internal/tools/provider/provider.go Start() is atomic by design -- 'Any failure rolls the complete family back' -- so a single failing provider unregisters every other provider including the builtins, capabilityHost.Start returns the error, and cmd/archied/main.go:655 exits 1. Under compose restart or systemd Restart=on-failure that is a crash loop. Observed 2026-07-30 on carina: mcp.desktop-commander could not start (missing unzip in the runtime image, fixed in 455fd6f) and archied restarted continuously, losing chat, the gateway and every builtin tool over one third-party npm package.\n\nAtomicity is correct for a provider the daemon requires. It is wrong for an external MCP server pulled from a registry at runtime, which is exactly the category that should be allowed to be absent -- the same reasoning applied to the forge in 518a1a0, where a missing credential now disables the capability rather than the daemon.
Design
Needs a required/optional distinction on the registration or manifest, which does not exist today. Keep the all-or-nothing path for required providers; the atomic batch semantics for tool visibility (tools appear only after Start and Discover both succeed) should hold per-provider either way. Relates to the plugin engine rule in ARCHITECTURE.md: failure isolation is explicitly the owning manager's responsibility.
Acceptance Criteria
A provider can be declared optional. An optional provider that fails to start or discover is logged, excluded from the running set, and does not roll back the family or stop the daemon. A required provider keeps today's atomic rollback. Provider health reflects the degraded state so the failure is visible rather than silent. Regression test starts a family containing one failing optional provider and asserts the other providers' tools remain registered and Start returns nil.
Imported from beads archie-core-1786637499393-366-5c34a02a · GitHub #387 · type bug · priority P2 · created 2026-08-08 · closed 2026-08-08
internal/tools/provider/provider.go Start() is atomic by design -- 'Any failure rolls the complete family back' -- so a single failing provider unregisters every other provider including the builtins, capabilityHost.Start returns the error, and cmd/archied/main.go:655 exits 1. Under compose restart or systemd Restart=on-failure that is a crash loop. Observed 2026-07-30 on carina: mcp.desktop-commander could not start (missing unzip in the runtime image, fixed in 455fd6f) and archied restarted continuously, losing chat, the gateway and every builtin tool over one third-party npm package.\n\nAtomicity is correct for a provider the daemon requires. It is wrong for an external MCP server pulled from a registry at runtime, which is exactly the category that should be allowed to be absent -- the same reasoning applied to the forge in 518a1a0, where a missing credential now disables the capability rather than the daemon.
### Design
Needs a required/optional distinction on the registration or manifest, which does not exist today. Keep the all-or-nothing path for required providers; the atomic batch semantics for tool visibility (tools appear only after Start and Discover both succeed) should hold per-provider either way. Relates to the plugin engine rule in ARCHITECTURE.md: failure isolation is explicitly the owning manager's responsibility.
### Acceptance Criteria
A provider can be declared optional. An optional provider that fails to start or discover is logged, excluded from the running set, and does not roll back the family or stop the daemon. A required provider keeps today's atomic rollback. Provider health reflects the degraded state so the failure is visible rather than silent. Regression test starts a family containing one failing optional provider and asserts the other providers' tools remain registered and Start returns nil.
---
_Imported from beads `archie-core-1786637499393-366-5c34a02a` · GitHub #387 · type bug · priority P2 · created 2026-08-08 · closed 2026-08-08_
sam
added the mediumbug labels 2026-10-27 13:38:11 +00:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
internal/tools/provider/provider.go Start() is atomic by design -- 'Any failure rolls the complete family back' -- so a single failing provider unregisters every other provider including the builtins, capabilityHost.Start returns the error, and cmd/archied/main.go:655 exits 1. Under compose restart or systemd Restart=on-failure that is a crash loop. Observed 2026-07-30 on carina: mcp.desktop-commander could not start (missing unzip in the runtime image, fixed in 455fd6f) and archied restarted continuously, losing chat, the gateway and every builtin tool over one third-party npm package.\n\nAtomicity is correct for a provider the daemon requires. It is wrong for an external MCP server pulled from a registry at runtime, which is exactly the category that should be allowed to be absent -- the same reasoning applied to the forge in 518a1a0, where a missing credential now disables the capability rather than the daemon.
Design
Needs a required/optional distinction on the registration or manifest, which does not exist today. Keep the all-or-nothing path for required providers; the atomic batch semantics for tool visibility (tools appear only after Start and Discover both succeed) should hold per-provider either way. Relates to the plugin engine rule in ARCHITECTURE.md: failure isolation is explicitly the owning manager's responsibility.
Acceptance Criteria
A provider can be declared optional. An optional provider that fails to start or discover is logged, excluded from the running set, and does not roll back the family or stop the daemon. A required provider keeps today's atomic rollback. Provider health reflects the degraded state so the failure is visible rather than silent. Regression test starts a family containing one failing optional provider and asserts the other providers' tools remain registered and Start returns nil.
Imported from beads
archie-core-1786637499393-366-5c34a02a· GitHub #387 · type bug · priority P2 · created 2026-08-08 · closed 2026-08-08