},\n): string[] {\n  const args: string[] = [\"-p\", prompt, \"--output-format\", \"json\"];\n  const permissionMode = config.permissionMode ?? \"bypassPermissions\";\n  appendClaudeCodeFlags(args, {\n    ...config,\n    permissionMode,\n  });\n  return args;\n}\n","/**\n * Process management for the Claude Code adapter.\n *\n * This module owns spawning, timeout, abort signal handling, and process-tree\n * teardown. The orchestrator (`index.ts`) consumes the returned handle —\n * reading stdout and waiting for completion — but doesn't worry about how\n * the process gets killed or how its config gets isolated.\n *\n * Why a separate module? Process management is the one part of the adapter\n * with real I/O complexity (process groups, signal escalation, temp-dir\n * lifecycle, env merging). Isolating it makes the orchestrator easy to read\n * and lets us swap the spawning logic if we later need to, e.g., wrap claude\n * in a sandbox runner.\n */\n\nimport { spawn, type ChildProcess } from \"node:child_process\";\nimport { mkdtemp, rm } from \"node:fs/promises\";\nimport { tmpdir } from \"node:os\";\nimport { join } from \"node:path\";\nimport type { Readable } from \"node:stream\";\n\nimport { buildArgs } from \"./flags\";\nimport type { ClaudeCodeAdapterConfig } from \"./types\";\n\n/** Default hard timeout per run. Tunable via config.timeoutMs. */\nconst DEFAULT_TIMEOUT_MS = 5 * 60 * 1000;\n\n/**\n * Grace period between SIGTERM and SIGKILL. Most processes shut down cleanly\n * within a few seconds; this gives them that chance while preventing CI from\n * hanging indefinitely on a stuck child.\n */\nconst KILL_GRACE_MS = 5_000;\n\n/**\n * Handle to a spawned `claude` process. The orchestrator drives it:\n *   - Read `stdout` (typically via parseStreamJson).\n *   - Await `done` to learn the exit state.\n *   - Await `stderrCollected` for diagnostic stderr.\n *   - Check `timedOut()` after exit to distinguish kill-by-timeout from\n *     normal termination.\n *   - Call `cleanup()` after all of the above to remove the temp config dir.\n */\nexport interface SpawnedClaude {\n  stdout: Readable;\n  done: Promise<{ exitCode: number | null; signal: NodeJS.Signals | null }>;\n  stderrCollected: Promise<string>;\n  timedOut: () => boolean;\n  cleanup: () => Promise<void>;\n}\n\n/**\n * Spawn `claude` in headless mode with isolated config and a process-group\n * lifecycle. See {@link SpawnedClaude} for how to consume the result.\n */\nexport async function spawnClaude(\n  config: ClaudeCodeAdapterConfig,\n): Promise<SpawnedClaude> {\n  const binary = config.binary ?? \"claude\";\n  const args = buildArgs(config);\n\n  const isolateConfig = config.isolateConfig !== false;\n\n  // Isolated runs use a fresh temp dir so plugins/settings don't leak between\n  // reps. Non-isolated runs inherit the caller's Claude login and plugins.\n  const tempConfigDir = isolateConfig\n    ? await mkdtemp(join(tmpdir(), \"harness-eval-\"))\n    : null;\n\n  const env: Record<string, string | undefined> = {\n    ...process.env,\n    ...config.env,\n  };\n  if (tempConfigDir) {\n    // Override after ...env so callers can't accidentally un-isolate.\n    env.CLAUDE_CONFIG_DIR = tempConfigDir;\n  }\n\n  const child = spawn(binary, args, {\n    cwd: config.cwd ?? process.cwd(),\n    env,\n    stdio: [\"ignore\", \"pipe\", \"pipe\"],\n    // detached: true means the child becomes the leader of its own process\n    // group. We exploit this to kill the entire group (including any MCP\n    // server subprocesses and tool processes) on timeout/abort.\n    detached: true,\n  });\n\n\n  let timedOut = false;\n  let killEscalation: NodeJS.Timeout | null = null;\n  const timeoutMs = config.timeoutMs ?? DEFAULT_TIMEOUT_MS;\n\n  const scheduleKillEscalation = () => {\n    if (killEscalation) clearTimeout(killEscalation);\n    killEscalation = setTimeout(\n      () => killTree(child, \"SIGKILL\"),\n      KILL_GRACE_MS,\n    );\n  };\n\n  const timeoutTimer = setTimeout(() => {\n    timedOut = true;\n    killTree(child, \"SIGTERM\");\n    scheduleKillEscalation();\n  }, timeoutMs);\n\n  // External cancellation uses the same kill path.\n  const onAbort = () => {\n    killTree(child, \"SIGTERM\");\n    scheduleKillEscalation();\n  };\n  config.signal?.addEventListener(\"abort\", onAbort, { once: true });\n\n\n  // Drain stderr eagerly so the OS-level buffer never fills and stalls the\n  // child (Node child processes will block on a full pipe).\n  const stderrChunks: string[] = [];\n  child.stderr?.setEncoding(\"utf8\");\n  child.stderr?.on(\"data\", (chunk: string) => {\n    stderrChunks.push(chunk);\n  });\n\n  const stderrCollected = new Promise<string>((resolve) => {\n    const finalize = () => resolve(stderrChunks.join(\"\"));\n    child.stderr?.on(\"end\", finalize);\n    // Errors during stderr capture shouldn't fail the whole run; we just\n    // return what we've buffered so far.\n    child.stderr?.on(\"error\", finalize);\n  });\n\n\n  const done = new Promise<{\n    exitCode: number | null;\n    signal: NodeJS.Signals | null;\n  }>((resolve) => {\n    let settled = false;\n    const finalize = (\n      exitCode: number | null,\n      signal: NodeJS.Signals | null,\n    ) => {\n      if (settled) return;\n      settled = true;\n      clearTimeout(timeoutTimer);\n      if (killEscalation) clearTimeout(killEscalation);\n      config.signal?.removeEventListener(\"abort\", onAbort);\n      resolve({ exitCode, signal });\n    };\n\n    child.on(\"close\", (code, signal) => finalize(code, signal));\n    // ENOENT and other spawn failures emit `error` — `close` may not follow.\n    child.on(\"error\", () => finalize(null, null));\n  });\n\n\n  const cleanup = async () => {\n    if (!tempConfigDir) return;\n    try {\n      await rm(tempConfigDir, { recursive: true, force: true });\n    } catch {\n      // Best-effort. A leftover temp dir is annoying but not catastrophic;\n      // we don't want to fail the run for it.\n    }\n  };\n\n  // stdout is guaranteed non-null because we passed `stdio: [..., \"pipe\", ...]`.\n  // The `!` is safe; the alternative would be a redundant runtime check that\n  // could never fire.\n  return {\n    stdout: child.stdout!,\n    done,\n    stderrCollected,\n    timedOut: () => timedOut,\n    cleanup,\n  };\n}\n\n/**\n * Kill the child's process group, then fall back to the bare PID if the\n * group is already gone. This catches MCP server subprocesses and tool\n * processes spawned by claude.\n *\n * Why both? On some platforms the process group dies before we get here\n * (the child itself already cleaned up), in which case `kill(-pid)` throws\n * ESRCH. The fallback handles that edge case without leaking zombies in\n * the common case.\n */\nfunction killTree(child: ChildProcess, signal: NodeJS.Signals): void {\n  if (child.pid === undefined) return;\n  try {\n    // Negative PID = process group. Requires the child was spawned with\n    // detached=true (which we do).\n    process.kill(-child.pid, signal);\n  } catch {\n    try {\n      child.kill(signal);\n    } catch {\n      // Nothing left to kill.\n    }\n  }\n}\n","/**\n * Claude Code adapter — public API.\n */\n\nimport { parseStreamJson } from \"../../parsers/stream-json\";\nimport { TrajectoryBuilder } from \"../../trajectory/builder\";\nimport type { StreamEvent } from \"../../types/stream\";\n\nimport { AdapterError } from \"../types\";\nimport { spawnClaude } from \"./process\";\nimport type {\n  AdapterDiagnostics,\n  ClaudeCodeAdapterConfig,\n  ClaudeCodeAdapterResult,\n  ParseErrorRecord,\n} from \"./types\";\nimport type { HarnessAdapter } from \"../types\";\n\nexport { AdapterError } from \"../types\";\nexport type {\n  AdapterDiagnostics,\n  AdapterResult,\n  ClaudeCodeAdapterConfig,\n  ClaudeCodeAdapterResult,\n  ClaudeCodeOptions,\n  ParseErrorRecord,\n  PermissionMode,\n} from \"./types\";\n\n/**\n * Run Claude Code in headless mode and return a trajectory.