dsh-code-server-app 0.3.58 → 0.3.63

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.en.md CHANGED
@@ -12,10 +12,12 @@
12
12
 
13
13
  A static profile plugin (npm package with host + client bundle) that ships the **VS Code server tree** from a [code-server](https://github.com/coder/code-server) release as a **platform-independent dependency package** (pack-time artifact `vendor/vscode` → `@jinsiyu/dshcs-vscode-server`, no install scripts, no postinstall). The code-server **Node service layer is replaced by the plugin's own `lib/launcher.mjs`**: it drives `<tree>/lib/vscode/out/server-main.js` (`loadCodeWithNls()` / `createServer()` / `handleRequest()` / `handleUpgrade()`) directly and re-adds the few HTTP endpoints code-server used to provide (`/healthz`, `/manifest.json`, `/_static/*`, `/proxy/:port`). The 16 native modules (node-pty / @vscode/sqlite3 / spdlog / …) come from `@jinsiyu/dshcs-*` sub-packages declared **directly on the plugin's dependency table** under their real names (os/cpu-gated per target), with the original import names restored by runtime junctions. VS Code's inner dependencies and the prebuilt native modules are **all installed by the package manager together with the plugin** — no global npm install, no `bin` configuration, no profile config changes, no second install command, **no argon2/C++ toolchain**.
14
14
 
15
- > Since 0.3.22 the ask panel renders the session's new content with **DSH's own Markdown renderer** (the same
16
- > renderer and design tokens as the DSH UI; new content only) and can **answer approval requests in place**
17
- > (writing outside the workspace / running commands). See "Working with DSH: the editor bridge" and section 21 of
18
- > `docs/analysis-code-server-as-dsh-plugin.md`.
15
+ > **The "ask DSH" dialog** renders the session's new content with **DSH's own Markdown renderer** and can
16
+ > **answer approval requests in place** (writing outside the workspace / running commands). The panel is a
17
+ > hand-written React component inside `lib/client.js`: it requires `react-dom/client` and
18
+ > `@deepseek-ai/dsh-client-ui-primitives` straight from the DSH page's module table (the very instance the UI
19
+ > uses), so typography, highlighting and math match the UI and a renderer version mismatch is impossible.
20
+ > There is **no build step anywhere on that chain**. See "Working with DSH: the editor bridge".
19
21
 
20
22
  ## UI carrier and required DSH version (0.2.3: right-sidebar DSH only)
21
23
 
@@ -48,7 +50,7 @@ A static profile plugin (npm package with host + client bundle) that ships the *
48
50
  the narrow-viewport handling). When the button is missing it keeps the current mode and logs one `console.warn` — panel
49
51
  rendering is never affected.
50
52
  - **Resident IDE (0.2.2, on by default)**: switching to another tab or collapsing the sidebar and coming back **no longer reloads** code-server — unsaved editor buffers, terminals and debug sessions all stay put (see "Why switching tabs no longer reloads" below).
51
- - The settings card has exactly **three settings**: "**Claim types**", "**Fullscreen on open**" and "Resident in background" — no other rows (0.2.7 removed the "Entry", "dependency install" and "environment check" rows).
53
+ - The settings card has exactly **four settings**: "**Claim types**", "**Fullscreen on open**", "Resident in background" and "**FIM completion (experimental, off by default)**" — no other rows (0.2.7 removed the "Entry", "dependency install" and "environment check" rows).
52
54
  Open the IDE from the **Code Server box** on the sidebar's guide page, or by clicking DSH's own produced-file chips / delivered-file previews / inline file names;
53
55
  diagnostics stay out of the UI — the `[code-server]` lines in the DSH host log are the place to look (`/api/code-server/status` still returns `env` for scripts).
54
56
  The old `windowedOpen` (open in a window) and `reserveComposer` were **removed in 0.2.6**: leftover keys in an old settings document neither fail nor apply (they are no longer part of the schema).
@@ -225,62 +227,70 @@ only the editor knows, and lets editor gestures drive the current session.
225
227
  | Direction | Capability | Mechanism |
226
228
  |---|---|---|
227
229
  | editor → agent | **unsaved buffers** (disk ≠ what the user sees), active file and selection, **language-server diagnostics** with `file:line`, source and code | agent tools `editor_context` / `editor_diagnostics`; plus a notice attached before writing a dirty file |
228
- | editor → DSH | select code → context menu **"DSH: ask about selection"** → an **ask panel** opens (carrying `file:line` and the selection); the question enters the current session as **user input**, and that session's **new content** is rendered in the panel by DSH's own Markdown renderer | extension command `dsh-code-server.askAboutSelection` (one of the **top two** editor context-menu items) + a webview panel + `POST /ask` + the `thread` field of `/sync` |
229
- | DSH → editor (approval) | when the agent wants to **write outside the workspace or run a command**, the approval request shows up as a card in the panel (tool, reason, countdown); "allow once" / "reject" takes effect immediately | the `approvals` field of `/sync` + `POST /approve` (the bridge's **only** non-read-only route; constraints under "Security model") |
230
+ | editor → DSH | select code → context menu **"DSH: ask about selection"** → an **ask dialog** opens in the DSH page (its title bar carries `file:line`); the question enters the current session as **user input**, and that session's **new content** is rendered in the dialog by DSH's own Markdown renderer | extension command `dsh-code-server.askAboutSelection` (one of the **top two** editor context-menu items) → bridge `POST /event {kind:'ask-open'}` → the client half's `POST /api/code-server/ask/send` |
231
+ | DSH → editor (approval) | when the agent wants to **write outside the workspace or run a command**, the approval request shows up as a card in the dialog (tool, reason, countdown); "allow once" / "reject" takes effect immediately | the `approvals` field of `/api/code-server/ask/state` + `POST /api/code-server/ask/approve` (the plugin's **only** write route; constraints under "Security model") |
230
232
  | agent → editor | the agent changed a file → a **native diff** opens (left = the **full pre-write text**, right = what is on disk now); if that buffer has unsaved changes you get a warning and **no overwrite** | the host reads `result.value.before` (the complete pre-write text) in `tools/post-execute` into a bounded snapshot cache → the `tools/result` event carries an opaque key → the extension fetches the text and opens the diff |
231
233
 
