Follow-up to cache-first registration (c95ec04). The cold-cache wall
clock is now bounded by the slowest individual plugin, not the sum:
- Fan out HTTP and stdio discovery together in goroutines rather than
running the stdio loop serially after HTTP.
- HTTP cold budget: 4s → 700ms (auth) / 500ms (plain). Honest dial
timeouts on unreachable endpoints fail fast; healthy third-party
endpoints respond well under the window. The outcome is still saved
as a negative cache, so a miss this run costs the next run ~0ms.
- Stdio cold budget: 4s → 1s. A local subprocess handshake is milliseconds.
With three user plugins (one pointing at TEST-NET-1, permanently
unreachable), cold `dws --help` drops from ~3.3s to ~0.8s and warm stays
at ~80ms. The savings compound linearly with plugin count.
Follow-up to the 4s startup cap (df01f36). Plugin command registration
now reads the tools snapshot directly from the cache store on the hot
path, and only falls through to an Initialize+ListTools RPC when no
snapshot exists.
- registerHTTPServer / registerStdioServer try cache.LoadTools first;
on hit they build Cobra commands synchronously with zero network I/O.
- Cold cache falls back to the existing synchronous discovery (already
bounded at 2-4s) and persists the outcome — including empty tool
lists — as a negative cache so the next invocation is fast regardless
of endpoint health.
- Cache entries are namespaced under "plugin:<name>:<server>" so they
are distinct from Market cache entries in `dws cache status`, and
refresh on-demand via `dws cache clean` / `dws cache refresh`; the
existing 7d ToolsTTL otherwise expires entries naturally.
Warm-cache `dws --help` with three user plugins (one unreachable) now
returns in ~80ms versus 3.7s with synchronous discovery, a >40x win
when endpoints are offline.
Issue #119: an unreachable third-party MCP plugin (e.g. blocked/firewalled
endpoint) blocks `dws --help` for ~10s on every CLI invocation, because
plugin discovery happens eagerly during command-tree construction.
Root cause was a stack of timeouts that multiplied under transport failure:
* `transport.Client.Initialize` loops three supported protocol versions on
every error — including dial timeouts and HTTP 5xx — even though those
failure modes are protocol-version-independent. Three loops × the dial
budget = 3× the worst-case startup cost.
* Default `DialContext.Timeout` was 10s, so a single dial against an
unroutable address (e.g. TEST-NET-1) burned the full plugin context.
* `registerHTTPServer` granted plugins with `AuthHeaders` a 10s outer
context (intended for slow third-party services), and `registerStdioServer`
granted every stdio plugin 10s. Either single misbehaving plugin therefore
stalled the entire CLI.
* Registry-side `defaultDiscoveryTimeout` (10s) and `perServerDiscoveryTimeout`
(5s) had the same shape on the LoadCatalog path.
Changes:
* `transport.Client.Initialize`: short-circuit when the underlying
`*CallError.Stage` is anything other than `CallStageJSONRPC`. Protocol
version negotiation is the only justification for retrying with another
version; transport/HTTP failures fail identically and should surrender.
* `transport.defaultTransport`: `DialContext.Timeout` 10s → 3s.
* `app.registerHTTPServer`: AuthHeaders timeout 10s → 4s.
* `app.registerStdioServer`: outer ctx timeout 10s → 4s.
* `cli.defaultDiscoveryTimeout`: 10s → 4s.
* `discovery.Service`: rename `perServerDiscoveryTimeout` 5s → 2s
(`defaultPerServerDiscoveryTimeout`); add `Service.PerServerTimeout`
override field for tests/callers needing a tighter or looser bound.
Tests:
* `TestInitializeShortCircuitsOnHTTPError` — asserts only ONE protocol
version is attempted on HTTP 5xx.
* `TestInitializeShortCircuitsOnDialFailure` — asserts Initialize returns
in <2s on a refused-connection address.
* Existing `TestInitializeNegotiatesProtocolVersion` continues to pass —
JSON-RPC-stage errors still trigger version fallback.
End-to-end measurement with a blackhole plugin
(endpoint=192.0.2.1, AuthHeaders set):
| Variant | median `dws --help` |
| ------------- | ------------------- |
| main | 10.2s |
| PR #121 alone | 9.1s |
| this branch | 3.72s (-63%) |
The `loader.go` and `discovery/service.go` timeout reductions overlap with
PR #121 by @utafrali; see PR description for attribution.
Refs: #119
Supersedes: #121
- Support file path reference in plugin.json cli field (e.g. "cli": "overlay.json")
in addition to inline JSON objects, resolving path relative to plugin root
- Add description field to CLIToolOverride for static command descriptions
as fallback when MCP tools/list is unavailable (e.g. upstream server offline)
- Fix plugin install on Windows: use cmd /C instead of sh -c for build commands
- Skip copying identical files during plugin install to avoid overwriting
locked executables (running stdio plugin processes on Windows)
- Add symlink skip and path traversal guard in copyDir for security
- Clean up stale files in destination during plugin upgrade via removeStaleFiles
🤖 Generated with [Qoder][https://qoder.com]
- Add `EnsureManaged` to the plugin updater to automatically install missing default plugins on startup.
- Translate the managed plugin removal error message to English for better consistency.
- Update tests to match the new error message.
Prevent plugins from hijacking built-in commands (auth, plugin, cache,
etc.) and detect duplicate names between plugins.
- Add reservedCommands set for protected built-in command names
- Add addPluginCommandsSafe with three conflict rules:
- Plugin vs reserved → reject with warning
- Plugin vs plugin → first wins, second rejected with warning
- Plugin vs Market → plugin wins (intentional override)
- Fix plugin create scaffold format string arg count
Add `build` field to plugin.json so stdio servers can be automatically
compiled to native binaries during install, eliminating runtime deps.
- Add BuildConfig struct to Manifest (command + output fields)
- Add BuildPlugin/runBuild in loader with output verification
- Auto-trigger build in InstallFromDir and InstallFromGit (rollback on failure)
- Add `dws plugin build <dir>` command for manual builds
- Update `plugin create` scaffold to include build template
Implement 'dws plugin config set/get/list/unset' commands that persist
plugin configuration (e.g. API keys) to ~/.dws/settings.json. Values
are automatically injected as environment variables at plugin load time,
so ${KEY} references in plugin.json headers/endpoints resolve without
manual 'export' each session.
Key changes:
- loader.go: add Get/Set/Unset/List/InjectPluginConfigEnv methods
- plugin_cmd.go: add 'plugin config' subcommand group (set/get/list/unset)
- root.go: call InjectPluginConfigEnv() before plugin loading
- loader_config_test.go: 9 unit tests covering all new functionality
User env vars take precedence over settings.json values.
Sensitive values are masked in 'plugin config list' output.
Add Auth Token Registry pattern (inspired by stdio_registry.go) to allow
plugins to declare per-server HTTP headers in plugin.json. This enables
third-party streamable-http MCP servers (e.g. Alibaba Cloud Bailian) to
use their own API keys independently from the default DingTalk OAuth token.
Changes:
- New: internal/app/auth_registry.go — per-productID auth credential store
- manifest.go: MCPServer gains Headers field for custom HTTP headers
- registry.go: ServerDescriptor gains AuthHeaders field
- converter.go: resolve and pass headers through ToServerDescriptors
- root.go: inject plugin auth at discovery time + auto-generate ToolOverrides
- runner.go: route plugin-auth servers to their own Bearer token at runtime
Multi-token isolation: each server is keyed by its CLI.ID (productID),
so different servers can use independent tokens without interfering with
each other or with the default DingTalk OAuth token.
- Fix schema-derived flags not being collected into tool call params
- Add registerHTTPServer for plugin streamable-http tool discovery
- Include dev plugins in ListInstalled output
- Isolate settings path for test environments
- Add stdio and HTTP MCP end-to-end integration tests
Stdio MCP servers discovered by the plugin loader now have their tools
automatically registered as CLI subcommands. The runner dispatches
stdio:// virtual endpoints to the local StdioClient subprocess instead
of the HTTP transport, enabling plugin tools to be invoked directly
from the command line (e.g. `dws hello greet --name Peter`).
