@capybara-hq/kernel 0.1.0-rc.0 → 0.1.0-rc.1

package/dist/log.js

@@ -1,9 +1,13 @@
+import { pino } from "pino";
+import { prettyFactory } from "pino-pretty";
 import { secretScalars } from "./config/fields.js";
+import { logContext } from "./log-context.js";
+import { createLokiStream } from "./log-loki.js";
 /** 日志级别，按严重程度递增。 */
 export const LOG_LEVELS = ["debug", "info", "warn", "error"];
 const LEVELS = LOG_LEVELS;
 const REDACTED = "[REDACTED]";
-const RESERVED = new Set(["ts", "level", "msg"]);
+const RESERVED = new Set(["time", "level", "msg"]);
 /** 部署方密钥的最短登记长度：过短的值打码会误伤普通文本。 */
 const MIN_SECRET_LEN = 4;
 /** 组织来源密钥的最短登记长度。 */
@@ -101,16 +105,64 @@ function expandSecret(value) {
     }
     return out;
 }
+/** Console format: `pretty` (pino-pretty, one line per record), `json`, or `auto` (pretty on a TTY). */
+export const LOG_FORMATS = ["auto", "pretty", "json"];
+/** Fields shown in the pretty prefix instead of the trailing object, in this order. */
+const PREFIX_FIELDS = ["reqId", "traceId", "jobId", "runId", "plugin"];
+/** pino-pretty's message format: `[role] [reqId=…] [plugin=…] msg`, each part only when present. */
+function prettyMessage(log, messageKey) {
+    const parts = [];
+    if (typeof log.role === "string")
+        parts.push(`[${log.role}]`);
+    for (const k of PREFIX_FIELDS) {
+        const v = log[k];
+        if (typeof v === "string" || typeof v === "number")
+            parts.push(`[${k}=${v}]`);
+    }
+    parts.push(String(log[messageKey] ?? ""));
+    return parts.join(" ");
+}
+function consoleStream(opts) {
+    if (opts.write) {
+        const write = opts.write;
+        // pino hands each record as one string ending in "\n"
+        return { write: (s) => write(s.endsWith("\n") ? s.slice(0, -1) : s) };
+    }
+    const format = opts.format ?? "auto";
+    if (!LOG_FORMATS.includes(format)) {
+        throw new Error(`invalid log.format "${format}" (expected ${LOG_FORMATS.join(", ")})`);
+    }
+    const out = opts.out ?? process.stdout;
+    const isTTY = opts.isTTY ?? process.stdout.isTTY === true;
+    const pretty = format === "pretty" || (format === "auto" && isTTY);
+    if (!pretty)
+        return { write: (s) => out.write(s) };
+    // In process (no worker thread): pino-pretty formats each line synchronously.
+    const format_ = prettyFactory({
+        colorize: isTTY,
+        colorizeObjects: isTTY,
+        singleLine: true,
+        // errors stay inside the one-line object instead of a multi-line stack block
+        errorLikeObjectKeys: [],
+        translateTime: "SYS:HH:MM:ss.l",
+        ignore: ["pid", "hostname", "role", "spanId", ...PREFIX_FIELDS].join(","),
+        messageFormat: (log, messageKey) => prettyMessage(log, messageKey),
+    });
+    return { write: (s) => out.write(format_(s)) };
+}
 /**
- * 创建结构化 JSON logger：每条日志写一行 `{ ts, level, msg, ...fields }`。
- * 字段不能覆盖 ts / level / msg；Error 序列化为 name / message / stack；
- * 经 `addSecrets` 登记的部署方密钥在消息与字段的任何字符串里都被打码；经 `addOrgSecrets` 登记的组织来源密钥
- * 只在字段里打码，不动顶层 msg；顶层 ts / level 从不打码。
- * `opts.orgSecretCapacity` 为组织来源密钥集合的容量（测试用）；`opts.onMatcherBuild` 在每次重建匹配器时调用（测试用）。
+ * Creates the kernel logger on pino. Each record is `{ time, level, msg, ...base, ...fields }` with
+ * `level` the label and `time` ISO 8601; fields cannot overwrite time / level / msg or the base
+ * fields; an Error becomes name / message / stack (and cause). Redaction happens before pino, so
+ * every output sees the same redacted record: deployment secrets (`addSecrets`) in msg and every
+ * field string, org secrets (`addOrgSecrets`) in fields only; level and time are never redacted.
