Skip to main content

​
Agent Context

This page explains context in the openboa Agent runtime. Use this page when you want to answer:
  • what the prompt contains
  • what the prompt does not contain
  • why the session is the truth instead of the context window
  • what retrieval candidates mean
  • what context pressure is for

​
The core rule

The session is durable truth. The prompt is a bounded view assembled for one wake. That is the most important context rule in the runtime.

​
Why this rule exists

Compacted summaries and prompt-local context are useful, but they are not good enough to be the only truth. Future wakes may need:
  • old user messages
  • prior shell evidence
  • earlier tool results
  • the lead-up to a blocked action
If those are only preserved as one summary, the runtime becomes brittle.

​
The context model

This means the prompt is assembled, used, and discarded. The durable objects remain.

​
What goes into prompt context

Typical prompt context includes:
  • recent conversation continuity
  • selected runtime notes
  • active outcome posture
  • current environment and resource posture
  • retrieval candidates for prior truth
What matters is not that everything is loaded. What matters is that the runtime can reopen what it needs.

​
Retrieval candidates are not truth

Retrieval candidates are hints. They exist to tell the model:
  • which prior session may matter
  • which memory source may matter
  • which exact reread tool is the right next step
They should not be confused with canonical truth. The canonical truth still lives in:
  • session events
  • runtime artifacts
  • durable memory surfaces

​
Reread versus trust

The Agent should prefer:
  1. find a relevant candidate
  2. reopen the underlying truth
  3. act using the reopened evidence
instead of trusting one compacted summary forever. That is why retrieval and session_get_events style reread exist.

​
Context pressure

The runtime also tracks context pressure. Context pressure answers questions like:
  • how much recent history fit into this wake
  • how much was dropped
  • whether protected continuity had to be preserved
  • whether the model should prefer reread instead of broad continuation

​
Current behavior under pressure

When pressure is high, the runtime can bias the Agent toward:
  • session_describe_context
  • retrieval and reread
  • read-first inspection
  • narrower next steps
instead of broad, assumption-heavy continuation.

​
Design rule

Do not treat context engineering and truth storage as the same problem. In openboa:
  • truth storage belongs to sessions and durable artifacts
  • context engineering belongs to the harness
That separation keeps the runtime flexible as models change.