This release fixes issues for SREs watching stability and regressions.
✓ No known CVEs patched in this version
Summary
AI summaryRouting tables now resolve the selected model and Linux Docker compose portability issues are fixed.
Full changelog
Maintenance release. No new agent-facing features — onboarding accuracy, Linux portability, and a few latent bugs from the 35B model swap.
Fixes
Smaller-model regression (latent since v5.7.0). When the model swap to qwen3.6:35b-a3b landed, four hardcoded routing tables were left pointing at the 35B model regardless of what the user actually picked in the setup wizard:
agents/model_manager.COMPLEXITY_MODELagents/scheduler.ROLE_MODELagents/adaptive_routerstatic defaultsapi/server.MODEL_ROUTES
A user picking llama3.2:3b or gemma3:4b in the wizard would have the scheduler route through a model they never pulled. All four now resolve from cfg.ollama.default_model at import. The VRAM fallback table also picked up rows for the 9B / 4B / 3B options.
Linux Docker compose portability. Two issues hit external Linux users:
host.docker.internalresolves automatically on Docker Desktop (Windows/macOS) but not on native Linux Docker. Addedextra_hosts: host.docker.internal:host-gatewayto the api and store services. Harmless on Win/Mac.- The setup wizard's GPU fallback now detects the specific
could not select device driver "nvidia"error and prints annvidia-container-toolkitinstall hint when falling back to CPU-only mode. The CUDA toolkit alone is not enough — Docker needs a separate package, which was a common confusion.
Ghost tools (#15). tools/dynamic/ could accumulate JSON specs whose paired .py implementation was missing — agents would see them in capability listings, call them, get null back, and burn cycles investigating. The 5-cycle dynamic-tool retire block in daemon.py now sweeps orphan JSON specs alongside its existing broken-tool cleanup.
Docs
README rewritten against the actual 5.7.x system:
- Default model VRAM requirement corrected (was 8GB+, actually 24GB+ for the default 35B). All four wizard options now shown in a table so users can pick honestly.
- Capability count corrected (21 → ~30).
- New Linux setup section covering Docker Desktop file sharing (#17) and
nvidia-container-toolkit(#18). - Sections added for the five-layer false-completion gate, mechanical capability gating, and the lessons system — these were the headline 5.6.x and 5.7.x changes.
CI
New docker-smoke-test job: stock Ubuntu runner, no GPU, no Ollama. Builds the image locally, brings the stack up, waits for /health, hits the dashboard, tears down. The smallest test that would have caught the compose issues that hit external users before release.
Issues addressed
- Closes #15 (ghost tools)
- #13 (skipped_agents drain) — confirmed already fixed in code, closed
- #17, #18 — README + wizard fixes; reporters can retest
Upgrade
Standard upgrade — no migrations, no config changes required.
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About Hollow
All releases →Related context
Related tools
Earlier breaking changes
- v5.7.32 Web dashboard removed; operator panel is canonical UI
Beta — feedback welcome: [email protected]