dsh-model-flatmap 0.0.0-stage → 1.0.1

package/cordis.patch.yml

@@ -0,0 +1,3 @@
+- insert:
+    - id: dsh-model-flatmap
+      name: dsh-model-flatmap

package/docs/DETAIL.md

@@ -0,0 +1,243 @@
+# Flat model picker
+
+Replaces the composer's model selector with a single flat panel. There are no
+second-level menus: the model grid, the reasoning levels of the current model,
+and the agent teamwork switch are all visible at once.
+
+## What it does
+
+- **One surface.** The trigger opens one panel capped at 320 px. The header,
+  and one adaptive footer holding the reasoning group and the teamwork row,
+  stay fixed; only the model grid scrolls. The two footer groups share a single
+  line when both fit with a 30 px gap, and otherwise fall back to one row each.
+- **Global search.** The search box filters across every provider and every
+  family. Groups that end up empty are not rendered, 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 has to land somewhere in the name, so `mod inte` finds
+  `ModelScope International` even though the two words are never contiguous.
+  A hit ranks by its weakest token, so a multi-word query never outranks a
+  plain prefix. A single-token query keeps the original ranking exactly.
+- **Provider-wide search.** A query that matches a *provider* — its id, its
+  catalog name, or the label its section draws — selects that provider as a
+  whole: every model it offers stays in the grid, not only the ones whose names
+  contain the query. Typing `dou` therefore surfaces `豆包(doubao)` with its
+  entire catalog. Models matched by their own name still rank first; the rest
+  follow in catalog order. Provider matching is word-wise too, and tokens may
+  land in different faces of the same provider — the id answers `mod`, the
+  localized name answers `inte`. It accepts prefixes and substrings only: an
+  ordered-subsequence pass would let a two-letter query drag every provider in,
+  which is noise rather than a match. This applies to both grouping modes.
+- **Two grouping modes.** A segment control switches the grid between
+  *By provider* and *By family*. Grouping and search are independent: the
+  segment only changes how the filtered result is organised.
+- **Model families.** A built-in dictionary recognises DeepSeek, Grok, Opus,
+  Haiku, Sonnet, Claude, Qwen, GLM, Kimi, GPT, Gemini, Llama, Mistral and
+  others. A model the dictionary does not cover is inferred from its id;
+  anything unjudgeable falls into *Others*, which always sorts last. Families
+  stay flat — no nesting.
+- **Registered brands keep their generations apart.** Merging `hy3` and `hy4`
+  into one `HY` family (DEC-004) was wrong: the generation has to be visible.
+  A registered brand (`FAMILY_BRANDS`, currently `HY`) resolves to
+  `BRAND + generation` when its id carries a generation digit, so `hy3` /
+  `hy3-x` → `HY3`, `hy4-preview` / `hy4free` / `hy4std` → `HY4`, and a bare
+  `hy` → `HY`. Generations rank next to their bare brand in numeric order.
+  Unregistered brands still collapse their version tail (`acme4-turbo` and
+  `acme-7b` are one `Acme`), and dictionary families are never split by
+  generation. A provider prefix does not hide the brand (`codebuddy-hy3` →
+  `HY3`).
+- **Family-grouped rows are provider-primary.** A family heading says what the
+  models *are*, so the family name is not repeated in a row. The row's primary
+  text is the provider name — the one thing a heading cannot tell you when a
+  family spans providers. The full model name and description stay in the row
+  tooltip. Only when one provider offers several models in the same family do
+  those rows carry a short secondary variant derived from the id tail (e.g.
+  `hy4free` → `free`, `hy3-x` → `x`); the plain base model carries no tag.
+  Colliding variants fall back to the unique model id. Under *By provider* the
+  heading already is the provider, so the primary text stays the model name and
+  no secondary tag is drawn.
+- **Adaptive columns.** The panel measures one group title and one model row,
+  subtracts the fixed sections from the 320 px cap, and picks the smallest
+  column count from 1 to 3 that fits. If even three columns overflow, it keeps
+  three and the grid scrolls. A group title never separates from its first
+  model, and the last column's deficit is balanced against its neighbour.
+- **Reasoning levels.** The current model's advertised levels render as a
+  caption ("推理等级" / "Reasoning effort") with a sparkle icon and a hint
+  tooltip, on the same line as a multi-segmentation control — the track wraps
+  to its own line only when the row cannot fit. The segments are radio
+  choices (`role="radiogroup"` / `role="radio"`, exactly one `aria-checked`
+  at a time, arrow keys move the selection) drawn with the header segment's
+  visual: a toned track and a floating active segment, so the active level is
+  unmistakable. A click submits through `session.selectModel` and the group
+  refreshes from the directory once it updates. A model without reasoning
+  metadata hides the whole group, and it costs no height.
+  The group shares a wrapping footer line with the teamwork row: when
+  reasoning + 30 px + teamwork fits the panel's inner width they sit side by
+  side, otherwise the footer wraps back to the original two-line stack. The
+  break is decided by CSS flex wrapping and the 30 px `column-gap`, never by a