232
234
  - The tools are only registered while the bridge is live (so the model never sees an unusable tool), and the
233
235
  system-prompt section renders only then too.
234
- - **Ask panel** (extension 0.2.0; the official renderer since 0.2.3; a **floating dialog over the DSH UI since
235
- 0.2.5**): the context-menu command no longer opens a panel inside the editor (that is always a tab or a column,
236
- never a dialog) — the plugin's client half pops up a draggable, resizable floating chat window in the DSH page
237
- itself (bottom-right, ✕ closes it), leaving the editor layout alone. Hosts without that capability fall back to
238
- the in-editor webview panel. The selection can still be changed while the dialog stays open.
239
- The two commands **remember their intent** (0.3.21): "ask about selection" carries a **line range + selection text**
240
- only when something is actually selected, while "ask about file" **never carries line numbers or a selection** — the
241
- cursor line is irrelevant to the question and only misleads the agent. With no selection, the selection command also
242
- degrades to the plain file.
243
- - **Injected context is collapsed** (0.3.24): the location line plus the selection code block the bridge adds to the
244
- message are split out into a collapsed Context row (click it to see the code), while the bubble keeps only the user's
245
- own words — the same treatment the DSH UI gives injected context.
246
- - **The panel renders exactly what DSH renders** (0.3.22): the panel bundles DSH's official Markdown renderer
247
- (`MarkdownText` from `@deepseek-ai/dsh-client-ui-primitives`) plus the official design tokens — the same
248
- micromark/mdast pipeline, the same incremental streaming parser, the same shiki highlighting (boot set:
249
- typescript / shellscript / json), KaTeX math and the same heading/table typography. Only **new content** is
250
- rendered (from the moment the panel subscribes); history is **not replayed** and there is no "load earlier".
251
- - **Thinking shows up like in DSH** (0.3.23): assistant reasoning becomes a Think row — **collapsed by default**,
236
+ - **Ask dialog**: the context-menu command only reports its intent to the host; the dialog itself is popped up by
237
+ the plugin's client half inside the DSH page — draggable, resizable (bottom-right, ✕ closes it), leaving the
238
+ editor layout alone. The selection can still be changed while the dialog stays open.
239
+ When the host cannot prove the dialog is alive (page not open / browser still running an old client) you get a
240
+ one-line notice telling you to open or refresh the Code Server tab — there is **no second ask UI** in the editor.
241
+ - **The two commands remember their intent**: "ask about selection" carries a **line range + selection text**
242
+ only when something is actually selected, while "ask about file" **never carries line numbers or a selection** —
243
+ the cursor line is irrelevant to the question and only misleads the agent. With no selection, the selection
244
+ command also degrades to the plain file. The host takes that context from its cached editor state; the extension
245
+ only reports the intent.
246
+ - **Follow-ups are delivered according to DSH's own setting**: `ui-conversation.busyEnter` (Settings → Conversation,
247
+ "Enter while busy") accepts `queue` (the default) or `steer`. Pressing Enter in the dialog is the same gesture as
248
+ pressing Enter in the main composer, so it reads the same value: `steer` ⇒ the host calls `agent.steer()` and the
249
+ follow-up is consumed at the **running turn's next step boundary** (answered within that turn); `queue` ⇒ the host
250
+ calls `agent.followup()` and the question becomes **its own later turn**, leaving the running one alone. If the
251
+ setting cannot be read (namespace unregistered / minimal composition) or the host has no `agent.steer`, delivery
252
+ falls back to `queue` — a setting never makes a question undeliverable. The panel's status row **says which one was
253
+ used** ("inserted into the current turn…" / "queued for the next turn…"), because while `queue` is in effect the
254
+ **DSH main UI cannot show that message yet**: it sits in the host-side pending queue (`next-turn`) and the main
255
+ client does not render pending queues (it joins the chat flow only once it becomes its own turn). The message is
256
+ not lost.
257
+ - **Injected context is collapsed**: the location line plus the selection code block the bridge adds to the
258
+ message are split out into a collapsed Context row (click it to see the code), while the bubble keeps only the
259
+ user's own words — the same treatment the DSH UI gives injected context.
260
+ - **The body is exactly what DSH renders**: it is handed to DSH's official Markdown renderer
261
+ (`MarkdownText` from `@deepseek-ai/dsh-client-ui-primitives`) — the same micromark/mdast pipeline, the same
262
+ incremental streaming parser, the same shiki highlighting (DSH's own lazily-loaded grammar set), KaTeX math and
263
+ the same heading/table typography. Only **new content** is rendered (from the moment the dialog subscribes);
264
+ history is **not replayed** and there is no "load earlier". If the official components cannot be resolved the
265
+ body degrades to plain-text `<pre>` instead of a blank panel.
266
+ - **Thinking shows up like in DSH**: assistant reasoning becomes a Think row — **collapsed by default**,
252
267
  showing its first line (or the latest line while streaming) and expanding on a row click, built from the official
253
268
  `DisclosureRow` plus the official think icon and typography language.
254
- - **Approvals are handled right in the panel** (0.3.22; window fixed in 0.3.23): while the panel is open, that
255
- session's approval requests ask the panel first (5 minutes by default). Clicking "allow once" / "reject" settles it
256
- immediately; **closing the panel** or letting the window expire hands the request back **unchanged** to the official
257
- path (the DSH UI shows the same card). 0.3.22's 8-second window was far too short for a human — the buttons went
258
- grey before anyone could click (reported as "the approval box stopped working"); the window is now 5 minutes and
259
- closing the panel hands off immediately instead of waiting it out.
269
+ - **Approvals are handled right in the dialog**: while it is open, that
270
+ session's approval requests ask the dialog first (5-minute window). Clicking "allow once" / "reject" settles it
271
+ immediately; **closing the dialog** or letting the window expire hands the request back **unchanged** to the official
272
+ path (the DSH UI shows the same card).
260
273
  **Nothing is ever auto-approved** — `allowed-once` can only come from a click, and there is no "always allow".
261
- - The question enters the DSH session as a **plain user message** (`source: { kind: 'user' }`, host 0.3.19): earlier
262
- versions used `{kind:'plugin'}`, which DSH renders as a *context update* — it did not look like something the user
263
- said. Provenance stays in the first line of the text: `From the editor: <file>[:<line>]`.
264
- - **Read-only with one constrained exception**: the bridge never writes files, applies edits, or runs commands; the
265
- single non-read-only route is `POST /approve`, which can only **answer an approval request that already exists**
266
- (see invariant 2 below). The agent's writes still go through its own `fs` tools; the bridge only *knows about* them
267
- and carries your answer back.
274
+ - The question enters the DSH session as a **plain user message** (`source: { kind: 'user' }`): provenance stays in
275
+ the first line of the text (`From the editor: <file>[:<line>]`), and the panel folds it into the Context row.
276
+ - **The bridge is completely read-only**: it never writes files, applies edits, or runs commands — all four routes
277
+ (`/health`, `/sync`, `/old`, `/event`) are reads. The only route that can change state is the DSH-same-origin
278
+ `POST /api/code-server/ask/approve`, which can only **answer an approval request that already exists**
279
+ (see invariant 2 below). The agent's writes still go through its own `fs` tools; the bridge only *knows about*
280
+ them and carries your answer back.
268
281
  - Status bar shows `$(plug) DSH` while connected (click it for the log in the "DSH Editor Bridge" output channel).
269
- - **The extension ships as a built-in** (fixed in 0.3.12): `dshcs-editor-bridge` is installed into
270
- `<tree>/lib/vscode/extensions/` next to `dshcs-open-file`. 0.3.0–0.3.11 installed it as a *user* extension
271
- instead, and the VS Code server marks any extension that sits in the user extensions folder but in no profile
272
- manifest as removed (`.obsolete`, log line `Marked extension as removed`) and then skips it forever — re-marked on
273
- every start, so **the bridge never reported any state**. To turn the bridge off use the plugin setting
274
- `editorBridge=false` (no mount, no tools) rather than uninstalling the extension from the Extensions view.
282
+ - **The extension ships as a built-in**: `dshcs-editor-bridge` is installed into
283
+ `<tree>/lib/vscode/extensions/` next to `dshcs-open-file` — an extension left in the *user* extensions folder is
284
+ marked `.obsolete` (log line `Marked extension as removed`) by the VS Code server and skipped forever.
285
+ To turn the bridge off use the plugin setting `editorBridge=false` (no mount, no tools).
275
286
 
