Reflect the review feedback on the reader list: assistantStreamFirstTokenTime and the has-visible readers stop at the first qualifying member, while lastAssistantStreamChunk, assistantStreamChunks, and joinAssistantStreamText scan the whole stream; assembleAssistantStream feeds a BlockAssembler one joined delta per run. Record-level readers trust the static record type; expandAssistantStream is the validating path.
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.
The migrating first open measured 1,963 ms on the CI runner under the
128 MB heap limit; 4,000 ms keeps a 2x margin while the repeated-snapshot
implementation still fails by heap exhaustion and would need ~10 s.
A records-proportional fold still joins 20x more text in the large window,
which measured about 2.5x the small fold; the per-delta replay measured
about 11x. A 5x bound separates the two on a noisy runner.
The migrating first open of the synthetic log costs about 1 s on the
reference machine under the 128 MB heap limit and about twice that on the
CI runner, so 3,000 ms keeps headroom while staying far below the ~5 s
(~10 s on CI) of the repeated-snapshot implementation.
scripts/test-proxy-environment.spec.ts pins the Vitest configs that declare
setupFiles so the proxy isolation setup is never dropped; the benchmark
lane declares it too.
Add a benchmark lane (`vitest.bench.config.ts`, `pnpm run test:bench`,
gate mode `ci-bench`) and a required `node 24 / benchmarks` CI job that
runs it alone. Benchmarks synthesize their input in-process from fixed
parameters and fail on documented budgets:
- `open-generation.bench.ts`: a 200-turn released-v0 log with 500 text
and 125 reasoning deltas per reply (127,400 events, ~2.8 MB) encoded
through the frozen v0 codec; the migrating first `open()` must finish
within 2,000 ms in a child process capped at 128 MB of old space, and a
fresh process must open the published current generation within 500 ms.
- `conversation-fold.bench.client.ts`: 200 replies whose compact streams
hold 2,000 text + 500 reasoning deltas each, folded through every Chat
Definition by the real assembler; the fold must finish within 150 ms
and stay within 3x the fold of the same window with 100 deltas per
reply.
On this commit both gates fail: the migration exhausts the 128 MB heap
(4.8 s and 696 MB peak RSS without the cap; the pre-stack decode of the
same bytes took 34 ms and 168 MB) and the fold scales 11x with the delta
count. The stacked fixes bring both paths to O(records).
Replace the external ArrayBuffer view and manual UTF-16 scan with native Koffi decoding through a pointer-width buffer. Preserve the original address for CoTaskMemFree.
Exercise production resultPath with real Koffi decoding of live UTF-16 buffers. Cover zero-low-byte BMP characters, surrogate pairs, NUL termination, and strings exceeding 32 KiB; test four- and eight-byte BigInt pointer preservation and the original release address. A negative-control mutation confirms truncating a 64-bit address fails.
Keep the bilingual Agent Note focused on implementation rationale and required memory ownership. Preserve master archive seals during rebase.
Co-authored-by: zhanglunet <1160377+zhanglunet@users.noreply.github.com>
Tianyi Cui