dsh-frostfin

dsh-frostfin (Moonglow Frostfin Whale): Kimi Code as DSH Agent Loop

English | 中文

"A wondrous aquatic creature altered by the influence of kuuvahki." — in-game archive entry, Moonglow Frostfin Whale

Kimi Code's brain, DeepSeek Harness's body.

frostfin is a DSH loop plugin: it replaces the driver of a DSH session wholesale with Kimi Code itself — connected directly over ACP (Agent Client Protocol), with no second agent paraphrasing in between. You get two things that used to be mutually exclusive:

• From Kimi: Kimi Code's own agent loop — planning, tool orchestration and thinking modes tuned for the kimi-for-coding model family. Not a single component is replaced, and upgrades ride along;
• From DSH: the entire ecosystem — Web UI, themes, session management, trajectory view, community plugins, and the Cordis substrate where "everything is a plugin, and every plugin is revertible".

The strengths of both, in one project.

The name

Layer one: the whale and the moon. DeepSeek's totem is a whale; Moonshot AI's Chinese name (月之暗面, "the dark side of the moon") comes from Pink Floyd's album The Dark Side of the Moon. What this plugin does, literally, is: let DeepSeek's whale be transformed by the power of the moon.

Layer two: the Frostfin Whale actually exists — in Genshin Impact. The base species is a whale from the seas of Nod-Krai, added to the bestiary in version 6.0 "Luna I"[^1][^2], and it breached the surface in that chapter's main-story cutscene[^6]. Version 6.7 "Luna VIII" (July 2026) introduced its lunar variant — the Moonglow Frostfin Whale[^3], whose habitat the in-game bestiary places on the moon map "Frost Moon", with the entry: "A wondrous aquatic creature altered by the influence of kuuvahki."[^4] (Kuuvahki — 月矩力 — is the moon's power in Nod-Krai lore.) An ordinary aquatic creature, altered by the moon's power — we scoured the dictionary and found no more precise metaphor: the whale is still that whale, the sea is still that sea — but the tides now answer to the moon.

Layer three: the easter egg closes the loop. On the far side of "Frost Moon", the Moonglow Frostfin Whale's home map, lies the region "Dark Side of the Moon" — both arrived in the same 6.7 "Luna VIII" release[^3][^5]. Light up that region's map and the achievement that pops is called "Any Colour You Like"[^5] — a track on The Dark Side of the Moon. The joke miHoYo buried and the source of Moonshot AI's company name are the same 1973 record. Three chains of tribute, closing a loop half a century back.

The bestiary also records a folk saying from Nod-Krai: "anyone who sees a pink Frostfin Whale will have their wish come true."[^4] May everyone who installs this plugin have their wish come true.

On precision: "altered by the influence of kuuvahki" describes the lunar variant, not the base species — so the full brand is Moonglow Frostfin Whale, the entry's true owner; the package name dsh-frostfin keeps the family name: short, searchable, early in the alphabet.

• dsh-moontide — moon pulls the tides, tides carry the whale; we later learned 6.7's moon map really does have a sea called the Moontide Sea[^5]
• dsh-moonwhale — the two mascots fused directly; the most literal
• dsh-moonsea — real astronomical term for a lunar mare on the far side; the grandest

[^1]: Bestiary: Frostfin Whale · Bilibili Genshin Wiki (habitat: Nod-Krai; introduced in Luna I) [^2]: "Luna I" version update notes (official, full text) · Bilibili Genshin Wiki (entry 9: new wildlife includes the Frostfin Whale) [^3]: "Luna VIII" version update notes (official, full text) · Bilibili Genshin Wiki (entry 8: new wildlife — the Moonglow Frostfin Whale) [^4]: Bestiary: Moonglow Frostfin Whale · Bilibili Genshin Wiki (habitat: Frost Moon; introduced in Luna VIII; source of the pink-whale folk saying) [^5]: Achievement set "Unfettered Crescent" · Bilibili Genshin Wiki (the achievement "Any Colour You Like": "Light up the Dark Side of the Moon map"; the same set mentions the Moontide Sea in several entries) [^6]: Archon Quest "Where the Moon Rises" · Bilibili Genshin Wiki (cutscene line: "a Frostfin Whale breaches the surface")

How it works

you ──► DSH Web UI (themes / ecosystem / session management / approvals)
          │  session events
    frostfin (a DSH plugin occupying the agent-loop slot)
          │  ACP (JSON-RPC over stdio)
    kimi acp subprocess ── all thinking, tool calls, replies

In DSH's architecture the agent loop itself is a replaceable plugin (ctx.agents.setFactory). frostfin registers its own factory, bridges every DSH session to a kimi acp subprocess, and translates the ACP event stream into DSH session events in real time — so typewriter-effect streaming, history replay, and approval dialogs are all native DSH experiences.

