Segments
A segment is one fresh model session.
The segment planner is in skills/auwai/scripts/segments.py.
Planning flow
A chunk larger than the budget becomes an oversized segment by itself.
Budget
output_reserve = min(32,000, max(8,000, context // 8))
read_budget = context - fixed_prompt - output_reserve - safety
The current default fixed prompt is 20,000 tokens. The default safety margin is 2,000 tokens. The read budget is at least 2,000 tokens.
| Context | Output reserve | Read budget |
|---|---|---|
| 64K | 8,192 | 35,344 |
| 128K | 16,384 | 92,688 |
| 256K | 32,000 | 208,144 |
These are planning values. The fixed prompt should be measured for the target harness before a production run.
Packing
The planner packs chunks across partition boundaries.
It does not start a new segment at every leaf.
It flushes a segment when the next chunk would exceed the budget.
A chunk larger than the budget gets its own segment and is marked oversized.
Synthesis
synthesis=final adds one synthesis segment at the end.
synthesis=partition adds one synthesis segment per leaf.
synthesis=none adds none.
The final synthesis is the default because it is the smallest reset that still writes the map.
A synthesis segment has no chunk bodies. It reads the partition map and the findings from the ledger. It writes the updated map and any decisions.
Plan output
auwai fullread plan writes the plan to SQLite.
The plan records the context, fixed prompt, output reserve, safety margin,
read budget, read segment count, synthesis count, turns, and compactions.
auwai fullread status reports pending, active, and done segments.
auwai fullread verify reports unread first-party chunks.
Current limits
The fixed prompt is a constant, not a measured per-harness value. The planner does not model prompt caching. The synthesis cadence is a single choice per plan. A model can still meander inside a segment.