276
287
  ### The channels (since 0.3.13 over **local IPC**: a Windows named pipe / unix socket)
277
288
 
278
289
  ```
279
- extension → host POST /code-server-bridge/sync one round trip: push editor state (+ which session the panel watches) + take events and thread deltas
280
- extension → host POST /code-server-bridge/ask push an editor question into the current session
281
- extension → host POST /code-server-bridge/approve answer an approval request that **already exists** (the only non-read-only route)
290
+ extension → host POST /code-server-bridge/sync one round trip: push editor state + take events and the capability bit
282
291
  extension → host GET /code-server-bridge/health unauthenticated liveness probe
283
- extension → host POST /code-server-bridge/event extension reports open/close etc. (host log tail)
292
+ extension → host GET /code-server-bridge/old fetch one "pre-write text" snapshot (events carry an opaque key)
293
+ extension → host POST /code-server-bridge/event report intent: open the dialog / open·close a file etc. (host log tail)
284
294
  host → extension <extensionsDir>/.dshcs-bridge/bridge.json endpoint + token, re-read every 5s
285
295
  (the same content is also written **next to the built-in extension** in
286
296
  `<tree>/lib/vscode/extensions/.dshcs-bridge/` — the env var is only injected when the host
@@ -289,14 +299,19 @@ host → extension <extensionsDir>/.dshcs-bridge/bridge.json endpoint + toke
289
299
  ```
290
300
 
291
301
  Requests use `http.request({ socketPath })` (`fetch` has no socket support) and **no port is ever opened**.
302
+ All four routes are **read-only**; questions and approval answers do not go through the bridge but through the
303
+ DSH-same-origin `/api/code-server/ask/*` (called by the plugin's client half inside the DSH page, under DSH's own
304
+ cookie/Origin checks).
292
305
 
293
- The three `/sync` fields the panel actually consumes (0.3.22):
306
+ The four state fields the dialog actually consumes (`GET /api/code-server/ask/state?rev=N`; an unchanged revision
307
+ returns a single number):
294
308
 
295
309
  | Field | Content | How the panel uses it |
296
310
  |---|---|---|
297
- | `thread` | **new content** entries (user / assistant / tool / approval) of the session the panel watches; bounded: ≤120 entries per session, ≤8000 chars per body, ≤4 watched sessions | assistant bodies go to the official renderer; tools and approvals become compact summary rows |
298
- | `approvals` | pending approval requests `[{id, toolName, reason, at}]` (≤4) | renders the card with a countdown; a click posts `/approve` |
299
- | `approvalHoldMs` / `uiVersion` | the approval window (300000 ms = 5 minutes by default) / the DSH UI version | countdown basis; a renderer-version mismatch is surfaced in the panel |
311
+ | `entries` | **new content** entries (user / assistant / tool / approval) of the session the dialog watches; bounded: ≤120 entries per session, ≤8000 chars per body, ≤4 watched sessions | assistant bodies go to the official renderer; tools and approvals become compact summary rows |
312
+ | `approvals` | pending approval requests `[{id, toolName, reason, at}]` (≤4) | renders the card with a countdown; a click posts `/ask/approve` |
313
+ | `approvalHoldMs` | the approval window (300000 ms = 5 minutes by default) | countdown basis |
314
+ | `contextText` / `mode` | the title line (from the host's cached editor state) plus the ask intent | title text; `mode` decides whether line numbers / the selection travel with the question |
300
315
 
301
316
  > **Why not HTTP (settled in 0.3.13, all three measured)**
302
317
  > 1. **Desktop has no HTTP surface at all**: the renderer calls `host.fetch()` through Electron IPC
@@ -339,10 +354,15 @@ Regression: the "push → take → clear, twice" case in `scripts/test-bridge-ro
339
354
  The token lives in `<extensionsDir>/.dshcs-bridge/bridge.json`, **readable by any process of the same local
340
355
  user**, so:
341
356
 
342
- 1. **`/code-server-bridge/*` is read-only, with `/approve` as the single exception.** No route writes files,
357
+ 1. **`/code-server-bridge/*` is read-only, with two bounded exceptions.** No route writes files,
343
358
  edits documents, runs commands, or spawns processes. A leaked token is therefore bounded to "sees information
344
359
  that is in the editor" and **can never** become arbitrary file writes or command execution. A whitelist
345
360
  assertion in `scripts/test-bridge-routes.mjs` guards this.
361
+ `/complete` (0.3.61, **experimental FIM completion, off by default**) is the one route that makes a **model
362
+ call**: it takes two strings (the text before/after the cursor) and returns one string. It still writes
363
+ nothing and runs nothing, is disabled unless the setting is on (403 otherwise), and is bounded by
364
+ length/rate/concurrency/timeout caps — so the worst case for a leaked token grows only to "spends a little
365
+ completion budget and reads back one completion".
346
366
  `/old` (added in 0.3.55) lives under the same invariant: it only reads the bounded cache of "pre-write copies of
347
367
  the last few agent writes" (≤8 entries, ≤1 MB each, ≤4 MB total, 5-minute TTL) by **opaque key**, 404s when it
348
368
  is gone, takes no path argument (so it cannot read arbitrary files) and does not consume (repeat polls get the
@@ -355,7 +375,7 @@ user**, so:
355
375
  (d) when no panel is watching, the panel is closed, or the window (5 minutes by default) expires, the request goes
356
376
  **back to the official
357
377
  path** — never auto-approved (DSH's `approval/request` itself fails closed; this bridge can only keep
358
- "nobody answered" as "nobody answered"). `pnpm test:webview` asserts these four plus the host-side whitelist.
378
+ "nobody answered" as "nobody answered"). `pnpm test:ask-dialog` asserts these four plus the host-side whitelist.
359
379
  3. **Any request carrying `Origin` gets 403.** Browsers always send one (including a sandboxed iframe's literal
360
380
  `Origin: null`); the Node extension host never does. Origin is checked **before** the token — otherwise the
361
381
  bridge would be a "did you guess the token right" oracle for a web page.
@@ -379,6 +399,64 @@ as stated for the loopback port.
379
399
  Diagnostics: `GET /api/code-server/status` exposes
380
400
  `bridge: { enabled, live, toolsRegistered, supported, url, file }` — **never the token** (that only exists in the file).
381
401
 
