@odla-ai/harness 0.3.0 → 0.5.0

package/dist/{chunk-C5VQI2IF.js → chunk-FVOMJKWK.js}

@@ -1,6 +1,6 @@
 import {
   HARNESS_PROTOCOL_VERSION
-} from "./chunk-3QP4VDQS.js";
+} from "./chunk-LNQNFGQC.js";
 
 // src/protocol.ts
 var CONTROL = /[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f]/;
@@ -90,4 +90,4 @@ export {
   encodeAgentInput,
   makeHarnessEvent
 };
-//# sourceMappingURL=chunk-C5VQI2IF.js.map
+//# sourceMappingURL=chunk-FVOMJKWK.js.map

package/dist/{chunk-GKDKIU4P.js → chunk-K76I2TCQ.js}

@@ -1,10 +1,10 @@
 import {
   encodeAgentInput,
   parseAgentOutput
-} from "./chunk-C5VQI2IF.js";
+} from "./chunk-FVOMJKWK.js";
 import {
   HARNESS_PROTOCOL_VERSION
-} from "./chunk-3QP4VDQS.js";
+} from "./chunk-LNQNFGQC.js";
 
 // src/container.ts
 import { execFile, spawn } from "child_process";
@@ -548,4 +548,4 @@ export {
   stageWorkspacePair,
   safeWorkspaceLabel
 };
-//# sourceMappingURL=chunk-GKDKIU4P.js.map
+//# sourceMappingURL=chunk-K76I2TCQ.js.map

package/dist/{chunk-3QP4VDQS.js → chunk-LNQNFGQC.js}

@@ -6,4 +6,4 @@ export {
   HARNESS_PROTOCOL_VERSION,
   DEFAULT_AI_ROUTE
 };
-//# sourceMappingURL=chunk-3QP4VDQS.js.map
+//# sourceMappingURL=chunk-LNQNFGQC.js.map

