Harness is a plugin that speaks gRPC A2A

"Harness is a plugin that speaks grpc A2A"

"NOT BE ENTIRELY DEPENDENT ON ONE HARNESS", "WHY ISNT PI SEPARATE" ([[ADR-048-Node-Agent]]).

"I see multiple harnesses existing side by side, as in various Claude Code sessions on my laptop all also being under the umbrella of the QNTX node on it, and also the ones on the node on my desktop and the other ones on the node on a server"

Status

Proposed.

Decision

Each harness (Claude Code, Pi, Codex, agy) is a plugin that provides a service, as an LLMProvider does ([[ADR-001-domain-plugin-architecture]]): "A plugin that provides a service registers it with core. Other plugins can consume these services through ServiceRegistry without knowing which plugin implemented the .proto." The .proto is A2A's own A2AService, served over gRPC (A2A §10), a2a-go's a2agrpc.NewHandler.

"Core mediates everything" ([[ADR-003-plugin-communication]]). Core speaks to a harness with a2a-go's client and a2agrpc.WithGRPCTransport, as it speaks to a peer's agent and to the agent a card names (server/a2a_client.go). A plugin that wants an agent, a Rooms plugin among them ([[ADR-057-Agent-Rooms]]), asks core, not the harness.

A harness is added, configured, swapped and built as every plugin is ([[ADR-002-plugin-configuration]]), in any language gRPC speaks: Pi's own included.

One harness service

One harness service, core's, with Claude Code, Pi and Codex its backends, as LLMServer has its providers ([[ADR-014-llm-as-plugin-provided-service]]): "Yes. Do this"

What holds there holds here:

Side by side

How a session on a machine is told

"SESSION A ON MY LAPTOP HAS A NAME" "THAT NAME IS NOT TAKEN" "GROUND IS THE DELIVERY NETWORK" "AND IT HAPPENS VIA UG" "BECAUSE UG SEES THE STATUSLINE OF QNTX" "AGENTS CAN TALK TO EACH OTHER BY THEIR NAME" "AND GROUND DELIVERS IT VIA UG AND STATUSLINE" "A SESSION THATS SENDING TO ANOTHER SESSION BY NAME IS TOLD ABOUT THE MESSAGE BEING RECEIVED OR NOT, AND WE DONT INVENT HERE, WE TAKE THE A2A CONVENTIONS, WE TAKE THE SDK SERIOUSLY" "EACH SESSION HAS ITS OWN GIT IDENTITY" "SESSIONS HAVE THEIR OWN MAIL INBOX" "AGENTS HAVE THEIR OWN INBOX" "AN UNAVAILABLE AGENT MAY BE EMAILED"

"Messages from QNTX can end up in ground and we need no new path"

QNTX writes a message into ground's db, and ground hands it to the session:

  1. The node leaves the message on the status row as news, for the person, naming the session it is for (server/statusline/news.go). It stays there two minutes (NewsHold). CI verdicts are left this way today (server/ci_pulse.go).
  2. ug polls the status row and writes each item it has not seen into ground's db: one immediate:news row in the context session:<id>, or project:<repo> when the node names no session (newsPass in ground ug/tmux.d, and ug/sql.d). It does so while ug tmux draws (ground UNDERGROUND.md, Limitation).
  3. sky reads what is addressed to its session or its project and wakes the session with it. A delivered:<id> row per session makes it once (ground source/sky.d, source/immediate.d).

The node knows which session to name because every row a ground hook writes goes to the node as well, in a session:<id> context (ground source/stream.d). am ground answers the sessions started and left (server/am_ground.go).

A deferred message is the same row with an after gate (ground CLAUDE.md).

A session's key and its name

"I have two nodes running at the moment" "Why wouldn’t the agents talk to each other through A2A?" "Why can’t we make Node Agents aware of each other?" "As Peer Node" "The condition is no handwritten A2A, use spec faithfully, be opportunistic, make sure discovery is implemented in a way that we can move forwards and make it ship to sentry, so next time I come back when I have it deployed and then I can ask you"

Agents talk to each other through A2A: first two nodes, then their Node Agents as peers (a2a Peers, server/a2a_peer_client.go), on a2a-go's server and client and nothing hand-written (server/a2a_route.go, server/a2a_client.go). A session is the next agent in that line.

"The true way forwards is A2A" "if every session is its own addresable A2A via QNTX so Alder would be addressable" "I WANT TO MAKE EVERY SESSION BE ABLE TO NAME ITSELF" "session's having their own name" "and sessions naming themselves" "key per session"

"I HAVE 4 SESSIONS ON MY LAPTOP" "EACH OF THEM SHOULD BE ABLE TO GIVE THEMSELVES A NAME THAT HAS NOT BEEN TAKEN" "I HAVE A DESKTOP CALLED K" "ON K I EXPECT EACH SESSION TO ALSO HAVE ITS OWN KEY AND NAME" "I EXPECT K SESSIONS TO BE ABLE TO TALK TO SESSIONS ON MY LAPTOP IF CONNECTIVITY ALLOWS FOR IT" "I WANT A SESSION TO BE ABLE TO GIVE ITSELF AN AVAILABLE NAME, AMONGST ALL PARTICIPANTS IN THE CLUSTER" "SO, ONE ALDER" "ONE SLATE" "ONE OTEL"

Where it is built:

What the repository already answers

Related