\n */\nexport async function runClaudeCode(\n  config: ClaudeCodeAdapterConfig,\n): Promise<ClaudeCodeAdapterResult> {\n  const startTs = Date.now();\n  const spawned = await spawnClaude(config);\n\n  const builder = new TrajectoryBuilder();\n  const rawEvents: StreamEvent[] = [];\n  const parseErrors: ParseErrorRecord[] = [];\n\n  try {\n    for await (const result of parseStreamJson(spawned.stdout)) {\n      if (result.ok) {\n        builder.consume(result.event);\n        rawEvents.push(result.event);\n      } else {\n        parseErrors.push({\n          line: result.rawLine,\n          error: result.error.message,\n        });\n      }\n    }\n\n    const [{ exitCode, signal }, stderr] = await Promise.all([\n      spawned.done,\n      spawned.stderrCollected,\n    ]);\n\n    const diagnostics: AdapterDiagnostics = {\n      exitCode,\n      signal,\n      stderr,\n      parseErrors,\n      timedOut: spawned.timedOut(),\n      durationMs: Date.now() - startTs,\n    };\n\n    let view;\n    try {\n      view = builder.build();\n    } catch (err) {\n      const message = err instanceof Error ? err.message : String(err);\n      throw new AdapterError(\n        `harness produced no usable trajectory: ${message}`,\n        diagnostics,\n      );\n    }\n\n    return { view, diagnostics, rawEvents };\n  } finally {\n    await spawned.cleanup();\n  }\n}\n\nexport const claudeCodeAdapter: HarnessAdapter<ClaudeCodeAdapterConfig> = {\n  id: \"claude-code\",\n  run: runClaudeCode,\n};\n"],"mappings":";;;;;;;AAuMA,SAAgB,aAAa,GAAsC;CACjE,OAAO,EAAE,SAAS,YAAa,EAAsB,YAAY;AACnE;AAEA,SAAgB,cAAc,GAAuC;CACnE,OAAO,EAAE,SAAS,YAAa,EAAuB,YAAY;AACpE;AAEA,SAAgB,mBAAmB,GAA4C;CAC7E,OAAO,EAAE,SAAS;AACpB;AAEA,SAAgB,cAAc,GAAuC;CACnE,OAAO,EAAE,SAAS;AACpB;AAEA,SAAgB,SAAS,GAAkC;CACzD,OAAO,EAAE,SAAS;AACpB;AAEA,SAAgB,YAAY,GAAiC;CAC3D,OAAO,EAAE,SAAS;AACpB;AAEA,SAAgB,eAAe,GAAoC;CACjE,OAAO,EAAE,SAAS;AACpB;AAEA,SAAgB,kBAAkB,GAAuC;CACvE,OAAO,EAAE,SAAS;AACpB;;;;;;;;;;;;;;AC9FA,SAAgB,YAAY,UAAiC;CAC3D,IAAI,CAAC,SAAS,WAAW,OAAO,GAAG,OAAO;CAC1C,MAAM,QAAQ,SAAS,MAAM,IAAI;CACjC,IAAI,MAAM,SAAS,GAAG,OAAO;CAC7B,OAAO,GAAG,MAAM,GAAG,IAAI,MAAM;AAC/B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC7FA,IAAa,oBAAb,MAA+B;CAC7B,OAAmC;CACnC,iBAAwC;CAExC,QAAiC,CAAC;CAClC,eAAmC,CAAC;;;;;CAMpC,+BAA8C,IAAI,IAAI;CAEtD,UAAiC,CAAC;CAElC,aAAmC;CACnC,eAAuB;CACvB,kBAA0B;CAC1B,gBAAwB;CACxB,kBAA0B;CAC1B,iBAAyB;CACzB,gBAAwB;;;;;;;CAQxB,QAAQ,OAA0B;EAChC,IAAI,aAAa,KAAK,GAAG;GACvB,KAAK,OAAO;IACV,WAAW,MAAM;IACjB,OAAO,MAAM;IACb,KAAK,MAAM;IACX,gBAAgB,MAAM;IACtB,gBAAgB,MAAM,SAAS,CAAC;IAChC,aAAa,MAAM,eAAe,CAAC,EAAA,CAAG,KAAK,OAAO;KAChD,MAAM,EAAE;KACR,QAAQ,EAAE;IACZ,EAAE;GACJ;GACA,KAAK,iBAAiB,KAAK,IAAI;GAC/B;EACF;EAEA,IAAI,MAAM,SAAS,YAAY,MAAM,YAAY,aAAa;GAC5D,KAAK,QAAQ,KAAK;IAChB,UAAU,KAAK,iBAAiB,KAAK,IAAI,IAAI,KAAK,iBAAiB;IACnE,KAAK;GACP,CAAC;GACD;EACF;EAEA,IAAI,mBAAmB,KAAK,GAAG;GAC7B,KAAK,uBAAuB,KAAK;GACjC;EACF;EAEA,IAAI,cAAc,KAAK,GAAG;GACxB,KAAK,kBAAkB,KAAK;GAC5B;EACF;EAEA,IAAI,SAAS,KAAK,GAAG;GACnB,KAAK,iBAAiB;GACtB,KAAK,gBAAgB,MAAM;GAC3B,KAAK,aAAa,MAAM,SAAS;GACjC,KAAK,eAAe,MAAM,kBAAkB;GAC5C,KAAK,kBAAkB,MAAM,eAAe;GAC5C,KAAK,gBAAgB,MAAM,aAAa;GACxC,KAAK,kBAAkB,MAAM,UAAU;GACvC;EACF;CAGF;;;;;;;;CASA,QAAwB;EACtB,IAAI,KAAK,SAAS,MAChB,MAAM,IAAI,MACR,4JAEF;EAGF,MAAM,WAAW,KAAK,MAAM,KAAK,MAAM,SAAS;EAKhD,MAAM,kBAAkB,KAAK,MAC1B,KAAK,MAAM,EAAE,IAAI,CAAC,CAClB,QAAQ,MAAM,EAAE,SAAS,CAAC,CAAC,CAC3B,KAAK,MAAM,CAAC,CACZ,KAAK;EAER,OAAO;GACL,MAAM,KAAK;GACX,WAAW,KAAK;GAChB,OAAO,KAAK;GACZ,eAAe,mBAAmB,KAAK;GACvC,iBAAiB,UAAU,cAAc;GACzC,OAAO;IACL,aAAa,KAAK,YAAY,gBAAgB;IAC9C,cAAc,KAAK,YAAY,iBAAiB;IAChD,cAAc,KAAK;IACnB,YAAY,KAAK;IAEjB,UAAU,KAAK,iBAAiB,KAAK,MAAM;GAC7C;GACA,SAAS,KAAK;GAGd,SAAS,KAAK,kBAAkB,CAAC,KAAK;EACxC;CACF;CAIA,uBACE,OACM;EACN,MAAM,YAAY,KAAK,MAAM;EAC7B,MAAM,aAAuB,CAAC;EAC9B,MAAM,oBAAgC,CAAC;EAEvC,KAAK,MAAM,SAAS,MAAM,QAAQ,SAAS;GACzC,IAAI,YAAY,KAAK,GAAG;IACtB,WAAW,KAAK,MAAM,IAAI;IAC1B;GACF;GACA,IAAI,eAAe,KAAK,GAAG;IACzB,MAAM,OAAiB;KACrB,MAAM,MAAM;KACZ,WAAW,YAAY,MAAM,IAAI;KACjC,QAAQ,MAAM;KACd,MAAM,MAAM;KACZ,QAAQ;KACR,SAAS;KACT;KACA,WAAW,KAAK,aAAa;IAC/B;IACA,KAAK,aAAa,KAAK,IAAI;IAC3B,KAAK,aAAa,IAAI,MAAM,IAAI,IAAI;IACpC,kBAAkB,KAAK,IAAI;IAC3B;GACF;EAIF;EAEA,KAAK,MAAM,KAAK;GACd;GACA,MAAM,WAAW,KAAK,EAAE,CAAC,CAAC,KAAK;GAC/B,WAAW;GACX,YAAY,MAAM,QAAQ,eAAe;EAC3C,CAAC;CACH;CAEA,kBACE,OACM;EACN,MAAM,UAAU,MAAM,QAAQ;EAK9B,IAAI,OAAO,YAAY,UAAU;EAEjC,KAAK,MAAM,SAAS,SAAS;GAC3B,IAAI,CAAC,kBAAkB,KAAK,GAAG;GAE/B,MAAM,OAAO,KAAK,aAAa,IAAI,MAAM,WAAW;GACpD,IAAI,CAAC,MAIH;GAGF,KAAK,SAAS,MAAM;GACpB,KAAK,UAAU,MAAM,YAAY;GACjC,KAAK,aAAa,OAAO,MAAM,WAAW;EAC5C;CACF;AACF;;;;;;;;;AAUA,eAAsB,gBACpB,QACyB;CACzB,MAAM,UAAU,IAAI,kBAAkB;CACtC,WAAW,MAAM,SAAS,QACxB,QAAQ,QAAQ,KAAK;CAEvB,OAAO,QAAQ,MAAM;AACvB;;;;;;;;;;;;;AC1NA,gBAAuB,gBACrB,QACyC;CACzC,IAAI,SAAS;CAGb,OAAO,YAAY,MAAM;CAEzB,WAAW,MAAM,SAAS,QAAQ;EAChC,UAAU;EAKV,IAAI;EACJ,QAAQ,aAAa,OAAO,QAAQ,IAAI,OAAO,IAAI;GACjD,MAAM,OAAO,OAAO,MAAM,GAAG,UAAU,CAAC,CAAC,KAAK;GAC9C,SAAS,OAAO,MAAM,aAAa,CAAC;GACpC,IAAI,KAAK,WAAW,GAAG;GACvB,MAAM,aAAa,IAAI;EACzB;CACF;CAKA,MAAM,WAAW,OAAO,KAAK;CAC7B,IAAI,SAAS,SAAS,GACpB,MAAM,aAAa,QAAQ;AAE/B;;;;;;;;;AAUA,SAAS,aAAa,MAA2B;CAC/C,IAAI;EAEF,OAAO;GAAE,IAAI;GAAM,OADL,KAAK,MAAM,IACF;GAAG,SAAS;EAAK;CAC1C,SAAS,KAAK;EACZ,OAAO;GACL,IAAI;GACJ,OAAO,eAAe,QAAQ,MAAM,IAAI,MAAM,OAAO,GAAG,CAAC;GACzD,SAAS;EACX;CACF;AACF;;;;;;;;;AC/BA,IAAa,eAAb,cAAkC,MAAM;CAGpB;CAFlB,YACE,SACA,aACA;EACA,MAAM,OAAO;EAFG,KAAA,cAAA;EAGhB,KAAK,OAAO;CACd;AACF;;;AC/DA,SAAS,mBAAmB,MAAgB,MAAc,QAAyB;CACjF,IAAI,CAAC,QAAQ;CACb,KAAK,MAAM,SAAS,QAClB,KAAK,KAAK,MAAM,KAAK;AAEzB;AAEA,SAAS,iBACP,MACA,MACA,OACM;CACN,IAAI,UAAU,KAAA,GAAW;CACzB,IAAI,OAAO,UAAU,WAAW;EAC9B,IAAI,OAAO,KAAK,KAAK,IAAI;EACzB;CACF;CACA,KAAK,KAAK,MAAM,OAAO,KAAK,CAAC;AAC/B;;AAGA,SAAgB,sBACd,MACA,QACM;CACN,mBAAmB,MAAM,gBAAgB,OAAO,UAAU;CAC1D,mBAAmB,MAAM,gBAAgB,OAAO,UAAU;CAC1D,mBAAmB,MAAM,aAAa,OAAO,OAAO;CAEpD,iBAAiB,MAAM,gBAAgB,OAAO,SAAS;CACvD,iBAAiB,MAAM,WAAW,OAAO,KAAK;CAC9C,iBAAiB,MAAM,qBAAqB,OAAO,cAAc;CACjE,iBAAiB,MAAM,YAAY,OAAO,MAAM;CAChD,iBAAiB,MAAM,WAAW,OAAO,KAAK;CAC9C,iBAAiB,MAAM,oBAAoB,OAAO,aAAa;CAC/D,iBAAiB,MAAM,WAAW,OAAO,KAAK;CAC9C,iBAAiB,MAAM,cAAc,OAAO,QAAQ;CACpD,iBAAiB,MAAM,qBAAqB,OAAO,cAAc;CACjE,iBAAiB,MAAM,eAAe,OAAO,QAAQ;CACrD,iBAAiB,MAAM,oBAAoB,OAAO,YAAY;CAC9D,iBAAiB,MAAM,mBAAmB,OAAO,YAAY;CAC7D,iBAAiB,MAAM,wBAAwB,OAAO,gBAAgB;CACtE,iBAAiB,MAAM,0BAA0B,OAAO,kBAAkB;CAC1E,iBACE,MACA,+BACA,OAAO,sBACT;CACA,iBAAiB,MAAM,WAAW,OAAO,KAAK;CAC9C,iBAAiB,MAAM,gBAAgB,OAAO,SAAS;CAEvD,IAAI,OAAO,gBAAgB,OAAO,aAAa,SAAS,GACtD,KAAK,KAAK,kBAAkB,OAAO,aAAa,KAAK,GAAG,CAAC;CAG3D,IAAI,OAAO,mBAAmB,OAAO,gBAAgB,SAAS,GAC5D,KAAK,KAAK,qBAAqB,OAAO,gBAAgB,KAAK,GAAG,CAAC;CAGjE,iBAAiB,MAAM,uBAAuB,OAAO,eAAe;CACpE,iBAAiB,MAAM,yBAAyB,OAAO,iBAAiB;CACxE,iBAAiB,MAAM,4BAA4B,OAAO,oBAAoB;CAC9E,iBAAiB,MAAM,4BAA4B,OAAO,oBAAoB;CAC9E,iBAAiB,MAAM,UAAU,OAAO,IAAI;CAC5C,iBAAiB,MAAM,eAAe,OAAO,QAAQ;CACrD,iBACE,MACA,wCACA,OAAO,+BACT;CACA,iBACE,MACA,kCACA,OAAO,0BACT;AACF;;;;;;;AAQA,SAAgB,UAAU,QAA2C;CACnE,MAAM,OAAiB;EACrB;EACA,OAAO;EACP;EACA;EACA;CACF;CAEA,sBAAsB,MAAM,MAAM;CAElC,OAAO;AACT;;AAGA,SAAgB,eACd,QACA,SAAiD,CAAC,GACxC;CACV,MAAM,OAAiB;EAAC;EAAM;EAAQ;EAAmB;CAAM;CAC/D,MAAM,iBAAiB,OAAO,kBAAkB;CAChD,sBAAsB,MAAM;EAC1B,GAAG;EACH;CACF,CAAC;CACD,OAAO;AACT;;;;;;;;;;;;;;;;;;AC1FA,MAAM,qBAAqB,MAAS;;;;;;AAOpC,MAAM,gBAAgB;;;;;AAuBtB,eAAsB,YACpB,QACwB;CACxB,MAAM,SAAS,OAAO,UAAU;CAChC,MAAM,OAAO,UAAU,MAAM;CAM7B,MAAM,gBAJgB,OAAO,kBAAkB,QAK3C,MAAM,QAAQ,KAAK,OAAO,GAAG,eAAe,CAAC,IAC7C;CAEJ,MAAM,MAA0C;EAC9C,GAAG,QAAQ;EACX,GAAG,OAAO;CACZ;CACA,IAAI,eAEF,IAAI,oBAAoB;CAG1B,MAAM,QAAQ,MAAM,QAAQ,MAAM;EAChC,KAAK,OAAO,OAAO,QAAQ,IAAI;EAC/B;EACA,OAAO;GAAC;GAAU;GAAQ;EAAM;EAIhC,UAAU;CACZ,CAAC;CAGD,IAAI,WAAW;CACf,IAAI,iBAAwC;CAC5C,MAAM,YAAY,OAAO,aAAa;CAEtC,MAAM,+BAA+B;EACnC,IAAI,gBAAgB,aAAa,cAAc;EAC/C,iBAAiB,iBACT,SAAS,OAAO,SAAS,GAC/B,aACF;CACF;CAEA,MAAM,eAAe,iBAAiB;EACpC,WAAW;EACX,SAAS,OAAO,SAAS;EACzB,uBAAuB;CACzB,GAAG,SAAS;CAGZ,MAAM,gBAAgB;EACpB,SAAS,OAAO,SAAS;EACzB,uBAAuB;CACzB;CACA,OAAO,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;CAKhE,MAAM,eAAyB,CAAC;CAChC,MAAM,QAAQ,YAAY,MAAM;CAChC,MAAM,QAAQ,GAAG,SAAS,UAAkB;EAC1C,aAAa,KAAK,KAAK;CACzB,CAAC;CAED,MAAM,kBAAkB,IAAI,SAAiB,YAAY;EACvD,MAAM,iBAAiB,QAAQ,aAAa,KAAK,EAAE,CAAC;EACpD,MAAM,QAAQ,GAAG,OAAO,QAAQ;EAGhC,MAAM,QAAQ,GAAG,SAAS,QAAQ;CACpC,CAAC;CAGD,MAAM,OAAO,IAAI,SAGb,YAAY;EACd,IAAI,UAAU;EACd,MAAM,YACJ,UACA,WACG;GACH,IAAI,SAAS;GACb,UAAU;GACV,aAAa,YAAY;GACzB,IAAI,gBAAgB,aAAa,cAAc;GAC/C,OAAO,QAAQ,oBAAoB,SAAS,OAAO;GACnD,QAAQ;IAAE;IAAU;GAAO,CAAC;EAC9B;EAEA,MAAM,GAAG,UAAU,MAAM,WAAW,SAAS,MAAM,MAAM,CAAC;EAE1D,MAAM,GAAG,eAAe,SAAS,MAAM,IAAI,CAAC;CAC9C,CAAC;CAGD,MAAM,UAAU,YAAY;EAC1B,IAAI,CAAC,eAAe;EACpB,IAAI;GACF,MAAM,GAAG,eAAe;IAAE,WAAW;IAAM,OAAO;GAAK,CAAC;EAC1D,QAAQ,CAGR;CACF;CAKA,OAAO;EACL,QAAQ,MAAM;EACd;EACA;EACA,gBAAgB;EAChB;CACF;AACF;;;;;;;;;;;AAYA,SAAS,SAAS,OAAqB,QAA8B;CACnE,IAAI,MAAM,QAAQ,KAAA,GAAW;CAC7B,IAAI;EAGF,QAAQ,KAAK,CAAC,MAAM,KAAK,MAAM;CACjC,QAAQ;EACN,IAAI;GACF,MAAM,KAAK,MAAM;EACnB,QAAQ,CAER;CACF;AACF;;;;;;;;;;;;;;ACxKA,eAAsB,cACpB,QACkC;CAClC,MAAM,UAAU,KAAK,IAAI;CACzB,MAAM,UAAU,MAAM,YAAY,MAAM;CAExC,MAAM,UAAU,IAAI,kBAAkB;CACtC,MAAM,YAA2B,CAAC;CAClC,MAAM,cAAkC,CAAC;CAEzC,IAAI;EACF,WAAW,MAAM,UAAU,gBAAgB,QAAQ,MAAM,GACvD,IAAI,OAAO,IAAI;GACb,QAAQ,QAAQ,OAAO,KAAK;GAC5B,UAAU,KAAK,OAAO,KAAK;EAC7B,OACE,YAAY,KAAK;GACf,MAAM,OAAO;GACb,OAAO,OAAO,MAAM;EACtB,CAAC;EAIL,MAAM,CAAC,EAAE,UAAU,UAAU,UAAU,MAAM,QAAQ,IAAI,CACvD,QAAQ,MACR,QAAQ,eACV,CAAC;EAED,MAAM,cAAkC;GACtC;GACA;GACA;GACA;GACA,UAAU,QAAQ,SAAS;GAC3B,YAAY,KAAK,IAAI,IAAI;EAC3B;EAEA,IAAI;EACJ,IAAI;GACF,OAAO,QAAQ,MAAM;EACvB,SAAS,KAAK;GAEZ,MAAM,IAAI,aACR,0CAFc,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,KAG7D,WACF;EACF;EAEA,OAAO;GAAE;GAAM;GAAa;EAAU;CACxC,UAAU;EACR,MAAM,QAAQ,QAAQ;CACxB;AACF;AAEA,MAAa,oBAA6D;CACxE,IAAI;CACJ,KAAK;AACP"}
