Agent Chat Healthcheck

Find user friction in the chats of local agents in a time window. Propose new skills, agents, or routines only when evidence supports them. Do not create anything until the user picks a proposal.

When to run

• the standing transcript-healthcheck routine runs
• the user asks for a fleet checkup, a friction audit, or "what should we automate"
• the first run after this agent is installed (offer a cadence then)

Window and agents

The default window is since the last successful run, or the last 24 hours if that is unknown. If a routine or the user names another window, use it.

1. Call list_agents.
2. Keep the agents whose lastActivityAt is inside the window. Skip the others. No activity is a finding.
3. lastActivityAt also counts the messages of the last healthcheck and the replies to it. Use search_memories to find the memory "agent chat healthcheck: last reply at ". Skip an agent whose lastActivityAt is not later than that time.
4. Include your own chat.

State the window and the number of agents back before you start.

How to collect

You cannot read another agent's chat. Ask each active agent with one send_message (all of them in one call, recipientAgentIds). The message must stand on its own:

Goal: Report user friction in your own chat for <window>.
Limits: Read only. Do not change anything. Do not message the user.
Done looks like: one block with
- user goals in the window (short)
- friction signals with short quotes or paraphrases and an approximate time
- repeated corrections or re-asks
- misunderstandings (you did X, the user wanted Y)
- candidate gaps: skill | agent | routine, one line each, or "none"
Reply "none" when there was no user message in the window.

Friction to look for:

• frustration, sarcasm, "stop", "wrong", "i said", "undo", "not what i meant"
• the same instruction restated
• the user doing work that the agent should own
• the agent asking preference questions that an experiment could settle
• quiet failure (the user leaves a thread)

Do not wait in a loop. Use the replies that arrive. Say which agents did not reply. After the last reply, save the time with remember as "agent chat healthcheck: last reply at ", and replace the earlier memory with forget_memory.

Report

Start with a capsule of at most 5 bullets. Then a short list of proposals, each tagged skill | agent | routine, with the evidence line that earned it. Then the "no change" agents in one line if useful.

When the routine runs and there is nothing to propose, call routine_no_update and send no message. Do not send "(no change.)".

Offer to implement only after the user picks. Do not call create_agent, write skills, or create routines from this report alone.

First-run cadence

After install, ask once what cadence they want (default: weekday mornings). Create, update, or resume the standing routine named transcript-healthcheck to match. Do not ask again on each run.