+  JS width threshold, so it stays correct across panel resizes, font loads and
+  both themes. The single grey rule belongs to the footer container, so it sits
+  above the whole footer block and stays put in both states; neither group
+  draws a divider of its own (DEC-016).
+  The fixed height budget measures the footer container once, not its two
+  children, so a shared line is not double-counted against the grid.
+- **Selected model.** The current model reads as a solid deep-blue band with
+  white, bold text — no checkmark on the right. Selection lives on the row
+  itself (`aria-pressed="true"`), so nothing is reserved at the trailing edge:
+  unselected rows keep their full width for the name, the provider tag, and the
+  badge, and the measurement probe no longer mirrors a tail column. The blue
+  is `--dsw-static-deepseek-600` and the text is
+  `--dsw-static-neutral-bluish-00`, both from the theme-invariant static
+  palette, so the band looks the same in light and dark themes. Hover skips the
+  selected row so the neutral hover wash never covers the highlight.
+- **Badges.** A model's catalog badge (for example a third-party rate
+  multiplier such as `×0.1`) renders inside its row. A model with no badge
+  renders no empty slot. On the selected row the badge switches to a
+  translucent white chip so it stays legible on the blue fill.
+- **Out-of-balance models are struck through.** When a model call fails because
+  the account has no usable balance, credit, or quota left, that model's row
+  text is drawn with a `line-through` decoration, and the strike stays until a
+  later call for the same model returns something other than an
+  insufficient-balance failure. The row is not recoloured, so the strike reads
+  the same on the selected row's blue band and in both themes, and it does not
+  compete with the error strip's red. Only the row's text (the name and a
+  family variant tag) is struck; the badge chip and the pending spinner are
+  not.
+  - **What counts.** The Host already classifies an exhausted account into the
+    provider-neutral `QUOTA` / `ACCOUNT_QUOTA` codes and maps a provider HTTP
+    `402` to `QUOTA`. The panel accepts those codes, a structured `status: 402`
+    (or an `HTTP_402` code), and — for a failure the Host flattened to another
+    code with the provider's wording kept — the provider text itself
+    (`insufficient balance` / `insufficient credits` / `quota exceeded` /
+    `out of credits` / `not enough quota` / `no credits remaining` /
+    `额度已用尽` / `余额不足` / `预算已用完`, among others).
+  - **A `429` is decided by its body, not by the status.** Several providers
+    answer an exhausted balance with HTTP **429** rather than 402 — for example
+    `429: {"message":"insufficient balance","request_id":"…"}`, which the Host
+    classifies as `RATE_LIMIT`. A `429` therefore neither strikes nor clears on
+    its status alone: the message decides, so a 429 carrying balance wording
+    strikes while a genuine rate limit (`Rate limit exceeded`,
+    `访问量过大`, `使用量已超出频率限制`) does not. The rule is pinned against
+    real recorded failures (see the `REAL_INSUFFICIENT` / `REAL_TRANSIENT`
+    tables in the render tests, harvested from local session history).
+  - **What does not count.** A transient limit leaves the account usable, so
+    `RATE_LIMIT` with rate-limit wording, `TIMEOUT`, `TRANSPORT`,
+    `CONTEXT_WINDOW_EXCEEDED`, a `SERVER` failure, and a credential that simply
+    cannot use the model never strike a row. Nor does a message that only
+    *mentions* 额度 — a context-overflow failure says “与网络、账号、额度都无关”
+    and a transport failure merely suggests waiting for quota to recover; both
+    are excluded, and the tests pin them.
+  - **Which model.** The failing call's model is read from the `request/header`
+    event that immediately precedes the failing `turn/end`, so the mark lands
+    on the route (`provider/model`) the request was actually dispatched to —
+    not on the currently selected model, which may have changed since.
+  - **When the strike ends.** A `turn/end` for the same route that proves the
+    model answered clears it: a completed turn, a `max-tokens` turn, or a turn
+    that failed for a different reason. A cancellation (`aborted`), a `blocked`
+    turn, and the synthetic `interrupted` / `forked` closers clear nothing —
+    the model never answered, so they say nothing about the account.
+  - **The marks are global, not per conversation.** Whether a model has quota
+    left is a fact about the model route and its account, so the state is NOT
+    scoped to the Session a panel happens to render for. One module-level store
+    per plugin generation observes the durable feed of **every** Session the
+    Session list names, and every seat reads that one store. A failure recorded
+    in any Session therefore strikes the same model in every Session and
+    workspace, including in a Session whose feed this page never opened.
+  - **It survives a reload.** The marks are persisted to `localStorage`
+    (origin-scoped, like the grouping choice), so a fresh page load starts from
+    the last known state instead of an empty one. Reads and writes degrade
+    silently when storage is unavailable, and records older than 7 days are
+    dropped so a stale strike cannot live forever.
+  - **The newest observation wins.** A Session feed carries no cross-Session
