Time to ship · about 5 minutes
Live context model.
Aura maintains an ambient context digest from local history, startup project briefing, and explicit feeder lookups. This is what she can use before dispatching a worker.
The context system is designed to answer many questions without dispatching a worker. When you ask about repo state, recent work, or a read-only source fact, she first uses the feeder lane. If the digest is thin, Aura can request a short feeder lookup and wait for that context instead of creating a coding task.
Context sources.
The current feeder uses three proven input lanes:
Startup briefing — repo structure, workspace metadata, top-level docs, git state, project memory, and recent Aura history are prepared when the feeder starts.
History tail — the feeder tails .aura/history.jsonl for new voice, tool, task, and callback events during the call.
Feeder lookup — when Aura needs read-only source or status context, she can request a focused lookup that returns safe local facts or waits for broader feeder context.
The feeder model.
A background model — the context feeder, implemented in crates/aura-context-feeder — turns those inputs into a compact digest. The Claude feeder path uses a long-running Claude subagent. The Codex feeder path uses a launch prefill plus hot and research models configured under codex.hot_model and codex.research_model.
Read-only and invisible
The feeder is strictly read-only. It cannot write files, run commands, or trigger dispatches. Its activity is invisible in the normal chat view — you only see feeder output when it surfaces an answer to a question you asked. The activity log behind the toggle shows feeder events for debugging.
Working set and token budget.
The working set is the portion of the digest and lookup context that fits within the token budget for the current call. Aura favors recent voice turns, recent task results, and source facts connected to the current question. If the budget is tight, older context is summarized or left out.
The token budget is fixed per model. Aura does not expand it dynamically — if a question requires more context than the budget allows, she says so and offers to dispatch a worker read instead.
Digging deeper.
The feeder implementation lives in crates/aura-context-feeder. aura doctor reports the configured feeder models, Codex app-server availability, and local history event count.