🛰️ dsh-lsp-actions
- 1024 store channel:
npm i -g dsh1024 once, then dsh1024 plugin --profile web add dsh-lsp-actions (counts toward the deepseek1024.com install ranking).
The LSP action surface for DeepSeek Harness — real language servers, real feedback, and the IDE integration backend for editors.
Diagnostics, formatting, completion, code actions, symbols, signature help, inlay hints, and rename for your agent's editor loop — plus the stable editor action protocol (lsp.actions.*) that lets any editor consume them directly.
English · 简体中文 · Español · Português · हिन्दी
Compatibility
| Surface | Status |
|---|
| Harness | DeepSeek Harness dsh-v0.1.6-alpha.2 (compat declared for >=0.1.2-rc.1 <0.2.0 || >=0.1.5-alpha.1 <0.2.0 || >=0.1.6-0 <0.2.0); the plugin writes no session events itself - the host records the standard tool/call + tool/result events. The released seam still carries only the four legacy operations, so the plugin keeps its own client and records that vintage on the first action call. Verified 2026-09-18 (dual typecheck rulers + full test suite + self-contained/artifacts gates). |
| Node | ^22.19.0 || >=24.0.0 |
| Platforms | All (pure host; subprocess + filesystem, no network) |
| Model | Any (tools are model-agnostic; the plugin never calls a model) |
What you get
dsh-lsp-actions mounts as a single host row (id: lsp-actions, name: dsh-lsp-actions, inject: [tools, fs, subprocess]). The official DeepSeek Harness ctx.lsp seam covers navigation (go-to-definition, references, implementation, hover); this plugin completes the action surface — the feedback loop an agent needs while it writes and fixes code:
- Eight
lsp_* tools — diagnostics, formatting, completion, code actions, symbols, signature help, inlay hints, and rename, all served by the same language servers your IDE uses.
- Editor action protocol v1 — a stable JSON-RPC surface (
lsp.actions.list / lsp.actions.run / lsp.events) that lets any editor (VS Code first) consume those capabilities directly.
- Real-server verification — a real
typescript-language-server run is part of the test suite (self-contained, CI on Node 22/24 across Linux, Windows, and macOS), not just mocks.
Quick start
# 1. install the bundle into your profile
dsh plugin --profile web add "github:PerryLink/dsh-lsp-actions#main"
# or from npm (published releases)
dsh plugin --profile web add dsh-lsp-actions
# 2. restart and verify the row
dsh --profile web --dump-config | grep -A3 'id: lsp-actions'
Install & uninstall
- git channel (latest
main): dsh plugin --profile web add "github:PerryLink/dsh-lsp-actions#main" — the prepare script builds (tsc --noEmitOnError && node scripts/fix-dts.mjs).
- npm channel (published releases):
dsh plugin --profile web add dsh-lsp-actions.
- tarball channel:
pnpm pack in this repo, then dsh plugin --profile web add ./dsh-lsp-actions-<version>.tgz.
- uninstall:
dsh plugin --profile web remove dsh-lsp-actions (or remove the row from the profile patch).
Configuration
All tunables are Schemastery Config fields (changeable from cordis.yml). An id-targeted override replaces the whole row — restate every key you need. cordis.patch.yml documents each key inline.
| Key | Default | Meaning |
|---|
servers | {} | Named language servers; an empty table activates no servers |
editor.enabled | false | Serve the editor action protocol over JSON-RPC stdio (headless backend only) |
editor.requestTimeoutMs | 60000 | Per-run timeout budget (ms) for the editor protocol |
editor.diagnosticsCacheMaxFiles | 64 | Bounded LRU diagnostics-cache size (files) |
maxDiagnostics | 200 | Diagnostics cap per result |
maxCompletionItems | 20 | Completion-items cap per result |
maxCodeActions | 50 | Code-actions cap per result |
maxSymbols | 100 | Symbol-results cap |
maxSignatures | 10 | Signature-help cap |
maxInlayHints | 200 | Inlay-hints cap |
maxResultChars | 16000 | Rendered-result cap (chars) |
maxDocumentBytes | 4000000 | Document-read cap (bytes) |
timeoutMs | 60000 | Per-call timeout, enforced by the official timeout policy |
Each servers entry is an LspServerEntry: command (executable resolved on PATH at load) and extensionToLanguage (".ts" → typescript) are required; optional fileGlobs, projectMarkers, args, env, initializationOptions, configuration, formattingOptions, maxMessageBytes, maxStderrBytes, killGraceMs, shutdownTimeoutMs, diagnosticsSettleMs, diagnosticsDebounceMs, and idleTimeoutMs (0 = keep the server process alive) tune the built-in stdio client.
