fix(lsp): align operations and harden lifecycle

This commit is contained in:
Tianyi Cui
2026-07-21 13:29:40 +08:00
parent 2a55f6b684
commit 2fd995bf3c
46 changed files with 680 additions and 366 deletions

View File

@@ -9,7 +9,7 @@ Namespace plugin (`name` / `inject` / `Config` / `apply`, no default export).
- Resolves every server-local setting before registration; an invalid mapping or registration conflict rolls back earlier entries, so a failed load leaves no provider routes.
- Lazily single-flights one server process per `(server id, canonical workspace realpath)`. A crash fails the active query without replay; a later query may replace the process.
- Uses a compatibility-first **transient-open** sequence per query: canonicalize and read the source with Node APIs, `textDocument/didOpen` (version 1, full text), the requested request, then `textDocument/didClose` in `finally`. Documents close after each call, so the first version needs no `didChange`, content cache, or document LRU.
- Serializes queries through one abortable per-instance queue so a cancellation that fails to stop the server can terminate it without killing unrelated work; distinct instances run in parallel.
- Serializes each source-read/open/query/close lifecycle through one abortable per-workspace queue so queued calls read current source only when their turn starts; distinct workspaces run in parallel.
- Reads sources through Node filesystem APIs in the subprocess's host namespace — NOT `ctx.fs`, and emits no `fs/observed`: only the LSP result is model-visible, so a query does not satisfy read-before-write policy.
## Configuration
@@ -28,13 +28,13 @@ The `servers` record key is the stable provider id reserved on `ctx.lsp`; each v
| `maxStderrBytes` | `1000000` | Largest stderr tail retained for diagnostics. |
| `maxDocumentBytes` | `4000000` | Largest source file this host will open. |
| `shutdownTimeoutMs` | `5000` | Graceful `shutdown`/`exit` budget before escalation. |
| `killGraceMs` | `2000` | SIGTERM→SIGKILL grace after graceful shutdown fails. |
| `killGraceMs` | `2000` | Grace for request cancellation and for SIGTERM→SIGKILL escalation. |
`servers` must contain at least one entry, and every id must be non-empty. All executables resolve at load after credential scrubbing; a bad later entry prevents every provider from registering. Processes launch lazily on the first matching query.
`servers` must contain at least one entry, and every id must be non-empty. Timer budgets must be positive integers no greater than Node's `2_147_483_647` ms timer limit. All executables resolve at load after credential scrubbing; a bad later entry prevents every provider from registering. Processes launch lazily on the first matching query.
## Protocol behavior
Initialization advertises `general.positionEncodings: ['utf-16']`, `workspace: { workspaceFolders: true, configuration: true }`, `textDocument.hover.contentFormat: ['markdown', 'plaintext']`, and `linkSupport: true` for definition and implementation, with no dynamic registration. The server's returned capabilities are authoritative: an unsupported operation, or synchronization without transient open/close, fails the query. An omitted server `positionEncoding` defaults to `utf-16`; any other value is a protocol error. The client answers `workspace/configuration` from static config, accepts lifecycle bookkeeping requests, and rejects `workspace/applyEdit` — it never applies edits or runs commands. Navigation maps `Location` directly and `LocationLink` from `targetUri` + `targetSelectionRange`; hover normalization takes `MarkupContent.value`, preserves string `MarkedString`s, renders language-tagged values as fenced code, and joins arrays with one blank line.
Initialization advertises `general.positionEncodings: ['utf-16']`, `workspace: { workspaceFolders: true, configuration: true }`, `textDocument.hover.contentFormat: ['markdown', 'plaintext']`, and `linkSupport: true` for definition and implementation, with no dynamic registration. The server's returned capabilities are authoritative: an unsupported operation, or synchronization without transient open/close, fails the query. An omitted server `positionEncoding` defaults to `utf-16`; any other value is a protocol error. The client answers `workspace/configuration` from static config, accepts lifecycle bookkeeping requests, and rejects `workspace/applyEdit` — it never applies edits or runs commands. Navigation maps `Location` directly and `LocationLink` from `targetUri` + `targetSelectionRange`; hover normalization takes valid `MarkupContent.value`, preserves string `MarkedString`s, renders language-tagged values as fenced code, and joins arrays with one blank line. Missing results, malformed ranges or positions, and malformed hover encodings fail as structured `LSP_MALFORMED_RESPONSE` errors.
## Security boundary
@@ -50,6 +50,6 @@ No direct invalidation; `dsh-tool-lsp` owns request-prefix changes.
## Known Limitations and Deferred Work
- **Trusted host-local only** — no sandbox confinement, no private cache/temp write contract; supporting untrusted binaries or restricted/remote/virtual workspaces requires a later process/filesystem contract and a different provider ([seam Agent Note](../../../.agents/notes/implemented/architecture/2026-07-15-lsp-capability-seam.md)). Containment resolves `realpath`, then opens the source through one handle with `O_NOFOLLOW` (final-component symlink guard) and a bounded read; a concurrent mutator that swaps an *ancestor* directory for a symlink between the resolve and the open is an accepted residual TOCTOU under this trusted-deployment model, not closed with non-portable `openat` segment walks.
- **Trusted host-local only** — no sandbox confinement, no private cache/temp write contract; supporting untrusted binaries or restricted/remote/virtual workspaces requires a later process/filesystem contract and a different provider ([seam Agent Note](../../../.agents/notes/implemented/architecture/2026-07-15-lsp-capability-seam.md)). Containment resolves `realpath`, then opens the source through one handle with `O_NOFOLLOW | O_NONBLOCK` (final-component symlink guard plus nonblocking rejection of FIFOs) and a bounded read; a concurrent mutator that swaps an *ancestor* directory for a symlink between the resolve and the open is an accepted residual TOCTOU under this trusted-deployment model, not closed with non-portable `openat` segment walks.
- **Transient-open compatibility floor** — servers whose synchronization omits open/close (or advertise `None`) are unsupported even if closed-document queries would work; the pinned TypeScript e2e establishes one compatibility floor, not a cross-language claim.
- **Per-server/workspace serialization latency** — parallel agents sharing one server and workspace queue behind one process; long-lived workspace processes consume memory until disposal.