Repository navigation
Replies: 2 comments
|
I have a concrete use case for this: DeepSeek Harness (dsh). I already run its ACP server through T3 Code (orchestrator v2), but it has no official ACP Registry entry. My workaround is to select another registered agent and override its executable with the dsh ACP server. That works in my setup, but T3 displays the borrowed registry entry’s icon, making the provider’s identity misleading. A manually configured ACP provider with its own command, arguments, display name, and a neutral ACP icon would solve this. It would also avoid inheriting launch defaults from an unrelated registry entry. EnvironmentT3 Code: 0.0.46-nightly.20261004.2644 @juliusmarminge would registry-independent ACP configuration fit the direction of the current ACP integration? |
|
Note Grok responding on behalf of Julius. This landed in #16021 (merged today). Provider settings now have a local-command ACP source: you give it a display name, an executable, a literal argument array, and environment overrides, and T3 spawns it with |
Uh oh!
There was an error while loading. Please reload this page.
T3 Code already has the two hard parts for this use case: a reusable ACP session runtime and a
ProviderDriver/ provider-instance seam. A downstream that wants to expose another standards-compatible ACP process currently still has to add a first-party driver and aBUILT_IN_DRIVERSregistration.Would you be open to a small, product-neutral generic ACP driver?
Proposed boundary:
acpdriver;The first slice would remain stdio-only. Remote HTTP/WebSocket transport, marketplace/catalog behavior, product identity, and vendor-specific protocol extensions would stay out of scope.
The observable contract would be: a configured ACP command can create and resume a session, stream the existing canonical updates, handle approval/cancel, and fail as an unavailable provider instance when the process cannot start. No new session UI should be necessary.
If this direction fits the project, I can prepare a focused PR limited to the generic driver/config schema, its built-in registration, and one public-boundary round-trip test. Is there a preferred shape for the optional authentication resolver, or should the first contribution leave authentication entirely to the configured command wrapper?
All reactions