OpenClaw Weekly: OOM Hardening, Model Fallback Fixes & Test Migration (June 29 – July 5)
- Commits summarized
- 1,759
- Period
- Jun 29 - Jul 05, 2026
- Published
- Jul 06, 2026
OpenClaw Weekly Development Update
June 29 – July 5, 2026
Highlights from 200 commit summaries (1759 total commits in period)
This week saw a major push to harden OpenClaw against out-of-memory (OOM) conditions by bounding HTTP response reads across dozens of integrations, alongside meaningful improvements to model fallback behavior, diagnostics, and test infrastructure. If you operate OpenClaw at scale, run it on memory-constrained hosts, or maintain provider integrations, the OOM-bound work and model-fallback fixes are directly relevant to your deployments.
🛡️ System-Wide OOM Hardening
A coordinated effort to cap unbounded HTTP response body reads that could exhaust gateway memory. The work follows a consistent pattern: replace bare res.json() / res.text() calls with bounded helpers (readProviderJsonResponse, readResponseWithLimit, etc.) in provider and integration code.
AI/LLM Provider Endpoints
- Z.AI: Bound endpoint-probe error body reads (411a5e63)
- Inworld: Bound TTS audio, voices, and error response reads (4c8470c0)
- Ollama: Bound cloud-auth 401 JSON response reads (a82cff88)
- Mistral: Bound streaming response bodies at 16 MiB (e4e4b016)
- OpenAI Completions: Bound SSE response reads via
buildGuardedModelFetch(ff820d39) - xAI: Bound OAuth response reads and video response body reads (d5aca1d6, 46e11907)
- Anthropic: Bound usage error response reads (16 MiB cap) (615558f6)
- DashScope: Bound video generation JSON response reads (238398e3)
Video & Media Generation
- Runway: Bound video create/poll response reads (63b08938)
- OpenAI: Bound video create-submit response reads (f0e2f7b4)
- Together / Pixverse: Bound video response reads (ca1bc587)
- Fal: Bound music/video generation response reads (ce1217a4)
- Vydra: Bound control response reads (74a9beb8)
- CLI: Bound and redact generated video downloads (561c713b)
Messaging & Communication
- Google Meet: Bound Drive document export reads (2001b15f)
- Signal: Bound container REST response reads (dd6143f6)
- QA Matrix: Bound homeserver response reads (5f86c3a9)
- APNs: Bound relay response body (89b5a879)
- Google: Bound OAuth response body reads (a6aaba76)
- GitHub Copilot: Bound login JSON response reads (748bea34)
Provider Usage & Auth
- Provider usage: Bound usage response body reads (db2786bd)
- Chutes OAuth: Bound plugin and core helper JSON response reads (5723222b, f83cdec2)
- Comfy: Bound JSON response reads via
readProviderJsonResponse(bf66b4e1) - WHAM: Bound usage probe JSON response reads (0ce10d77)
- Proxy Capture: Bound captured response bodies to prevent OOM (c1d17c8d)
Tool Argument Validation
- LLM tool argument coercion: Stringified JSON arrays/objects are now parsed (with a 64KB size guard) instead of rejected, fixing MCP tool calls from MiMo, Ollama, and similar providers. (9202dbb1)
🔄 Model Failover & Fallback Improvements
- Fallback on model_not_found: The failover chain now allows
model_not_foundto trigger configured fallback models instead of treating it as terminal. Previously, a decommissioned model silently defeated the entire fallback chain. (f6115dd5) - Upstream provider errors trigger fallback: Provider
upstream_error/ LLM request failures now correctly trigger the fallback chain. (cfd8ba8f) - **Clearer