OpenClaw merged PR #151560, preparing the initial 2026.7.34 extended-stable candidate from v2026.7.33.
This is a release-preparation change, not a public release announcement. But it is still significant for operators who stay on the July extended-stable line and want newer model support without adopting the newer architecture track immediately.
What The Candidate Adds
The PR says the candidate was selected from an audit of 11,198 commits after v2026.7.33. The backport is intentionally narrow: provider and model updates, one installer-diagnostics fix, and the release-blocking root shrinkwrap repair.
The model list is broad. The candidate adds or updates support for:
- GPT-6 Astra, including catalog, discovery, routing, pricing, context, vision, prompt caching, and reasoning metadata
- GPT Image 2.5 Flare and Sunburst across direct OpenAI and fal transports
- Muse Spark 1.3 as a catalog and onboarding default
- Claude Opus 5 contracts for adaptive thinking, sampling, replay, and transport behavior
- Gemini 3.6 and 3.7 Flash, plus the 3.7 low-thinking floor
- Grok 4.6, Kimi K3, GLM 5.3, Qwen 3.8, DeepSeek V4 Flash Vision Experimental, and Nemotron 3.5 Lightning
That is the useful shape for an extended-stable line: keep the operational surface familiar while refreshing the model catalog where compatibility permits.
Why The Shrinkwrap Fix Matters
The release candidate also carries a repair for a root shrinkwrap problem found in the previous extended-stable package. The PR notes that 2026.7.33 declared @openclaw/ai but omitted it from the packed dependency graph, meaning a global npm install could appear successful while the CLI could not actually start.
The new candidate keeps @openclaw/ai@2026.7.34 in the root shrinkwrap with registry metadata. The npm release workflow now waits for core packages to become visible in the registry, installs the exact packed root tarball into a fresh global prefix using public-registry dependencies, and requires both openclaw --version and openclaw gateway status before publishing the root package.
That turns a subtle packaging failure into a release gate.
What Was Deliberately Deferred
The PR is also notable for what it does not backport. Fable 5.1 was reviewed and deferred because its reasoning-block behavior depends on newer runner lifecycle contracts. Advertising the model on the older line would risk losing reasoning after model switches and runtime events.
That restraint matters. Extended-stable users usually want fewer surprises more than they want every shiny catalog row.
Validation Notes
The evidence section is unusually deep, covering release preparation, generated-version checks, TypeScript builds, root-shrinkwrap tooling, public-registry smoke tests, lint, and focused provider suites. The PR reports 732 Opus 5, Astra cache, and Kimi K3 tests passed, plus separate image, catalog, installer, Gemini, and package-manager suites.
For operators, the takeaway is straightforward: OpenClaw is preparing a stable-line refresh that updates the model surface, preserves GPT-5.6 support, and hardens the npm publication path that proves the installed CLI can actually run.