feat(kimi): new adapter for kimi.com (21 commands)

[LeoLin990405]

Summary

Greenfield web adapter for Kimi (Moonshot AI's web chat at kimi.com). Ships 21 commands across 4 layers, mirroring the Grok adapter's structure.

Coverage

Layer	Commands
Connectivity	status
Chat lifecycle	new history detail read send ask
Per-message	copy-message regenerate react share
Sidebar / global UI	mode sidebar-toggle view-all-history settings account model
Browser-side storage	storage-keys storage-get cookies idb-list

Special: mode switches Kimi's 7 work modes

mode ppt|docs|deep-research|websites|sheets|agent-swarm|code — navigates
to Kimi's specialized work surfaces.

E2E verified

• status → logged in as "登月者"
• history → 5 chats with proper UUIDs
• account → "登月者 Allegretto"
• mode → all 7 modes listed
• storage-keys → access_token, _tea*, _gcl_ls
• cookies → 8 JS-visible cookies (httpOnly auth deliberately not shown)
• idb-list → AiTopia DB
• model (current) → "K2.6 思考"

Selector strategy

Kimi labels its icons via <svg role="img" name="X"> (Copy/Refresh/Like/Dislike/Share_a/Send/LeftBar/Down_b). The clickBySvgNameScript helper finds the svg, walks up to the nearest clickable ancestor, and dispatches the full pointer chain.

Known limitation

Per-message commands (copy-message / regenerate / react / share)
need the page to be on a chat URL already; follow-up will add --conv <id>
to allow navigation first (mirroring Grok's pattern). Workaround:
opencli browser kimi open <chat-url> first.

Test plan

• CI passes
• Verify on a fresh Kimi install (history returns EmptyResult cleanly)
• Verify mode <name> navigates correctly without auth issues

[LeoLin990405]

Kimi deep-audit summary (3 commits, 26 commands)

Same deep-audit pass landed in PR #1801 / #1798 / #1799 / #1800 now applied to Kimi.

Final coverage: 26 commands across 4 layers

Layer	Commands
Connectivity	status
Chat lifecycle	new history detail read send ask
Per-message	copy-message regenerate react (like/dislike) share
Modes (7 specialized work surfaces)	mode (ppt/docs/deep-research/websites/sheets/agent-swarm/code)
Sidebar / global UI	sidebar-toggle view-all-history settings account model upgrade sign-out dismiss-banner
Templates / history	templates [--mode] history-rename
Browser-side storage	storage-keys storage-get cookies idb-list

E2E verified

• status → logged in as "登月者" with full profile
• history → 5 chats with proper UUIDs
• account → "登月者 Allegretto"
• mode → all 7 modes routable
• templates --mode deep-research → 4 cards across 3 categories (社科人文/商业财经/科技创新)
• storage-keys (52 LS keys) / cookies (8 visible) / idb-list (AiTopia)
• model → "K2.6 思考"
• upgrade → clicked
• all write-class commands print clean --yes false dry-runs

Known structural limitation (documented)

Per-message commands (copy-message/regenerate/react/share/read) and detail can navigate to a chat URL via --conv <id> but Kimi's React app does not trigger the messages fetch on programmatic navigation. The .message-list element mounts but has clientHeight: 0 and zero children; no /v1/messages requests appear in performance entries. Auth (access_token in localStorage, 8 cookies) is present and correct; this is purely Kimi's intentional lazy-load gate, likely keyed on user interaction events the headless navigation doesn't emit.

Workaround: Manually navigate to the chat URL in your real Chrome tab first, then run per-message commands without --conv.

What's NOT wrapped (and why)

• Per-conv More 3-dot menu: opens a Vue teleported popover (.popover-menu is just the trigger; actual items mount elsewhere) and the click event doesn't surface a stable selector. Listed conversation rows + history-rename cover the common case.
• Voice input: low CLI value (audio).
• Grammarly extension: third-party.

[LeoLin990405]

Update — render bug FIXED + per-message commands fully functional (commit `46535a10`)

Diagnosed via DOM probe + network audit. The previous "known limitation" (per-message commands return empty with --conv) is RESOLVED:

Root cause (dual bug):

1. Kimi's React app only triggers the /api/messages fetch when the URL has ?chat_enter_method=history (a query-param sentinel). Plain /chat/<id> loads the title via cookies but leaves message container at clientHeight=0.
2. Messages render in .chat-content-list as .chat-content-item-user / .chat-content-item-assistant, NOT .message-list (legacy scaffolding).

Fix: Append ?chat_enter_method=history in maybeNavigateConv() + detail(); update readKimiTurns() to enumerate .chat-content-item and .segment children.

E2E verified live on kimi.com:

opencli kimi read --conv "19e7475c..."  →  User: "电犀牛是啥"
                                            Assistant: "搜索网页 电犀牛 电动车 品牌 2026 15 个结果..."
                                            (full multi-turn output)
opencli kimi detail "19e7475c..."        →  same complete result
opencli kimi copy-message --conv         →  real assistant text extracted

Final state: all 26 commands functional, including per-message ones. The "known limitation" in earlier summary can be deleted from PR description.