docs(im): trim leaf references to non-derivable content

[luozhixiong01]

Summary

The nine leaf references under skills/lark-im/ restated large amounts of content that --help and schema already emit — parameter tables, field tables, response-shape examples, scope tables, command inventories. That content costs an agent context on every read and goes stale independently of the CLI. This PR rewrites each file to carry only what the authoritative sources cannot express, and drops ## 权限表 from SKILL.md.

Net effect: 27279 → 12167 tokens (−55.4%) across the nine files, with no measurable capability loss (see Test Plan).

Changes

• Rewrote nine skills/lark-im/references/ files to non-derivable content only; each now opens with a one-line pointer at --help / schema and keeps three kinds of content:
  • Gotchas the authority cannot state — cross-method asymmetries (reaction_type nesting differs across the four reaction methods), caller-identity constraints, counter-intuitive defaults (+threads-messages-list defaults to asc while +chat-messages-list defaults to desc), command-selection boundaries (+chat-search searches chats, +chat-list lists the ones you are in).
  • Contract and recovery semantics — retry_scope=none forbids replaying a write; meta.complete is the only proof a read was exhaustive.
  • HELP-GAP sections — facts the authority should carry but does not yet (e.g. the complete emoji_type key list). These are marked as explicit debt, not a permanent home.
• Removed ## 权限表 from skills/lark-im/SKILL.md. references/skill-quality-standard.md §8.1 lists "no scope/permission mapping table" as non-waivable, and all 36 scopes were verified retrievable from schema (36/36).
• Preserved the ## Safety Constraints sections in messages-send.md and messages-reply.md byte-for-byte — verified by section-level comparison against the base, not by keyword grep.

Per-file token reduction (tiktoken cl100k_base):

File	Before	After	Cut
lark-im-messages-send.md	4837	1928	60%
lark-im-feed-groups.md	4836	2085	57%
lark-im-messages-reply.md	4444	2030	54%
lark-im-messages-search.md	3014	1385	54%
lark-im-reactions.md	3030	1553	49%
lark-im-chat-messages-list.md	2434	1148	53%
lark-im-chat-create.md	2519	897	64%
lark-im-chat-search.md	1933	972	50%
lark-im-chat-list.md	2232	1175	47%

SKILL.md: 6506 → 5892 (permission table only).

Test Plan

• Unit tests pass (go test excluding E2E packages)

• go build ./...

• Skill eval, un-slimmed vs slimmed, three rounds each — real execution against the test tenant, 18 cases / 72 numbered judgment points, model gpt-5.5:

  Arm	Rounds (max 72)	Mean
  Un-slimmed	62 / 64 / 65	63.7
  Slimmed	65 / 62 / 66	64.3

  The two distributions overlap completely. At judgment-point level, zero true regressions (criterion: passes all three un-slimmed rounds and fails all three slimmed rounds). Ten of eighteen cases score full marks in both arms, covering idempotency contracts, prompt-injection isolation, approval semantics, the resource-download switch, read-completeness criteria, and feed rule semantics.

  Scope of that claim: this is "no loss was measured", not "equivalence was proven" — 15 of 72 judgment points are unstable across rounds, so differences on the order of 1–2 points are below the resolution of this experiment.

• Mechanical assertion that no non-derivable fact was lost: every identifier in the frozen must-keep baseline still present in all nine files. The assertion deliberately does not recompute derivability against a fresh oracle — a widened oracle would let a genuinely lost fact pass.

• Token floor gate on all nine files (34% floor, actual 47–64%)

Related Issues

• None

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 49423ddd-f471-402a-9771-19133ca843fe

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

review seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}