Routing is deterministic and configuration-driven. Every file takes the first match from: (1) a server entry whose fileGlobs match the path; (2) the nearest ancestor directory — walking up to the workspace root — that holds a project config file listed in some entry's projectMarkers, matched against the entries that also map the file's extension; (3) the entry whose extensionToLanguage maps the file's extension, in config order. A workspace holding apps/node-app/{package.json,tsconfig.json,src/main.ts} and apps/deno-app/{deno.json,src/main.ts} therefore serves each main.ts from its own project's server, with no path rule to maintain and nothing to change when a project is added, renamed, or moved:
servers:
typescript:
command: typescript-language-server
args: ["--stdio"]
extensionToLanguage: { ".ts": typescript, ".tsx": typescriptreact }
projectMarkers: ["package.json", "tsconfig.json"]
deno:
command: deno
args: ["lsp"]
extensionToLanguage: { ".ts": typescript, ".tsx": typescriptreact }
projectMarkers: ["deno.json", "deno.jsonc"]
A project marker only decides among servers that already map the file's extension (it never widens an entry's file types); the walk never leaves the workspace root, and a project config in one directory never applies to a sibling directory. Step (1) outranks a project marker: an entry whose fileGlobs match the path still wins, so drop project-level globs when you adopt projectMarkers — a lingering **/*.ts glob keeps claiming those files. These rules govern the built-in stdio client — a mounted ctx.lsp seam provider decides its own routing.
Tools & surfaces
| Surface | Kind | Notes |
|---|
lsp_diagnostics | tool | <file> — compiler/analyzer errors, warnings, and hints with severity, range, message, and source server (read-only) |
lsp_format | tool | <file> [range?] — formats a file/selection through the language server and applies it, returning the diff (writes via fs/write-intent) |
lsp_completion | tool | <file> <line> <character> — completion suggestions at a cursor position, including the insertion text (read-only) |
lsp_code_action | tool | <file> [range?] [only?] — server-verified quickfixes/refactorings with their edits, for a range or the first diagnostic (reference-only) |
lsp_symbols | tool | <query?> <file_path?> — workspace-wide symbol search by name, or one file's outline (read-only) |
lsp_signature | tool | <file> <line> <character> — signature help (parameters and documentation) inside a call (read-only) |
lsp_inlay_hints | tool | <file> [range?] — type annotations and parameter-name hints from the server (read-only) |
lsp_rename | tool | <file> <line> <character> <new_name> — server-verified rename, applied workspace-wide with per-file diffs (writes via fs/write-intent) |
lsp.actions.* | protocol | Editor action protocol v1: lsp.actions.list / lsp.actions.run / lsp.events over JSON-RPC |
examples/vscode/ | extension | UI-only VS Code extension plus the headless backend composition it connects to |
Editor action protocol v1
When editor.enabled: true is set in a dedicated headless composition, dsh-lsp-actions serves a stable editor protocol over newline-delimited JSON-RPC 2.0 (the same wire framing as the official SDK/ACP transports):
| Method | What it does |
|---|
lsp.actions.list | Returns the lsp-actions/v1 protocol version, the action catalog (diagnostics.get, completion.get, quickfix.apply, format — each flagged writes), and the addressable DSH sessions |
lsp.actions.run | Executes one action with a structured { requestId, action, status, result | error } envelope; errors carry the stable LSP_ACTION_* codes |
lsp.events | Subscribes to the streamed lsp.event notifications: diagnostics.updated, action.status, file.changed, sessions.changed |
All write actions (quickfix.apply, format) go through the official permission presets and approval: a read-only session is refused with LSP_ACTION_READ_ONLY before any server round-trip, edits ride the fs/write-intent waterfall, and the sandbox_permissions + justification escalation pair resolves through the official approveEscalation ask (fail-closed when no answerer can decide). Full wire spec, bilingual: docs/editor-protocol.md · docs/editor-protocol.zh-CN.md.
