Lark Feishu Ops

Lark Feishu Ops

Use this as the single main-agent entry for Feishu/Lark platform operations. Keep product, technical, and business judgment in the parent thread. Treat embedded lark-cli guidance as authoritative and load it only for the requested domain.

Read references/feishuops-protocol.md after dispatch chooses FeishuOps, or when the exact compact request, lifecycle, evidence, continuity, or waiting contract is needed.

Preflight

Run the read-only doctor before platform use or when runtime state may have drifted:

python3 ../../scripts/lark_feishu_ops_doctor.py --json
python3 ../../scripts/lark_feishu_ops_doctor.py --repo <repo> --json

Surface update, official-guidance sync, global-skill unload, and installed-cache refresh actions to the user; do not execute them without explicit authorization. Treat ~/.agents/skills as shared: Codex-only unload must block without mutation there, and relocation or all-Agent removal requires separate approval. Missing npx alone is not a runtime blocker when lark-cli and embedded skills are healthy. Doctor is no-write by default: it may read an existing daily update cache but persists a fresh check only with explicit --write-update-cache authorization.

Dispatch Policy

Main-agent direct lark-cli is allowed only when all of these are true:

• The action is in the explicit read allowlist.
• The operation is read-only, bounded, and easy to validate.
• The request is single-domain, profile-stable, and requires no side effect.
• No auth/profile mutation, raw OpenAPI, broad pagination, large download, or high-risk confirmation is required.
• The user did not explicitly ask for FeishuOps, subagents, or delegated execution.

Caller hints never downgrade derived risk. Unknown, raw, write, and high-risk-write actions are not direct eligible. A blocked/unmapped guidance source also denies direct routing.

Escalate to FeishuOps when the request explicitly asks for delegation or writes, sends, creates, updates, deletes, is cross-domain, multi-step, raw-OpenAPI-heavy, permission-sensitive, large, or needs bounded follow-up work. If the user explicitly requested FeishuOps/subagent routing, do not silently fall back to direct execution.

FeishuOps Protocol

After dispatch chooses FeishuOps:

1. Read references/feishuops-protocol.md.
2. Run the agent-context helper to normalize risk and select trusted cli_embedded_skill or focused CLI fallback metadata.
3. Spawn fresh FeishuOps work with fork_turns: none; pass only the compact request and minimum context capsule.
4. Confirm any active-agent reuse against current Codex runtime state; persisted registry state is never authoritative.
5. Keep one platform operation bounded. Return discovered resources as result.next_resources unless expansion was explicitly requested.
6. Return question-focused evidence to the parent and leave final judgment there.

All current embedded guidance is discovered dynamically with lark-cli skills list/read. Never trust a caller-provided skill path, inject_as, or replacement command. Missing guidance uses focused help/schema; unknown domains stay blocked unless raw OpenAPI is explicitly authorized. Meeting domains prefer embedded lark-meeting, with the legacy domain skills used only on older CLI inventories. Preserve identity/profile and no-cache rules.

Continuity

Continuity is schema-v2 metadata only: bounded identifiers, resource type/revision, affinity, identity/profile, risk, timestamps/expiry, freshness, and provenance classifications. It excludes request/evidence bodies, command arguments, tokens, contacts, mail, table rows, excerpts, and raw provenance.

Limits are 64 snapshots, 32 resource references per snapshot, 256 UTF-8 bytes per retained string, 32 KiB per file, 30 minutes active idle, and 24 hours maximum TTL. Sensitive domains default to no-cache; cache_policy: disabled opts out; purge removes persisted metadata. require_refetch: true overrides cached freshness.

Operator Boundaries

• Preserve identity/profile and confirmation requirements from lark-cli. Omitted values remain unknown; apply prepared cli_execution arguments, validate returned identity/profile, and never retry user work as bot.
• Do not guess Feishu IDs, expose secrets, or silently expand resource scope.
• Do not globally load all lark-* skills into the parent context.
• Do not treat project-local agent metadata as automatic Codex registration.
• Do not replace repository-specific source snapshot/freeze tooling.
• Do not refresh plugin cache, unload global skills, change auth/profile, or mutate Feishu resources without the matching authorization.

Validation

After development-source changes, run the focused tests and validators from the repository root. Release generation and installed-cache refresh remain separate, main-agent-owned actions.

Common Mistakes

• Treating any unrecognized action as a safe read.
• Injecting a local SKILL.md path instead of CLI-embedded guidance.
• Spawning fresh FeishuOps work with inherited thread history.
• Reusing registry metadata without current runtime confirmation.
• Persisting evidence bodies instead of bounded continuity metadata.
• Confusing update or cache recommendations with authorization.