Also fixes a bug where plugin endpoints registered via AppendDynamicServer
were overwritten by the subsequent SetDynamicServers call in
loadDynamicCommands, and a context lifecycle bug where stdio subprocesses
were killed immediately after startup due to a short-lived timeout context.
- Wire updater into loadPlugins to check managed plugin updates on
CLI startup with 10s timeout and best-effort semantics
- Add StdioClients() to Plugin for creating stdio transport clients
with DWS_PLUGIN_ROOT/DWS_PLUGIN_DATA variable expansion
- Start and initialize stdio MCP server subprocesses during plugin load
- Reorder loadPlugins: update check → load → http inject → stdio start → hooks
- Add updater.go for managed plugin auto-update (version check, download,
Y/n prompt, zip extraction with zip-slip protection)
- Add stdio.go transport for local MCP server subprocess communication
via stdin/stdout JSON-RPC 2.0
- Implement InstallFromGit in loader.go with URL parsing for HTTPS/SSH
- Wire up --git flag in plugin install command
- Switch all plugin CLI output to English (headers, status, messages)
- Add tests for parseGitURL and promptUpdate
Introduce the plugin system for DWS CLI, enabling official and third-party
plugin management. This includes plugin manifest parsing/validation, loader
with managed/user directory-based identity, MCP server conversion and
injection into the dynamic routing registry, pipeline hook adapter for
shell-based hooks, and CLI commands (list/install/info/enable/disable/
remove/validate).
- Add ExitCoder interface for edition-specific error types to provide custom exit codes
- Add RawStderrError interface for errors that bypass CLI formatting and output raw content to stderr
- Update ExitCode() to resolve exit codes via ExitCoder interface before falling back to default
- Add ClassifyToolResult hook in edition.Hooks for custom MCP tool result error classification
- Invoke ClassifyToolResult in runner before default business-error detection
- Handle RawStderrError in printExecutionError to pass raw JSON through to desktop runtime
- Add unit tests for ExitCoder and RawStderrError interfaces
- Add skill find (keyword search) and skill get (download to temp dir)
using mcp.dingtalk.com endpoints; keep skill add (aihub download)
- Add hidden skill search hint for old usage
- List utility commands in root help; keep skill visible (not hidden)
- Extend tests for new subcommands and root help
Made-with: Cursor
EnvironmentLoader.Load() previously returned (empty Catalog, nil) on all
failure paths, making it impossible for callers to distinguish "no services
available" from "discovery failed due to auth/network issues".
Changes:
- Add CatalogDegraded error type with three reasons: unauthenticated,
market_unreachable, runtime_all_failed
- Add auth pre-check: return DegradedUnauthenticated immediately when no
token is available, avoiding doomed MCP connections
- Schema command now handles CatalogDegraded gracefully: outputs hint to
stderr and includes degraded/reason/hint fields in JSON output
- Runner preserves graceful degradation by ignoring CatalogDegraded errors
- Skip empty-products cache to prevent stale cache from masking auth errors
- Edition-aware hint text (open-source vs embedded/wukong)
Made-with: Cursor
Extract resolveAccessTokenFromDir from getCachedRuntimeToken so both
MCP and non-MCP clients (e.g. A2A gateway) share the same OAuth +
legacy fallback logic including host compatibility hooks.
- Add ResolveAuxiliaryAccessToken in internal/app for overlay use;
reuses process-level token cache when configDir matches the edition
default, avoiding repeated Keychain access.
- Simplify pkg/runtimetoken to a thin delegate to the shared impl.
- Refactor getCachedRuntimeToken to call resolveAccessTokenFromDir,
removing duplicated provider/manager setup code.
Made-with: Cursor
- Add token validation in executeInvocation to reject requests early
- Return clear Chinese error message: '未登录，请先执行 dws auth login'
- Include actionable hint and suggested command in error response
- Avoid wasting network resources on HTTP 400 from MCP gateway
Also includes:
- Add DWS_PERF_TIMING env for CLI execution timing analysis
- Add TimingCollector to track cmd_init, auth_token, mcp_call durations
Test changes:
- Add --token flag to tests that require authentication
- Add TestRuntimeRunnerRejectsUnauthenticatedRequest test case
修雨