Versioning and the backward-compatibility promise
- The protocol is versioned —
lsp.actions.list returns protocol: "lsp-actions/v1", version: 1. v1 is frozen: field names, action ids, event kinds, and error codes stay stable forever.
- Evolution is additive only: new actions, fields, and event kinds arrive without a version bump; existing semantics never change in place; a breaking change ships under a new
protocol version, which servers may serve side by side.
- Clients must ignore unknown fields, unknown event kinds, and unknown actions, and route on the stable error
code, never on message text.
Error codes
Every failure carries a stable code; models and callers route on the code, never on message text.
| Code | Meaning |
|---|
LSP_ACTION_UNAVAILABLE | No server entry and no seam provider handles this file |
LSP_ACTION_UNSUPPORTED | The server (or seam provider) does not advertise the operation |
LSP_ACTION_SERVER_FAILED | The server failed (with its stderr tail); startup failures retry once |
LSP_ACTION_MALFORMED_RESPONSE | The server sent a structurally invalid payload |
LSP_ACTION_CONFLICT | The file changed since it was read, or the edits overlap / go out of bounds / leave the workspace |
LSP_ACTION_READ_ONLY | The session's sandbox mode forbids the formatting/rename write |
LSP_ACTION_WORKSPACE_REQUIRED | The calling session has no workspace cwd to root the server in |
LSP_ACTION_NO_SYMBOL | The server found no renameable symbol at the cursor position |
LSP_ACTION_UNKNOWN | Editor protocol: unknown action id, or no code action matched title/index |
LSP_ACTION_INVALID_ARGS | Editor protocol: malformed action parameters |
LSP_ACTION_APPROVAL_UNAVAILABLE | Editor protocol: the approval path could not grant a wider sandbox mode (fail-closed) |
LSP_PROTOCOL_VERSION_UNSUPPORTED | Editor protocol: the declared protocol version is not supported |
VS Code extension
examples/vscode/ ships a UI-only extension (sidebar with the DSH sessions, the diagnostics list, one-click quickfix apply, open-at-range, and format) plus the headless backend composition (backend/cordis.yml) it connects to over ACP-style JSON-RPC. The extension implements zero LSP logic — every capability and every byte written belongs to the plugin. Install steps, settings, and the demo-gif recording script are in examples/vscode/README.md.
Editor demo
Permissions & data
- Permissions: formatting and rename ride the official permission presets and approval — the
fs/write-intent waterfall and the sandbox_permissions / justification escalation pair resolved through ctx.approval. The plugin declares fs:read, fs:write, subprocess:spawn, and network:none in its workshop manifest.
- Data: nothing is stored on disk; tool results live only in the session log (no cross-session persistence). The editor protocol keeps one bounded in-memory LRU diagnostics cache, freshness-stamped and never persisted across restarts.
- No network: the plugin makes no network requests; it talks to language servers over local subprocess stdio.
Security boundaries
- Read-only by default. Six of the eight tools are reference-only; only
lsp_format and lsp_rename mutate, and they do so as real write/edit mutations.
- Official seams, not re-implemented. Every byte goes through the
fs/write-intent waterfall (observation → guarded write → observation) and the per-call sandbox policy; escalation matches the official write/edit tools.
- Fail loud, fast, structured. Empty
servers + no ctx.lsp seam → LSP_ACTION_UNAVAILABLE; read-only sessions → LSP_ACTION_READ_ONLY before any server round-trip; command forms are reported and never executed.
- Conflicts never clobber. A file changed on disk after it was read fails with
LSP_ACTION_CONFLICT; lsp_rename pre-flights every edited file before the first write.
- Bounded work. Result caps, byte caps, and the platform's timeout policy bound every call; the diagnostics cache is a bounded LRU.
- Nothing cached on the model path. Tool results live only in the session log; the diagnostics cache never persists across restarts.
- Bad servers fail loudly. A missing executable fails at load; a server that dies at startup fails the call with
LSP_ACTION_SERVER_FAILED plus its stderr tail (after one fresh-spawn retry).
- Prompt hygiene. The plugin injects no persona or prompt prose into the session system prompt — its model-facing surface is the eight tool schemas.
