$ the-wire · showcase
Cron ticks leave the main thread, gateway startup drops 20s
By RepoJournal · Filed · About OpenClaw · Composed from the cited sources · methodology
![openclaw-clownfish[bot]](https://avatars.githubusercontent.com/in/3535277?v=4)
A day of latency work across openclaw and crabbox: cron session state moves to workers, gateway startup maintenance runs after readiness, and the coordinator stops letting one stalled provider call park every lease.
perf(cron): serve cron tick session state from workers openclaw/openclaw
A real scheduled cron turn was running 360 successful SQLite statements on the Gateway main thread, most of them from the native session-patch snapshot and transaction path used by cron preparation and continuation persistence. The patch routes session patches through the existing agent writer, including live CLI settlement guards, so runtime callers of the cron current-entry reader now await t...
perf(gateway): move startup maintenance after readiness openclaw/openclaw
Mandatory config, schema, permission, session, model, and channel admission still happens before ready; nonessential maintenance now runs afterward through the existing tracked lifetime. On the latest matched synthetic rig, startup went from 64.113 s to 43.876 s with refreshed metadata and 49.597 s to 36.868 s warm.
perf(gateway): avoid full display refresh for session member lists openclaw/openclaw
The members response only needs creator identities and membership evidence, but opening session sharing waited on every unrelated session display row refreshing after a model catalog change. The reproduced 20-member catalog-refresh case improved from 4,541 ms to 29.3 ms median with identical response bytes, and the response contract, per-request authority checks, fresh stored-entry validation, ...
fix(coordinator): bound lease-create admission with a retryable 503 deadline openclaw/crabbox
On 2026-10-03 a token-bound POST /v1/leases with Prefer: respond-async got no response for the full 31-minute client timeout while GET /v1/leases/<slug> returned 404 the whole time, meaning admission never committed a lease record. AWS creates take the legacy admission path in createLeaseObserved, and the fix bounds lease-create admission with a retryable 503 deadline.
fix(coordinator): bound provider I/O held under the lifecycle mutex openclaw/crabbox
Stalled provider calls held under the coordinator-wide runExclusive mutex could park every queued lease create indefinitely. The change adds provider-neutral ambient operation deadlines in provider-deadline.ts plus a retryable ProviderRequestTimeoutError mapped to HTTP 503 provider_request_timeout with Retry-After, and gives AWS transport a default 60-second per-call budget covering credentials...
fix(logging): collect initialization logs for long release names openclaw/openclaw-enterprise
Initialization Job Pods are created with generateName <release>-initialization-, and the API server cuts that prefix to 58 characters before appending 5 random ones, so past a 42-character release name the Pod name no longer contains -initialization- and the bundled Collector's glob dropped the migration and bootstrap logs. The glob now matches *-init*, what survives for any release name Helm a...