package/dist/chunk-LNQNFGQC.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/types.ts"],"sourcesContent":["import type { ChatInput, OracleResponse } from \"@odla-ai/ai\";\n\n/** Current JSONL protocol version exchanged between a runner and an agent container. */\nexport const HARNESS_PROTOCOL_VERSION = 1 as const;\n\n/** Default control-plane route used for model inference requested by coding agents. */\nexport const DEFAULT_AI_ROUTE = \"coding\" as const;\n\n/** Lifecycle state reported for a harness task and its active attempt. */\nexport type HarnessTaskStatus =\n  | \"queued\"\n  | \"running\"\n  | \"cancel_requested\"\n  | \"completed\"\n  | \"failed\"\n  | \"cancelled\";\n\n/** Lifecycle state of one execution attempt for a task. */\nexport type HarnessAttemptStatus = HarnessTaskStatus;\n\n/** Trusted or untrusted participant that emitted a harness event. */\nexport type HarnessActor = \"operator\" | \"runner\" | \"agent\" | \"model\" | \"system\";\n\n/** Resource and isolation limits enforced while an untrusted task executes. */\nexport interface HarnessPolicy {\n  network: \"none\";\n  timeoutMs: number;\n  maxOutputBytes: number;\n  maxPatchBytes: number;\n}\n\n/** Immutable task instructions and execution policy delivered with a lease. */\nexport interface HarnessTaskSpec {\n  taskId: string;\n  attemptId: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  policy: HarnessPolicy;\n  parentAttemptId?: string | null;\n  checkpointSeq?: number | null;\n}\n\n/** Time-bound assignment authorizing a runner to execute one task attempt. */\nexport interface HarnessLease {\n  protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n  leaseId: string;\n  generation: number;\n  expiresAt: number;\n  task: HarnessTaskSpec;\n}\n\n/** Runner-supplied event before the control plane assigns sequence and task metadata. */\nexport interface HarnessEventInput {\n  eventId: string;\n  kind: string;\n  actor: HarnessActor;\n  payload: unknown;\n  createdAt: number;\n}\n\n/** Persisted, ordered event associated with a specific task attempt. */\nexport interface HarnessEvent extends HarnessEventInput {\n  seq: number;\n  taskId: string;\n  attemptId: string;\n}\n\n/** List-view metadata for a task and its current attempt. */\nexport interface HarnessTaskSummary {\n  taskId: string;\n  attemptId: string;\n  appId: string;\n  env: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  status: HarnessTaskStatus;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** List-view metadata for one attempt, including retry ancestry and runner ownership. */\nexport interface HarnessAttemptSummary {\n  attemptId: string;\n  taskId: string;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  status: HarnessAttemptStatus;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** Complete task view including attempts, events, result, and generated patch. */\nexport interface HarnessTaskDetail extends HarnessTaskSummary {\n  attempts: HarnessAttemptSummary[];\n  events: HarnessEvent[];\n  patch: string | null;\n  result: unknown;\n}\n\n/** Public control-plane view of a registered harness runner. */\nexport interface HarnessRunnerView {\n  runnerId: string;\n  appId: string;\n  env: string;\n  name: string;\n  createdAt: number;\n  lastSeenAt: number | null;\n  revokedAt: number | null;\n}\n\n/** Credential-free normalized model request sent from an agent through the runner. */\nexport interface HarnessInferenceRequest {\n  requestId: string;\n  /** Exact initial, follow-up, or resume command whose budget this call consumes. */\n  interactionId?: string;\n  call: ChatInput;\n}\n\n/** Normalized model response plus auditable provider, policy, and token metadata. */\nexport interface HarnessInferenceResponse {\n  requestId: string;\n  response: OracleResponse;\n  receipt: {\n    provider: string;\n    model: string;\n    policyVersion: number;\n    inputTokens: number;\n    outputTokens: number;\n    /**\n     * USD charged for this call, priced against the model the control plane\n     * actually resolved.\n     *\n     * ABSENT when the live catalog has no price for that model — never zero.\n     * An unpriced call is unknown spend, and reporting it as free is what let\n     * a goal's `maxUsd` look enforced while nothing enforced it. The runtime\n     * cannot compute this itself: it asks for `brokered` and only the control\n     * plane knows which model answered.