OpenClaw Weekly: Oct 26 – Nov 1, 2026 — Performance Wins, Targeted Fixes & the Deslop Test Campaign
- Commits summarized
- 4,568
- Period
- Oct 26 - Nov 01, 2026
- Published
- Nov 02, 2026
OpenClaw Weekly — Oct 26 – Nov 1, 2026
Who should care: this issue is for anyone running gateway- or worker-heavy deployments (the allocation and CPU wins this week are large), Docker self-hosters, Tlon/Urbit integrators, and maintainers of provider or UI test suites — the test-reduction campaign touched dozens of packages.
Scope note: 4,568 commits landed in the period; this issue highlights the 200 commit summaries we have. Treat it as a highlights report from the available summaries, not an exhaustive changelog.
Performance
A cluster of allocation- and CPU-focused PRs landed this week across agents, gateway, sessions, and the worker.
- agents: avoid redundant media-selection array copies (#183954,
8ebe214) — drops repeated array copies in media selection; adds owned-mutation test coverage. - gateway: prepare session pin eligibility once per sort (#183967,
e256b65) — moves pin-eligibility work out of the hot sort path. - agents: reuse catalog identity groups in status replies (#183973,
f729278) — reuses precomputed identity groups instead of rebuilding them per reply. - gateway: reduce repeated sampling, proxy and signing work (#183975,
cbe47c7) — removes intermediate sampling arrays and repeated IP/CIDR parsing; native Ed25519 yields byte-identical Reef request signatures without caching private keys. - providers: remove redundant request preparation copies (#183959,
d6685f2). - gateway: reuse owned native control transfer lists (#183982,
09042d1). - gateway: avoid repeated identifier normalization (#183971,
98e18e5) — reuses canonical session keys, parsed media-request ownership, and trimmed skill names within synchronous lookups. - sessions: reuse placement facts during maintenance (#183974,
b430641) — traverses the placement owner's retained index and newer postimages instead of rebuilding and deep-cloning the inventory per preserve capture. Owner probes: preservation allocation −65.6%, CPU −79.9%; pin eligibility allocation −24.1%, CPU −10.7%. - worker: count transcript batch bytes incrementally (#183970,
286b334) — counts the envelope and each JSON message once while preserving exact control/media budgets and full-frame transport admission. Owner measurements: CPU −95%, sampled allocation −97%.
Bug fixes
- agents: retry malformed tool calls after visible progress (#183807,
889968f) — when a model produces visible progress followed by a malformed tool call, the turn now gets the same single bounded retry it would get without prior progress; a second malformed call still ends the turn. - Tlon: keep upload filenames and encode custom-S3 public URLs (#141326,
6cb193f) — Content-Disposition filenames are preserved on outbound image uploads, and public URLs for custom-S3 object names containing#or?are now encoded (previously 404). - Docker: honor
OPENCLAW_GATEWAY_PORTin healthcheck when config load fails (#172765,787a43e) — the official image healthcheck now falls back to the environment port when config fails before port resolution, instead of probing the wrong place and reporting a healthy Gateway as unhealthy. - gateway: accept short-session navigation hints (#183148,
7a93079) — existing session-key and composer-focus hints are validated together and preserved through reloads, while duplicate or malformed query parameters are rejected.
Test-reduction campaign ("deslop")
Three large batches (d573, d574, d575) plus follow-ups landed this week, removing low-value tests, sharing fixtures, and consolidating duplicate assertions while preserving boundary, lifecycle, and regression coverage. Per campaign policy, validation is static (oxlint, oxfmt, git diff --check) with runtime proof deferred to hosted CI.
- Batch d573 — providers (#183948,
a5468d6) — zalo, music-generation, vault, plugin-sdk-impact, radius, zai, llama-cpp, baseten, oauth, logs, levanto, lmstudio, venice, worker-runtime, agents, copilot. Most shards hit roughly 19–25% line reductions while keeping distinct cases intact. - Batch d574 — core, plugins, UI (#183952,
d8a6c85) — elevenlabs, whatsapp, agent-core, system-agent, approvals, workboard, wizard, qa-lab, cli, onnx, perplexity, line, UI dashboards, cli-runner, and more. - Batch d575 — providers & UI (#183969,
82109fa) — mattermost, routing, x, UI browser lifecycle, hooks, daemon-cli, bonjour, agents-home, flows, cron, terminal-core, zoom-meetings, cloudflare, team-reports.
Representative outcomes, all with assertions preserved: onnx 494→414 lines, cloudflare 396→297, agent-core 469→351, system-agent 468→351, approvals 468→351. Follow-up commits retained boundary proof and repaired fixture types within each batch.
Refactors & test infrastructure
- refactor: trust native timer methods in runtime helpers (#183292,
36cc2e4) — removes 30 redundant native-timer probes (31 conditional operators) across lifecycle, channel, and transport helpers; no public API or intended behavior change. - refactor: deslop gateway, approval and config test fixtures (#183963,
09104d5) — slims fixture layers without removing cases. - test(control-ui): align reload fixtures with subscribed history (#183832,
b4bd4b6) — the mock Gateway subscription now buildsinitialHistoryfrom the startup-response owner when requested, matching the real Gateway contract since #182846, so reload benchmarks no longer measure a stale request chain. - test(ui): check catalog action visibility without freezing opacity (#183972,
12a569d) — keeps hover, keyboard focus, touch, and pointer-event checks while accepting the approved icon color treatment from #183401; all 7 catalog browser tests pass.
Highlights from 200 of 4,568 commits, 2026-10-26 → 2026-11-01. Full history: github.com/openclaw/openclaw.