402
+ ## FIM completion (experimental, since 0.3.61, **off by default**)
403
+
404
+ When you pause while typing, a grey continuation appears after the cursor (Tab accepts, Esc discards).
405
+ It is **off by default**; turn it on with that row in the settings card — it takes effect immediately
406
+ (no host restart, no reinstall).
407
+
408
+ | Item | Value |
409
+ |---|---|
410
+ | Endpoint | DeepSeek **FIM (Beta)**: `POST https://api.deepseek.com/beta/completions`, params `prompt` (prefix) + `suffix` (suffix) |
411
+ | Model | `deepseek-flash` (the model the official FIM doc uses — also this deployment's default) |
412
+ | Credential | Reuses the `DEEPSEEK_API_KEY` DSH already has: `ctx.get('credentials').resolve(...)`, the same path the official adapter takes (falling back to the launch environment) |
413
+ | Measured latency | **112–416 ms** (non-streaming; streaming was *slower*, hence non-streaming on purpose) |
414
+ | When it fires | After a ≥250 ms pause and only past the gates: non-empty selection / non-`file` document / empty context / document >20k lines — those **never send a request** |
415
+
416
+ **How it is wired in (option A)**: FIM speaks the Completions API, while `ctx.llm.stream(GenerateOptions)`
417
+ only knows `messages` (no `prompt`/`suffix`; `purpose` is a closed union of `'compaction' | 'session-title'`).
418
+ So the plugin **registers its own LLM adapter route** `dshcs-fim`: the prefix/suffix travel inside a
419
+ `messages` envelope with a fixed marker (`dshcs-fim/1 `), and the adapter decodes it before hitting that
420
+ endpoint. The call therefore still goes **through DSH's LLM service** — cancellation, timeouts, terminal
421
+ chunks and stable error codes all follow the service contract (instead of the plugin bypassing it with a
422
+ raw fetch). Evidence and measurements: B6/B7 of `docs/analysis-continuedev-reuse.md`.
423
+
424
+ **Three knobs (0.3.62, all in the settings card, effective immediately)**:
425
+
426
+ | Setting | Default | Effect |
427
+ |---|---|---|
428
+ | Pause in ms | `250` | How long typing must stop before a request. Range 100–3000 ms (clamped on save); the endpoint round trip measured **112–416 ms**, so the pause *is* the perceived latency |
429
+ | Allow multi-line | on | Off ⇒ the host returns the first line only; an empty first line means no completion. Turn it off to be less intrusive |
430
+ | Disable by glob | empty | In these files **no request is sent at all**. Semantics: `*` does not cross directories, `**` does, a pattern without `/` matches the basename, a pattern with `/` matches any path suffix (so `vendor/**` and `src/*.ts` work at any depth), a trailing `/` means `/**`. Example: `*.md;vendor/**;**/dist/**` |
431
+
432
+ > Both sides implement these semantics independently (the extension is a static file shipped with the package
433
+ > and cannot import host code); `scripts/test-fim.mjs` pins their equivalence with one shared (pattern, path)
434
+ > corpus.
435
+
436
+ **Safety and bounds** (this is the only capability of the plugin that sends content out):
437
+
438
+ - **Still read-only**: the extension never writes files and never runs commands — it only *proposes* text;
439
+ insertion happens when you press Tab;
440
+ - The bridge gains its fifth route `POST /complete` (**the only route that makes a model call**) only while
441
+ the setting is on; with it off the route answers 403 and no request is made;
442
+ - Bounded: prefix/suffix ≤6000/2000 chars (120/40 lines, whichever comes first), output ≤2000 chars,
443
+ 4 s timeout, 120 ms minimum interval, concurrency 1, 60 calls/minute; over-limit requests are rejected
444
+ (409/429) and the extension simply shows no completion for that beat;
445
+ - One more layer in the extension: turning the setting off disposes the provider immediately (no requests at
446
+ all), and identical context hits a local 2-minute cache;
447
+
448
+ **Usage is visible in exactly two places**: these calls are **not session requests**, so they are not written
449
+ to the session log and DSH's own token accounting (per-turn usage, context pressure, telemetry) **excludes
450
+ them**. Usage therefore shows up in ① the IDE status bar's DSH item — a `$(zap) 1.2k` counter when on, with
451
+ input/output/cache-read/last-latency/last-error on hover, and ② the `fim` field of
452
+ `GET /api/code-server/status` (same snapshot).
453
+
454
+ **Known limits**: the model occasionally invents an insertion where nothing is needed (2/2 reproduced at a
455
+ cursor position that needed none); the filter pipeline strips code fences and control markers, but "should
456
+ this be completed at all" stays your call — Esc or typing on makes it disappear. A more reliable shape needs
457
+ a faster completion route, or DSH making completion a first-class request (at which point only the adapter's
458
+ data call changes; the caller stays as it is).
459
+
382
460
  ## Legacy DSH (unsupported since 0.2.3)
383
461
 
384
462
  **Behaviour**: when `sidebarRightTabs` / `sidebarRight` cannot be found, the plugin registers a single settings card:
@@ -438,10 +516,11 @@ Diagnostics: `GET /api/code-server/status` exposes
438
516
  crash/exit updates status live; after a DSH host restart the plugin **adopts** a still-running instance (verifies pid + `/healthz`), without duplicate start or killing unrelated processes;
439
517
  - `node_modules` and the pack-time artifact `vendor/` are git-ignored; after cloning, follow
440
518
  "Install the plugin (script-free install; code-server bundled)" below — `pnpm install` → `pnpm run vendor:vscode` →
441
- `pnpm run build:webview` → `pnpm pack` + `dsh plugin --profile web add` (the client half has no build step).
519
+ `pnpm pack` + `dsh plugin --profile web add` (after 0.3.58/0.3.59 the client half **and** the ask panel are
520
+ committed hand-written source: there is no build step anywhere on that chain).
442
521
 
443
522
  > Verified locally (BM: Windows 11 ARM64): the whole tree/dependency chain hangs directly off the plugin's
444
- > dependency table — the tree package `@jinsiyu/dshcs-vscode-server` (currently 4.137.0, a 50.8 MB tarball),
523
+ > dependency table — the tree package `@jinsiyu/dshcs-vscode-server` (currently 4.138.0, a 50.9 MB tarball),
445
524
  > the pure-JS inner dependencies plus the 8 platform-independent repacks in `dependencies`, and the 8
446
525
  > platform-specific repacks (win32-arm64 / win32-x64) in `optionalDependencies` with their own os/cpu gates;
447
526
  > the original names are restored by junctions created at runtime (`lib/native.js`)
@@ -452,13 +531,12 @@ Diagnostics: `GET /api/code-server/status` exposes
452
531
 
