[P2] Docker runtime config loads MCP servers whose prolog-mcp and gitnexus commands are absent #109

New issue

Open

opened 2026-09-06 04:15:56 +00:00 by nsaspy · 1 comment

nsaspy commented

2026-09-06 04:15:56 +00:00

Owner

Summary

The live Docker runtime repeatedly configures the prolog and gitnexus MCP servers even though neither command exists in the image.

Exact live evidence (2026-09-06)

Since the current container started, logs contain:

9 occurrences of Command 'prolog-mcp' not found.
9 occurrences of Command 'gitnexus' not found.

The failures occur during initial preload and again when scheduled task contexts initialize. The application continues and reports healthy, so agents silently operate with fewer tools than their configuration promises.

This is separate from the Prolog-RLM in-process runtime, which reports ready through its own smoke test.

Impact

Chats/tasks lose configured Prolog and repository-index tools.
Repeated initialization noise hides actionable model/runtime failures.
Health remains green despite broken configured capabilities.

Acceptance criteria

Either include the configured commands in the Docker image/runtime PATH or do not activate those MCP server entries in that deployment.
Validate every enabled stdio MCP command at startup.
Surface one durable, structured capability error instead of repeating the same missing-command traceback for each context.
Make health/readiness expose configured-but-unavailable mandatory MCP servers.
Add a container smoke test that resolves and starts every enabled stdio MCP command without network/provider spend.

## Summary The live Docker runtime repeatedly configures the prolog and gitnexus MCP servers even though neither command exists in the image. ## Exact live evidence (2026-09-06) Since the current container started, logs contain: - 9 occurrences of Command 'prolog-mcp' not found. - 9 occurrences of Command 'gitnexus' not found. The failures occur during initial preload and again when scheduled task contexts initialize. The application continues and reports healthy, so agents silently operate with fewer tools than their configuration promises. This is separate from the Prolog-RLM in-process runtime, which reports ready through its own smoke test. ## Impact - Chats/tasks lose configured Prolog and repository-index tools. - Repeated initialization noise hides actionable model/runtime failures. - Health remains green despite broken configured capabilities. ## Acceptance criteria - Either include the configured commands in the Docker image/runtime PATH or do not activate those MCP server entries in that deployment. - Validate every enabled stdio MCP command at startup. - Surface one durable, structured capability error instead of repeating the same missing-command traceback for each context. - Make health/readiness expose configured-but-unavailable mandatory MCP servers. - Add a container smoke test that resolves and starts every enabled stdio MCP command without network/provider spend.

nsaspy added the

bug

label

2026-09-06 04:15:56 +00:00

nsaspy commented

2026-09-06 10:02:11 +00:00

Author

Owner

the mcp server stuff can be found in my dotfiles.
https://git.starintel.actor/nsaspy/dotfiles

the mcp server stuff can be found in my dotfiles. https://git.starintel.actor/nsaspy/dotfiles