+ * The console gets pino-pretty or JSON lines per `opts.format`; `opts.write` replaces the console
+ * with a JSON line sink.
  */
 export function createLogger(level = "info", opts = {}) {
     let current = parseLevel(level);
-    const write = opts.write ?? ((line) => process.stdout.write(`${line}\n`));
+    const base = opts.base ?? {};
+    const reserved = new Set([...RESERVED, ...Object.keys(base)]);
     const orgCapacity = opts.orgSecretCapacity ?? ORG_SECRET_CAPACITY;
     /** 部署方密钥（数量由配置决定，不淘汰）。 */
     const secrets = new Set();
@@ -136,12 +188,10 @@ export function createLogger(level = "info", opts = {}) {
         return deployMatcher(s);
     };
     const isSecret = (v) => secrets.has(v) || orgSecrets.has(v);
+    /** Redacts and serializes the fields to JSON (errors expanded, cycles cut). */
     const serialize = (entry) => {
         const seen = new WeakSet();
-        return JSON.stringify(entry, function (key, value) {
-            if (this === entry && RESERVED.has(key)) {
-                return key === "msg" && typeof value === "string" ? redactMsg(value) : value;
-            }
+        return JSON.stringify(entry, (_key, value) => {
             if (typeof value === "string")
                 return redact(value);
             if ((typeof value === "number" || typeof value === "bigint") && isSecret(String(value))) {
@@ -167,21 +217,42 @@ export function createLogger(level = "info", opts = {}) {
             return value;
         });
     };
+    const loki = opts.loki ? createLokiStream(opts.loki) : undefined;
+    const sink = pino({
+        level: "debug",
+        base,
+        messageKey: "msg",
+        timestamp: pino.stdTimeFunctions.isoTime,
+        formatters: { level: (label) => ({ level: label }) },
+        // request / job / trace fields of the current async context (log-context.ts)
+        mixin: () => ({ ...logContext() }),
+        // fields arrive already serialized: no pino serializers, no depth or size limits
+        serializers: { err: (e) => e },
+        depthLimit: Number.MAX_SAFE_INTEGER,
+        edgeLimit: Number.MAX_SAFE_INTEGER,
+    }, 
+    // level filtering is ours (setLevel); every stream takes everything that got this far
+    loki
+        ? pino.multistream([
+            { level: "debug", stream: consoleStream(opts) },
+            { level: "debug", stream: loki },
+        ])
+        : consoleStream(opts));
     const emit = (lvl, msg, fields) => {
         if (LEVELS.indexOf(lvl) < LEVELS.indexOf(current))
             return;
-        const entry = { ts: new Date().toISOString(), level: lvl, msg };
+        const entry = {};
         for (const [k, v] of Object.entries(fields ?? {}))
-            if (!RESERVED.has(k))
+            if (!reserved.has(k))
                 entry[k] = v;
-        let line;
+        let obj;
         try {
-            line = serialize(entry);
+            obj = JSON.parse(serialize(entry));
         }
         catch {
-            line = serialize({ ts: entry.ts, level: lvl, msg, logError: "unserializable fields" });
+            obj = { logError: "unserializable fields" };
         }
-        write(line);
+        sink[lvl](obj, redactMsg(msg));
     };
     return {
         debug: (msg, fields) => emit("debug", msg, fields),
@@ -192,6 +263,9 @@ export function createLogger(level = "info", opts = {}) {
             return current;
         },
         redact: (text) => redact(text),
+        flush: async () => {
+            await loki?.flush();
+        },
         setLevel(l) {
             current = parseLevel(l);
         },