453
532
  ```powershell
454
533
  cd C:\Users\User\Desktop\dsh-code-server-app
455
- pnpm install # dev deps (esbuild + the official-renderer bundling deps); allowBuilds is explicit → no postinstall runs
456
- pnpm run build:webview # ask panel: official Markdown renderer + panel shell → webview/thread.{js,css} (not committed; must be built first)
534
+ pnpm install # one dev dependency left (@deepseek-ai/schemastery); allowBuilds is explicit → no postinstall runs
457
535
  pnpm run vendor:check # optional: show the bundled tree version vs the latest code-server release
458
536
  pnpm run vendor:vscode # ① produce vendor/vscode (the trimmed VS Code tree, ~197MB)
459
537
  pnpm run repack:build -- --target win32-arm64,win32-x64 --pack # ② one script builds every sub-package
460
538
  pnpm run publish:repacks # ③ publish every @jinsiyu/* sub-package (default dist-tag: next)
461
- pnpm pack # ④ → dsh-code-server-app-<version>.tgz (~750KB, including the panel renderer assets)
539
+ pnpm pack # ④ → dsh-code-server-app-<version>.tgz
462
540
  pnpm run publish:plugin # ⑤ publish the plugin itself (default dist-tag: next)
463
541
  # once the user has restarted dsh web and confirmed it works, promote latest:
464
542
  pnpm run promote -- <version>
@@ -473,13 +551,11 @@ pnpm run promote -- <version>
473
551
  > so their dist-tags do not affect resolution, but they default to `next` as well.
474
552
  > Inspect the current tags with `npm dist-tag ls dsh-code-server-app`.
475
553
 
476
- > `build:webview` bundles DSH's **official** Markdown renderer and design tokens into the panel assets
477
- > (~1.34MB: 996KB JS + 87KB CSS + 254KB KaTeX fonts), so it needs a local DSH deployment: the script reads the
478
- > `@deepseek-ai/dsh-web-frontend` version from that deployment and compares it with the renderer version pinned in
479
- > devDependencies — a mismatch **fails the build** (unless `--allow-version-mismatch`). Same convention as
480
- > `lib/client.js`: the artifacts are not committed and `prepack` rebuilds them.
481
- > Rationale (why not an iframe, where the tokens come from, size trade-offs) is section 21 of
482
- > `docs/analysis-code-server-as-dsh-plugin.md`.
554
+ > **There is no build step on this chain**: `lib/client.js` — including the "ask DSH" dialog panel — is committed
555
+ > hand-written source, and so is the extension (plain JS). `prepack` is down to `vendor:vscode`, the published
556
+ > package carries no frontend bundle, and `devDependencies` holds exactly one entry (`@deepseek-ai/schemastery`,
557
+ > used by the tests).
558
+ > The per-version analysis of the panel and the DSH page lives in `docs/analysis-code-server-as-dsh-plugin.md`.
483
559
 
484
560
  `repack:build` (`scripts/vendor-repacks.mjs`) is the **single script that produces every sub-package**:
485
561
 
@@ -492,7 +568,7 @@ pnpm run promote -- <version>
492
568
  | Goal | Command |
493
569
  |---|---|
494
570
  | **Build from the latest upstream release** | `pnpm run vendor:latest` (= `--force`): pulls `code-server@latest`'s tree into `vendor/vscode`; afterwards you **must** re-run `repack:build` and republish every sub-package |
495
- | **Pin a version** | `pnpm run vendor:vscode -- --version 4.137.0` |
571
+ | **Pin a version** | `pnpm run vendor:vscode -- --version 4.138.0` |
496
572
  | **Snapshot from an existing tree** | `pnpm run vendor:vscode -- --from <code-server dir>` (seconds) |
497
573
  | **Rebuild every sub-package** | `pnpm run repack:build -- --target win32-arm64,win32-x64 --pack` (without `--from` it npm-installs and compiles the source tree itself — slow) |
498
574
  | **Rebuild only the tree + dependency table** | `node scripts/vendor-repacks.mjs --reuse --target win32-arm64,win32-x64 --pack` (reuses the natives already in `repack/build`; also rewrites `lib/vendored.json` and the plugin dependency table) |
@@ -513,7 +589,7 @@ Both workflows live in `.github/workflows/`, and the regression list exists exac
513
589
 
514
590
  | Workflow | Trigger | What it does |
515
591
  |---|---|---|
516
- | `ci.yml` | push to `main` / PR / manual | `ubuntu-latest` + `windows-latest` matrix: `pnpm install --frozen-lockfile` → `build:webview` → `pnpm test` (the whole suite) → `vendor:check` (report only) → upload the panel assets |
592
+ | `ci.yml` | push to `main` / PR / manual | `ubuntu-latest` + `windows-latest` matrix: `pnpm install --frozen-lockfile` → `pnpm test` (the whole suite; since 0.3.59 there is **no build step in front of it**) → `vendor:check` (report only) |
517
593
  | `release.yml` | push a `v<version>` tag / manual (rehearsal, never publishes) | prepares `vendor/vscode` **at the version pinned in `dependencies`** → builds → full suite → `pnpm pack` → verifies the tarball manifest → **really installs it twice** (windows-latest proves the 16 win32 sub-packages, ubuntu-latest the 10 Linux ones: each deploys a real DSH, installs via the official path `dsh plugin --profile web add <tgz>`, then runs the `test:installed` + `dump-config` assertions; both legs must pass before anything is published) → publishes to npm **`next`** → creates a GitHub Release with the tgz attached |
518
594
  | `linux-repack-probe.yml` | push to this file / manual | **feasibility probe (never publishes; superseded by the Linux legs of `repacks.yml`)**: on Linux, builds the platform-specific repack packages per target (`linux-x64` → `ubuntu-latest`, `linux-arm64` → `ubuntu-24.04-arm`) and reports which modules really produce a `.node` and which are Windows-only. It runs the existing `vendor-repacks.mjs` itself; all writes happen in a copy of the repo under `$RUNNER_TEMP`. **Note**: it emits one notice per module, which hits GitHub's ~20-annotations-per-check-run cap and leaves only the tail; for the full verdict use the Linux legs of `repacks.yml` (one summary line per target) |
519
595
  | `repacks.yml` | manual (`publish` and `probe_oidc` both default to **false**, the four `build_*` legs default to **true**) / push to this file / push `.github/oidc-probe.enabled` | **builds and publishes the platform-specific sub-packages** (`@jinsiyu/dshcs-*`): one host-architecture runner per target (`win32-x64` → `windows-latest`, `win32-arm64` → `windows-11-arm`, `linux-x64` → `ubuntu-latest`, `linux-arm64` → `ubuntu-24.04-arm`); by default it only builds and uploads `repack/tgz/*.tgz`, and only publishes to npm (default `next`) when `publish` is checked. Ownership and ordering (**five legs, disjoint sets**): the `independent` leg runs **first** (windows-latest; it produces the **VS Code tree package + the 8 platform-independent repacks**, which are the same artifact for all four targets and are therefore published only once); the four platform-specific legs `needs: independent`, build with `--skip-independent` and publish with `--only <their own target>` ⇒ a broken base layer blocks the rest (no half-published state) and no package name is ever published twice. **Auth**: with no `NPM_TOKEN` it uses OIDC (per-package trust entries, all with workflow `repacks.yml` — see below). The Linux legs additionally verify that the `lib/vendored.json` / `package.json` they generate match the committed ones (the platform policy is meant to be host-independent). A `probe-oidc` job additionally does a **staged-only** probe of that OIDC route, so the channel can be proven without publishing anything real |