+    sequence number, but every durable event keeps the Host's wall-clock `time`.
+    Each route keeps the time of its last decision, and an observation older
+    than the stored one is ignored. This is what makes the global state safe:
+    replaying an older window (a history replacement, or a second Session whose
+    history is older) cannot resurrect a strike that a newer call already
+    cleared, and a cleared route keeps its watermark for the same reason.
+  - **Where it comes from.** The store folds each Session's durable event feed
+    (`sessions.binding(id).eventSource`), which is the same source the Chat
+    reads its quota notice from, and re-attaches when the `sessions` service is
+    replaced. No Host change is involved. A history replacement refolds that
+    Session from scratch; `prepend` pages in older history and cannot change a
+    mark, so it is ignored.
+- **Agent teamwork.** The switch keeps the host's semantics: the same label,
+  the same confirmation before disabling it mid-conversation, and no optimistic
+  update — the state comes from the `agentTeam` projection. The row lives in
+  the shared footer line described under *Reasoning levels*, immediately right
+  of the reasoning group when both fit on one line.
+
+## Data and interaction
+
+Everything goes through public Client APIs:
+
+- `ctx.modelDirectories.directoryFor(sessionId)` supplies the catalog, the
+  current selection, and the pending selection; `load()` refreshes it and
+  `select()` submits.
+- The directory store is subscribed through `useSyncExternalStore`, so a
+  catalog update — including one caused by a selection — repaints the panel.
+- `ctx.sessions.binding(sessionId).eventSource` is the Session's contiguous
+  durable event window. A single module-level store folds the feed of **every**
+  Session the Session list names (`ctx.sessions.list`) to derive the global
+  per-model insufficient-balance marks, persists them to `localStorage`, and
+  every seat subscribes through `useSyncExternalStore` so a mark transition
+  repaints the panel while every other event stays quiet.
+- `ctx.get('remote').agentTeams.setEnabled()` writes the teamwork selection.
+- `ctx.locale` supplies the copy, resolves catalog badge text, and prefers the
+  host's own `agent-team` wording when that namespace is registered.
+
+No host component, stylesheet, or DOM is read or restyled.
+
+## Layout behaviour
+
+- Window resize, a grouping switch, and search input all re-plan the columns,
+  debounced, and the plan runs once more after fonts finish loading so the
+  first frame's measurements do not stick.
+- The panel opens above the trigger and flips below when there is no room,
+  staying inside the viewport.
+- An unbound Session, an empty catalog, a failed provider, and a search with no
+  hits each have their own state. A directory that has not resolved yet still
+  renders a trigger rather than blanking the composer.
+
+## Styling
+
+Colours, radii, spacing, and elevation come from the `--dsw-alias-*`,
+`--dsw-static-*`, `--dsw-elevation-*`, `--dsw-menu-*`, and `--dsw-radius-*`
+tokens, so light and dark themes both follow the host palette. The selected
+row's blue is deliberately drawn from the static palette (`--dsw-static-*`),
+which is theme-invariant: the alias `--dsw-alias-state-business-primary`
+flips to a light blue with dark text in the dark theme, which would break the
+requested deep-blue-with-white-text treatment there. The panel uses the same menu surface
+material as the host's own menus (menu fill, backdrop blur, prominent
+elevation, large radius), and hover, focus, selected, and disabled states follow
+the shipped picker's visual language.
+
+Each provider / family label is deliberately set apart from the model rows
+beneath it: it draws as a filled, rounded band that floats on
+`--dsw-alias-bg-layer-1` with `--dsw-elevation-soft` — the same recipe as the
+active segment. That fill is opaque white in light themes and near-black in
+dark, so it steps clearly away from the translucent panel surface (and the
+transparent model rows) in both themes, with no hard-coded colour. The bare
+`--dsw-alias-menu-group-header-fill` token was only a few levels from the
+surface and read as almost the same background as the rows (see DEC-005). The
+label keeps smaller but bold type with a stronger label colour. The model rows
+stay 13 px regular text on a transparent background, so the band reads as a
+section header rather than another selectable row.
+
+The band also carries a low-key frame: a hairline border
+(`--dsw-alias-border-l2`) and a short brand-blue left edge
+(`--dsw-alias-state-business-primary`, the panel's own accent). That marks each
+group start clearly without boxing the group's rows — a group's rows may flow
+across columns, so only its label is framed (see DEC-003).
+
+## Limitations
+
+- The panel replaces the composer seat only; the `/model` popup is untouched.
+- The panel does not host the `conversation.input.model.label` and
+  `conversation.input.model.options` child slots. The shipped registration keeps
+  those declarations alive, so other plugins' contributions stay valid; their
+  content simply is not drawn inside this panel.
+- Addressed subagent sessions hide the seat, matching the shipped selector.
+- The badge comes from the catalog entry, so a provider that publishes no badge
+  shows none.

