OpenClaw 2026.9.8 Cuts Codex Memory and Repairs Skill Refresh

2026.9.8 Trims the Cost of a Codex Fleet

The OpenClaw 2026.9.8 release notes describe less duplicate memory when many native Codex agents run at once. OpenClaw now waits until a conversation list is needed before starting the background process for that list and shares settings where compatible; memory for saved shell environments is also bounded. This is a concrete capacity improvement, not a claim that every model turn becomes cheaper or faster. It targets the resident processes and bookkeeping that accumulate around concurrent work.

For a team running parallel coding tasks, memory pressure often arrives before the model is saturated. An extra background process per agent can turn a plausible worker count into swap, stalled browsers and missed handoffs. The official Codex harness reference explains the embedded app-server route. A useful test compares idle and active memory at a fixed agent count, then checks that delayed conversation discovery still returns the right history when a human opens it.

Read-Only Skill Copies Can Refresh Again

The release fixes refreshing skills copied from a read-only installation into sandboxed workspaces and Claude CLI sessions on Unix systems. Previously, inherited file restrictions could block a legitimate refresh. The project does not say that all copied skills should be automatically rewritten; it says the copied folders no longer prevent the refresh operation. That distinction matters when a customized workspace skill and a bundled source share a name.

The skills reference explains loading, precedence, gating and environment injection. Before refreshing, identify the source of a skill and whether the local copy is owned by the project or by the installed package. Preserve local edits if the project owns them; let the package update replace package-owned copies according to the documented rules. Then inspect which version is actually available in the target session. A successful filesystem copy is not proof that the running agent loaded it.

Local Models and Background Work Get Platform-Specific Repairs

For managed llama.cpp on Windows, 2026.9.8 can add missing Microsoft runtime files beside the local model server so it can start. If startup still fails, the notes direct operators to rerun setup or configure a compatible server manually. The fix does not make every workstation model compatible, and it does not establish a benchmark score. It addresses a specific missing-prerequisite failure on a clean host.

The same release repairs Anthropic conversations that could become stuck or finish before background commands, agents or workflows had been collected. These are different symptoms with the same consequence: the visible answer can be disconnected from the work that was delegated. A local-model evaluation should test startup and one real inference turn. A background-work evaluation should retain a task receipt and confirm the final answer includes its result, not just an early acknowledgment.

The exec tool reference is a useful reminder that shell execution is a mutating surface, even when other filesystem tools are disabled. Background work should therefore be bounded by the same authority and filesystem policy as foreground commands. A retry that produces the right answer after changing the wrong directory is not a successful test. Use disposable paths and compare the intended files before accepting the result.

Security Practice: Verify Local HTTPS Through Its Managed Proxy

The release notes say local HTTPS connection checks now follow managed proxy settings and verify that the server certificate matches the configured identity on every connection. This is subtle but important: a loopback address or a familiar hostname does not make an unverified TLS connection trustworthy. When testing a Gateway behind an authenticated proxy, compare the effective endpoint and trusted certificate rather than disabling verification to make a health check turn green.

A second privacy fix prevents secrets in long, unusually formatted log entries from remaining visible when OpenClaw runs on Bun. That is not a promise that arbitrary logs are safe to publish. Operators should still minimize logging of sensitive inputs, inspect log access controls and test redaction with non-secret canaries. The Gateway security guidance treats auditing and hardening as layers, not one switch.

For readers choosing between release channels, note the precise claim in the 2026.9.8 notes: GPT-6.1 Sol support was still incomplete in that release, although preparatory changes landed. Do not infer model availability from a nearby extended-stable release or a partially populated model list. The right release is the one whose required routes are proven in your own environment.