Lark Event

⚙️ Cowork / Claude Desktop 执行环境说明（自动注入）

本技能依赖本地 lark-cli（@larksuite/cli，可用 command -v lark-cli 定位）及其 ~/.lark-cli 登录态（应用密钥存于 macOS keychain）。

在 Cowork 中运行任何 lark-cli 命令时，必须在本地 macOS 上执行（使用 Desktop Commander 的 start_process / interact_with_process，或其它本地 shell 工具），不要用隔离的 Linux 沙箱 mcp__workspace__bash——沙箱里没有 lark-cli、也读不到 keychain。 执行前确保 npm 全局 bin 目录（npm prefix -g 输出目录下的 bin）在 PATH 中。

（在 Claude Code 中可忽略本说明，lark-cli 在本机 shell 直接可用。）

Lark Events

Prerequisite: Read ../lark-shared/SKILL.md first for authentication, --as user/bot switching, Permission denied handling, and safety rules.

Core commands

Common flags

Examples

# Default: stream every event for the key (no filter, no projection)
lark-cli event consume im.message.receive_v1 --as bot

# Grab one sample event to inspect payload shape
lark-cli event consume im.message.receive_v1 --max-events 1 --timeout 30s --as bot

# Run for 10 minutes then auto-exit
lark-cli event consume im.message.receive_v1 --timeout 10m --as bot

# Consume multiple EventKeys concurrently (one shape per process, no dispatcher)
lark-cli event consume im.message.receive_v1          --as bot > receive.ndjson &
lark-cli event consume im.message.reaction.created_v1 --as bot > reaction.ndjson &
wait

Call flow

1. lark-cli event list --json → pick a legal key
2. lark-cli event schema <key> --json → read resolved_output_schema + jq_root_path to determine field paths
3. lark-cli event consume <key> [--jq '<expr>'] → consume

Subprocess contract

Ready marker

event consume's stderr emits a fixed line [event] ready event_key=<key>. Parent processes should block on stderr until this line appears, then start reading stdout. Do not fall back to sleep.

stdin EOF = graceful exit

event consume treats stdin close as a shutdown signal (wired for AI subprocess callers). Bounded runs are exempt: when --max-events or --timeout is set (> 0), stdin EOF is ignored and the run exits only via its own bound, timeout, or SIGTERM. For unbounded runs, < /dev/null / nohup / systemd's default StandardInput=null will cause an immediate graceful exit (stderr reason: signal). To keep an unbounded run alive:

• Feed stdin a source that never EOFs: < <(tail -f /dev/null)
• Or run bounded: --max-events N / --timeout D

Exit codes & reason

On exit, the last stderr line is [event] exited — received N event(s) in Xs (reason: ...).

Startup and runtime failures emit a structured JSON envelope on stderr: {"ok":false,"error":{"type","subtype","param","message","hint",...}} (the envelope may also carry top-level identity / _notice siblings). Parse error.type / error.subtype to branch (e.g. missing_scope carries a missing_scopes list), error.param to find the offending flag, and error.hint for the recovery action — do not regex-match message text.

Orchestrators should treat reason: limit/timeout/signal (all exit 0) as "business completion" and non-zero as "failure".

Never kill -9

Avoid kill -9 on consume processes: for EventKeys with a PreConsume hook (those that register server-side subscriptions via OAPI), kill -9 skips the OAPI unsubscribe and leaks server-side subscriptions (symptoms: "subscription already exists" on restart, duplicate event delivery). Prefer SIGTERM or closing stdin.

One consume, one EventKey (multi-key = multi-shell)

The command takes exactly one positional argument; k1,k2 and wildcards are unsupported. Listening to N keys means N subprocesses — this is intentional:

• One shape per process stdout; no dispatcher logic required in the AI
• Fault isolation (one key failing doesn't affect others)
• Independent --as / --jq / --max-events / --timeout per key

All N consumers share a single bus daemon (UDS local IPC), so the overhead is small

Writing jq via schema

event schema <key> --json is the source of truth for writing --jq. Four things to look at:

(1) Where fields start — see jq_root_path

• Value "." → fields are at the top level, write .chat_id
• Value ".event" → fields are inside a V2 envelope, write .event.chat_id

(2) Field list and types — see resolved_output_schema.properties.<name>

Each field carries type / description, and some also have format. Snippet (from event schema im.message.receive_v1 --json):

{
  "chat_id":     {"type":"string", "format":"chat_id",      "description":"Chat ID, prefixed with oc_"},
  "sender_id":   {"type":"string", "format":"open_id",      "description":"Sender open_id, prefixed with ou_"},
  "create_time": {"type":"string", "format":"timestamp_ms", "description":"Send time as ms-epoch string"}
}

(3) Field semantics — see the format tag

Lark-defined semantic tags (not JSON Schema's standard format). Common values: open_id / chat_id / message_id / timestamp_ms / email. Purpose: distinguish "same string type, different meanings" fields so you can reverse-lookup via API or convert formats.

(4) Decoded state — read the field's description

event consume runs Process hooks that may pre-decode some payload fields (flattening V2 envelopes, rendering .content to plain text, etc.) — behavior differs from raw OAPI. Always read the field's description before writing jq, especially for generic field names like content / data / body / payload.

Why it matters: blindly applying fromjson to an already-decoded text field makes jq error on every event and silently drop it — the consumer looks alive but emits nothing, with only a single WARN line buried on stderr. (This is the general behavior: any jq runtime error skips the event with a one-line WARN; the loop does not abort.)

Don't shortcut the schema: when projecting event schema --json with jq, do not strip .description from properties — that's the field that tells you whether a field is already decoded. Dump the full property objects, not just keys.

Aside: --param's valid parameters also live in the schema — the params section lists name / type / required / enum / default / description; section missing = this key accepts no --param.

Topic index