How this differs from existing approaches: projects like kimi-tide attach Kimi at the model or tool layer — DSH's main loop is unchanged and Kimi is the object being called; frostfin swaps the loop itself for Kimi Code — the one you're talking to in DSH is kimi.

Features

Every item below is covered by automated tests (unit + end-to-end, without the real kimi) and verified in a real browser:

• Loop bridge: DSH sessions driven by a kimi acp subprocess — streaming replies, thinking blocks, tool calls and plan blocks all render natively in DSH, with turn/step event discipline aligned to the native loop event-for-event
• Approval bridge: kimi's tool approval requests land on DSH's native approval dialog (first in the ecosystem) — the dialog's command preview lines up with the tool card's callId, and "allow" grants exactly that one call
• Question channel: kimi's AskUserQuestion pops a plugin-built multiple-choice modal — ACP has no question RPC, so kimi reuses the permission channel for questions, but DSH's approval dialog can't return which option was picked; questions therefore get their own path, independent of the permission policy. Skip/cancel maps to kimi's "user didn't answer" semantics — answers are never fabricated
• Image input: paste an image into the composer and it truly reaches kimi (bytes read from DSH's attachment store → base64 → ACP image block; kimi gates formats and compresses). If the bytes can't be read, a text placeholder stands in — images are never silently dropped
• Session lifecycle: resume across DSH restarts, self-healing kimi process crashes (the next prompt reconnects), attach any existing kimi session (/frostfin-attach or one click in the "Moonglow Frostfin Whale" tab, history replayed into the DSH log); permission mode and thinking level are remembered per kimi session and replayed automatically after a process restart
• Mode dispatch: adds the "Moonglow Frostfin Whale" mode and makes it the default — it runs on kimi, the standard mode runs the native loop, and neither disturbs the other; once a session is created, its driver is locked in and never silently swapped
• Model layer: the model selector shows kimi's real model list and selecting switches live; models configured in DSH (e.g. DeepSeek) are synced into kimi's config automatically, so kimi can run them directly
• Status bar & slash commands: the dock below the composer shows kimi's model / thinking level / permission mode / context usage / cwd & git branch in real time; /frostfin-mode, /frostfin-thinking, /frostfin-plan, /yolo, /auto switch directly; kimi's built-in /compact /status /usage /mcp /tasks /help pass through and execute as-is; in the command menu, plugin commands carry a [frostfin] tag and kimi commands a [kimi] tag, instantly distinguishable from DSH host commands
• Goal & plan: DSH's /goal drives kimi sessions (a host-level turn driver; the card can pause/edit/delete); kimi's plan mode works via /frostfin-mode plan (engine-enforced read-only, not a prompt convention)
• Revertibility: uninstalling undoes every registration (including the managed block synced into kimi's config.toml — your kimi config is restored verbatim); session bindings and the model cache under ~/.frostfin/ are deliberately kept so a reinstall picks up where you left off — to wipe them: rm -rf ~/.frostfin

Which permission mode matters

In frostfin sessions you only need to care about kimi's permission mode (/frostfin-mode <default|plan|auto|yolo>; shortcuts /yolo, /auto, /frostfin-plan). The two DSH-side switches can mostly be ignored, with one exception each:

• The composer's "Workspace Write" selector bounds DSH's native tools; kimi's tools run in kimi's own process, out of that sandbox's reach. Which to pick: stay on the default workspace-write (read-only is equally harmless — both are no-ops for kimi); the only one to avoid is danger-full-access — it sets the session's approval policy to never, injects an English policy notice into kimi's conversation, and, when the plugin is configured with permission: 'ask', auto-rejects every subsequent approval request (no dialog ever appears).
• DSH's approval policy (ask / never) only enters the answer chain when the plugin is configured with permission: 'ask'. Precision on the kimi side: auto mode self-approves everything (AskUserQuestion is even denied by the engine); yolo self-approves almost everything, but accesses to sensitive files (.env, SSH keys, credentials) and git control directories still prompt — those two ask policies run ahead of yolo's auto-approve in the chain.

Likewise, DSH's /plan only works for native-loop sessions (a prompt convention kimi never sees); for kimi sessions use /frostfin-plan — engine-enforced read-only: with plan mode active, kimi's permission policy chain denies Write/Edit and other mutating tools outright (the plan file excepted), regardless of what the model wants.

