refactor(e2b): group remote providers

packages/e2b/lsp-e2b/README.md

@@ -0,0 +1,47 @@
+# @deepseek-ai/dsh-lsp-e2b
+
+English | [中文](README.zh.md)
+
+Generic E2B language-server backend for [`ctx.lsp`](../../lsp/lsp/README.md). It runs configured stdio servers and reads their source documents inside the shared `ctx.e2b` sandbox; the provider registry, normalized query results, queues, and protocol connection state remain on the host.
+
+## Plugin and configuration
+
+The `lsp-e2b` plugin injects `e2b`, `lsp`, and the concrete `dsh-subprocess-e2b` service. `servers` is a non-empty provider-id table:
+
+| Server key | Default | Meaning |
+|---|---|---|
+| `command` | required | Remote executable, absolute or resolved on the sandbox PATH at load. |
+| `args` | `[]` | Remote server arguments. |
+| `env` | `{}` | Explicit environment entries passed through the subprocess adapter. |
+| `extensionToLanguage` | required | Lowercase leading-dot extension to LSP language id. |
+| `initializationOptions` / `configuration` | `null` / `null` | Static initialize options and `workspace/configuration` answer. |
+| `maxMessageBytes` | `16000000` | Largest LSP message accepted from the server. |
+| `maxStderrBytes` | `1000000` | Retained raw server stderr tail. |
+| `maxDocumentBytes` | `4000000` | Largest remote source opened for one query. |
+| `shutdownTimeoutMs` | `5000` | Graceful protocol-shutdown budget. |
+| `killGraceMs` | `2000` | Request-cancel and TERM-to-KILL grace. |
+
+Provider ids and commands are non-empty; numeric bounds are positive safe integers, and timer values cannot exceed Node's maximum timer delay. Setup uploads one owner-private proxy under `ctx.e2b.runtimeRoot`, resolves Node and every configured server executable remotely, then registers all providers atomically.
+
+## Remote protocol and filesystem
+
+E2B command callbacks are text, while LSP is byte-framed. The installed proxy therefore base64-frames raw server stdout, stderr, and stdin as newline-delimited ASCII JSON; the host validates and decodes every frame before handing bytes to the shared `LspInstance` protocol engine. `initialize.processId` is `null` because host and server do not share a process namespace.
+
+One language-server process is pooled per provider and canonical remote workspace. Queries serialize per workspace but different workspaces run concurrently. Each query canonicalizes the remote workspace and source with `realpath`, rejects paths outside that workspace, requires a regular file, enforces the size bound before and after reading, decodes strict UTF-8, and uses the ordinary transient `didOpen` / request / `didClose` lifecycle. A transport failure disposes the instance and retries the read-only query once on a fresh remote process.
+
+The subprocess adapter owns process groups and escalation, so cancellation and disposal await remote server quiescence. The host owns LSP request ids, pending requests, provider queues, and normalized results.
+
+## Model Experience
+
+Indirectly, through `@deepseek-ai/dsh-tool-lsp`, which exposes normalized semantic navigation and hover results without changing its model-facing schema.
+
+#### KV Cache effect
+
+No direct invalidation; `dsh-tool-lsp` owns request-prefix changes.
+
+## Known Limitations and Deferred Work
+
+- **Configured servers only** — this package does not install language servers, select presets, or synchronize a host workspace into E2B.
+- **Host protocol state is not reconnectable** — retaining a sandbox does not restore provider queues, JSON-RPC requests, subprocess handles, or document lifecycle state.
+- **SDK output retention remains** — ASCII framing preserves protocol bytes, but E2B and the subprocess adapter still retain callback output in host memory.
+- **Sandbox policy is template-owned** — this provider adds no volume, snapshot, credential, or network-policy layer.