+{"version":3,"file":"claude-code-DZ4Vkgp6.js","names":[],"sources":["../src/types/stream.ts","../src/types/trajectory.ts","../src/trajectory/builder.ts","../src/parsers/stream-json.ts","../src/adapters/types.ts","../src/adapters/claude-code/flags.ts","../src/adapters/claude-code/process.ts","../src/adapters/claude-code/index.ts"],"sourcesContent":["/**\n * Discriminated union of events emitted by Claude Code's\n * `--output-format stream-json` mode.\n *\n * The format is NDJSON (one JSON object per line on stdout). Each line has\n * a required `type` field and often a `subtype` for further disambiguation.\n *\n * Source notes: the stream-json schema is not formally documented as of mid-2026.\n * These types are derived from:\n *   - https://code.claude.com/docs/en/headless\n *   - https://github.com/anthropics/claude-code/issues/24612 (event-types tracking issue)\n *   - https://takopi.dev/reference/runners/claude/stream-json-cheatsheet/\n *   - The `@anthropic-ai/claude-agent-sdk` TypeScript declaration files,\n *     which are the de-facto source of truth.\n *\n * When adding new event types, prefer extending the union here rather than\n * branching on `any` in callers. Unknown events should be tolerated silently\n * by the builder (the schema evolves and we don't want CI to break on a new\n * event type we haven't modelled yet).\n */\n\n/** Top-level discriminated union of stream-json events. */\nexport type StreamEvent =\n  | SystemInitEvent\n  | SystemRetryEvent\n  | SystemPluginInstallEvent\n  | SystemCompactBoundaryEvent\n  | SystemUnknownEvent\n  | AssistantMessageEvent\n  | UserMessageEvent\n  | ResultEvent;\n\n// system events\n\n/** Emitted once at session start. Carries the session-level metadata. */\nexport interface SystemInitEvent {\n  type: \"system\";\n  subtype: \"init\";\n  session_id: string;\n  cwd: string;\n  model: string;\n  permissionMode?: string;\n  apiKeySource?: string;\n  /** Names of tools available in the session (built-in + MCP). */\n  tools: string[];\n  /** MCP servers configured for this session, with connection status. */\n  mcp_servers: McpServerStatus[];\n}\n\nexport interface McpServerStatus {\n  name: string;\n  status: \"connected\" | \"disconnected\" | \"error\" | string;\n}\n\n/** Emitted when the API rate-limits us or otherwise asks for a retry. */\nexport interface SystemRetryEvent {\n  type: \"system\";\n  subtype: \"api_retry\";\n  session_id: string;\n  /** Implementation-defined retry payload (delay, reason, etc). */\n  [key: string]: unknown;\n}\n\n/** Emitted while marketplace plugins are installing pre-session. */\nexport interface SystemPluginInstallEvent {\n  type: \"system\";\n  subtype: \"plugin_install\";\n  session_id: string;\n  [key: string]: unknown;\n}\n\n/** Emitted when Claude Code compacts the context window mid-session. */\nexport interface SystemCompactBoundaryEvent {\n  type: \"system\";\n  subtype: \"compact_boundary\";\n  session_id: string;\n  [key: string]: unknown;\n}\n\n/**\n * Catch-all for `type: \"system\"` events we haven't modelled.\n *\n * Keeps the union exhaustive while tolerating schema evolution. Callers should\n * either explicitly handle a known subtype or fall through to ignore.\n */\nexport interface SystemUnknownEvent {\n  type: \"system\";\n  subtype: string;\n  session_id?: string;\n  [key: string]: unknown;\n}\n\n// conversational events\n\n/** One assistant turn. The `message` field mirrors the Anthropic Messages API shape. */\nexport interface AssistantMessageEvent {\n  type: \"assistant\";\n  session_id: string;\n  message: AssistantMessage;\n}\n\nexport interface AssistantMessage {\n  id: string;\n  type: \"message\";\n  role: \"assistant\";\n  content: ContentBlock[];\n  model?: string;\n  stop_reason?: StopReason | null;\n  usage?: Usage;\n}\n\n/**\n * A user-role message in the stream.\n *\n * In stream-json these are usually *synthetic* — the harness injects them to\n * feed tool results back into the conversation after dispatching a tool. The\n * very first user message (the prompt) is also emitted here for completeness.\n */\nexport interface UserMessageEvent {\n  type: \"user\";\n  session_id: string;\n  message: UserMessage;\n}\n\nexport interface UserMessage {\n  role: \"user\";\n  /** String for the initial prompt, array of blocks when carrying tool results. */\n  content: ContentBlock[] | string;\n}\n\n// content blocks\n\nexport type ContentBlock = TextBlock | ToolUseBlock | ToolResultBlock;\n\nexport interface TextBlock {\n  type: \"text\";\n  text: string;\n}\n\nexport interface ToolUseBlock {\n  type: \"tool_use\";\n  /** Unique id assigned by the model; used to match tool_result back to this call. */\n  id: string;\n  /** Tool name. MCP tools follow the convention `mcp__<server>__<tool>`. */\n  name: string;\n  /** Arguments the model passed. Schema is per-tool. */\n  input: unknown;\n}\n\nexport interface ToolResultBlock {\n  type: \"tool_result\";\n  /** The id of the corresponding tool_use block. */\n  tool_use_id: string;\n  /** Tool output. May be plain text or further content blocks for richer tools. */\n  content: string | ContentBlock[];\n  is_error?: boolean;\n}\n\n// result envelope\n\n/** Emitted once at session end. Carries aggregate usage and cost. */\nexport interface ResultEvent {\n  type: \"result\";\n  subtype: \"success\" | \"error\";\n  session_id: string;\n  total_cost_usd: number;\n  is_error: boolean;\n  duration_ms: number;\n  duration_api_ms?: number;\n  num_turns: number;\n  /** The final text the harness returned, if any. */\n  result?: string;\n  usage?: Usage;\n}\n\n// shared scalars\n\n/**\n * Reasons the model can stop a turn. Open-ended string union because new\n * stop reasons appear over time.\n */\nexport type StopReason =\n  | \"end_turn\"\n  | \"tool_use\"\n  | \"max_tokens\"\n  | \"stop_sequence\"\n  | (string & {});\n\nexport interface Usage {\n  input_tokens: number;\n  output_tokens: number;\n  cache_creation_input_tokens?: number;\n  cache_read_input_tokens?: number;\n}\n\n// type guards\n\n/** Type guards. Prefer these over manual `e.type === \"...\"` checks at call sites. */\n\nexport function isSystemInit(e: StreamEvent): e is SystemInitEvent {\n  return e.type === \"system\" && (e as SystemInitEvent).subtype === \"init\";\n}\n\nexport function isSystemRetry(e: StreamEvent): e is SystemRetryEvent {\n  return e.type === \"system\" && (e as SystemRetryEvent).subtype === \"api_retry\";\n}\n\nexport function isAssistantMessage(e: StreamEvent): e is AssistantMessageEvent {\n  return e.type === \"assistant\";\n}\n\nexport function isUserMessage(e: StreamEvent): e is UserMessageEvent {\n  return e.type === \"user\";\n}\n\nexport function isResult(e: StreamEvent): e is ResultEvent {\n  return e.type === \"result\";\n}\n\nexport function isTextBlock(b: ContentBlock): b is TextBlock {\n  return b.type === \"text\";\n}\n\nexport function isToolUseBlock(b: ContentBlock): b is ToolUseBlock {\n  return b.type === \"tool_use\";\n}\n\nexport function isToolResultBlock(b: ContentBlock): b is ToolResultBlock {\n  return b.type === \"tool_result\";\n}\n","/**\n * TrajectoryView — the assertion-friendly projection of a Claude Code session.\n *\n * The view is derived from the stream of {@link StreamEvent} values produced by\n * the harness, but is optimized for the queries that the assertion DSL needs to\n * express:\n *\n *   - did tool X get called?  (look at `toolCalls`)\n *   - did tool A come before tool B?  (compare `turnIndex` / `callIndex`)\n *   - was a tool called with arguments matching predicate P?  (`toolCalls[i].args`)\n *   - did the agent answer without using any tool?  (`toolCalls.length === 0`)\n *\n * The view is reconstructable from the raw events (lossless w.r.t. assertions),\n * but operating on it directly is dramatically simpler than walking event\n * streams or OTel span trees.\n *\n * Design notes:\n *  - `turnIndex` and `callIndex` are the right primitives for ordering.\n *    Wall-clock timestamps from the stream are unreliable for sub-second\n *    ordering and parallel tool dispatch.\n *  - Parallel tool calls (multiple `tool_use` blocks in one assistant message)\n *    share a `turnIndex` but have distinct `callIndex` values in emission order.\n *  - `namespace` is precomputed so assertions like `called(pattern: \"mcp__api__*\")`\n *    can do a cheap string check.\n */\n\nimport type { StopReason } from \"./stream\";\n\nexport interface TrajectoryView {\n  /** Session metadata, captured from the `system/init` event. */\n  meta: SessionMeta;\n\n  /** Every tool call, in global emission order. */\n  toolCalls: ToolCall[];\n\n  /** Each assistant turn: text content + any tool calls emitted in that turn. */\n  turns: AssistantTurn[];\n\n  /** All assistant text concatenated across turns. Useful for `response_contains`. */\n  finalResponse: string;\n\n  /** Stop reason of the *last* assistant turn. */\n  finalStopReason: StopReason | null;\n\n  /** Aggregate usage and cost from the result event. */\n  usage: UsageSummary;\n\n  /** Retry events observed during the run (rate limits, transient errors). */\n  retries: RetryRecord[];\n\n  /** Whether the result envelope indicated success. */\n  success: boolean;\n}\n\nexport interface SessionMeta {\n  sessionId: string;\n  model: string;\n  cwd: string;\n  permissionMode?: string;\n  /** Tool names the harness reported as available at session start. */\n  availableTools: string[];\n  /** MCP servers configured for the session, with connection status. */\n  mcpServers: { name: string; status: string }[];\n}\n\nexport interface ToolCall {\n  /** Fully-qualified tool name, e.g. `\"mcp__api__search_skills\"` or `\"Bash\"`. */\n  name: string;\n\n  /**\n   * Namespace prefix for MCP-style names (`\"mcp__api\"`), or null for built-ins.\n   * Precomputed via {@link namespaceOf} for cheap pattern matching.\n   */\n  namespace: string | null;\n\n  /** The `tool_use` block's `id`; matches a later `tool_result.tool_use_id`. */\n  callId: string;\n\n  /** Args the model emitted on this call. Tool-specific schema. */\n  args: unknown;\n\n  /** Tool result, or null if no result was observed (e.g. process killed). */\n  result: unknown | null;\n\n  /** Whether the tool reported an error in its result. */\n  isError: boolean;\n\n  /**\n   * Which assistant turn produced this call. Parallel calls within a single\n   * assistant message share a `turnIndex`.\n   */\n  turnIndex: number;\n\n  /** Index in the global ordered tool-call sequence. */\n  callIndex: number;\n}\n\nexport interface AssistantTurn {\n  turnIndex: number;\n  /** Text emitted in this turn (may be empty if turn was tool-only). */\n  text: string;\n  /** Tool calls emitted in this turn, in their block order. */\n  toolCalls: ToolCall[];\n  /** Stop reason reported by the model for this turn. */\n  stopReason: StopReason | null;\n}\n\nexport interface UsageSummary {\n  inputTokens: number;\n  outputTokens: number;\n  totalCostUsd: number;\n  durationMs: number;\n  numTurns: number;\n}\n\nexport interface RetryRecord {\n  /** ms since session start (approximate; the stream doesn't include precise ts). */\n  offsetMs: number;\n  /** Raw payload from the `system/api_retry` event for diagnostics. */\n  raw: unknown;\n}\n\n// helpers\n\n/**\n * Extract the MCP namespace prefix from a tool name.\n *\n * Claude Code formats MCP tool names as `mcp__<server>__<tool>`. The namespace\n * is the first two segments joined: `mcp__<server>`. Returns null for non-MCP\n * tool names (built-ins like `Bash`, `Read`, `Edit`).\n *\n * @example\n *   namespaceOf(\"mcp__api__search_skills\") // \"mcp__api\"\n *   namespaceOf(\"Bash\")                     // null\n */\nexport function namespaceOf(toolName: string): string | null {\n  if (!toolName.startsWith(\"mcp__\")) return null;\n  const parts = toolName.split(\"__\");\n  if (parts.length < 3) return null;\n  return `${parts[0]}__${parts[1]}`;\n}\n","/**\n * TrajectoryBuilder — consumes a stream of {@link StreamEvent} values and\n * produces a {@link TrajectoryView}.\n *\n * State machine: the builder is a small, tolerant state machine. Invariants:\n *\n *   - Exactly one `system/init` event opens the session. The builder requires\n *     it to be present before `build()`.\n *   - Each `assistant` event begins a new turn. Text blocks accumulate into\n *     the turn's text; `tool_use` blocks become `ToolCall` records.\n *   - `user` events with `tool_result` blocks deliver tool results back. We\n *     match them to pending calls by `tool_use_id`.\n *   - One `result` event closes the session and carries aggregate usage.\n *\n * The builder is *tolerant of partial streams*: a process killed mid-run\n * produces a coherent (but flagged) view. Tool calls without matching results\n * keep `result: null`. The `success` flag reflects whether a successful result\n * event was actually observed.\n *\n * Why a class (not a reducer)?\n *   The internal `pendingCalls` map is mutable by design — we modify ToolCall\n *   objects in place when results arrive, so other parts of the view (which\n *   hold references to the same objects) see the update for free. A reducer\n *   would force a deep copy per result event, which is wasteful and would\n *   complicate identity-based queries.\n */\n\nimport {\n  isAssistantMessage,\n  isResult,\n  isSystemInit,\n  isTextBlock,\n  isToolResultBlock,\n  isToolUseBlock,\n  isUserMessage,\n  type StreamEvent,\n  type Usage,\n} from \"../types/stream\";\nimport {\n  namespaceOf,\n  type AssistantTurn,\n  type RetryRecord,\n  type SessionMeta,\n  type ToolCall,\n  type TrajectoryView,\n} from \"../types/trajectory\";\n\nexport class TrajectoryBuilder {\n  private meta: SessionMeta | null = null;\n  private sessionStartTs: number | null = null;\n\n  private turns: AssistantTurn[] = [];\n  private allToolCalls: ToolCall[] = [];\n\n  /**\n   * tool_use_id → ToolCall, for matching results back to calls.\n   * Entries are removed once a result is observed.\n   */\n  private pendingCalls: Map<string, ToolCall> = new Map();\n\n  private retries: RetryRecord[] = [];\n\n  private finalUsage: Usage | null = null;\n  private finalCostUsd = 0;\n  private finalDurationMs = 0;\n  private finalNumTurns = 0;\n  private finalResultText = \"\";\n  private sawResultEvent = false;\n  private resultIsError = false;\n\n  /**\n   * Consume one event. Safe to call with events in stream order.\n   *\n   * Unknown event types are silently ignored — the schema evolves and we\n   * don't want CI to break on a new event type we haven't modelled.\n   */\n  consume(event: StreamEvent): void {\n    if (isSystemInit(event)) {\n      this.meta = {\n        sessionId: event.session_id,\n        model: event.model,\n        cwd: event.cwd,\n        permissionMode: event.permissionMode,\n        availableTools: event.tools ?? [],\n        mcpServers: (event.mcp_servers ?? []).map((s) => ({\n          name: s.name,\n          status: s.status,\n        })),\n      };\n      this.sessionStartTs = Date.now();\n      return;\n    }\n\n    if (event.type === \"system\" && event.subtype === \"api_retry\") {\n      this.retries.push({\n        offsetMs: this.sessionStartTs ? Date.now() - this.sessionStartTs : 0,\n        raw: event,\n      });\n      return;\n    }\n\n    if (isAssistantMessage(event)) {\n      this.handleAssistantMessage(event);\n      return;\n    }\n\n    if (isUserMessage(event)) {\n      this.handleUserMessage(event);\n      return;\n    }\n\n    if (isResult(event)) {\n      this.sawResultEvent = true;\n      this.resultIsError = event.is_error;\n      this.finalUsage = event.usage ?? null;\n      this.finalCostUsd = event.total_cost_usd ?? 0;\n      this.finalDurationMs = event.duration_ms ?? 0;\n      this.finalNumTurns = event.num_turns ?? 0;\n      this.finalResultText = event.result ?? \"\";\n      return;\n    }\n\n    // Unknown event: ignored. See class doc.\n  }\n\n  /**\n   * Finalize the view. Call after consuming the last event from the stream.