package/dist/net/public-address.js

@@ -6,8 +6,9 @@ import { BlockList, isIP } from "node:net";
  * checked addresses only (see `safe-fetch.ts`), so a DNS answer that changes between check and
  * connect (DNS rebinding, e.g. `10.0.0.1.nip.io`) cannot reach a private address.
  *
- * The blocked ranges are those of `plugins/sandbox-docker/egress-proxy/proxy.mjs`, ported here
- * because the kernel must not import from a plugin. IPv6 addresses that embed an IPv4 address
+ * The blocked ranges are those of the sandbox-docker plugin's egress proxy
+ * (`plugins/sandbox-docker/egress-proxy/proxy.mjs` in capybara-hq/plugins), ported here because the
+ * kernel must not import from a plugin. IPv6 addresses that embed an IPv4 address
  * (v4-mapped `::ffff:0:0/96`, v4-compatible `::/96`, NAT64 `64:ff9b::/96`, 6to4 `2002::/16`) are
  * judged by the embedded IPv4 address; anything unparseable is refused. One addition over proxy.mjs:
  * SIIT IPv4-translated addresses (`::ffff:0:a.b.c.d`, `::ffff:0:0:0/96`, RFC 2765) are also judged by

package/dist/runtime/maintenance.js

@@ -1,5 +1,6 @@
 import { Queue, Worker } from "bullmq";
 import { sql } from "kysely";
+import { withJobContext } from "../log-context.js";
 import { bullConnection } from "./run-queue.js";
 /**
  * 内核维护队列（不含 `:`）：等待超时任务（`wait-timeout`，见 waits.ts）与周期性维护（输出存档清理、
@@ -104,7 +105,7 @@ export class Maintenance {
         if (this.worker || this.closed)
             return;
         this.workerConn = bullConnection(this.d.redisUrl, this.d.log);
-        this.worker = new Worker(MAINTENANCE_QUEUE_NAME, (job, token) => this.process(job, token), {
+        this.worker = new Worker(MAINTENANCE_QUEUE_NAME, (job, token) => withJobContext(MAINTENANCE_QUEUE_NAME, job, () => this.process(job, token)), {
             connection: this.workerConn,
             concurrency: 5,
         });

package/dist/runtime/run-worker.js

@@ -4,6 +4,7 @@ import { sql } from "kysely";
 import { normalizeStoredDefinition } from "../agents/definition.js";
 import { effectiveDefinition, NO_SELECTION, readSelection, sameModel, sameSelection, } from "../conversations/selection.js";
 import { ApiError } from "../errors.js";
+import { withJobContext } from "../log-context.js";
 import { pinnedFailureText, recordSkillsCatalog } from "../skills/catalog.js";
 import { RunError } from "./model-call.js";
 import { ProviderListError } from "./providers.js";
@@ -60,7 +61,7 @@ export class RunWorker {
         const runtime = this.d.config().runtime;
         this.conn = bullConnection(this.d.redisUrl, this.d.log);
         for (const queue of [RUN_QUEUE_NAME, SUBAGENT_RUN_QUEUE_NAME]) {
-            const worker = new Worker(queue, (job, token) => this.process(job, token), {
+            const worker = new Worker(queue, (job, token) => withJobContext(queue, job, () => this.process(job, token)), {
                 // the worker name is recorded on the job (BullMQ processedBy): which replica ran it
                 name: `${hostname()}-${process.pid}`,
                 connection: this.conn,

package/dist/runtime/usage-report.js

@@ -2,6 +2,7 @@ import { DelayedError, Job, Queue, UnrecoverableError, Worker } from "bullmq";
 import { sql } from "kysely";
 import { enabledExtensionsQuery } from "../extensions/enabled.js";
 import { deepFreeze, errMsg } from "../extensions/hooks.js";
+import { withJobContext } from "../log-context.js";
 import { bullConnection } from "./run-queue.js";
 /** Queue of `usage.report` deliveries (no `:` in the name). */
 export const USAGE_REPORT_QUEUE = "cap-kernel-usage-report";
@@ -256,7 +257,7 @@ export class UsageReporter {
         // replica that never enqueued (m-b)
         this.getQueue();
         this.workerConn = bullConnection(this.d.redisUrl, this.d.log);
-        this.worker = new Worker(USAGE_REPORT_QUEUE, (j, t) => this.process(j, t), {
+        this.worker = new Worker(USAGE_REPORT_QUEUE, (j, t) => withJobContext(USAGE_REPORT_QUEUE, j, () => this.process(j, t)), {
             connection: this.workerConn,
             concurrency: this.d.config().hooks.usageReportConcurrency,
             lockDuration: jobLockMs,

