Skip to content

telemetry.md

Latest commit

 

History

History
92 lines (65 loc) · 7.63 KB

telemetry.md

File metadata and controls

92 lines (65 loc) · 7.63 KB
Raw

Telemetry

All automatic telemetry uploads are disabled by default and each requires its own opt-in: TUI usage events (telemetry.enabled), runtime performance metrics (telemetry.metrics), and automatic error diagnostics (telemetry.diagnostics). Opting in to usage events does not authorize the other two channels. Login, model requests, update checks, and user-submitted feedback are separate product actions and are not controlled by these switches.

Turn it on or off

Add this to the active profile's config.yaml, then restart MCode. Builds from this repository and the published npm CLI @minimax-ai/code@0.4.12 default to ~/.minimax/config.yaml (or ~/.minimax-<profile>/config.yaml when a profile is selected). Overrides can change this path; see Accounts and data to locate it. The telemetry controls below describe this repository's source build and are not a guarantee of feature parity with that npm release.

telemetry:
  enabled: true      # TUI usage events (anonymous, described below)
  metrics: false     # Runtime counters, gauges, and histograms
  diagnostics: false # Account-linked TUI and LLM error diagnostics

Remove a setting or set it to false to turn that channel off. Either environment variable below turns all channels off and takes precedence over the config file:

MCODE_DISABLE_TELEMETRY=1 mcode
DO_NOT_TRACK=1 mcode

Inspect the effective setting of every channel with mcode telemetry status. Run mcode telemetry preview to see a representative decoded usage-event request. Preview does not send a request. When usage telemetry is disabled, preview shows request: null, independently of the metrics and diagnostics settings.

Usage-event data sent

The HTTP body is application/x-www-form-urlencoded with two fields:

Field Value
data Base64-encoded JSON envelope described below
ext crc=<number>, an integrity checksum of the encoded data

The decoded JSON envelope contains:

Field Value
identities.$identity_cookie_id Fresh random ID for this event
distinct_id The same fresh random ID
lib.$lib js
lib.$lib_method code
lib.$lib_version MCode version
properties Common and event-specific fields listed below
type track
event Event name from the table below
time Unix timestamp in milliseconds

Every properties object includes:

Field Value
surface tui
os Node.js platform name, such as darwin, linux, or win32
region cn or en
build_env dev, test, staging, or prod
app_version MCode version

Event-specific fields are limited to:

  • tui_launch: launch_type (cold, hot).
  • login_click, logout_click: no extra fields.
  • login_result: source (agent_web, agent_desktop, openplatform, mcode_tui, mcode_cli); result_type (1 success, 2 failure); fail_reason (empty for success, 1 server, 2 network, 3 cancelled, 4 other, 5 OAuth/authorization); login_type (google, mobile, wechat, apple, minimax_sso, minimax_oauth).
  • btw_session_lifecycle: phase (opened, closed); duration_bucket (not_applicable, under_1m, 1m_to_5m, 5m_to_30m, over_30m); exit_reason (empty, ctrl_c, ctrl_d, navigation, replaced).
  • chat_send: chat_type (chat, agent_team, claw, hermes, IM); is_first_message (0, 1); is_attachment (text, attachment).
  • slash_command_menu_view, at_command_menu_view: chat_type (values above).
  • slash_command_click: chat_type; command_type (skill, new_chat, summarize, plan_mode, goal_mode, other).
  • at_command_click: chat_type; command_type (plugins, goal_mode, plan_mode, file, directory).

MCode does not send account IDs, device IDs, workspace paths or names, session IDs, model names, prompts, responses, filenames, command text, plugin names, or credentials through this channel. The receiver requires an identity-shaped envelope, so the client creates a fresh random event ID for each request. It is never persisted or reused and cannot link two events on its own. MCode does not add an account authorization header to these requests.

As with any network request, the receiving server can observe transport metadata such as the source IP address. The client does not add that value to the event payload. mcode telemetry preview displays the decoded envelope.

Usage-event destinations

The destination depends on region and build environment:

  • China production: https://data.hailuoai.com/meerkat-reporter/api/report?project=MiniMaxAgent
  • Global production: https://data.hailuo.ai/meerkat-reporter/api/report?project=MiniMaxAgent
  • China non-production: https://bigdata-test.xingyeai.com/meerkat-reporter/api/report?project=MiniMaxAgent
  • Global non-production: https://bigdata-test.talkie-ai.com/meerkat-reporter/api/report?project=MiniMaxAgent

The client keeps pending events only in memory and does not write them to disk. This repository does not define or verify server-side retention. Keep telemetry disabled when that policy does not meet your requirements.

Runtime performance metrics

telemetry.metrics: true enables the built-in cloud metrics transport: metric names, timestamps, counter/gauge/histogram values, and low-cardinality labels (runtime owner and mode, version, and per-instrument dimensions such as model, tool, or outcome). No account credential is attached. Production destinations are https://agent.minimax.cn/matrix/api/v1/metrics/batch (China) and https://agent.minimax.io/matrix/api/v1/metrics/batch (global). Metrics stay in memory; when the channel is disabled, no cloud reporter is created.

Automatic error diagnostics

telemetry.diagnostics: true authorizes both TUI incident reports and LLM request-failure reports. Both additionally require a signed-in account: the transport uses the account's Bearer token and a user_id query parameter, so these reports are account-linked even though their contents are minimized and encrypted. The minimization schemas are described in TUI capability coverage. Both use /minimax-cloud/api/v1/observability/desktop-errors/batch on the regional MiniMax host. When the channel is disabled, TUI incidents are written as local-only files (7 days / 200 files) that are never uploaded, and LLM failure reports are dropped before buffering.

Server-side retention for any channel is not defined or verified by this repository. Login, model requests, update checks, and user-submitted feedback have separate network behavior described in TUI capability coverage.