Decode four layers in order.
Select .jsonl.zstd or plaintext .jsonl from the configured persistence root.
Decode the header frame and every checksummed append frame. The file is a frame concatenation.
Classify ordinary events and all three packed storage-row types.
Expand each packed row, then require one contiguous logical seq stream.
One row can represent many events.
texts[] preserves every streamed fragment.
The same envelope carries thinking-model output.
args[] shares one call id and optional name.
{ type: "text-chunks", seq0: 10, time0: 1787150994000,
data: { turn: 1, step: 1, index: 0,
dt: [21, 1, 0, 20],
texts: ["He", "llo", ", ", "wor", "ld"] } }This physical row reconstructs five assistant/chunk events with sequence numbers 10 through 14. Filtering rows without seq drops all five.
Run every post-header value through decodeStorageRecord. Do not independently guess a durable prerelease format from the generated event catalog.
Minimum reader gates.
- Read plaintext and concatenated Zstandard artifacts.
- Validate the header and reject unsupported format versions first.
- Expand all three packed-row types without joining fragments early.
- Preserve packed, unpacked, and mixed-layout equivalence.
- Fail loudly on malformed packed rows and committed sequence gaps.
- Inspect a quiesced copy; never repair the only live artifact.
Keep the complete format map.
The canonical guide includes exact reconstruction rules, unsafe parser examples, a read-only inspection procedure, eleven acceptance gates, and pinned rc.8 sources.
Read the complete guide