\n   *\n   * Throws if no `system/init` was observed — at that point we have no model,\n   * no session id, and no available-tools list, which means assertions like\n   * \"called any mcp__api__* tool\" can't even be evaluated meaningfully.\n   */\n  build(): TrajectoryView {\n    if (this.meta === null) {\n      throw new Error(\n        \"TrajectoryBuilder.build() called before any system/init event was observed. \" +\n          \"The harness may have failed to start, or the stream was truncated before init.\",\n      );\n    }\n\n    const lastTurn = this.turns[this.turns.length - 1];\n\n    // Prefer the assistant text we accumulated turn-by-turn over the\n    // `result.result` field, because the latter is sometimes a summary\n    // and the former is exactly what the model said.\n    const accumulatedText = this.turns\n      .map((t) => t.text)\n      .filter((t) => t.length > 0)\n      .join(\"\\n\\n\")\n      .trim();\n\n    return {\n      meta: this.meta,\n      toolCalls: this.allToolCalls,\n      turns: this.turns,\n      finalResponse: accumulatedText || this.finalResultText,\n      finalStopReason: lastTurn?.stopReason ?? null,\n      usage: {\n        inputTokens: this.finalUsage?.input_tokens ?? 0,\n        outputTokens: this.finalUsage?.output_tokens ?? 0,\n        totalCostUsd: this.finalCostUsd,\n        durationMs: this.finalDurationMs,\n        // Fall back to observed turn count if the result event was missing.\n        numTurns: this.finalNumTurns || this.turns.length,\n      },\n      retries: this.retries,\n      // Successful = saw a non-error result envelope. Streams that ended without\n      // a result event are reported as unsuccessful regardless of tool outcomes.\n      success: this.sawResultEvent && !this.resultIsError,\n    };\n  }\n\n  // private handlers\n\n  private handleAssistantMessage(\n    event: Extract<StreamEvent, { type: \"assistant\" }>,\n  ): void {\n    const turnIndex = this.turns.length;\n    const textChunks: string[] = [];\n    const toolCallsThisTurn: ToolCall[] = [];\n\n    for (const block of event.message.content) {\n      if (isTextBlock(block)) {\n        textChunks.push(block.text);\n        continue;\n      }\n      if (isToolUseBlock(block)) {\n        const call: ToolCall = {\n          name: block.name,\n          namespace: namespaceOf(block.name),\n          callId: block.id,\n          args: block.input,\n          result: null,\n          isError: false,\n          turnIndex,\n          callIndex: this.allToolCalls.length,\n        };\n        this.allToolCalls.push(call);\n        this.pendingCalls.set(block.id, call);\n        toolCallsThisTurn.push(call);\n        continue;\n      }\n      // tool_result blocks don't appear in assistant messages — those arrive\n      // via user messages. If one does appear, ignore it; we'd rather drop\n      // an unexpected block than crash the eval.\n    }\n\n    this.turns.push({\n      turnIndex,\n      text: textChunks.join(\"\").trim(),\n      toolCalls: toolCallsThisTurn,\n      stopReason: event.message.stop_reason ?? null,\n    });\n  }\n\n  private handleUserMessage(\n    event: Extract<StreamEvent, { type: \"user\" }>,\n  ): void {\n    const content = event.message.content;\n\n    // The very first user message carries the prompt as a plain string. We\n    // already know the prompt (the caller passed it to the adapter), so we\n    // ignore this case — there's nothing assertion-relevant in it.\n    if (typeof content === \"string\") return;\n\n    for (const block of content) {\n      if (!isToolResultBlock(block)) continue;\n\n      const call = this.pendingCalls.get(block.tool_use_id);\n      if (!call) {\n        // Unmatched result: ignore. Can happen if events arrive out of order\n        // or the corresponding tool_use was emitted in an earlier run that\n        // we're resuming. Either way, dropping is safer than throwing.\n        continue;\n      }\n\n      call.result = block.content;\n      call.isError = block.is_error ?? false;\n      this.pendingCalls.delete(block.tool_use_id);\n    }\n  }\n}\n\n/**\n * Convenience: drain an async iterable of events through a fresh builder.\n *\n * Suitable when you have the full event stream and just want the view.\n * For interactive/incremental scenarios (e.g. surfacing partial state in a UI)\n * instantiate {@link TrajectoryBuilder} directly and call `consume()` /\n * `build()` yourself.\n */\nexport async function buildTrajectory(\n  events: AsyncIterable<StreamEvent>,\n): Promise<TrajectoryView> {\n  const builder = new TrajectoryBuilder();\n  for await (const event of events) {\n    builder.consume(event);\n  }\n  return builder.build();\n}\n","/**\n * Line-buffered NDJSON parser for Claude Code's `--output-format stream-json`.\n *\n * Claude Code emits one JSON object per line on stdout. The parser:\n *   - buffers across chunk boundaries (a single JSON line may arrive in two reads)\n *   - skips empty lines (defensive — shouldn't occur, but harmless if it does)\n *   - emits a discriminated `ParseResult` per line so callers can decide whether\n *     a malformed line should abort the run or just be logged.\n *\n * Why a generator (and not a Transform stream)?\n *   The eval adapter consumes events sequentially and synchronously updates a\n *   builder. Async iteration is the simplest interface for that pattern and\n *   composes cleanly with `for await` in the adapter. A Transform would force\n *   the builder into event-handler style.\n */\n\nimport type { Readable } from \"node:stream\";\nimport type { StreamEvent } from \"../types/stream\";\n\n/**\n * Result of attempting to parse a single line.\n *\n * Successful parses yield `{ ok: true }` with the typed event and the raw line\n * (kept for diagnostics and OTel `events.attributes.raw`). Failed parses yield\n * `{ ok: false }` with the parse error and the raw line — callers can log,\n * skip, or fail the run as they see fit.\n */\nexport type ParseResult =\n  | { ok: true; event: StreamEvent; rawLine: string }\n  | { ok: false; error: Error; rawLine: string };\n\n/**\n * Parse a readable stream of NDJSON into a sequence of typed stream-json events.\n *\n * @example\n *   const child = spawn(\"claude\", [\"-p\", prompt, \"--output-format\", \"stream-json\", \"--verbose\"]);\n *   for await (const result of parseStreamJson(child.stdout)) {\n *     if (result.ok) builder.consume(result.event);\n *     else console.warn(\"malformed stream line:\", result.rawLine, result.error);\n *   }\n */\nexport async function* parseStreamJson(\n  stream: Readable,\n): AsyncGenerator<ParseResult, void, void> {\n  let buffer = \"\";\n  // The Node child_process stdout is a binary stream by default. Setting the\n  // encoding here means `for await (const chunk of stream)` yields strings.\n  stream.setEncoding(\"utf8\");\n\n  for await (const chunk of stream) {\n    buffer += chunk as string;\n\n    // Drain every complete line currently in the buffer before reading more.\n    // Multiple JSON objects can arrive in one chunk (e.g. when the harness\n    // emits a burst of events at session start).\n    let newlineIdx: number;\n    while ((newlineIdx = buffer.indexOf(\"\\n\")) !== -1) {\n      const line = buffer.slice(0, newlineIdx).trim();\n      buffer = buffer.slice(newlineIdx + 1);\n      if (line.length === 0) continue;\n      yield tryParseLine(line);\n    }\n  }\n\n  // Flush any trailing content that arrived without a final newline. Stream-json\n  // typically ends with a newline-terminated `result` event, but a killed\n  // process may not flush, so we still try to emit what we have.\n  const trailing = buffer.trim();\n  if (trailing.length > 0) {\n    yield tryParseLine(trailing);\n  }\n}\n\n/**\n * Parse a single line. Extracted as a helper so the generator stays readable.\n *\n * Note: we do not validate the event structure beyond `JSON.parse`. Runtime\n * validation (e.g. zod) is overkill here — the schema is stable enough at\n * runtime, and the TrajectoryBuilder is tolerant of missing fields. Adding\n * validation would be premature.\n */\nfunction tryParseLine(line: string): ParseResult {\n  try {\n    const event = JSON.parse(line) as StreamEvent;\n    return { ok: true, event, rawLine: line };\n  } catch (err) {\n    return {\n      ok: false,\n      error: err instanceof Error ? err : new Error(String(err)),\n      rawLine: line,\n    };\n  }\n}\n","/**\n * Generic harness adapter contract.