The +chat-search row in lark-im SKILL.md described the search as
"by keyword and/or member open_ids", which doesn't match the real
flag names (--query, --member-ids). Naming them inline avoids
agents guessing --keyword from the prose, matching the style
already used by +chat-messages-list.
Change-Id: Ife8668d9b13ee66711bc4e81a7b2bcc7f05d9586
Add IM flag shortcut commands to lark-cli, enabling users to create, list, and cancel bookmarks on messages and threads via +flag-create, +flag-list, and +flag-cancel.
Change-Id: I8f87f0eadf83fb59b024a3b9fe67b23d363abe0a
- Assemble applinks via net/url to ensure proper encoding
- Normalize message position values across more numeric types
- Avoid leaking null message_app_link; assemble when missing
- Update unit tests to assert URL semantics and cover edge cases
Change-Id: Ic473cb563c8a648c4f6677c32b25b9f371a0f84e
Adds a new top-level safety section "数据真实性与操作合规" to the
lark-mail skill via the canonical generation pipeline:
- skill-template/domains/mail.md (source) — adds the section to the
domain introduction file that gen-skills.py renders into SKILL.md.
- skills/lark-mail/SKILL.md (regenerated product) — produced by
`make gen-skills project=mail` from larksuite-cli-registry against
the modified mail.md source.
Why both files: skills/lark-mail/SKILL.md is auto-generated from
skill-template/domains/mail.md + registry-conf/skill-meta.yaml +
output/from_meta/mail.json. Editing only SKILL.md would be reverted on
the next `make gen-skills` run because SKILL.md has no AUTO-GENERATED
markers and falls into the "no markers -> overwrite whole file" branch
in scripts/gen-skills.py.
The section adds 3 hard constraints on agent behavior:
- empty result is a valid answer; do not fabricate IDs or placeholders
- explicit action preview before destructive write operations
(delete / trash / batch_trash / cancel_scheduled_send / rules.*)
- reversible modifications (label / read state / folder move) are
exempt from the preview requirement
Addresses recurring evaluation failures (c03/c04/c06/c09/c14/c19~c24/c40)
where the agent fabricated IDs or auto-executed destructive operations.
The --as flag displayed (default "bot"), (default "user"), or
(default "auto") in help text, but ResolveAs() never uses the cobra
default — it resolves identity via credential config and auto-detect.
The displayed default misled users into thinking a fixed identity was
used when --as was omitted.
Set cobra default to empty string so no (default ...) suffix appears.
Also remove "auto" from visible options since --as auto is equivalent
to omitting --as entirely.
Change-Id: I51ba550a6697eb3675a29f5cee4d0010e0a1cc16
Users who install or upgrade lark-cli via make install, go install, or
direct binary download end up with a binary but no AI agent skills,
degrading agent UX. This PR adds a startup-time skills version drift
notice (injected into JSON envelope _notice.skills, mirroring the
existing _notice.update pattern) and unifies lark-cli update's skills
sync across all three branches (npm / manual / already-latest) with
stamp-based dedup, so any explicit update invocation keeps skills in
sync regardless of how the binary was installed.
Changes:
- new internal/skillscheck package: notice (StaleNotice + atomic
pending), stamp (~/.lark-cli/skills.stamp), skip (CI / DEV /
non-release / LARKSUITE_CLI_NO_SKILLS_NOTIFIER opt-out), check
(synchronous Init)
- cmd/root.go: rename setupUpdateNotice -> setupNotices, compose
output.PendingNotice returning {update?, skills?}; capture
build.Version locally before spawning the async update goroutine
- cmd/update/update.go: add runSkillsAndStamp helper with stamp-based
dedup; rewire the three branches through shared applySkillsResult /
emitSkillsTextHints helpers; add skills_status block to --check JSON
output as a pure report (no side effects)
- internal/update: export IsRelease(version) bool / IsCIEnv() bool
for cross-package reuse; refresh UpdateInfo.Message to append
', run: lark-cli update' so both notices recommend the same fix
- AGENTS.md: add Notification Opt-Outs section documenting
LARKSUITE_CLI_NO_UPDATE_NOTIFIER and LARKSUITE_CLI_NO_SKILLS_NOTIFIER
- internal/binding/types.go: bump default exec-provider timeout from
5s to 10s (out-of-scope flake fix for TestResolveExecRef_JSONResponse
under heavy parallel test load)
AI agents running inside OpenClaw / Hermes were routinely creating a parallel
app via `config init --new` instead of binding to the agent's existing app,
because every "not configured" hint and several deny errors hard-coded
`config init` regardless of workspace. Once bound, the same agents could
silently grant themselves user identity (impersonation) without the user
ever seeing a risk message in chat.
