Mental model

Auwai separates the repository from the model session.

Workspace
  Repo
    Partition
      Segment
        Chunk

The shape

A workspace is the root of the same tree. A repository can also stand alone.

The four objects

Chunk: a stable line range with an id, path, symbol, kind, language, and token count.
Partition: a directory, module, or service node with a token count and a kind tag.
Segment: one fresh model session. A read segment carries chunks. A synthesis segment carries the map and findings.
Ledger: the SQLite state that records what is indexed, mapped, read, stale, and known.

The three budgets

Index budget: CPU and disk. It is not a model context.
Read budget: the usable context for one segment. It is large during a first pass and small during maintenance.
Carry budget: the room reserved for the map, findings, and checkpoint that survive a context reset.

The read budget is computed from the model window, the fixed prompt, the output reserve, and a safety margin.

The session rule

A segment is a new session. The model does not receive the previous raw transcript. It receives a brief from the database:

the partition and its map;
the coverage count;
the chunk ids and one-line labels;
the stop condition.

The model reads the segment, records findings, and stops. The next segment starts with a fresh context and the updated ledger.

The coverage rule

A chunk is covered when it is read or explicitly skipped with a reason. The ledger distinguishes:

indexed: the chunk exists in the index;
mapped: the chunk has a partition and a label;
read: the chunk body was returned in a segment;
with finding: at least one finding is attached;
stale: the source changed after the read.

A full-read certificate reports the first-party chunks that are read or skipped and the chunks that remain.

Coverage state

read means the body was delivered. Finding means at least one finding is attached. Stale is an overlay, not a replacement for the earlier state.