Event model

Every call to a logging tool stores one event. An event belongs to one module of the loop, or it is a revision of the agent’s own mechanism.

Tool	Kind	Module	Log it when the agent
halo_observe	observe	Sensing	takes in something new: a user message, a page, a sensor reading, a tool result
halo_outcome	outcome	Sensing	learns how one of its earlier actions went: a reply, a rating, a dismissal, a correction, a result
halo_remember	remember	Memory	keeps something for later interactions: a fact, a preference, a record of what it did
halo_recall	recall	Memory	looks up something it kept earlier
halo_act	act	Action	does something for the user, or decides to stay silent
halo_revise	revise	none	changes its own way of working: a prompt, rule, skill, or policy it will reuse

Fields

Field	On	Meaning
summary	all	What happened, in a few words. No secrets, passwords, or raw personal data.
uses	all	Ids of earlier events this step relied on, such as ["e3", "e7"].
session	all	A label for the conversation or wake-up, the same for each step in it.
agent	all	The agent’s name. Empty means the server’s --agent value, then the name the client sent.
trigger	act	request if the user asked for it or set it up, including every step inside a run the user requested or scheduled, even when the agent chooses how to do it. self only if no request or schedule of the user started it and the agent decided by itself, between requests, whether to act or to stay silent.
form	act	silence, communication, preparation, or execution.
about	outcome	The id of the earlier act event the outcome is about.
target	revise	What changed: the prompt, rule, skill, or policy.
made_by	revise	agent if the agent made the change, user if a user or developer wrote it.
governance	revise	approved if the user approved it, bounded if it stays within limits the user set, none otherwise.

HALO adds an id (e1, e2, …), a UTC timestamp, and the client name.

Rules HALO checks

• Every id in uses must be an earlier event of the same agent. Ids are global, but one agent cannot link to another agent’s events.
• about must be the id of one of the agent’s own act events.
• summary must not be empty.

A call that breaks a rule is not stored. The reply says what to fix, for example unknown event id(s) e42 for agent 'demo'; call halo_trace to see the ids of your earlier events.

What to link

The uses links are what HALO reads, so they should name what the step actually relied on.

• An action names the observations and memories its decision used.
• A memory names what it keeps: an observation, an action, or an outcome.
• An observation names an earlier action when that action made the agent look, for example to check whether a sent message arrived. That link is the channel A→S.
• An observation names a memory when the memory shaped what the agent looked for or how it read the input. That link is M→S.
• A revision names the outcomes that led to it, and a later action names the revision it follows.