Changes:
- Introduce `core.NotConfiguredError` / `NoActiveProfileError` /
`reconfigureHint` helpers that branch on `CurrentWorkspace()`. In agent
workspaces they point at `lark-cli config bind --help` (a help page, not
a ready-to-run command) so AI must read the binding workflow and confirm
identity preset with the user before acting. In local terminals they
preserve the previous `config init --new` guidance.
- Migrate every `config init` hint that should be workspace-aware:
RequireConfigForProfile, default credential provider, credential provider
fallback, secret-resolve mismatch, config show, strict-mode entry-point
errors, default-as, profile use/rename/remove, auth list, doctor's
config_file check (which now also wraps the OS-level "no such file"
noise into the user-shaped "not configured" message).
- Refuse `config init` when run inside an OpenClaw / Hermes workspace by
default; add `--force-init` for the rare case the user genuinely wants
a parallel app. Without this guard, hint fixes were undone the moment
AI ignored them.
- Rewrite the strict-mode deny errors in cmd/auth/login.go, cmd/prune.go,
and internal/cmdutil/factory.go. The previous "AI agents are strictly
prohibited from modifying this setting" terminated AI reasoning while
providing no real gate. New errors point at `config strict-mode --help`
with the legitimate confirmation flow and explicitly note that switching
does NOT require re-bind. Integration test envelopes updated.
- Tighten `config bind --help` and `config strict-mode --help` to encode
the user-confirmation discipline directly: identity preset semantics
(bot-only vs user-default), "DO NOT switch without explicit user
confirmation", and a cross-reference clarifying that `config bind` is
for changing the underlying app while `config strict-mode` is the
policy-only switch (resolves an ambiguity an audit run found).
- Surface user-identity (impersonation) risk at every config write that
newly grants it, by reusing the canonical IdentityEscalationMessage
string from bind_messages.go:
- `noticeUserDefaultRisk` fires on flag-mode bind landing on
user-default, including the first-time case `warnIdentityEscalation`
misses (it requires a previous bot lock).
- `setStrictMode` warns when transitioning bot → user or bot → off
(newly permits user identity); stays quiet on narrowing changes
and on off → user (off already permitted user).
- Add tests: notconfigured_test.go (workspace branches),
init_guard_test.go (refuse + --force-init bypass), bind_warning_test.go
(user-default warning fires; bot-only does not), strict_mode_warning_test.go
(5 transitions covering both warn and no-warn paths).
Two follow-ups intentionally deferred: the keychain master-key hint at
internal/keychain/keychain.go:42 still suggests `config init` because the
keychain package can't import core (would be circular); fixing requires
either parameterizing the hint via callback or extracting workspace into
its own package. The lark-shared skill doc still tells AI to run
`config init` for first-time setup; updating the skill is in scope for
a follow-up PR.
Change-Id: I02273e044d9e061d211ceaa4f3ed5a3fb28325b3
* fix(auth): handle missing scopes and device flow improvements
* fix: remove redundant error return in login scope handler
* test(auth): rename test for zero interval default case
* fix: increase device code polling timeout from 180 to 600 seconds
* feat(base): support batch record get and delete
* fix(base): address batch record PR feedback
* docs(base): refine record skill routing
* refactor(base): use batch record get and delete only
* refactor(base): share record selection normalization
* docs(base): clarify record get field projection help
* feat(drive): pre-flight per-text-element byte limit for +add-comment
The open-platform comment API returns an opaque [1069302] Invalid or
missing parameters whenever a single reply_elements[i] text exceeds
its implicit byte budget. The error does not name which element failed
or that length is the cause, so callers resort to binary-search
debugging.
Empirically: Chinese text up to ~80 chars (~240 bytes) lands; ~130
chars (~390 bytes) fails. Set the pre-flight limit to 300 bytes which
sits safely inside the known-good zone.
- parseCommentReplyElements now rejects any text element whose UTF-8
byte length exceeds 300, with an ExitError naming the element index
(#N, 1-based) and both the rune and byte counts, plus an ErrWithHint
recommending the correct remediation (split into multiple text
elements — the comment UI renders them as one contiguous comment).
- The previous 1000-rune check is removed: it was too lenient (a
Chinese text under that cap would still fail server-side).
- skills/lark-drive/references/lark-drive-add-comment.md documents
the per-element limit and the correct split pattern so agents
avoid constructing oversized single elements upstream.
Addresses Case 12 in the 踩坑列表 doc.
