Skip to content

perf(cli): reduce autosave write amplification for a growing session #2480

Backlog

Description

@spboyer

Summary

Shell autosave synchronously serializes and rewrites the growing persisted message history. The persistence shellMessages array is independent of the UI's 200-message cap, so trimming rendered messages does not bound this work.

Proposed primary owner: state-storage-dev.

Baseline evidence

Revision: 47082a81799710c3e457b594123d190c00596ebf, observed 2026-11-12.

Code: autosave call sites in shell/index.ts:1319-1359, saveSession at session-store.ts:64-75, and the separate UI cap in components/App.tsx:70-80.

Independent repeat, 256-character synthetic messages:

Messages in active session Bytes rewritten/save p50/save p95/save
200 75,159 0.296 ms 0.378 ms
2,000 750,159 1.742 ms 2.144 ms
10,000 3,750,159 8.378 ms 9.390 ms

The fixture measured one overwrite of an existing session at each size, including normal session-directory work. It did not measure complete-session physical disk writes or crash/power-loss durability. Earlier baseline passes measured approximately 8.9-9.5 ms at 10,000 messages; no before/after benefit is claimed.

Environment: Apple M4 arm64, 16 GiB, Darwin 25.6, Node 24.15.0; three rounds of ten sequential observations/case, two warmups, warm filesystem caches, setup/imports outside timing. Production source methods ran via Node's TypeScript transformation/relative-import hook, with the session store's SDK bindings supplied by direct reexports of the exact production FSStorageProvider/safeTimestamp modules.

Limit: dependency-feed E404 prevented locked restore and packaged build/tests. These are function-level persistence diagnostics, not packaged shell throughput or total-session memory results. Shared reproduction fixtures/raw evidence are required before optimization acceptance.

Scope

Parent: github/squad#2481.

One focused change to how the existing shell autosave persists a growing session. Latest-session lookup is separate (github/squad#2475). No UI redesign, history deletion, broad storage/backend migration or unrelated context compaction.

Measurable acceptance and tests

For an optimization, use at least three interleaved built-baseline/change rounds with at least 30 observations per case per round; publish raw observations, per-round p50/p95, exact SHA/environment and correctly labeled CPU/I/O/memory evidence. The budgets below are proposed acceptance targets, not already-demonstrated gains. If a safe material win cannot be reproduced, record rejection/defer rather than broaden the implementation to chase the target.

  • Fixed trace: append 2,000 messages of 256 characters, requesting the existing autosave after each append and a final flush. Also retain the 200/2,000/10,000-message single-save controls.
  • Record cumulative logical bytes written, number of writes, per-save/flush p50/p95 and retained pending-write heap. Physical disk writes are separate and must not be inferred from logical bytes.
  • Target at least 50% fewer cumulative logical bytes for the growing trace than the same built baseline, without dropping any message. Report CPU/latency trade-offs and smaller-size controls.
  • A round-trip after final flush must restore all 2,000 messages in order, preserving content/timestamps and external stateDir support.
  • Test exit/final flush, injected write failure and restart after interruption. Define which writes are acknowledged/durable and preserve or strengthen existing recovery; never hide lost pending messages behind successful output.
  • Bound pending-write memory if coalescing/deferred persistence is chosen. Do not prescribe both an append format and coalescing; choose one measured technique.
  • Check in the portable trace and publish before/after raw samples/byte counts. If the safe bounded change cannot meet the target, record the result rather than broaden to a new persistence architecture.

Related work

github/squad#1476 concerns coordinator token/context growth, not this synchronous JSON persistence path. This issue is separate from finding the latest saved session among archives.

Methodology: dotnet/aspnetcore#70219 and microsoft/waza#846.

Activity

  1. changed the title [-]perf(cli): reduce full-history rewrite amplification on shell autosave[/-] [+]perf(cli): reduce autosave write amplification for a growing session[/+] on Nov 12, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions