fix(mcp): re-read server config from disk on every respawn#71116
Open
dportier1021-crypto wants to merge 1 commit into
Open
fix(mcp): re-read server config from disk on every respawn#71116dportier1021-crypto wants to merge 1 commit into
dportier1021-crypto wants to merge 1 commit into
Conversation
MCPServerTask.run() captured the server config once at first connect and reused that snapshot for every respawn in the reconnect loop. Credentials rotated into config.yaml by external tooling (e.g. a pre-run script refreshing BW_SESSION) never reached respawned children — the only escape was a full process restart. The loop now re-reads _load_mcp_config() (mtime-aware via load_config()) at the top of every (re)spawn iteration, covering auto-reconnect, parked revival, and lazy recycle paths. Falls back to the last known config if the refresh fails or the server was removed from config. Regression test: respawn after simulated transport death picks up the refreshed env marker; removed-from-config server keeps last known config without crashing.
Contributor
There was a problem hiding this comment.
Pull request overview
This PR fixes MCP server respawn behavior so that when an MCP stdio/HTTP child process is restarted (e.g., after being killed to force a reconnect), the server’s configuration is re-read from disk each spawn iteration, allowing externally-rotated credentials in config.yaml (like BW_SESSION) to take effect without requiring a full process restart or manual /reload-mcp.
Changes:
- Refresh MCP server config from disk at the top of each respawn iteration in
MCPServerTask.run(). - Update in-memory task config used for transport startup to reflect the refreshed on-disk config.
- Add regression tests covering refreshed env propagation on respawn and resilience when a server is removed from config.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
tools/mcp_tool.py |
Re-reads MCP server config on every (re)spawn iteration before launching transport. |
tests/tools/test_mcp_config_refresh.py |
Adds tests asserting respawn picks up refreshed config and tolerates server removal. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+3109
to
+3123
| try: | ||
| fresh_config = _load_mcp_config().get(self.name) | ||
| if isinstance(fresh_config, dict) and fresh_config: | ||
| if fresh_config != config: | ||
| logger.info( | ||
| "MCP server '%s': config changed on disk, " | ||
| "(re)spawning with fresh config", | ||
| self.name, | ||
| ) | ||
| config = fresh_config | ||
| self._config = fresh_config | ||
| except Exception: | ||
| # Never let a config-refresh failure break the reconnect | ||
| # loop — fall back to the last known config. | ||
| pass |
Comment on lines
+32
to
+34
| # Keep backoff sleeps negligible. | ||
| monkeypatch.setattr(mcp_tool, "_MAX_BACKOFF_SECONDS", 0.01) | ||
| monkeypatch.setattr(mcp_tool, "_CONNECT_RETRY_BASE_BACKOFF_SEC", 0.01) |
alt-glitch
added
type/bug
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Symptom
An MCP stdio server whose credentials are rotated in
config.yamlby external tooling never receives the new values. Concrete case: a cron pre-run script unlocks a Bitwarden vault and writes a freshBW_SESSIONintomcp_servers.bitwarden.env.BW_SESSION, then kills the stalemcp-server-bitwardenchild so the gateway respawns it. The respawned child still comes up with the old token — verified live viaps ewwon three independent long-lived hosts (gateway, webui, desktop-serve), each holding a different stale token, none matching the just-written config value. The only escapes were a full process restart or/reload-mcp.Root cause
MCPServerTask.run()capturesconfigonce (self._config = config) and the reconnect loop'swhile True:respawns via_run_stdio(config)/_run_http(config)using that same snapshot forever._load_mcp_config()(and the mtime-awareload_config()beneath it) is only called fromdiscover_mcp_tools()/ status paths — never on the respawn path. Killing the child therefore can't propagate rotated env vars, because the auto-reconnect reuses the first-connect snapshot.Fix
Re-read
_load_mcp_config()at the top of every (re)spawn iteration inside the reconnect loop, and adopt the fresh config when it differs. This one site covers every respawn path (auto-reconnect after failure, parked revival, lazy stdio recycle, auth-recovery rebuild) because they all funnel through this loop.load_config()is mtime-aware, so the refresh is cheap when the file is unchanged. If the refresh raises or the server was removed from config, the loop falls back to the last known config — a config-refresh failure can never break reconnection.Test
tests/tools/test_mcp_config_refresh.py(2 tests):MARKERone → two)Existing MCP suites still pass:
test_mcp_tool.py,test_mcp_bridge_single_failure.py,test_mcp_reload_rev.py— 231 passed.