Full read
A full read covers every first-party chunk. It is a batch job made of fresh model sessions.
Flow
plan, next, read, done, and verify are local operations. The model
session is the caller's responsibility. The knowledge service is optional and
receives findings and coverage through sync.
Prepare
./auwai index --repo /path/to/repo
./auwai survey --repo /path/to/repo
survey prints the scope, the partition tree, and the segment plan.
Check the read budget, the segment count, and any oversized segment.
The window comes from [budget] context_window in config.toml.
Plan
./auwai fullread plan --repo /path/to/repo
The plan is stored in SQLite.
--fixed-prompt, --output-reserve, --safety, and --synthesis adjust the planning values.
Run one segment
./auwai fullread next --repo /path/to/repo
./auwai fullread read --repo /path/to/repo
next prints the brief and the chunk ids.
read returns the bodies for the next pending segment, marks the segment active, and marks the chunks read.
Record what the segment taught you:
./auwai record ID --repo /path/to/repo --finding "..." --purpose "short label"
./auwai fullread done --repo /path/to/repo
Use one record call per chunk or a batch file.
The finding is what survives the next context reset.
record does not mark a chunk read; only fullread read and read do.
fullread done refuses a read segment whose chunks are still unread.
Synthesis
The final plan uses one synthesis segment by default.
fullread next returns it after the read segments.
fullread read prints the partition findings and the map instruction.
fullread done closes the segment.
Use --synthesis partition when each leaf needs its own map update.
Use --synthesis none only when the caller writes the map outside Auwai.
Verify
./auwai fullread verify --repo /path/to/repo
The command lists every first-party chunk that is still unread. It exits nonzero when the pass is incomplete. The coverage count is a read count, not a finding count.
Resume
The plan and segment states are in SQLite.
After an interruption, run fullread status and fullread next.
An active segment is returned before a pending one.
If a source file changed, refresh and plan again.
Current limits
fullread read prints every chunk body in the segment.
A large segment can overflow the terminal or the harness capture.
The command does not call a model and does not record findings for you.
The model session wrapper is still the harness's responsibility.