package/dist/tools/builtins/sandbox/common.d.ts

@@ -37,9 +37,11 @@ export declare const SYNC_MARKER = "/workspace/attachments/.capybara-sync.json";
 /** 附件文件名的上限（UTF-8 字节，常见文件系统单个文件名上限 255 字节，留出 `-N` 后缀的余量）。 */
 export declare const ATTACHMENT_NAME_MAX_BYTES = 200;
 /**
- * 沙箱工具时限里额外留出的余量（毫秒）：创建 / 恢复沙箱，以及命令超时后停命令的兜底——E2B 杀进程脚本没跑完时
- * 重启沙箱（杀进程最多约 20 秒，加 pause + 冷启动 connect 最多 120 秒，约 160 秒）。首次创建时拉取镜像不在其内
- * （部署时应预先拉好 `sandbox.image`，见 plugins/sandbox-docker/README.md）。
+ * Extra margin (ms) on top of a sandbox tool's time limit: creating / resuming the sandbox, and the
+ * fallback for stopping a command after its timeout (when the E2B kill script does not finish, the
+ * sandbox is restarted: up to about 20 s to kill, plus pause and a cold connect of up to 120 s, about
+ * 160 s). Pulling the image on first creation is not included (deployments pre-pull
+ * `sandbox.image`; see the sandbox-docker plugin's README in capybara-hq/plugins).
  */
 export declare const SANDBOX_SETUP_MARGIN_MS = 180000;
 /**

package/dist/tools/builtins/sandbox/common.js

@@ -22,9 +22,11 @@ const MARKER_MAX_BYTES = 1024 * 1024;
 /** 附件文件名的上限（UTF-8 字节，常见文件系统单个文件名上限 255 字节，留出 `-N` 后缀的余量）。 */
 export const ATTACHMENT_NAME_MAX_BYTES = 200;
 /**
- * 沙箱工具时限里额外留出的余量（毫秒）：创建 / 恢复沙箱，以及命令超时后停命令的兜底——E2B 杀进程脚本没跑完时
- * 重启沙箱（杀进程最多约 20 秒，加 pause + 冷启动 connect 最多 120 秒，约 160 秒）。首次创建时拉取镜像不在其内
- * （部署时应预先拉好 `sandbox.image`，见 plugins/sandbox-docker/README.md）。
+ * Extra margin (ms) on top of a sandbox tool's time limit: creating / resuming the sandbox, and the
+ * fallback for stopping a command after its timeout (when the E2B kill script does not finish, the
+ * sandbox is restarted: up to about 20 s to kill, plus pause and a cold connect of up to 120 s, about
+ * 160 s). Pulling the image on first creation is not included (deployments pre-pull
+ * `sandbox.image`; see the sandbox-docker plugin's README in capybara-hq/plugins).
  */
 export const SANDBOX_SETUP_MARGIN_MS = 180_000;
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@capybara-hq/kernel",
-  "version": "0.1.0-rc.0",
+  "version": "0.1.0-rc.1",
   "description": "Capybara kernel: config, database, plugin host and HTTP API of the Capybara agent platform",
   "license": "Apache-2.0",
   "repository": {
@@ -47,6 +47,8 @@
     "jose": "^6.2.12",
     "kysely": "^0.29.6",
     "pg": "^8.23.1",
+    "pino": "10.4.0",
+    "pino-pretty": "13.2.0",
     "reflect-metadata": "0.2.2",
     "rxjs": "^7.8.2",
     "semver": "^7.8.5",
@@ -62,7 +64,7 @@
     "@types/pg": "^8.23.1",
     "@types/semver": "^7.8.0",
     "@types/yauzl": "3.4.0",
-    "@capybara-hq/testing": "0.1.0-rc.0"
+    "@capybara-hq/testing": "0.1.0-rc.2"
   },
   "scripts": {
     "build": "tsc -b"