\n *\n * Every harness adapter produces a {@link TrajectoryView} plus process\n * diagnostics. The runner and assertion engine depend only on these types —\n * not on any specific harness implementation.\n */\n\nimport type { TrajectoryView } from \"../types/trajectory\";\n\n/** Base config every adapter must accept. */\nexport interface BaseAdapterConfig {\n  prompt: string;\n  model?: string;\n  timeoutMs?: number;\n  signal?: AbortSignal;\n  env?: Record<string, string>;\n  cwd?: string;\n}\n\n/** Suite-level config: generic fields plus adapter-specific nested blocks. */\nexport type SuiteConfig = Partial<BaseAdapterConfig> & {\n  /** Claude Code adapter options (when `adapter` is `claude-code`). */\n  claudeCode?: Record<string, unknown>;\n};\n\n/** Generic harness adapter interface. */\nexport interface HarnessAdapter<\n  TConfig extends BaseAdapterConfig = BaseAdapterConfig,\n> {\n  readonly id: string;\n  run(config: TConfig): Promise<AdapterResult>;\n}\n\n/** Successful adapter run. */\nexport interface AdapterResult {\n  view: TrajectoryView;\n  diagnostics: AdapterDiagnostics;\n}\n\n/** Process-level diagnostics from any adapter. */\nexport interface AdapterDiagnostics {\n  exitCode: number | null;\n  signal: NodeJS.Signals | null;\n  stderr: string;\n  parseErrors: ParseErrorRecord[];\n  timedOut: boolean;\n  durationMs: number;\n}\n\nexport interface ParseErrorRecord {\n  line: string;\n  error: string;\n}\n\n/**\n * Thrown when the harness fails to produce a usable trajectory.\n *\n * Most commonly this means the process failed before emitting a usable\n * session init event. Inspect `diagnostics.stderr` for the cause.\n */\nexport class AdapterError extends Error {\n  constructor(\n    message: string,\n    public readonly diagnostics: Partial<AdapterDiagnostics>,\n  ) {\n    super(message);\n    this.name = \"AdapterError\";\n  }\n}\n","/**\n * Build CLI args for Claude Code judge subprocesses (JSON output, not stream-json).\n *\n * Shared flag assembly for harness runs (`buildArgs`) and LLM grading judges\n * (`buildJudgeArgs`).\n */\n\nimport type { ClaudeCodeAdapterConfig, ClaudeCodeOptions } from \"./types\";\n\n/** Append repeated `--flag value` pairs for array config fields. */\nfunction pushRepeatableFlag(args: string[], flag: string, values?: string[]): void {\n  if (!values) return;\n  for (const value of values) {\n    args.push(flag, value);\n  }\n}\n\n/**\n * Append an optional CLI flag. Boolean `true` emits the flag alone; other\n * scalars emit `--flag value`.\n */\nfunction pushOptionalFlag(\n  args: string[],\n  flag: string,\n  value: string | number | boolean | undefined,\n): void {\n  if (value === undefined) return;\n  if (typeof value === \"boolean\") {\n    if (value) args.push(flag);\n    return;\n  }\n  args.push(flag, String(value));\n}\n\n/** Append Claude Code CLI flags shared by harness runs and grading judges. */\nexport function appendClaudeCodeFlags(\n  args: string[],\n  config: ClaudeCodeOptions & { model?: string },\n): void {\n  pushRepeatableFlag(args, \"--plugin-dir\", config.pluginDirs);\n  pushRepeatableFlag(args, \"--plugin-url\", config.pluginUrls);\n  pushRepeatableFlag(args, \"--add-dir\", config.addDirs);\n\n  pushOptionalFlag(args, \"--mcp-config\", config.mcpConfig);\n  pushOptionalFlag(args, \"--model\", config.model);\n  pushOptionalFlag(args, \"--permission-mode\", config.permissionMode);\n  pushOptionalFlag(args, \"--effort\", config.effort);\n  pushOptionalFlag(args, \"--agent\", config.agent);\n  pushOptionalFlag(args, \"--fallback-model\", config.fallbackModel);\n  pushOptionalFlag(args, \"--tools\", config.tools);\n  pushOptionalFlag(args, \"--settings\", config.settings);\n  pushOptionalFlag(args, \"--setting-sources\", config.settingSources);\n  pushOptionalFlag(args, \"--max-turns\", config.maxTurns);\n  pushOptionalFlag(args, \"--max-budget-usd\", config.maxBudgetUsd);\n  pushOptionalFlag(args, \"--system-prompt\", config.systemPrompt);\n  pushOptionalFlag(args, \"--system-prompt-file\", config.systemPromptFile);\n  pushOptionalFlag(args, \"--append-system-prompt\", config.appendSystemPrompt);\n  pushOptionalFlag(\n    args,\n    \"--append-system-prompt-file\",\n    config.appendSystemPromptFile,\n  );\n  pushOptionalFlag(args, \"--debug\", config.debug);\n  pushOptionalFlag(args, \"--debug-file\", config.debugFile);\n\n  if (config.allowedTools && config.allowedTools.length > 0) {\n    args.push(\"--allowedTools\", config.allowedTools.join(\",\"));\n  }\n\n  if (config.disallowedTools && config.disallowedTools.length > 0) {\n    args.push(\"--disallowedTools\", config.disallowedTools.join(\",\"));\n  }\n\n  pushOptionalFlag(args, \"--strict-mcp-config\", config.strictMcpConfig);\n  pushOptionalFlag(args, \"--include-hook-events\", config.includeHookEvents);\n  pushOptionalFlag(args, \"--no-session-persistence\", config.noSessionPersistence);\n  pushOptionalFlag(args, \"--disable-slash-commands\", config.disableSlashCommands);\n  pushOptionalFlag(args, \"--bare\", config.bare);\n  pushOptionalFlag(args, \"--safe-mode\", config.safeMode);\n  pushOptionalFlag(\n    args,\n    \"--allow-dangerously-skip-permissions\",\n    config.allowDangerouslySkipPermissions,\n  );\n  pushOptionalFlag(\n    args,\n    \"--dangerously-skip-permissions\",\n    config.dangerouslySkipPermissions,\n  );\n}\n\n/**\n * Build the argument vector for spawning `claude`.\n *\n * Order matters only for flags that take values — value flags must come\n * after their flag name. Everything else is order-independent.\n */\nexport function buildArgs(config: ClaudeCodeAdapterConfig): string[] {\n  const args: string[] = [\n    \"-p\",\n    config.prompt,\n    \"--output-format\",\n    \"stream-json\",\n    \"--verbose\",\n  ];\n\n  appendClaudeCodeFlags(args, config);\n\n  return args;\n}\n\n/**\n * Build args for an LLM judge subprocess (`--output-format json`).\n *\n * Defaults permission mode to `bypassPermissions` so the judge does not\n * block on tool permission prompts during single-shot JSON grading.\n */\nexport function buildJudgeArgs(\n  prompt: string,\n  config: ClaudeCodeOptions & { model?: string } = {},\n): string[] {\n  const args: string[] = [\"-p\", prompt, \"--output-format\", \"json\"];\n  const permissionMode = config.permissionMode ?? \"bypassPermissions\";\n  appendClaudeCodeFlags(args, {\n    ...config,\n    permissionMode,\n  });\n  return args;\n}\n","/**\n * Process management for the Claude Code adapter.\n *\n * This module owns spawning, timeout, abort signal handling, and process-tree\n * teardown. The orchestrator (`index.ts`) consumes the returned handle —\n * reading stdout and waiting for completion — but doesn't worry about how\n * the process gets killed or how its config gets isolated.\n *\n * Why a separate module? Process management is the one part of the adapter\n * with real I/O complexity (process groups, signal escalation, temp-dir\n * lifecycle, env merging). Isolating it makes the orchestrator easy to read\n * and lets us swap the spawning logic if we later need to, e.g., wrap claude\n * in a sandbox runner.\n */\n\nimport { spawn, type ChildProcess } from \"node:child_process\";\nimport { mkdtemp, rm } from \"node:fs/promises\";\nimport { tmpdir } from \"node:os\";\nimport { join } from \"node:path\";\nimport type { Readable } from \"node:stream\";\n\nimport { buildArgs } from \"./flags\";\nimport type { ClaudeCodeAdapterConfig } from \"./types\";\n\n/** Default hard timeout per run. Tunable via config.timeoutMs. */\nconst DEFAULT_TIMEOUT_MS = 5 * 60 * 1000;\n\n/**\n * Grace period between SIGTERM and SIGKILL. Most processes shut down cleanly\n * within a few seconds; this gives them that chance while preventing CI from\n * hanging indefinitely on a stuck child.\n */\nconst KILL_GRACE_MS = 5_000;\n\n/**\n * Handle to a spawned `claude` process. The orchestrator drives it:\n *   - Read `stdout` (typically via parseStreamJson).\n *   - Await `done` to learn the exit state.\n *   - Await `stderrCollected` for diagnostic stderr.\n *   - Check `timedOut()` after exit to distinguish kill-by-timeout from\n *     normal termination.\n *   - Call `cleanup()` after all of the above to remove the temp config dir.