@@ -612,8 +688,8 @@ pnpm test:apply # apply() under a stub ctx
612
688
  pnpm test:claim-types # claim-type syntax and defaults
613
689
  pnpm test:bridge-routes # bridge route whitelist (read-only + /approve + /old) / Origin-vs-token order / token header agreement
614
690
  pnpm test:edit-snapshot # pre-write snapshots: value.before from tools/post-execute, session-cwd path resolution, triple-bounded cache, /old's 400-404-200
615
- pnpm test:bridge-extension # extension-side pure logic (dirty buffers, diagnostics, diff old-side priority, delivery, panel state)
616
- pnpm test:webview # panel bundle: official renderer + tokens, version match, the four /approve constraints
691
+ pnpm test:bridge-extension # extension-side pure logic (dirty buffers, diagnostics, diff old-side priority, delivery, ask intent + "dialog unavailable" notice)
692
+ pnpm test:ask-dialog # ask-dialog wiring: no artifacts/build chain left, the host's four ask routes, the extension only reporting editor state, the four approval constraints, the bridge's safety invariants
617
693
  pnpm test:launcher-routes # launcher HTTP surface (spawns a real process; slow)
618
694
  pnpm test:workspace-switch # switching workspaces does not restart the process
619
695
  pnpm test:workspace-cwd # "current workspace directory" resolution (DSH 0.1.6-alpha.2 sessionId vs. the older current)
@@ -726,9 +802,6 @@ one itself (with no `NPM_TOKEN` it uses OIDC):
726
802
  probe versions (`npm stage list`, then `npm stage reject <id>`; needs 2FA on your machine) — do **not**
727
803
  approve, since approving is what would turn a probe into a real version. Delete the sentinel file to
728
804
  return to "no automatic probe".
729
- - **Optional** repository variable `DSH_UI_VERSION` = the version of `@deepseek-ai/dsh-web-frontend` in the current
730
- deployment: when set, `release.yml` enforces that the panel renderer matches the deployed UI (the local
731
- `build:webview` always checks this; a runner has no DSH deployment).
732
805
 
733
806
  Things you must know:
734
807
 
@@ -741,8 +814,9 @@ Things you must know:
741
814
  assertion). Packing happens only in `release.yml`, after
742
815
  `node scripts/vendor-vscode-server.mjs --version <pinned>`.
743
816
  - **Release gates** (any failure stops the run; `next` is never advanced): tag ≠ `package.json.version`, the
744
- version already exists on npm, the tree version does not match (`test:vendored`), the suite fails, or
745
- `DSH_UI_VERSION` mismatches.
817
+ version already exists on npm, the tree version does not match (`test:vendored`), the suite fails, or the
818
+ tarball manifest / the two real-install legs disagree. (0.3.59 dropped one gate: the panel renderer no longer
819
+ needs to match the deployed DSH UI — it *is* that instance.)
746
820
  - `@deepseek-ai/schemastery` is a **devDependency** (pinned to 3.18.2, the version the deployment uses):
747
821
  `lib/index.js` normally takes it from the DSH deployment (in production, the copy hoisted inside the
748
822
  profile), and a clean clone / CI runner has no DSH at all — without this devDependency the `apply`-style
@@ -857,7 +931,7 @@ Costs and rules (read before editing `lib/client.js`):
857
931
  | Rule | Why | Enforced by |
858
932
  |---|---|---|
859
933
  | No top-level `import`/`export`/`await` | syntax errors in a classic script ⇒ the whole client half fails to load (empty UI) | `pnpm test:client-entry` E1 + the harness really loading it (E3) |
860
- | `require(...)` may only name `react` / `react/jsx-runtime` | the DSH module table is frozen; anything else throws "unknown module" | E1 |
934
+ | `require(...)` may only name DSH module-table seed words (`react` / `react/jsx-runtime` / `react-dom/client` / `@deepseek-ai/dsh-client-ui-primitives`) | the DSH module table is frozen; anything else throws "unknown module" | E1 (and, when a local DSH install exists, it cross-checks the real `staticModules` list word by word) |
861
935
  | Every section's top-level identifiers share one scope | after inlining, a `var`/`function` collision is a **silent overwrite** (real hit: `state` existed in both the surface and the plugin body — the former is now `surfaceState`) | no automatic guard ⇒ grep before adding a top-level name |
862
936
  | The claim-types section is a **copy** | a classic script cannot reach the host module `lib/claim-types.js` | E2 compares both over a sample table |
863
937
 
@@ -865,8 +939,35 @@ Size: ~114 KB uncompressed (was a 49.7 KB minified artifact) — a one-time down
865
939
  unchanged. Test hook: the entry exports `__internals` only when `window.__dshcsTestHooks === true` (used by
866
940
  `test-workspace-cwd.mjs` / `test-sidebar-fullscreen.mjs` to call pure functions); DSH never sets that flag.
867
941
 
868
- > The ask-panel webview is **still a build artifact** (`thread.{js,css}` is untracked and produced by
869
- > `pnpm run build:webview`); removing that build is the next phase.
942
+ ### Why the "ask DSH" dialog has no build step
943
+
944
+ The dialog's panel (conversation stream, collapsible thinking rows, approval cards, input box) lives **inside
945
+ `lib/client.js`** — hand-written, committed, with no artifact, no `/ask/bundle`, no injected `<script>` and no fake
946
+ `acquireVsCodeApi`.
947
+
948
+ Why it can work that way: the dialog already runs **inside the DSH page**, and the shell's module table
949
+ (`staticModules` of `dsh-web-frontend`) freezes `react` / `react/jsx-runtime` / `react-dom` / `react-dom/client` /
950
+ `@deepseek-ai/dsh-client-ui-primitives` / … — so the panel simply requires them:
951
+
952
+ - the renderer, the design tokens, the KaTeX styles and the shiki grammars **all come from the page** (the same
953
+ instance the DSH UI uses) ⇒ typography matches the UI and a version mismatch is impossible;
954
+ - the panel creates its own React root via `require('react-dom/client')` (its container is the dialog's own div);
955
+ - opening the dialog has no "fetch + parse + execute" step to wait for — there is no artifact.