* fix(drive): correct +add-comment hint to match actual escape coverage
`escapeCommentText` only expands `<` and `>` (each → 4 bytes via
`<` / `>`); `&` is intentionally left as-is. Both the over-limit
hint and the inline comment in `parseCommentReplyElements` previously
claimed `&` was also escaped, with a "4-5 bytes each" range that
implicitly assumed `&` (5 bytes) — a string of 300 `&` chars
would actually fit in the budget, but a user reading the hint would
think otherwise and pre-emptively split it.
Code:
- Hint string ends with `Note: '<' and '>' are HTML-escaped and
counted in their escaped form (4 bytes each).` (was: included `&`
and "4-5 bytes")
- Inline comment above the budget check now matches:
`escapeCommentText only expands '<' and '>' (each becomes 4 bytes:
< / >); '&' is intentionally left as-is.`
Tests (regression):
- New `300 ampersands accepted (escapeCommentText leaves '&' as-is)`
subtest pins that 300 `&` chars stay within budget. Without the fix
this also passed (function was always correct), but the hint was
lying — the test pins the budget contract loud and clear.
- New `TestParseCommentReplyElementsHintMatchesEscape` asserts the
hint string itself: must mention `'<' and '>'` / `4 bytes`, must NOT
mention `'&'` / `&` / `4-5 bytes`. Catches a future drift if
`escapeCommentText` is changed without updating the hint, or
vice-versa.
The skill md (`skills/lark-drive/references/lark-drive-add-comment.md`)
already had the right wording (`每个 < 或 > 占 4 字节`), so it was the
in-Go strings that drifted; this commit aligns code with doc.
* fix(drive): rewrite +add-comment length cap to match real server behavior
The original PR set a 300-byte per-element pre-flight check, justified
by the empirical pattern "~80 Chinese chars succeeds, ~130 fails". A
fresh round of probing the live `/open-apis/drive/v1/files/{token}/
new_comments` endpoint with a real docx shows that pattern does not
reproduce, and the actual contract is very different:
- 10000 ASCII / 10000 Chinese / 10000 '<' (escaped to 40000 bytes)
in a single text element: all OK
- 10001 of any of the above in a single text element: [1069302]
- 5000 + 5000 across two text elements (total 10000): OK
- 5000 + 5001 across two text elements (total 10001): [1069302]
- 4000 + 4000 + 4000 across three (total 12000): [1069302]
Two consequences:
1. The cap is *10000 runes total across all reply_elements text*, not
300 bytes per element. The old check rejected legitimate input
anywhere from ~100 to 10000 Chinese chars (≈100x too aggressive).
2. The hint that recommended "split the content across multiple
{\"type\":\"text\",\"text\":\"...\"} elements" was actively wrong —
splitting doesn't bypass a total cap. A user told to split a
10001-char message into 5000+5001 hits the same opaque [1069302].
This commit:
- Replaces `maxCommentTextElementBytes = 300` with
`maxCommentTotalRunes = 10000`. The constant's doc comment records
the probe matrix above so future maintainers know how it was
derived.
- Switches the measurement from `len(escapeCommentText(input.Text))`
to `utf8.RuneCountInString(input.Text)`. Server counts raw runes;
byte width and post-escape form are irrelevant. The escape itself
still happens — `<` and `>` still get rendered literally — but it
no longer participates in the length check.
- Tracks a running `totalRunes` across the whole reply_elements array
and bails at the first element that pushes the cumulative total
over the 10000-rune budget, with index reporting that points at the
offending element.
- Rewrites the over-cap hint to (a) name the actual 10000-rune budget,
(b) explicitly say splitting does NOT help, (c) drop the wrong
"comment UI still renders them as one contiguous comment" framing
that implied splitting was a workaround.
- Adds a `TestParseCommentReplyElementsHintForbidsSplitAdvice`
watchdog that fails if any future drift puts the discredited split
advice back into the hint.
Tests: 11 cases on TestParseCommentReplyElementsTextLength covering
single-element boundary (ASCII / Chinese / angle brackets at exactly
10000 and at 10001), multi-element total cap (5000+5000 OK, 5000+5001
rejected with index pointing at element #2), early-element-overshoot
indexing (first element at 10001 reports index #1, not the trailing
element), and mention_user not double-counting toward the cap.
Skill md updated: removes the 300-byte / "split into multiple
elements" advice; documents the 10000-rune total cap with a note that
the schema currently advertises 1-1000 chars and is out of date,
plus a procedure for re-probing if the server-side limit ever moves.
Manual API verification: rebuilt binary and posted comments at
boundary lengths — all OK cases (100 / 5000 / 10000 chars, 5000+5000
split) accepted by server; over-cap cases (10001 / 10100 single, and
5000+5001 split) rejected by the new pre-flight before reaching the
network.