\n     */\n    costUsd?: number;\n  };\n}\n\n/** Bounded, content-free session activity emitted by a Code runtime. Message\n * bodies are projected separately into the app's owner-private odla-db chat. */\nexport type CodeSessionEventData =\n  | { type: \"message\"; actor: \"agent\" | \"system\"; body: string }\n  | { type: \"diagnostic\"; level: \"error\"; message: string }\n  | { type: \"thinking\"; available: true; durationMs: number }\n  | { type: \"tool\"; phase: \"started\"; tool: HarnessToolName }\n  | { type: \"tool\"; phase: \"completed\"; tool: HarnessToolName; ok: boolean; durationMs: number }\n  | {\n      type: \"usage\"; provider: string; model: string;\n      inputTokens: number; outputTokens: number; durationMs: number;\n      interactionId?: string; interactionTokens?: number; interactionMaxTokens?: number;\n      /** USD for this call; absent when the model is unpriced, never zero. */\n      costUsd?: number;\n      /** Cumulative USD for this owner interaction, when every call in it was\n       *  priced. Absent the moment one was not, so a partial total can never be\n       *  mistaken for the whole. */\n      interactionCostUsd?: number;\n    }\n  | {\n      type: \"status\"; status: \"running\" | \"idle\" | \"failed\" | \"checkpointed\";\n      durationMs?: number;\n    };\n\n/** Registry-assigned cursor and timestamp for an owner-visible Code event. */\nexport type CodeSessionEvent = CodeSessionEventData & {\n  eventId: string; sequence: number; createdAt: number;\n};\n\n/** Closed set of effects an agent container may request from its trusted broker. */\nexport type HarnessToolName =\n  | \"sandbox.read\"\n  | \"sandbox.list\"\n  | \"sandbox.search\"\n  | \"sandbox.overview\"\n  | \"sandbox.where_is\"\n  | \"sandbox.who_imports\"\n  | \"sandbox.who_touches\"\n  | \"sandbox.apply_patch\"\n  | \"sandbox.run_recipe\";\n/** Correlated, structured tool request emitted by an untrusted agent container. */\nexport interface HarnessToolRequest {\n  requestId: string;\n  tool: HarnessToolName;\n  input: Record<string, unknown>;\n}\n/** Bounded tool result returned to an agent after trusted policy evaluation. */\nexport interface HarnessToolResponse {\n  requestId: string;\n  ok: boolean;\n  content: string;\n  details?: Record<string, unknown>;\n}\n\n/** Validated JSONL message emitted by an untrusted agent container. */\nexport type HarnessAgentOutput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"event\";\n      kind: string;\n      payload?: unknown;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.request\";\n      requestId: string;\n      call: ChatInput;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.request\" } & HarnessToolRequest)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.complete\";\n      status: \"completed\" | \"failed\" | \"cancelled\";\n      result?: unknown;\n    };\n\n/** JSONL command or inference result written by the trusted runner to an agent. */\nexport type HarnessAgentInput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"task.start\";\n      task: HarnessTaskSpec;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.response\";\n      requestId: string;\n      response: OracleResponse;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.response\" } & HarnessToolResponse)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.cancel\";\n      reason: string;\n    };\n\n/** Terminal attempt report submitted by a runner to the control plane. */\nexport interface HarnessCompletion {\n  status: \"completed\" | \"failed\" | \"cancelled\";\n  result?: unknown;\n  patch?: string;\n  error?: string;\n}\n\n/** Operations a credentialed runner may perform against the harness control plane. */\nexport interface HarnessControlPlane {\n  lease(workspaces: string[]): Promise<HarnessLease | null>;\n  heartbeat(attemptId: string, leaseId: string): Promise<{ cancelRequested: boolean; expiresAt: number }>;\n  appendEvents(attemptId: string, leaseId: string, events: HarnessEventInput[]): Promise<void>;\n  infer(attemptId: string, leaseId: string, request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n  complete(attemptId: string, leaseId: string, completion: HarnessCompletion): Promise<void>;\n}\n\n/** Trusted inference bridge used to keep model credentials outside agent containers. */\nexport interface HarnessAiConnection {\n  infer(request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n}\n\n/** Trusted tool boundary. Implementations must evaluate CaMeL policy before effects. */\nexport interface HarnessToolBroker {\n  execute(\n    context: { lease: HarnessLease; workspaceDir: string; signal?: AbortSignal },\n    request: HarnessToolRequest,\n  ): Promise<HarnessToolResponse>;\n}\n"],"mappings":";AAGO,IAAM,2BAA2B;AAGjC,IAAM,mBAAmB;","names":[]}