package/icon.svg

@@ -0,0 +1,11 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 64 64" width="64" height="64" role="img" aria-label="Flat model picker">
+  <title>Flat model picker</title>
+  <rect x="4" y="8" width="56" height="48" rx="10" fill="none" stroke="currentColor" stroke-width="3" opacity="0.85"/>
+  <rect x="11" y="15" width="20" height="5" rx="2.5" fill="currentColor" opacity="0.55"/>
+  <rect x="35" y="15" width="18" height="5" rx="2.5" fill="currentColor" opacity="0.25"/>
+  <rect x="11" y="25" width="20" height="5" rx="2.5" fill="currentColor" opacity="0.55"/>
+  <rect x="35" y="25" width="18" height="5" rx="2.5" fill="currentColor" opacity="0.25"/>
+  <rect x="11" y="35" width="20" height="5" rx="2.5" fill="currentColor" opacity="0.55"/>
+  <rect x="35" y="35" width="18" height="5" rx="2.5" fill="currentColor" opacity="0.25"/>
+  <rect x="11" y="45" width="42" height="4" rx="2" fill="currentColor" opacity="0.18"/>
+</svg>

package/index.js

@@ -0,0 +1,13 @@
+/**
+ * Host half of the flat model picker bundle.
+ *
+ * The plugin is client-side and runs in both the Web GUI and the Desktop app: it
+ * replaces the composer's `conversation.input.model` seat with a flat panel rendered
+ * in the client UI, and reads the Session's model directory, projections and
+ * `session.selectModel` through public Client APIs.
+ * The Host therefore contributes no service, tool or event.
+ */
+export const name = 'dsh-model-flatmap'
+export const inject = []
+
+export function apply() {}

package/locale/en.json

@@ -0,0 +1,7 @@
+{
+  "meta": {
+    "title": "Flat model picker",
+    "description": "Replaces the composer model selector with one flat panel: search, provider or model-family grouping, every model and every reasoning level visible at once, and the agent teamwork switch on the same surface. No second-level menus.",
+    "detail": "docs/DETAIL.md"
+  }
+}

package/locale/zh.json

@@ -0,0 +1,7 @@
+{
+  "meta": {
+    "title": "模型平铺选择面板",
+    "description": "用一块平铺面板替换输入框的模型选择器：搜索、按 Provider 或模型家族分组，所有模型与全部推理等级一次可见，智能体团队协作开关也在同一面板内，没有任何二级菜单。",
+    "detail": "docs/DETAIL.md"
+  }
+}

package/package.json

@@ -1,6 +1,28 @@
 {
   "name": "dsh-model-flatmap",
-  "version": "0.0.0-stage",
-  "stub": true,
-  "description": "Temporary package placeholder for staged publishing"
-}
+  "version": "1.0.1",
+  "description": "Flat, two-level-free model picker for the DSH composer: every model and reasoning level on one surface.",
+  "type": "module",
+  "exports": {
+    ".": "./index.js",
+    "./client": "./client.js",
+    "./package.json": "./package.json",
+    "./locale/*.json": "./locale/*.json",
+    "./icon": "./icon.svg"
+  },
+  "files": [
+    "index.js",
+    "client.js",
+    "cordis.patch.yml",
+    "locale/*.json",
+    "docs",
+    "icon.svg"
+  ],
+  "dsh": {
+    "bundle": { "patch": "./cordis.patch.yml" },
+    "client": {
+      "platform": "web",
+      "immediately": true
+    }
+  }
+}