\n */\nexport interface SpawnedClaude {\n  stdout: Readable;\n  done: Promise<{ exitCode: number | null; signal: NodeJS.Signals | null }>;\n  stderrCollected: Promise<string>;\n  timedOut: () => boolean;\n  cleanup: () => Promise<void>;\n}\n\n/**\n * Spawn `claude` in headless mode with isolated config and a process-group\n * lifecycle. See {@link SpawnedClaude} for how to consume the result.\n *\n * **Kill sequence:** timeout and abort both follow the same two-step path:\n * `SIGTERM` to the process group, then `SIGKILL` after {@link KILL_GRACE_MS}\n * if the group is still alive. This avoids leaving MCP/tool subprocesses\n * running while still giving claude a chance to flush stream-json output.\n *\n * @param config - Adapter options; `timeoutMs`, `signal`, and `isolateConfig`\n *   control lifecycle and config isolation.\n */\nexport async function spawnClaude(\n  config: ClaudeCodeAdapterConfig,\n): Promise<SpawnedClaude> {\n  const binary = config.binary ?? \"claude\";\n  const args = buildArgs(config);\n\n  const isolateConfig = config.isolateConfig !== false;\n\n  // Isolated runs use a fresh temp dir so plugins/settings don't leak between\n  // reps. Non-isolated runs inherit the caller's Claude login and plugins.\n  const tempConfigDir = isolateConfig\n    ? await mkdtemp(join(tmpdir(), \"harness-eval-\"))\n    : null;\n\n  const env: Record<string, string | undefined> = {\n    ...process.env,\n    ...config.env,\n  };\n  if (tempConfigDir) {\n    // Override after ...env so callers can't accidentally un-isolate.\n    env.CLAUDE_CONFIG_DIR = tempConfigDir;\n  }\n\n  const child = spawn(binary, args, {\n    cwd: config.cwd ?? process.cwd(),\n    env,\n    stdio: [\"ignore\", \"pipe\", \"pipe\"],\n    // detached: true means the child becomes the leader of its own process\n    // group. We exploit this to kill the entire group (including any MCP\n    // server subprocesses and tool processes) on timeout/abort.\n    detached: true,\n  });\n\n\n  // `timedOut` is set only by the hard timeout timer, not by abort — callers\n  // use it to distinguish \"ran too long\" from user cancellation or normal exit.\n  let timedOut = false;\n  let killEscalation: NodeJS.Timeout | null = null;\n  const timeoutMs = config.timeoutMs ?? DEFAULT_TIMEOUT_MS;\n\n  /**\n   * Arm (or re-arm) the SIGKILL fallback. Each SIGTERM attempt gets its own\n   * grace window so a slow shutdown doesn't leave orphaned MCP servers.\n   */\n  const scheduleKillEscalation = () => {\n    if (killEscalation) clearTimeout(killEscalation);\n    killEscalation = setTimeout(\n      () => killTree(child, \"SIGKILL\"),\n      KILL_GRACE_MS,\n    );\n  };\n\n  const timeoutTimer = setTimeout(() => {\n    timedOut = true;\n    killTree(child, \"SIGTERM\");\n    scheduleKillEscalation();\n  }, timeoutMs);\n\n  // AbortSignal cancellation mirrors timeout kills but does not flip `timedOut`.\n  const onAbort = () => {\n    killTree(child, \"SIGTERM\");\n    scheduleKillEscalation();\n  };\n  config.signal?.addEventListener(\"abort\", onAbort, { once: true });\n\n\n  // Drain stderr eagerly so the OS-level buffer never fills and stalls the\n  // child (Node child processes will block on a full pipe).\n  const stderrChunks: string[] = [];\n  child.stderr?.setEncoding(\"utf8\");\n  child.stderr?.on(\"data\", (chunk: string) => {\n    stderrChunks.push(chunk);\n  });\n\n  const stderrCollected = new Promise<string>((resolve) => {\n    const finalize = () => resolve(stderrChunks.join(\"\"));\n    child.stderr?.on(\"end\", finalize);\n    // Errors during stderr capture shouldn't fail the whole run; we just\n    // return what we've buffered so far.\n    child.stderr?.on(\"error\", finalize);\n  });\n\n\n  // Resolve once the process exits or fails to spawn. Guard against double\n  // settlement because both `close` and `error` can fire in edge cases.\n  const done = new Promise<{\n    exitCode: number | null;\n    signal: NodeJS.Signals | null;\n  }>((resolve) => {\n    let settled = false;\n    const finalize = (\n      exitCode: number | null,\n      signal: NodeJS.Signals | null,\n    ) => {\n      if (settled) return;\n      settled = true;\n      // Tear down timers/listeners so a late timeout cannot SIGKILL a reused PID.\n      clearTimeout(timeoutTimer);\n      if (killEscalation) clearTimeout(killEscalation);\n      config.signal?.removeEventListener(\"abort\", onAbort);\n      resolve({ exitCode, signal });\n    };\n\n    child.on(\"close\", (code, signal) => finalize(code, signal));\n    // ENOENT and other spawn failures emit `error` — `close` may not follow.\n    child.on(\"error\", () => finalize(null, null));\n  });\n\n\n  const cleanup = async () => {\n    if (!tempConfigDir) return;\n    try {\n      await rm(tempConfigDir, { recursive: true, force: true });\n    } catch {\n      // Best-effort. A leftover temp dir is annoying but not catastrophic;\n      // we don't want to fail the run for it.\n    }\n  };\n\n  // stdout is guaranteed non-null because we passed `stdio: [..., \"pipe\", ...]`.\n  // The `!` is safe; the alternative would be a redundant runtime check that\n  // could never fire.\n  return {\n    stdout: child.stdout!,\n    done,\n    stderrCollected,\n    timedOut: () => timedOut,\n    cleanup,\n  };\n}\n\n/**\n * Kill the child's process group, then fall back to the bare PID if the\n * group is already gone. This catches MCP server subprocesses and tool\n * processes spawned by claude.\n *\n * **Signal escalation:** callers typically invoke this first with `SIGTERM`,\n * then again with `SIGKILL` after {@link KILL_GRACE_MS}. The group kill is\n * essential — a bare `child.kill()` would leave MCP servers running.\n *\n * **Platform edge case:** when the group leader exits first, `kill(-pid)`\n * throws `ESRCH`. The single-PID fallback covers that without failing the\n * adapter run.\n *\n * @param child - Spawned process handle from {@link spawn}.\n * @param signal - POSIX signal to deliver (`SIGTERM` or `SIGKILL` in practice).\n */\nfunction killTree(child: ChildProcess, signal: NodeJS.Signals): void {\n  if (child.pid === undefined) return;\n  try {\n    // Negative PID targets the entire process group (requires detached spawn).\n    process.kill(-child.pid, signal);\n  } catch {\n    try {\n      // Group already reaped — try the leader PID directly.\n      child.kill(signal);\n    } catch {\n      // Process fully gone; nothing to do.\n    }\n  }\n}\n","/**\n * Claude Code adapter — public API.\n */\n\nimport { parseStreamJson } from \"../../parsers/stream-json\";\nimport { TrajectoryBuilder } from \"../../trajectory/builder\";\nimport type { StreamEvent } from \"../../types/stream\";\n\nimport { AdapterError } from \"../types\";\nimport { spawnClaude } from \"./process\";\nimport type {\n  AdapterDiagnostics,\n  ClaudeCodeAdapterConfig,\n  ClaudeCodeAdapterResult,\n  ParseErrorRecord,\n} from \"./types\";\nimport type { HarnessAdapter } from \"../types\";\n\nexport { AdapterError } from \"../types\";\nexport type {\n  AdapterDiagnostics,\n  AdapterResult,\n  ClaudeCodeAdapterConfig,\n  ClaudeCodeAdapterResult,\n  ClaudeCodeOptions,\n  ParseErrorRecord,\n  PermissionMode,\n} from \"./types\";\n\n/**\n * Run Claude Code in headless mode and return a trajectory.\n */\nexport async function runClaudeCode(\n  config: ClaudeCodeAdapterConfig,\n): Promise<ClaudeCodeAdapterResult> {\n  const startTs = Date.now();\n  const spawned = await spawnClaude(config);\n\n  const builder = new TrajectoryBuilder();\n  const rawEvents: StreamEvent[] = [];\n  const parseErrors: ParseErrorRecord[] = [];\n\n  try {\n    for await (const result of parseStreamJson(spawned.stdout)) {\n      if (result.ok) {\n        builder.consume(result.event);\n        rawEvents.push(result.event);\n      } else {\n        parseErrors.push({\n          line: result.rawLine,\n          error: result.error.message,\n        });\n      }\n    }\n\n    const [{ exitCode, signal }, stderr] = await Promise.all([\n      spawned.done,\n      spawned.stderrCollected,\n    ]);\n\n    const diagnostics: AdapterDiagnostics = {\n      exitCode,\n      signal,\n      stderr,\n      parseErrors,\n      timedOut: spawned.timedOut(),\n      durationMs: Date.now() - startTs,\n    };\n\n    let view;\n    try {\n      view = builder.build();\n    } catch (err) {\n      const message = err instanceof Error ? err.message : String(err);\n      throw new AdapterError(\n        `harness produced no usable trajectory: ${message}`,\n        diagnostics,\n      );\n    }\n\n    return { view, diagnostics, rawEvents };\n  } finally {\n    await spawned.cleanup();\n  }\n}\n\n/** Registered {@link HarnessAdapter} for Claude Code headless runs. */\nexport const claudeCodeAdapter: HarnessAdapter<ClaudeCodeAdapterConfig> = {\n  id: \"claude-code\",\n  run: runClaudeCode,\n};\n"],"mappings":";;;;;;;AAuMA,SAAgB,aAAa,GAAsC;CACjE,OAAO,EAAE,SAAS,YAAa,EAAsB,YAAY;AACnE;AAEA,SAAgB,cAAc,GAAuC;CACnE,OAAO,EAAE,SAAS,YAAa,EAAuB,YAAY;AACpE;AAEA,SAAgB,mBAAmB,GAA4C;CAC7E,OAAO,EAAE,SAAS;AACpB;AAEA,SAAgB,cAAc,GAAuC;CACnE,OAAO,EAAE,SAAS;AACpB;AAEA,SAAgB,SAAS,GAAkC;CACzD,OAAO,EAAE,SAAS;AACpB;AAEA,SAAgB,YAAY,GAAiC;CAC3D,OAAO,EAAE,SAAS;AACpB;AAEA,SAAgB,eAAe,GAAoC;CACjE,OAAO,EAAE,SAAS;AACpB;AAEA,SAAgB,kBAAkB,GAAuC;CACvE,OAAO,EAAE,SAAS;AACpB;;;;;;;;;;;;;;AC9FA,SAAgB,YAAY,UAAiC;CAC3D,IAAI,CAAC,SAAS,WAAW,OAAO,GAAG,OAAO;CAC1C,MAAM,QAAQ,SAAS,MAAM,IAAI;CACjC,IAAI,MAAM,SAAS,GAAG,OAAO;CAC7B,OAAO,GAAG,MAAM,GAAG,IAAI,MAAM;AAC/B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC7FA,IAAa,oBAAb,MAA+B;CAC7B,OAAmC;CACnC,iBAAwC;CAExC,QAAiC,CAAC;CAClC,eAAmC,CAAC;;;;;CAMpC,+BAA8C,IAAI,IAAI;CAEtD,UAAiC,CAAC;CAElC,aAAmC;CACnC,eAAuB;CACvB,kBAA0B;CAC1B,gBAAwB;CACxB,kBAA0B;CAC1B,iBAAyB;CACzB,gBAAwB;;;;;;;CAQxB,QAAQ,OAA0B;EAChC,IAAI,aAAa,KAAK,GAAG;GACvB,KAAK,OAAO;IACV,WAAW,MAAM;IACjB,OAAO,MAAM;IACb,KAAK,MAAM;IACX,gBAAgB,MAAM;IACtB,gBAAgB,MAAM,SAAS,CAAC;IAChC,aAAa,MAAM,eAAe,CAAC,EAAA,CAAG,KAAK,OAAO;KAChD,MAAM,EAAE;KACR,QAAQ,EAAE;IACZ,EAAE;GACJ;GACA,KAAK,iBAAiB,KAAK,IAAI;GAC/B;EACF;EAEA,IAAI,MAAM,SAAS,YAAY,MAAM,YAAY,aAAa;GAC5D,KAAK,QAAQ,KAAK;IAChB,UAAU,KAAK,iBAAiB,KAAK,IAAI,IAAI,KAAK,iBAAiB;IACnE,KAAK;GACP,CAAC;GACD;EACF;EAEA,IAAI,mBAAmB,KAAK,GAAG;GAC7B,KAAK,uBAAuB,KAAK;GACjC;EACF;EAEA,IAAI,cAAc,KAAK,GAAG;GACxB,KAAK,kBAAkB,KAAK;GAC5B;EACF;EAEA,IAAI,SAAS,KAAK,GAAG;GACnB,KAAK,iBAAiB;GACtB,KAAK,gBAAgB,MAAM;GAC3B,KAAK,aAAa,MAAM,SAAS;GACjC,KAAK,eAAe,MAAM,kBAAkB;GAC5C,KAAK,kBAAkB,MAAM,eAAe;GAC5C,KAAK,gBAAgB,MAAM,aAAa;GACxC,KAAK,kBAAkB,MAAM,UAAU;GACvC;EACF;CAGF;;;;;;;;CASA,QAAwB;EACtB,IAAI,KAAK,SAAS,MAChB,MAAM,IAAI,MACR,4JAEF;EAGF,MAAM,WAAW,KAAK,MAAM,KAAK,MAAM,SAAS;EAKhD,MAAM,kBAAkB,KAAK,MAC1B,KAAK,MAAM,EAAE,IAAI,CAAC,CAClB,QAAQ,MAAM,EAAE,SAAS,CAAC,CAAC,CAC3B,KAAK,MAAM,CAAC,CACZ,KAAK;EAER,OAAO;GACL,MAAM,KAAK;GACX,WAAW,KAAK;GAChB,OAAO,KAAK;GACZ,eAAe,mBAAmB,KAAK;GACvC,iBAAiB,UAAU,cAAc;GACzC,OAAO;IACL,aAAa,KAAK,YAAY,gBAAgB;IAC9C,cAAc,KAAK,YAAY,iBAAiB;IAChD,cAAc,KAAK;IACnB,YAAY,KAAK;IAEjB,UAAU,KAAK,iBAAiB,KAAK,MAAM;GAC7C;GACA,SAAS,KAAK;GAGd,SAAS,KAAK,kBAAkB,CAAC,KAAK;EACxC;CACF;CAIA,uBACE,OACM;EACN,MAAM,YAAY,KAAK,MAAM;EAC7B,MAAM,aAAuB,CAAC;EAC9B,MAAM,oBAAgC,CAAC;EAEvC,KAAK,MAAM,SAAS,MAAM,QAAQ,SAAS;GACzC,IAAI,YAAY,KAAK,GAAG;IACtB,WAAW,KAAK,MAAM,IAAI;IAC1B;GACF;GACA,IAAI,eAAe,KAAK,GAAG;IACzB,MAAM,OAAiB;KACrB,MAAM,MAAM;KACZ,WAAW,YAAY,MAAM,IAAI;KACjC,QAAQ,MAAM;KACd,MAAM,MAAM;KACZ,QAAQ;KACR,SAAS;KACT;KACA,WAAW,KAAK,aAAa;IAC/B;IACA,KAAK,aAAa,KAAK,IAAI;IAC3B,KAAK,aAAa,IAAI,MAAM,IAAI,IAAI;IACpC,kBAAkB,KAAK,IAAI;IAC3B;GACF;EAIF;EAEA,KAAK,MAAM,KAAK;GACd;GACA,MAAM,WAAW,KAAK,EAAE,CAAC,CAAC,KAAK;GAC/B,WAAW;GACX,YAAY,MAAM,QAAQ,eAAe;EAC3C,CAAC;CACH;CAEA,kBACE,OACM;EACN,MAAM,UAAU,MAAM,QAAQ;EAK9B,IAAI,OAAO,YAAY,UAAU;EAEjC,KAAK,MAAM,SAAS,SAAS;GAC3B,IAAI,CAAC,kBAAkB,KAAK,GAAG;GAE/B,MAAM,OAAO,KAAK,aAAa,IAAI,MAAM,WAAW;GACpD,IAAI,CAAC,MAIH;GAGF,KAAK,SAAS,MAAM;GACpB,KAAK,UAAU,MAAM,YAAY;GACjC,KAAK,aAAa,OAAO,MAAM,WAAW;EAC5C;CACF;AACF;;;;;;;;;AAUA,eAAsB,gBACpB,QACyB;CACzB,MAAM,UAAU,IAAI,kBAAkB;CACtC,WAAW,MAAM,SAAS,QACxB,QAAQ,QAAQ,KAAK;CAEvB,OAAO,QAAQ,MAAM;AACvB;;;;;;;;;;;;;AC1NA,gBAAuB,gBACrB,QACyC;CACzC,IAAI,SAAS;CAGb,OAAO,YAAY,MAAM;CAEzB,WAAW,MAAM,SAAS,QAAQ;EAChC,UAAU;EAKV,IAAI;EACJ,QAAQ,aAAa,OAAO,QAAQ,IAAI,OAAO,IAAI;GACjD,MAAM,OAAO,OAAO,MAAM,GAAG,UAAU,CAAC,CAAC,KAAK;GAC9C,SAAS,OAAO,MAAM,aAAa,CAAC;GACpC,IAAI,KAAK,WAAW,GAAG;GACvB,MAAM,aAAa,IAAI;EACzB;CACF;CAKA,MAAM,WAAW,OAAO,KAAK;CAC7B,IAAI,SAAS,SAAS,GACpB,MAAM,aAAa,QAAQ;AAE/B;;;;;;;;;AAUA,SAAS,aAAa,MAA2B;CAC/C,IAAI;EAEF,OAAO;GAAE,IAAI;GAAM,OADL,KAAK,MAAM,IACF;GAAG,SAAS;EAAK;CAC1C,SAAS,KAAK;EACZ,OAAO;GACL,IAAI;GACJ,OAAO,eAAe,QAAQ,MAAM,IAAI,MAAM,OAAO,GAAG,CAAC;GACzD,SAAS;EACX;CACF;AACF;;;;;;;;;AC/BA,IAAa,eAAb,cAAkC,MAAM;CAGpB;CAFlB,YACE,SACA,aACA;EACA,MAAM,OAAO;EAFG,KAAA,cAAA;EAGhB,KAAK,OAAO;CACd;AACF;;;;AC3DA,SAAS,mBAAmB,MAAgB,MAAc,QAAyB;CACjF,IAAI,CAAC,QAAQ;CACb,KAAK,MAAM,SAAS,QAClB,KAAK,KAAK,MAAM,KAAK;AAEzB;;;;;AAMA,SAAS,iBACP,MACA,MACA,OACM;CACN,IAAI,UAAU,KAAA,GAAW;CACzB,IAAI,OAAO,UAAU,WAAW;EAC9B,IAAI,OAAO,KAAK,KAAK,IAAI;EACzB;CACF;CACA,KAAK,KAAK,MAAM,OAAO,KAAK,CAAC;AAC/B;;AAGA,SAAgB,sBACd,MACA,QACM;CACN,mBAAmB,MAAM,gBAAgB,OAAO,UAAU;CAC1D,mBAAmB,MAAM,gBAAgB,OAAO,UAAU;CAC1D,mBAAmB,MAAM,aAAa,OAAO,OAAO;CAEpD,iBAAiB,MAAM,gBAAgB,OAAO,SAAS;CACvD,iBAAiB,MAAM,WAAW,OAAO,KAAK;CAC9C,iBAAiB,MAAM,qBAAqB,OAAO,cAAc;CACjE,iBAAiB,MAAM,YAAY,OAAO,MAAM;CAChD,iBAAiB,MAAM,WAAW,OAAO,KAAK;CAC9C,iBAAiB,MAAM,oBAAoB,OAAO,aAAa;CAC/D,iBAAiB,MAAM,WAAW,OAAO,KAAK;CAC9C,iBAAiB,MAAM,cAAc,OAAO,QAAQ;CACpD,iBAAiB,MAAM,qBAAqB,OAAO,cAAc;CACjE,iBAAiB,MAAM,eAAe,OAAO,QAAQ;CACrD,iBAAiB,MAAM,oBAAoB,OAAO,YAAY;CAC9D,iBAAiB,MAAM,mBAAmB,OAAO,YAAY;CAC7D,iBAAiB,MAAM,wBAAwB,OAAO,gBAAgB;CACtE,iBAAiB,MAAM,0BAA0B,OAAO,kBAAkB;CAC1E,iBACE,MACA,+BACA,OAAO,sBACT;CACA,iBAAiB,MAAM,WAAW,OAAO,KAAK;CAC9C,iBAAiB,MAAM,gBAAgB,OAAO,SAAS;CAEvD,IAAI,OAAO,gBAAgB,OAAO,aAAa,SAAS,GACtD,KAAK,KAAK,kBAAkB,OAAO,aAAa,KAAK,GAAG,CAAC;CAG3D,IAAI,OAAO,mBAAmB,OAAO,gBAAgB,SAAS,GAC5D,KAAK,KAAK,qBAAqB,OAAO,gBAAgB,KAAK,GAAG,CAAC;CAGjE,iBAAiB,MAAM,uBAAuB,OAAO,eAAe;CACpE,iBAAiB,MAAM,yBAAyB,OAAO,iBAAiB;CACxE,iBAAiB,MAAM,4BAA4B,OAAO,oBAAoB;CAC9E,iBAAiB,MAAM,4BAA4B,OAAO,oBAAoB;CAC9E,iBAAiB,MAAM,UAAU,OAAO,IAAI;CAC5C,iBAAiB,MAAM,eAAe,OAAO,QAAQ;CACrD,iBACE,MACA,wCACA,OAAO,+BACT;CACA,iBACE,MACA,kCACA,OAAO,0BACT;AACF;;;;;;;AAQA,SAAgB,UAAU,QAA2C;CACnE,MAAM,OAAiB;EACrB;EACA,OAAO;EACP;EACA;EACA;CACF;CAEA,sBAAsB,MAAM,MAAM;CAElC,OAAO;AACT;;;;;;;AAQA,SAAgB,eACd,QACA,SAAiD,CAAC,GACxC;CACV,MAAM,OAAiB;EAAC;EAAM;EAAQ;EAAmB;CAAM;CAC/D,MAAM,iBAAiB,OAAO,kBAAkB;CAChD,sBAAsB,MAAM;EAC1B,GAAG;EACH;CACF,CAAC;CACD,OAAO;AACT;;;;;;;;;;;;;;;;;;ACvGA,MAAM,qBAAqB,MAAS;;;;;;AAOpC,MAAM,gBAAgB;;;;;;;;;;;;;AA+BtB,eAAsB,YACpB,QACwB;CACxB,MAAM,SAAS,OAAO,UAAU;CAChC,MAAM,OAAO,UAAU,MAAM;CAM7B,MAAM,gBAJgB,OAAO,kBAAkB,QAK3C,MAAM,QAAQ,KAAK,OAAO,GAAG,eAAe,CAAC,IAC7C;CAEJ,MAAM,MAA0C;EAC9C,GAAG,QAAQ;EACX,GAAG,OAAO;CACZ;CACA,IAAI,eAEF,IAAI,oBAAoB;CAG1B,MAAM,QAAQ,MAAM,QAAQ,MAAM;EAChC,KAAK,OAAO,OAAO,QAAQ,IAAI;EAC/B;EACA,OAAO;GAAC;GAAU;GAAQ;EAAM;EAIhC,UAAU;CACZ,CAAC;CAKD,IAAI,WAAW;CACf,IAAI,iBAAwC;CAC5C,MAAM,YAAY,OAAO,aAAa;;;;;CAMtC,MAAM,+BAA+B;EACnC,IAAI,gBAAgB,aAAa,cAAc;EAC/C,iBAAiB,iBACT,SAAS,OAAO,SAAS,GAC/B,aACF;CACF;CAEA,MAAM,eAAe,iBAAiB;EACpC,WAAW;EACX,SAAS,OAAO,SAAS;EACzB,uBAAuB;CACzB,GAAG,SAAS;CAGZ,MAAM,gBAAgB;EACpB,SAAS,OAAO,SAAS;EACzB,uBAAuB;CACzB;CACA,OAAO,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;CAKhE,MAAM,eAAyB,CAAC;CAChC,MAAM,QAAQ,YAAY,MAAM;CAChC,MAAM,QAAQ,GAAG,SAAS,UAAkB;EAC1C,aAAa,KAAK,KAAK;CACzB,CAAC;CAED,MAAM,kBAAkB,IAAI,SAAiB,YAAY;EACvD,MAAM,iBAAiB,QAAQ,aAAa,KAAK,EAAE,CAAC;EACpD,MAAM,QAAQ,GAAG,OAAO,QAAQ;EAGhC,MAAM,QAAQ,GAAG,SAAS,QAAQ;CACpC,CAAC;CAKD,MAAM,OAAO,IAAI,SAGb,YAAY;EACd,IAAI,UAAU;EACd,MAAM,YACJ,UACA,WACG;GACH,IAAI,SAAS;GACb,UAAU;GAEV,aAAa,YAAY;GACzB,IAAI,gBAAgB,aAAa,cAAc;GAC/C,OAAO,QAAQ,oBAAoB,SAAS,OAAO;GACnD,QAAQ;IAAE;IAAU;GAAO,CAAC;EAC9B;EAEA,MAAM,GAAG,UAAU,MAAM,WAAW,SAAS,MAAM,MAAM,CAAC;EAE1D,MAAM,GAAG,eAAe,SAAS,MAAM,IAAI,CAAC;CAC9C,CAAC;CAGD,MAAM,UAAU,YAAY;EAC1B,IAAI,CAAC,eAAe;EACpB,IAAI;GACF,MAAM,GAAG,eAAe;IAAE,WAAW;IAAM,OAAO;GAAK,CAAC;EAC1D,QAAQ,CAGR;CACF;CAKA,OAAO;EACL,QAAQ,MAAM;EACd;EACA;EACA,gBAAgB;EAChB;CACF;AACF;;;;;;;;;;;;;;;;;AAkBA,SAAS,SAAS,OAAqB,QAA8B;CACnE,IAAI,MAAM,QAAQ,KAAA,GAAW;CAC7B,IAAI;EAEF,QAAQ,KAAK,CAAC,MAAM,KAAK,MAAM;CACjC,QAAQ;EACN,IAAI;GAEF,MAAM,KAAK,MAAM;EACnB,QAAQ,CAER;CACF;AACF;;;;;;;;;;;;;;AC/LA,eAAsB,cACpB,QACkC;CAClC,MAAM,UAAU,KAAK,IAAI;CACzB,MAAM,UAAU,MAAM,YAAY,MAAM;CAExC,MAAM,UAAU,IAAI,kBAAkB;CACtC,MAAM,YAA2B,CAAC;CAClC,MAAM,cAAkC,CAAC;CAEzC,IAAI;EACF,WAAW,MAAM,UAAU,gBAAgB,QAAQ,MAAM,GACvD,IAAI,OAAO,IAAI;GACb,QAAQ,QAAQ,OAAO,KAAK;GAC5B,UAAU,KAAK,OAAO,KAAK;EAC7B,OACE,YAAY,KAAK;GACf,MAAM,OAAO;GACb,OAAO,OAAO,MAAM;EACtB,CAAC;EAIL,MAAM,CAAC,EAAE,UAAU,UAAU,UAAU,MAAM,QAAQ,IAAI,CACvD,QAAQ,MACR,QAAQ,eACV,CAAC;EAED,MAAM,cAAkC;GACtC;GACA;GACA;GACA;GACA,UAAU,QAAQ,SAAS;GAC3B,YAAY,KAAK,IAAI,IAAI;EAC3B;EAEA,IAAI;EACJ,IAAI;GACF,OAAO,QAAQ,MAAM;EACvB,SAAS,KAAK;GAEZ,MAAM,IAAI,aACR,0CAFc,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,KAG7D,WACF;EACF;EAEA,OAAO;GAAE;GAAM;GAAa;EAAU;CACxC,UAAU;EACR,MAAM,QAAQ,QAAQ;CACxB;AACF;;AAGA,MAAa,oBAA6D;CACxE,IAAI;CACJ,KAAK;AACP"}