dsh-model-flatmap

dsh-model-flatmap

A flat, two-level-free model picker for the DeepSeek Harness (DSH) composer: every model and reasoning level on one surface, no second-level menus.

简体中文

What it is

dsh-model-flatmap is a client-side DSH plugin that runs in both the Web GUI and the Desktop app. It replaces the composer's conversation.input.model seat with a single flat panel rendered in the client UI. Instead of nested popovers, the whole catalogue is one searchable grid where the model grid, the current model's reasoning levels, and the agent-teamwork switch are all visible at once.

The plugin reads and writes through public Client APIs only — the per-Session model directory, the teamwork setting, and the host's locale. The host half (index.js) contributes no service, tool, or event; the UI lives in client.js, which registers with a negative priority so it shadows the shipped ModelSelect while active. Disabling the plugin restores the native control with no residue.

Screenshot

The panel in action — a flat grid of models, the current model's reasoning levels, and the agent-teamwork switch on one surface:

Features

• One surface. The trigger opens one panel capped at 320 px. The header, the reasoning row, and the teamwork row stay fixed; only the model grid scrolls.
• The trigger shows what the native one shows. Its name is not a string: the shipped picker renders the conversation.input.model.label slot there, and connection plugins use it for a promotion chip (夜间折扣), a rate, or a load mark (queued people / a rocket when idle). This plugin renders that slot's winning component on its own trigger — with the seats it needs — so the two modes show the same content. A missing or crashing occupant falls back to the catalog name plus the model's own model.badge.
• Global search. Filters across every provider and every family. Groups that end up empty are dropped, so no stray headings remain. Prefix matches rank first, then substring matches, then ordered character matches.
• Word-wise search. Whitespace splits the query into tokens, and every token must land somewhere in the name. A hit ranks by its weakest token.
• Provider-wide search. A query that matches a provider (its id, catalog name, or section label) keeps that provider's entire catalogue in the grid.
• Two grouping modes. A segment control switches the grid between By provider and By family. Grouping and search are independent.
• Model families. A built-in dictionary recognises DeepSeek, Grok, Opus, Haiku, Sonnet, Claude, Qwen, GLM, Kimi, GPT, Gemini, Llama, Mistral and others. Unrecognised models are inferred from their id; anything unjudgeable falls into Others (always sorts last). Families stay flat — no nesting.
• Generations are kept apart. A registered brand splits by generation: hy3 → HY3, hy4free / hy4std / hy4-preview → HY4, and a bare hy → HY. Unregistered brands still collapse their version tail; dictionary families (GPT, Qwen…) are never split.
• Rows are provider-primary, with a model tail. Under By family the heading already says what the models are, so the family name is not repeated in a row: the primary text is the provider name. What neither the heading nor the provider can say — which model of the family the row is, and the catalog's rate multiplier — is drawn as a quieter tail at the end of the row, with the family name and the provider label stripped off: Qwen3.8-27B under Qwen reads 3.8-27B, and Deepseek-V4.1-Flash · x0.01 under DeepSeek reads V4.1-Flash · x0.01 with the multiplier drawn smaller and bold. The full model name and description stay in the tooltip.
• Long names and long headings scroll on hover. A clipped model name or group heading runs one eased leftward pass after the pointer rests on it for 0.5 s, then stops; leaving resets it instantly. Only the text moves — the trailing badge and the title's band never do.
• Reasoning effort. The current model's reasoning levels render as an exclusive, keyboard-operable group ("推理等级" / "Reasoning effort") driven by radio semantics.
• Adaptive footer. The reasoning group and the agent-teamwork switch share one line when both fit with a 30 px gap, and fall back to one row each when they do not — the break is CSS flex wrapping, not a measured pixel threshold.
• Out-of-balance models are struck through — everywhere, for a day. A model whose last call failed on insufficient balance, credit, or quota keeps a line-through on its row text until a later call for that same model returns anything else, and until the mark has gone 24 hours without a new attempt — whichever comes first. A balance is topped up and a quota resets daily, so a strike nobody refreshed is stale information, not a warning. The strike follows the model the failing request actually used, and covers the Host's QUOTA / ACCOUNT_QUOTA codes, a provider 402, and a 429 whose body reports an empty balance — a plain rate limit never strikes. Because a balance is a property of the model route and its account, not of one conversation, the marks are global: a failure seen in any Session strikes the model in every Session and workspace, and the state is persisted to localStorage so it survives a reload. Every mark keeps the time of the observation that made it, so opening the list a day later draws that row plain. Observations are ordered by that same time, so the newest one wins and replaying older history cannot resurrect a strike that has already ended — by a later call or by age.
• Theme-aware. Styled with host theme tokens only; looks correct in both light and dark themes.
• Native mode. The panel's bottom-right Default style button hands the composer's model seat back to the host's own picker — the real shipped ModelSelect, not a reproduction of it. The plugin does this by stepping aside: it disposes its own conversation.input.model registration, so the host's entry underneath wins the slot and renders itself. Nothing is imported, copied, or re-mounted, which means an upgraded DSH shows its own newest picker and this mode can never drift out of date. The way back is a separated action band at the bottom of that picker's own menu — it reuses the conversation.input.model.options child slot the shipped ModelSelect declares, so it belongs to the panel it acts on and renders nothing at all while the flat panel is shown. It is drawn as an outlined button, not as one more menu row: the rows above it pick a value, this one performs an action. The choice is remembered across reloads. If the host ships no picker to hand over to, the button stays disabled and the panel stays — the composer never ends up without a model control.

See docs/DETAIL.md for the full behaviour spec and DECISIONS.md for the decision log behind the design.

Installation

This package is a DSH bundle. Install it from your DSH profile (the plugin manager), or drop it into the profile's bundle directory so the dsh loader picks up cordis.patch.yml.

# Example: install the bundle into a running profile (paths vary by install)
cp -r . "<profile>/bundles/dsh-model-flatmap"

Enable or disable the plugin from the profile's plugin list — no native control is touched when the bundle is off.

Project layout

Path	Purpose
index.js	Host half — bundle entry, contributes nothing at runtime.
client.js	Browser half — the entire flat picker UI.
cordis.patch.yml	Cordis patch applied by the dsh bundle loader.
locale/en.json	English copy and metadata.
locale/zh.json	Simplified-Chinese copy and metadata.
docs/DETAIL.md	Detailed behaviour specification.
test/	Render-path tests driven through a minimal hook runtime.

Development

The plugin is plain ES modules with no build step.

# Run the render-path tests (no browser required)
node --test test/

Tests read client.js directly and drive the real component through a minimal hook runtime to catch crashes, missing sections, and wrong column counts.

License

Released under the MIT License.