956
+
957
+ Costs and rules:
958
+
959
+ | Rule | Why | Enforced by |
960
+ |---|---|---|
961
+ | CSS injected into the page may only target `.dshcs-*` | the styles land in DSH's own document; touching `:root`/`body` would restyle the whole UI | `pnpm test:ask-panel` P3 (every selector + a ban on at-rules) |
962
+ | Every `var(--vscode-*)` needs a fallback | the DSH page has no `--vscode-*`; a bare `var()` is invalid at computed-value time ⇒ transparent buttons/inputs | P3 |
963
+ | A component with hooks may only be written as `React.createElement(Name, …)` | there is no JSX here; `Name({…})` puts the child's `useState` into the parent's hook chain, and a changed branch throws "Rendered more hooks than during the previous render" | P4 (source-level lookbehind regex) |
964
+ | Official components must be resolved as "function **or** `{$$typeof}` object" | `MarkdownText` is a `React.memo` product (an **object**); testing `typeof === 'function'` silently degraded every body to `<pre>` | P4's `askComponent` cases |
965
+ | Missing seed words / official components must degrade | the panel is the primary path; a blank panel means asking is broken | P5 (`<pre>` body, native `button`) + an error boundary |
966
+ | Three message routes (ask / approve / close) | panel and shell share one window (no postMessage), so messages must reach `/ask/send|approve|close` | P6 |
967
+
968
+ The extension side has no build step either: it is plain JS (`extension.js` + `lib/*.js`) and only reports intent.
969
+ Those two "no build step" claims are guarded by `pnpm test:client-entry` / `pnpm test:ask-panel` (the panel itself)
970
+ and `pnpm test:ask-dialog` (the wiring, plus "not one trace of the build chain may remain").
870
971
 
871
972
  ### Development: install from source (changes take effect immediately)
872
973
 
@@ -879,13 +980,12 @@ dsh plugin --profile web add C:\Users\User\Desktop\dsh-code-server-app
879
980
  > too — but the not-yet-published local `@jinsiyu/*` packages must either be published first, or the
880
981
  > `repack/tgz/*.tgz` files must be installed into the profile as `file:` dependencies.
881
982
  >
882
- > **Changing the client half**: edit `lib/client.js` directly (since 0.3.58 it is **hand-written source** —
883
- > there is no build step and no `src/**` intermediate layer; the format rules are in that file's header and are
884
- > enforced by `pnpm test:client-entry`). After installing into a profile a hard refresh picks it up
885
- > (a host restart may be needed; the bundle rev is hashed at host start).
886
- > **Changing the ask panel**: edit `assets/extensions/dshcs-editor-bridge/webview/src/*` then run
887
- > `pnpm run build:webview` (this half is still a generated artifact, not tracked; the IDE must be restarted once
888
- > to pick it up, because the extension host caches the webview resources).
983
+ > **Changing the client half / the ask dialog**: edit `lib/client.js` directly (it is **hand-written source**:
984
+ > no build step, no `src/**` layer; the format rules are in that file's header and are enforced by
985
+ > `pnpm test:client-entry` and `pnpm test:ask-panel`). After installing into a profile a hard refresh picks it up.
986
+ > **Changing the extension**: edit `assets/extensions/dshcs-editor-bridge/{extension.js,lib/*.js}` (plain JS, no
987
+ > build). The IDE side needs one restart to load the new extension code, because the extension host caches loaded
988
+ > extensions.
889
989
 
890
990
  ### Pack-machine environment (the user machine needs nothing)
891
991
 
@@ -919,7 +1019,7 @@ dsh plugin --profile web add C:\Users\User\Desktop\dsh-code-server-app
919
1019
  ### Upgrading the VS Code tree (upstream = a code-server release)
920
1020
 
921
1021
  - **The version is decided at pack time**: `pnpm run vendor:latest` (= `--force`) pulls the tree of the npm **latest**
922
- release; or use `pnpm run vendor:vscode -- --version 4.137.0` / `DSHCS_CODE_SERVER_VERSION`.
1022
+ release; or use `pnpm run vendor:vscode -- --version 4.138.0` / `DSHCS_CODE_SERVER_VERSION`.
923
1023
  With an existing `vendor/vscode`, a plain `pnpm pack` never upgrades (it is a no-op).
924
1024
  **Source-tree precedence (fixed in 0.2.13)**: an explicit `--from` uses that tree, while an explicit
925
1025
  `--force`/`--version` now **always goes to the registry** — before the fix a local source tree won
@@ -934,13 +1034,14 @@ dsh plugin --profile web add C:\Users\User\Desktop\dsh-code-server-app
934
1034
  - **A version bump means rebuilding and republishing the sub-packages** (all with the same script):
935
1035
  1. `pnpm run repack:build -- --target win32-arm64,win32-x64 --pack` → the new tree package
936
1036
  (`@jinsiyu/dshcs-vscode-server@<new version>`) and the natives rebuilt against the new inner dependencies (the
937
- script also rewrites the plugin's pure-JS `dependencies` and both aggregator versions);
938
- 2. `pnpm run republish:repacks` (`pnpm run publish:repacks`) → publish; then bump the plugin version → `pnpm pack`
1037
+ script also rewrites the plugin's pure-JS `dependencies`, writes the 16 repacks under their real names into
1038
+ `dependencies` / `optionalDependencies`, and rewrites `lib/vendored.json`);
1039
+ 2. `pnpm run publish:repacks` → publish; then bump the plugin version → `pnpm pack`
939
1040
  → publish the plugin.
940
1041
  - `productPath` (`<quality>-<commit>`, part of the client WebSocket path) is **computed from `lib/vscode/product.json`**,
941
1042
  so upgrading the tree needs no code change — but the routes are registered at activation, so restart `dsh web` afterwards.
942
1043
  - **No runtime auto-upgrade anymore**: nothing fetches latest at startup; the version is fully determined by the bundled artifact.
943
- - Bundled locally right now: the tree of `code-server@4.137.0` (VS Code 1.137.0, `productPath=stable-b11dabda…`).
1044
+ - Bundled locally right now: the tree of `code-server@4.138.0` (VS Code 1.138.0, `productPath=stable-59c988c7…`).
944
1045
 
945
1046
  ### Compatibility with the old install locations
946
1047
 
@@ -959,8 +1060,12 @@ persisted via the official settings domain (`settingsScope`, namespace `code-ser
959
1060
  | `claimExtensions` | `*` + three exclusion groups (preview-friendly / executables / Office; full lists in the "Claim types" section) | **Claim types** (0.2.11, replaces 0.2.5's `fileOpenScope`): decides by extension which files go to VS Code, semicolon-separated; `*` claims every other type, `!ext` excludes (exclusion wins). The default leaves the four categories DSH's preview renders well (markdown/html/images/PDF) to DSH and sends everything else to the IDE; an empty value claims nothing. **Scope (session vs absolute) is no longer distinguished** |
