Replace the Substrate turn transport with a kagent A2A client

[oldsj]

Native Claude Code and Codex turns now run on kagent harness agents, from the oldsj/kagent fork. Mainloop is an A2A client of the kagent gateway. There is no feature flag and no compatibility path.

This is slice (a) of the kagent migration. Workspaces, previews and agent-to-Mainloop tools follow in later PRs (see Known gaps).

What's in it

• Client (runtime/kagent_client.py): A2A JSON-RPC for turns (SendStreamingMessage, SubscribeToTask, GetTask, ListTasks, CancelTask) and SessionService over grpc-web for Session lifecycle.
• Delivery rules:
  • every message is recorded before it is sent and never replayed;
  • "send not accepted" retries the identical message within 30s;
  • any other uncertain send is resolved by looking up its messageId across all ListTasks pages, then re-reading the task so the reply is kept.
• Session flow (runtime/native_sessions.py):
  • only the pass that claims a recorded delivery sends it, so a cancel or a second pass can't send it again;
  • after a restart, recorded deliveries are sent once;
  • a kagent Session deleted by idle expiry is replaced under a fresh request id, with standing context resent.
• Atomic session creation: main and child sessions are created with their bindings in one transaction; the main session takes an advisory lock.
• Interim guard: POST /workspaces returns 409 until workspaces move to kagent.
• Events: new SSE event turn:updated (no reply text yet). Fixes enum encoding in SSEEvent.
• Removed: the journal, projection, contract and per-provider stream modules, the native_events and native_lineage tables, main-thread rotation and its settings, and the docs describing them.
• Docs: specs, architecture, README, and AGENTS.md/ROADMAP.md now name kagent as the workspace runtime.

Verification

• Backend suite against PostgreSQL 18: 332 tests, OK, none skipped. make lint and svelte-check clean.
• Live, against a kind cluster running the kagent fork, with Claude and Codex subscription agents through the Mainloop API: create, send and stream, restart reconciliation without a duplicate send, suspend and resume with context, delete, and a clear failure when kagent is unreachable.
• Fixtures only: the "send not accepted" retry, Session replacement, and recovery of a lost reply from an already-completed task.

Known gaps

• Agents can't call Mainloop tools yet. Delegation, notes, topics and reports don't work from kagent agents; children finish only through auto-report. The next slice adds an MCP tool channel with per-session identity.
• Workspaces: creation returns 409 until the workspace slice.
• No "stop current turn": a turn waiting for input blocks the main thread until the approvals slice adds it.
• Main-thread expiry: kagent's idle TTL is controller-wide (7 days). After that the main thread gets a fresh Session with standing context; a per-agent TTL is planned.
• Security: the kagent gateway runs unauthenticated. Use only on isolated clusters until gateway auth and NetworkPolicy land.