What We Shipped This Week — 2026-03-28/29

2026-03-29

Two days, eight merged PRs, ~4000 lines. Here’s what landed.

A2A Protocol — Milestone 1

Personas can now delegate tasks to each other over Talon’s internal queue. The A2A layer is fully durable — tasks persist in SQLite, survive restarts, and go through the same retry logic as regular messages.

Milestone 1 covers internal routing only: persona-to-persona delegation within a single daemon instance. The Hono HTTP server, Agent Cards, and external A2A endpoints are Milestone 2.

persona_send — MCP Host Tool

New host tool: persona_send. Agents call it to submit tasks to other personas. Supports fire-and-forget (await_reply: false) and synchronous delegation (await_reply: true, 25s cap, 2s poll interval).

The capability gate is persona.send:* — add it to a persona’s allow list to unlock the tool.

persona_list — Discovery Tool

Companion to persona_send. Returns all live-loaded personas (excluding the caller) with name, description, and skills. Description is extracted from the first line of the persona’s system prompt file.

Agents use this to decide who to delegate to before calling persona_send.

Queue Inflight Lock Bypass for A2A Collaboration

Fixed a deadlock that occurred under await_reply: true. When persona A submits a task to B and waits for the result, A holds its queue item’s inflight lock. If the queue processor refused to dequeue B’s item while A was running, everything stalled.

The fix: A2A collaboration items bypass the inflight lock check so they can be dequeued and processed while the source task is still in-flight.

Configurable Query Timeout

DEFAULT_QUERY_TIMEOUT_MS was 3 minutes. That’s too short for complex tasks. It’s now 10 minutes by default, configurable per persona via queryTimeoutMinutes in talond.yaml:

personas:
  - name: researcher
    queryTimeoutMinutes: 20

The 3-minute default was causing silent failures on longer agent runs.

input-required State in persona_send Polling

When a target persona enters input-required state (waiting for human confirmation), that state now surfaces in the persona_send polling response. Previously it would just keep polling until timeout.

Hop Count Propagation

Hop count propagates through the tool execution context. Each persona_send call increments it. Tasks rejected at MAX_HOPS (4) get a clear error message. The counter flows correctly through nested delegation chains.

Full Docs Site

Docs shipped at talond.dev. Starlight (Astro) site covering installation, configuration, channels, personas, skills, scheduling, background agents, and A2A. Including this blog.

That’s the week. Next up: A2A Milestone 2 (external HTTP endpoint + authentication) and M3 (runtime monitoring UI).