960
1061
  | `fullscreenOnOpen` | `true` | **Fullscreen on open** (0.2.9): opening the Code Server tab (including clicking a file) switches the right sidebar to fullscreen (fills the window); off keeps DSH's default push mode (side by side with the conversation). Only the moment of opening is affected — a manual "Exit fullscreen" is never fought back |
961
1062
  | `keepResident` | `true` | **Resident in background**: on, the host preloads the IDE into a parked surface right after start — switching tabs or collapsing the sidebar never reloads it and the first open needs no cold start; off loads it only when the panel is opened (saves memory) |
1063
+ | `fim` | `false` | **FIM completion (experimental, 0.3.61)**: when on, pausing while typing shows a grey inline continuation (Tab accepts, Esc discards). **Off by default** — it is the only capability of this plugin that sends content out (a bounded slice of code around the cursor, per completion), and those calls **do not enter DSH's token accounting**, so usage is visible only in the IDE status bar. See the "FIM completion" section |
1064
+ | `fimDebounceMs` | `250` | **FIM · pause in milliseconds** (0.3.62): how long typing must stop before one completion request is sent. Effective range **100–3000 ms**, clamped on save (same rule as the host); empty/invalid falls back to 250 |
1065
+ | `fimMultiline` | `true` | **FIM · allow multi-line completions** (0.3.62): off means the host returns the first line only (an empty first line = no completion this time). The model does invent insertions where none are needed, and multi-line amplifies that noise |
1066
+ | `fimDisableGlobs` | empty | **FIM · disable by glob** (0.3.62): semicolon/newline separated. `*` does not cross directories, `**` does, a pattern without `/` matches the basename, a pattern with `/` matches any path suffix, and a trailing `/` means `/**`. Examples: `*.md`, `vendor/**`, `**/dist/**`. **Checked on both sides**: the extension first (no request at all), the host again |
962
1067
 
963
- (Since 0.2.9 the card keeps only those three settings; `windowedOpen` and `reserveComposer` are gone — leftover keys in an old
1068
+ (Since 0.2.9 the card keeps only those settings (0.3.61 added FIM completion, 0.3.62 its three sub-rows); `windowedOpen` and `reserveComposer` are gone — leftover keys in an old
964
1069
  settings document neither fail nor apply. `serve` remains a key in the settings namespace (usable from a settings document) but
965
1070
  has **no card row** — see "Serving mode".)
966
1071
 
@@ -1093,6 +1198,22 @@ What remains on the plugin side:
1093
1198
  **built-in** extension (in `lib/vscode/extensions`), so users cannot remove it from the extensions panel, and the
1094
1199
  installer re-syncs it whenever its content changes.
1095
1200
 
1201
+ ## Third-party and license
1202
+
1203
+ This project is **MIT**-licensed (see [`LICENSE`](LICENSE)). Its experimental **FIM completion** feature was
1204
+ **designed with reference to** [continuedev/continue](https://github.com/continuedev/continue)
1205
+ (Apache License 2.0, Copyright 2023 Continue) — specifically the settings taxonomy of upstream
1206
+ `tabAutocompleteOptions` (`debounceDelay`, `useAutocompleteMultilineCompletions`, `disableInFiles`) and its
1207
+ debounce / cursor-window / filtering / bounded-cache approach.
1208
+
1209
+ **No source code, template strings or files from continuedev/continue are included here** (completion
1210
+ requests target DeepSeek's official FIM (Beta) endpoint; the prompt shape comes from DeepSeek's docs and
1211
+ local measurement). The standard Apache License 2.0 text ships as
1212
+ [`LICENSE-Apache-2.0.txt`](LICENSE-Apache-2.0.txt), and the per-location reference log is
1213
+ [`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md) — so Apache-2.0's obligations (notices, marking modified
1214
+ files, keeping an upstream NOTICE if one appears) are already satisfied should a code fragment ever be ported
1215
+ in. Apache-2.0 grants no trademark rights; this project does not use "Continue" as a name or in promotion.
1216
+
1096
1217
  ## Known limitations
1097
1218
 
1098
1219
  - **~~The editor bridge needs DSH to provide `webServer`~~ no longer true (fixed in 0.3.13)**: the bridge now runs
@@ -1105,16 +1226,12 @@ What remains on the plugin side:
1105
1226
  or by simply calling `editor_context` — 0.3.0–0.3.11 sat in the state "health says bridge:true, extension never
1106
1227
  loaded" (cause above: the user-level install was marked `.obsolete`).
1107
1228
  - **Bridged state can lag by up to 600 ms**, and the tools say "stale" rather than serving data older than 10 s.
1108
- - **The ask panel renders only "new content" (0.3.22)**: the subscription starts when the panel opens, the history
1229
+ - **The dialog renders only "new content"**: the subscription starts when the dialog opens, the history
1109
1230
  `records` from `follow`'s opening frame are discarded, and the panel has **no "load earlier"** (the history-paging
1110
1231
  API `sessionController.page()` is deliberately not called in this version). Switch to the DSH UI for older content.
1111
- - **Panel highlighting ships only DSH's boot grammar set** (typescript / shellscript / json): the rest of the
1112
- official grammars load lazily through `import()` (~1.6MB total), and the panel is a single-file IIFE with no
1113
- lazy loading, so those languages render as plain text (exactly like DSH's own first render, no errors).
1114
- For the full set: `node scripts/build-webview.mjs --all-grammars`.
1115
- - **Panel assets are pinned to the DSH version**: the renderer is bundled against the UI version of the deployed
1116
- DSH, so after upgrading DSH you must rebuild the panel (`pnpm run build:webview`; the build fails loudly on a
1117
- version mismatch). The panel also shows a mismatch notice at runtime instead of silently using the wrong renderer.
1232
+ - **Highlighting follows DSH's own lazily-loaded grammar set**: the panel uses the page's renderer instance, so
1233
+ there is no "the artifact only carries a few grammars" limitation.
1234
+ - **A renderer version mismatch is impossible**: the panel requires the very instance the UI uses.
1118
1235
  - **The approval window in the panel is 5 minutes**: while the panel is open, approvals ask the panel first (the card
1119
1236
  shows a countdown); **closing the panel** or letting the 5 minutes run out hands the request back to the DSH UI —
1120
1237
  after that, that request can **only** be answered there (the card disappears from the panel and the thread keeps an