Logging

Logging is runtime-owned observation. Agents can emit behavior logs through session.log, but the host chooses sinks, levels, redaction, and file destinations.

Core Sinks

import {
  ConsoleLogSink,
  MemoryLogSink,
  type HarnessLoggingConfig,
} from "@harness-kernel/core/runner/logging";

const logging: HarnessLoggingConfig = {
  sinks: [new ConsoleLogSink({ level: "info" })],
  redact: {
    keys: ["apiKey", "authorization"],
  },
};

Attach logging to the session store:

const store = await createHarnessSessionStore({
  agent: { definition: agent },
  providers: [new OpenAIProvider()],
  defaultModel: "openai/gpt-5.1",
  logging,
});

JSONL File Logs

import { JsonlFileLogSink } from "@harness-kernel/logging-file";

const logging = {
  sinks: [new JsonlFileLogSink({ path: ".harness-kernel/logs/runtime.jsonl", level: "info" })],
};

JsonlFileLogSink is an optional runtime module. It is separate from run storage; use it for operational logs.

Agent Logs

Tools, hooks, modes, and context providers can call:

session.log.info("Escalation recorded.", { ticketId: "T-123" });
session.log.warn("Missing customer account id.");
session.log.error(new Error("Tool failed."));

The host still owns where those records go.

Error Records

Runtime failure logs use the same canonical error shape as events and status, but logs keep the internal view after redaction. Public surfaces such as stream events and lastError are sanitized by errorPolicy; log sinks receive HarnessLogRecord.error with code, category, severity, recoverable, name, message, and stack when available.

Events are the session timeline. Logs are operational diagnostics. A model provider failure, for example, produces an ErrorEvent/RunFailedEvent for the timeline and ModelCallFailedLog/RunFailedLog for diagnostics.

Sandbox Logs

Sandbox activity is logged as runtime diagnostics:

• sandbox opened and closed;
• command execution started;
• command execution completed with exit code, signal, timeout flag, and duration;
• command execution failed before a result could be returned.

These records apply to every sandbox implementation, including NoopSandbox, LocalSandbox, and DockerSandbox. They are separate from tool timeline events: a BashTool call can produce ToolStartEvent/ToolEndEvent for the session timeline and sandbox logs for the host’s operational view.