Which to pick:

• Daily development: /yolo — self-approves almost everything, and only interrupts you for .env, SSH keys, credentials, and git control directories. Lowest friction with the safety net intact; the mode is remembered per session and survives DSH restarts.
• Review every step: default — every mutating tool pops a dialog. Note DSH's dialog only offers allow/reject — it can't express kimi's "approve for this session", so similar tools will ask each time; if that gets noisy, go back to yolo.
• Plan without touching anything: /frostfin-plan — engine-enforced read-only; the plan is presented for your approval.
• Unattended runs: /auto — self-approves everything and even denies the question tool (it won't bother you); pairs best with DSH's /goal for long-running tasks. Don't expose sensitive directories to it.

Installation

1. Install Kimi Code and log in

macOS / Linux (official script, no Node.js required):

curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash

Windows (PowerShell; install Git for Windows before first launch — kimi uses the bundled Git Bash as its shell environment):

irm https://code.kimi.com/kimi-code/install.ps1 | iex

Then run kimi once and log in with /login inside the TUI (Kimi Code OAuth or a Moonshot AI Open Platform API key). frostfin drives kimi through the kimi acp subprocess and reuses that login — no extra configuration.

2. Install DSH

Requires Node.js ≥ 22.19:

npx @deepseek-ai/dsh web   # Web UI defaults to http://127.0.0.1:3080

Pinned version: @deepseek-ai/dsh@0.1.0-rc.x (see "Status" below).

3. Install this plugin

# One-line install (prebuilt npm package — no build, no allowBuilds prompt)
npx @deepseek-ai/dsh plugin --profile web add dsh-frostfin

# Restart dsh web to take effect
npx @deepseek-ai/dsh web

From source (to review or hack on the code):

git clone https://github.com/pzc2004/dsh-frostfin.git
cd dsh-frostfin && pnpm install && pnpm build
npx @deepseek-ai/dsh plugin --profile web add .

Once installed: the new "Moonglow Frostfin Whale" mode becomes the default automatically; the conversation view gains a "月芒霜鳍鲸" tab (local and remote kimi sessions with one-click attach); no model configuration is needed — DSH's model gate is satisfied by a nominal route. The full feature guide (panel / remote / commands / permission modes / FAQ) lives at docs/guide.md (Chinese).

Running headless: for script/CI use without the Web UI, install into the headless profile instead — npx @deepseek-ai/dsh plugin --profile headless add dsh-frostfin. Only the host half takes effect there (kimi still drives the agent loop); the panel endpoints and input-row buttons do not appear.

Upgrade: npx @deepseek-ai/dsh plugin --profile web update dsh-frostfin, then restart dsh web. dsh plugin forwards its arguments verbatim to pnpm — so add dsh-frostfin@latest or a pinned add dsh-frostfin@0.2.0 works too.

Uninstall: npx @deepseek-ai/dsh plugin --profile web remove dsh-frostfin — uninstalling reverts every registration and removes the managed block synced into kimi's config.toml (your kimi config is restored verbatim). Session bindings and the model cache under ~/.frostfin/ are kept so a reinstall can resume; to wipe them completely, rm -rf ~/.frostfin.

Platform support

Fully developed and verified on macOS. Linux should behave identically (POSIX semantics are home turf). Windows is unverified: the known prerequisite is that Kimi Code itself needs Git Bash (KIMI_SHELL_PATH can point to it); the remaining risk points (process-termination semantics, host module resolution) await real testing — feedback from Windows users is welcome.

Roadmap

• Remote kimi sessions (shipped): Kimi Code sessions on servers are now managed from your local DSH — hosts are listed from ~/.ssh/config (same semantics as VS Code), one click brings up a remote kimi acp over ssh+tmux (disconnects and local shutdowns never kill the remote process; reconnects resume the very same session); remote sessions are shown as host → workspace → session with one-click attach. Prerequisites: tmux and kimi on the server with /login done (the panel tells you exactly what's missing).
• Next candidates: ControlMaster connection reuse, upstream asks for kimi ACP and DSH (see docs/upstream-kimi-acp.md and docs/upstream-dsh.md).

Status

Usable but early. Quality baseline: 82 automated tests (pure translation unit tests + end-to-end runs driving a scripted ACP child process, without the real kimi) plus real-browser verification (streaming replies, approval dialog, question modal, image understanding, status bar, slash commands). Pinned to @deepseek-ai/dsh@0.1.0-rc.6 and Kimi Code 0.36.x — both sides iterate fast, so treat upgrades as deliberate actions. See the design doc at docs/design-v0.1.md (written in the M1–M3 era; implementation notes for M4+ live in code comments).

License

The code is released under MIT. The Genshin Impact-related artwork in assets/ and in this document (bestiary images, map and region photos; the logo is a derivative work based on the in-game creature design) is © miHoYo / HoYoverse, shown only to illustrate the name's origin, and is NOT covered by the MIT license.