package/dist/{chunk-U2OFMIAW.js → chunk-UVGZHNLW.js}

@@ -4,10 +4,10 @@ import {
   stageWorkspace,
   stageWorkspacePair,
   verifyContainerEngineBoundary
-} from "./chunk-GKDKIU4P.js";
+} from "./chunk-K76I2TCQ.js";
 import {
   HARNESS_PROTOCOL_VERSION
-} from "./chunk-3QP4VDQS.js";
+} from "./chunk-LNQNFGQC.js";
 
 // src/workspace-digest.ts
 import { createHash } from "crypto";
@@ -205,6 +205,16 @@ function createCodeRuntimeControlClient(options) {
       }
       await call(`/registry/code/runtime/sessions/${validSessionId(sessionId)}/chat/events`, { eventId, event });
     },
+    recallMemories: async (sessionId, subjects, limit) => {
+      const response2 = await call(
+        `/registry/code/runtime/sessions/${validSessionId(sessionId)}/recall`,
+        { subjects: [...subjects], limit }
+      );
+      return Array.isArray(response2.memories) ? response2.memories : [];
+    },
+    rememberMemory: async (sessionId, memory) => {
+      await call(`/registry/code/runtime/sessions/${validSessionId(sessionId)}/remember`, memory);
+    },
     reportSessionFailure: async (sessionId, message2) => {
       if (!message2.trim() || message2.length > 2e3) throw new TypeError("invalid Code session failure");
       await call(`/registry/code/runtime/sessions/${validSessionId(sessionId)}/failure`, { message: message2 });
@@ -1362,6 +1372,9 @@ async function handleCodeRuntimeInference(input) {
     call: request.call
   });
   state.tokens += response2.receipt.inputTokens + response2.receipt.outputTokens;
+  const { costUsd } = response2.receipt;
+  if (costUsd === void 0) state.costKnown = false;
+  else state.costUsd += costUsd;
   await input.event({
     type: "usage",
     provider: response2.receipt.provider,
@@ -1371,7 +1384,9 @@ async function handleCodeRuntimeInference(input) {
     durationMs: Date.now() - startedAt,
     interactionId: command.commandId,
     interactionTokens: state.tokens,
-    interactionMaxTokens: metadata.maxTokensPerInteraction
+    interactionMaxTokens: metadata.maxTokensPerInteraction,
+    ...costUsd === void 0 ? {} : { costUsd },
+    ...state.costKnown ? { interactionCostUsd: state.costUsd } : {}
   }).catch(() => void 0);
   return {
     protocolVersion: HARNESS_PROTOCOL_VERSION,
@@ -2287,10 +2302,16 @@ async function startGoalPursuit(input) {
     attempt: async ({ prompt }) => {
       const result = await input.attempt(prompt);
       return {
-        // The runtime charges tokens through the control plane's own
-        // per-interaction reservation, so the goal budget bounds ATTEMPTS here
-        // and the token ceiling is enforced where the credential lives.
-        tokens: 0,
+        // What the attempt actually spent, so the runner's token_budget and
+        // cost_budget checks can be reached. This used to be a hardcoded 0 with
+        // no cost at all, which made maxTokens and maxUsd unreachable while
+        // callers reasonably read them as hard ceilings.
+        //
+        // costUsd is omitted rather than zeroed when any call in the attempt
+        // was unpriced: the runner only enforces a cost budget while the cost
+        // is known, and a zero would make it enforce against a lie.
+        tokens: result.tokens ?? 0,
+        ...result.costUsd === void 0 ? {} : { costUsd: result.costUsd },
         ...result.status === "failed" ? { error: result.error ?? "attempt failed" } : {}
       };
     },
@@ -2509,7 +2530,7 @@ var CodePiRuntimeEngine = class {
       recipeAuthorization: this.options.recipeAuthorization
     }, lease, metadata.role));
     const startedAt = Date.now();
-    const interaction = { tokens: 0, noticeEmitted: false };
+    const interaction = { tokens: 0, noticeEmitted: false, costUsd: 0, costKnown: true };
     const inference = createCodeRuntimeInference({
       command,
       metadata,
@@ -2547,7 +2568,11 @@ var CodePiRuntimeEngine = class {
       await this.#diagnostic(command, active, detail);
       await this.#failure(command, active, detail);
     }
-    return result;
+    return {
+      ...result,
+      tokens: interaction.tokens,
+      ...interaction.costKnown ? { costUsd: interaction.costUsd } : {}
+    };
   }
   /** Report every brokered effect as it starts and finishes. */
   #observed(command, active, broker) {
@@ -2641,4 +2666,4 @@ export {
   runGoal,
   CodePiRuntimeEngine
 };
-//# sourceMappingURL=chunk-U2OFMIAW.js.map
+//# sourceMappingURL=chunk-UVGZHNLW.js.map