---------
Co-authored-by: fangshuyu <fangshuyu@bytedance.com>
* feat(doc): expand callout type= shorthand into background-color and border-color
When users write <callout type="warning" emoji="📝"> without an explicit
background-color, the Feishu doc renders the block with no color. This
commit adds fixCalloutType() which maps the semantic type= attribute to
the corresponding background-color/border-color pair accepted by create-doc.
- warning → light-yellow/yellow
- info/note → light-blue/blue
- tip/success/check → light-green/green
- error/danger → light-red/red
- caution → light-orange/orange
- important → light-purple/purple
Explicit background-color or border-color attributes are always preserved.
The fix is applied via prepareMarkdownForCreate() in both +create and
+update paths, and also inside fixExportedMarkdown() for round-trip fidelity.
* refactor(doc): replace silent callout type→color injection with hint output
Per reviewer feedback (SunPeiYang996), silently rewriting user Markdown is
the wrong layer for this adaptation. The type→color mapping is not part of
the Feishu spec, and covert transforms make debugging harder.
Replace fixCalloutType() (which rewrote the Markdown) with WarnCalloutType()
which leaves the Markdown unchanged and instead writes a hint line to stderr
for each callout tag that has type= but no background-color, telling the user
the recommended explicit attributes to add:
hint: callout type="warning" has no background-color; consider: background-color="light-yellow" border-color="yellow"
Also fixes CodeRabbit feedback: the type= regex now accepts both single-quoted
and double-quoted attribute values (type='warning' and type="warning").
* fix(doc): harden background-color detection in WarnCalloutType
CodeRabbit flagged that the previous strings.Contains(attrs,
"background-color=") check missed forms like 'background-color =
"light-red"' with whitespace around the equals sign. Replace with a
regex that tolerates optional whitespace, and add a regression test.
* fix(doc): close real review gaps left over after rebase
PR #467's review thread had three substantive comments
(`fangshuyu-768`, 2026-04-21) that the prior reply messages claimed
were fixed in commit 7d4b556 — but that commit no longer exists on the
branch (lost in a rebase / squash), and the head still ships the
original buggy code. This commit makes the fixes real.
Three behavior fixes in shortcuts/doc/markdown_fix.go:
1. (#5) Tighten the type= and background-color= regex anchors. \b sits
at any word/non-word boundary, and `-` is a non-word char, so
`\btype=` also matched the suffix of `data-type=` — a tag like
`<callout data-type="warning">` would emit a bogus light-yellow
hint. Switched both regexes to `(?:^|\s)…` so a real attribute
separator is required. The same anchor on background-color closes
the symmetric case where a `data-background-color=` attribute
would silently suppress the real hint.
2. (#4) WarnCalloutType is now a fence-aware line walker. Previously
the regex ran over the entire markdown body, so a callout sample
inside a documentation code fence (```markdown … ```) would
generate a phantom stderr hint every time the docs mentioned the
feature. The walker tracks fence state via the existing
codeFenceOpenMarker / isCodeFenceClose helpers from
docs_update_check.go, which handle both backtick and tilde fences
per CommonMark §4.5.
3. (#3) Drop the ReplaceAllStringFunc-as-iterator pattern. The
previous code routed callout iteration through a rewrite primitive
whose rebuilt-string return value was discarded, then ran the same
regex a second time inside the callback to recover the capture
groups. New scanCalloutTagsForWarning helper uses
FindAllStringSubmatch — one pass, no thrown-away allocation,
intent matches the surface (read-only scan, not a mutator).
Tests: 5 new TestWarnCalloutType subtests pin each contract:
- data-type attribute does not trigger hint (#5)
- data-background-color does not suppress hint (#5, symmetric)
- callout inside backtick fence emits no hint (#4)
- callout inside tilde fence emits no hint (#4)
- callout after fence close still emits hint (#4, fence-state reset)
All 14 TestWarnCalloutType cases pass; go vet / golangci-lint
--new-from-rev=origin/main both clean.
* feat(base): add record read SOP guidance
1. Add a unified lark-base record read SOP for get/search/list routing, field projection, temporary view querying, pagination, matrix result binding, and link field reads.
2. Inline command-focused parameter guidance into +record-get, +record-search, and +record-list help, including examples, JSON shape, view scope, projection, and limit constraints.
3. Preserve base shortcut flag order in help output and add tests covering record read help guidance.
4. Remove the single-method record read skill references in favor of the unified SOP.
* test(base): remove stale record list fixture
* fix(base): scan record markdown output
* fix(base): fallback record markdown output
* fix(base): unify base token wording in shortcuts and skills
