dsh-input-limit
直接在 dsh web 对话页设置当前大模型的输入上限（上下文窗口）。
A community plugin for DeepSeek Harness
(dsh) that adds a small chip to the composer tool row — right next to the
model picker — showing the current model's input limit and letting you change
it on the spot, without digging through the Settings page.
What it does
- Shows the current model's effective input limit (context window) as a tiny
pill in the message input bar, e.g.
输入上限 128K (Input limit 128K).
A · 默认 tag means the value is inherited from the provider default.
Hovering shows the exact token count.
- Click the pill → a popover lets you type an exact token count (one unit —
e.g.
131072, or 131,072) with a live ≈ 128K equivalent shown while you
type, then Save, or Reset to default.
- Persists a per-model
contextWindow override into the provider's settings
section (llm-pi-ai / llm-deepseek) via the same typed RPCs the Settings
page uses. The change is live on the next request — it drives context
pressure display and auto-compaction, so a model whose real input limit is
smaller than the adapter default (e.g. 128K vs the pi-ai default 262144) is
accounted for correctly.
- Writes go through a fresh settings read and are retried once automatically
when the document moved in the meantime (
settings/conflict), so two open
editors never clobber each other silently.
Numbers use the binary scale model capacities are conventionally quoted in
(1K = 1024, 1M = 1024²): 131072 reads 128K, and the pi-ai default
262144 reads 256K.
It works for every configurable provider (pi-ai routes, the direct DeepSeek
adapter), hides itself when the current model/provider is not configurable, and
is a browser-only plugin: no custom host code, no server restart required
beyond loading the plugin itself.
Install
Requirements: dsh CLI (tested on 0.1.5-rc.2, works with any build that
ships the web UI plugin system), pnpm (the plugin installer drives it), and
a configurable provider (pi-ai or deepseek-official).
The web UI is the web profile. Installing into it and restarting dsh web
is all it takes.
From the Web UI (recommended)
Once the package is on npm: sidebar Plugins → Add plugin, type
dsh-input-limit, pick the registry (npm by default, npmmirror as the built-in
fallback), then Install → Enable now.
From npm
dsh plugin --profile web add dsh-input-limit
dsh web
From GitHub
dsh plugin --profile web add github:<you>/dsh-input-limit
dsh web
pnpm blocks build scripts of git-hosted dependencies by default; this repo
ships prebuilt lib/, so a blocked prepare is harmless. To rebuild on
install instead, add the key pnpm prints under allowBuilds in the profile's
pnpm-workspace.yaml and re-run.
From a local checkout / tarball
# from a directory that contains the plugin package
dsh plugin --profile web add ./dsh-input-limit
dsh web # restart (or re-open http://127.0.0.1:3080)
Uninstall
dsh plugin --profile web remove dsh-input-limit
Build from source
lib/ ships prebuilt in this repo, so installs from a checkout or GitHub work
without a toolchain. To rebuild or iterate:
pnpm install
pnpm run build # emits lib/index.mjs (host half) + lib/client.js (browser half)
pnpm run typecheck
pnpm run test
The browser bundle is emitted in the exact closure-factory format the dsh web
shell's __ModuleLoader__ consumes (platform modules externalized, everything
else inlined, CSS Modules compiled with lightningcss and injected as a
<style data-plugin> element) — the same contract the built-in UI plugins use.
Dependency-free sanity checks (no install needed, Node ≥ 22):
node tests/bundle-smoke.mjs (loads lib/client.js under a stub
__ModuleLoader__), plus node --experimental-strip-types tests/capacity.verify.mjs
and tests/provider.verify.mjs (logic + an end-to-end run against a mock rc.2
client wire — the Typert ctx.remote namespaces and the ctx.sessions object
layer — built from a real pi-ai route layout).
How it works
cordis.patch.yml mounts one plugin row; the host half (src/index.ts) is a
no-op whose presence makes the Host serve the browser half.src/client/index.ts registers a component into the composer's
conversation.input.right list seat. Services are read through ctx.get
inside the plugin fiber and the concrete settings/session faces are
detached there, so no component-time callback ever touches the Cordis
context proxy.src/client/provider.ts resolves the session's current model from its
durable modelSelection projection (ctx.sessions), maps it to the
provider's settings namespace via ctx.remote.settings.describe(), reads the
effective context window, and writes the override with
ctx.remote.settings.mutate(ns, ops, expectedRevision) (preserving the rest
of the models array). The forwarded settings/document-updated event and the
session's model projection refresh open chips live.
Compatibility notes
- Tested on
dsh 0.1.5-rc.2. The client plugin declares the dotted
services it dereferences (remote.settings, remote.session) in its
inject list — required by the Cordis context proxy; older/newer builds that
rename or drop these faces will fail the plugin load loudly rather than
misbehave silently.
Development & contributing
- Product copy is Chinese (
src/client/locales.ts ships a matching English
dictionary); code comments are English.
- Design deep-dive:
docs/DESIGN.md — architecture, data flow,
parser rules, and the Cordis pitfalls this plugin works around.
- Release process & official-requirements checklist:
docs/RELEASING.md.
- PRs welcome. Keep the client bundle's external list to the platform modules
(react/cordis/ui-slots/...); everything else must inline.
Publish to the community
To make the plugin discoverable as an official community plugin:
-
Create a GitHub repository (dsh-input-limit) and push this directory.
-
On the repository Settings → Topics, add the dsh-plugin topic
(topic listing) — this is what the
official README designates for plugin discoverability.
-
Optionally publish to npm:
pnpm publish # runs `prepare` → rebuilds lib/ and ships it
The full release gates, artifact definitions, compliance checklist, and
post-publish verification live in docs/RELEASING.md.
License
MIT
