Session format v2 embeds each attempt's compact stream in
assistant/message and assistant/attempt, but Host and client consumers
still expanded it into per-member TimedStreamChunk arrays and did
per-member work; expandAssistantStream materializes the full array before
find/toReversed/break can answer. Session Stats (the projection phase of
every Session open), the token meter's usage and provider-assembly folds,
the subagent output fold, and the Session Controller image lookup still
paid O(members) allocation and time per settlement.
The Chat and Trajectory definitions were already settled from
message.content on master; the remaining per-member folds stay.
dsh-llm now exports record-level readers (first token, visible content,
visible text, last raw chunk of a type, raw chunks of a type, joined
text, run-aware assembly, per-run first-token/first-visible times) that
scan the compact records once with early exit. Session Stats reads
assistantStreamFirstTokenTime, the token meter reads
lastAssistantStreamChunk(stream, 'usage') and assembles through
assembleAssistantStream, the subagent output fold appends
joinAssistantStreamText, and the Session Controller scans
assistantStreamChunks(stream, 'block-end').
expandAssistantStream is deliberately not memoized: retaining expansions
costs roughly ten times the compact stream for the Session's lifetime.
It remains the validating path at durable boundaries.
Synthetic 200-turn v0 migration benchmark, median of five: first-open
projection 28.0 ms -> 5.4 ms, first-open total 76.9 -> 50.0 ms, peak RSS
137.2 -> 94.9 MB; reopen projection 17.8 -> 5.6 ms; all phase budgets and
the 128 MB heap constraint keep passing.
`SESSION_FORMAT_VERSION` is the only value that names the Session format
generation. Two other version fields had moved to 2 alongside it:
- The `dsh_session_log` request extension's outer schema `version` returns
to 1 and its Session header projection keeps `seedLength`, derived from
the logical inherited cut, so the external wire is unchanged in this PR;
`sessionFormatVersion` alone identifies the embedded generation.
- The `sessionStats` projection unit's `stateVersion` returns to 1: the
projection cache binds every checkpoint to the format generation, so a
generation change discards old rows without a unit version bump.
Test doubles and the standalone Client fixture that build a current
header now spell `SESSION_FORMAT_VERSION` instead of the literal.
- stateVersion starts at 1: the package is new, no persisted rows predate it
- pendingCalls pairs by own key so a provider callId naming a prototype
property cannot fold toolMs to NaN on an unmatched crash-recovery result
- StatsLine folds the window fallback only when no sessionStats value is
served, and gates the token group on actual token activity instead of
steps, so failed-only sessions drop the zero-token group
- correct the crash-step counting semantics in the README and Agent Note:
recovery closes interrupted steps with a synthetic step/end on reload
- reword the window-scoped alternative as a plain rejected option and name
isTokenDelta's home beside the StreamChunk type
Tianyi Cui