plat 468

← Pulse platform issue index

PLAT-468 — Muse chats on a Mac start with no MCP bridge (Seatbelt refused Muse's read of the folders above its working folder)

Coordination	Value
State	fixed on main (provider 4dba54e, pinned e6cc149); needs a rebuild and backend restart
Date	2026-10-04
Owner	coding-agent-bridge
Related	PLAT-394 (Mac Seatbelt for all CLIs), PLAT-463

Source

The owner's Muse Builder chats for jobsearch and upwork could not call any platform tool (get_contract_upgrades, set_workflow_contract_version, decisions). The server built the bridge config (6 tools) and wrote mcpServers.api-bridge into Muse's settings, but no mcpbridge process ever started. Muse printed MCP configuration error in <runtime>/.mcp.json; MCP is disabled for this runtime. The jobsearch agent then hand-edited workflow.json to stamp contract 1.0.45.

Cause

Muse's own bootstrap log (~/.local/share/muse/sessions/<date>/<id>/cli-*.log) says mcp.config.resolve stage="startup" outcome="failed" reason="source_reservation" server_count=0. Muse walks up from its working folder looking for project sources and must read every folder on the way. Under the Mac Seatbelt the runtime folder lives inside the closed AgentWorks state area, whose parent folders were allowed metadata-only, so the read was refused, MCP startup failed and the chat had no bridge. It was reproduced outside the app by running the Muse TUI under the same kind of profile (fails), then adding a literal file-read-data grant on the four folders above the runtime folder (works; any three of the four still fail). Project rules (AGENTS.md), the path's space, the environment and an empty .mcp.json were each ruled out by experiment. Codex and Claude do not walk up this way and were unaffected.

Done

• multi-llm-provider-go internal/clisandbox/seatbelt.go: for muse-cli the ancestor folders get (allow file-read-data (literal ...)) next to the metadata grant. Names of siblings become listable; nothing inside the closed folders opens. Other CLIs' profile is unchanged.

• seatbelt_darwin_test.go: under the real sandbox, Muse can list the three folders above its runtime folder, Codex cannot, and Muse still cannot list or read another runtime.

• agent_go/go.mod pins multi-llm-provider-go at e6cc149 (includes 4dba54e). The first push of the provider fix did not bump this pin, so the server built after the owner's 19:57 restart still had the old sandbox and the relaunched Muse chats (muse resume) failed the same way. A provider change reaches the app only through this pin.

Left

• Takes effect on the next backend restart (the running server holds the old code).
• Chats started before the restart stay without a bridge; start a new chat or switch provider and back.
• Workflows migrated by hand-editing workflow.json while the bridge was missing (jobsearch) should be re-checked with get_contract_upgrades once a chat has the bridge. A Muse run with project rules and the real bridge was not re-run end to end after the fix; the reproduction used the same profile shape.