Architecture
Actions run official-seam-first and fall back to the plugin's own minimal stdio client:
lsp_diagnostics / lsp_format / lsp_completion / lsp_code_action /
lsp_symbols / lsp_signature / lsp_inlay_hints / lsp_rename
│
▼
ctx.lsp seam (extended: diagnostics / formatDocument / completion)
│ absent · legacy · no provider for this file
▼
built-in stdio client ← servers table (ctx.subprocess.spawn + JSON-RPC)
The seam extension is proposed upstream (upstream/lsp-action-seam.patch, PR description in upstream/PR-description.md). Once it lands, the plugin keeps working unchanged — the built-in client simply stops being used. The built-in client stays as the standalone fallback for the servers table. The editor protocol rides the same runner, the same write path, and the same permission machinery. Full research and design notes: docs/seam-extension-notes.md.
The action backend is behind a public provider interface — ActionRunner (exported from the package) — implemented by the bundled runner and open to third-party providers: the eight tools consume only that interface, never a concrete client.
Known limitations
- Resident documents. Each document opens once and stays open for the server instance's lifetime — a later request with changed source sends one full-document
didChange, and the process shutdown closes every document — so project-based servers like tsls serve textDocument/signatureHelp and document-free workspace/symbol instead of answering null on a freshly-closed document. The resident set is bounded by the workspace and cleared with the instance (idle eviction or plugin disposal).
- Range formatting requires the server's range provider. Servers that only advertise whole-document formatting fail range requests with
LSP_ACTION_UNSUPPORTED.
- Rename applies text edits only. Resource operations (create/delete/rename files) in a server's rename answer are refused with
LSP_ACTION_UNSUPPORTED, and edits outside the workspace fail as LSP_ACTION_CONFLICT before anything is written.
Development
pnpm install # node ^22.19 || >=24
pnpm run lint # oxlint over src/ and tests/
pnpm test # vitest: unit + fixture-server integration + editor-protocol e2e + real tsls e2e
pnpm run test:coverage # coverage gate
pnpm build # tsc --noEmitOnError && node scripts/fix-dts.mjs → lib/
pnpm run prepare # tsc --noEmitOnError && node scripts/fix-dts.mjs (runs on install)
pnpm run prepublishOnly # tsc --noEmitOnError && node scripts/fix-dts.mjs (runs before publish)
Topics
dsh, dsh-plugin, deepseek-harness, lsp, language-server, diagnostics, formatting, completion, code-action, symbols, signature-help, inlay-hints, rename, refactor, ide, editor, vscode, acp, json-rpc
Contributors
- @PerryLink — creator and maintainer: the LSP action client and server lifecycle, all eight tools, the editor action protocol, tests, CI, and the five-language docs.
PerryLink DSH Plugin Family
This project is one of the 40 DeepSeek Harness plugins maintained by PerryLink. If this one helps you, the others likely will too:
| Plugin | One-liner |
|---|
| dsh-auto-review | Second-model auto-review on the approval chain, fail-closed by default |
| dsh-background-agents | Durable background child agents with a Web UI sidebar, messaging and interrupt |
| dsh-budget | Cost governance for DeepSeek Harness: budgets, carbon, and latency in one panel. |
| dsh-checkpoint-rewind | Claude Code /rewind-equivalent: snapshots, session forks, one-shot restore |
| dsh-claude-move | Migrate Claude Code sessions, memory, skills and CLAUDE.md into DSH |
| dsh-click | Cross-platform native desktop control for DeepSeek Harness — Windows first. |
| dsh-composer-history | Terminal-style input history for the web composer: arrows, Ctrl+R search |
| dsh-data-quality | Dataset quality checks and citation cross-checks (the optional numeric bridge consumed here) |
| dsh-defend | Prompt-injection, jailbreak, and secret-leak defense for DeepSeek Harness. |
| dsh-doublecheck | Engineering-discipline guard: requirements grill, test gates, adversary review |
| dsh-draw | Unified static-image generation routing for DeepSeek Harness. |
Install from the DSH Desktop Market
All PerryLink plugins are browsable in the built-in DSH Desktop Market: Market → Sources → add source → paste https://perrylink-dsh-catalog.perrylink.workers.dev/catalog-source.json → select it. Installation still goes through the Market's npm-identity verification and your confirmation.
License
Apache License 2.0 © 2026 dsh-lsp-actions contributors
