This release includes 1 breaking change for platform teams planning a safe upgrade.
✓ No known CVEs patched in this version
ReleasePort's take
Light signalHollow v5.7.32 fixes goal execution stalls and validation hangs that blocked production workflows. The web dashboard is removed; migrate to the operator panel before upgrading.
Why it matters: Goals now complete via fixed progress cap. Validation completes in 9-22 seconds instead of hanging. Claude invocation now queues correctly. Migrate to operator panel before upgrade.
Summary
AI summaryGoals can complete, Claude invocation queue works, and validation no longer hangs.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Breaking | Medium |
Web dashboard removed; operator panel is canonical UI Web dashboard removed; operator panel is canonical UI Source: llm_adapter@2026-05-21 Confidence: high |
— |
| Feature | Medium |
Setup wizard surfaces compose-failure errors with platform fixes Setup wizard surfaces compose-failure errors with platform fixes Source: llm_adapter@2026-05-21 Confidence: high |
— |
| Feature | Medium |
Wiki shipped at https://ninjahawk.github.io/hollow-wiki/ Wiki shipped at https://ninjahawk.github.io/hollow-wiki/ Source: llm_adapter@2026-05-21 Confidence: low |
— |
| Bugfix | Medium |
Progress cap in autonomy_loop.execute_step fixed; goals now complete Progress cap in autonomy_loop.execute_step fixed; goals now complete Source: llm_adapter@2026-05-21 Confidence: high |
— |
| Bugfix | Medium |
Watchdog correctly backgrounds daemon and uvicorn in Docker Watchdog correctly backgrounds daemon and uvicorn in Docker Source: llm_adapter@2026-05-21 Confidence: high |
— |
| Bugfix | Medium |
invoke_claude force-included in planner candidates; receives work invoke_claude force-included in planner candidates; receives work Source: llm_adapter@2026-05-21 Confidence: low |
— |
| Bugfix | Medium |
Validation now bounded with ThreadPoolExecutor; completes in 9-22 seconds Validation now bounded with ThreadPoolExecutor; completes in 9-22 seconds Source: llm_adapter@2026-05-21 Confidence: low |
— |
| Bugfix | Medium |
Validator prefers real artifacts over evidence steps Validator prefers real artifacts over evidence steps Source: llm_adapter@2026-05-21 Confidence: low |
— |
| Bugfix | Medium |
hollow.py status no longer crashes on Windows cp1252 console hollow.py status no longer crashes on Windows cp1252 console Source: llm_adapter@2026-05-21 Confidence: low |
— |
| Bugfix | Medium |
Goal generator rotates stressor types; avoids repetition Goal generator rotates stressor types; avoids repetition Source: llm_adapter@2026-05-21 Confidence: low |
— |
| Bugfix | Medium |
invoke_claude forced into planner candidates; can receive work invoke_claude forced into planner candidates; can receive work Source: granite4.1:30b@2026-05-23-audit Confidence: low |
— |
Full changelog
Three issues that had been silently broken for the entire project lifetime are now fixed.
Goals can complete. The progress math in autonomy_loop.execute_step capped at 0.90 in 6 steps (6 × 0.15). The progress >= 1.0 gate that triggers commit was mathematically unreachable; every staged fs_write was rolling back at end-of-cycle. HVC delta bumped to 0.20. First persisting agent artifacts observed post-fix.
The invoke_claude queue can receive work. The planner candidate filter ran semantic search top-25 against goal text, and invoke_claude / propose_change / synthesize_capability almost never won against phrases like "investigate workspace." They were invisible to the planner. Now force-included in _ALWAYS_INCLUDE. The first non-empty claude_requests.jsonl in the project's history landed during testing.
Validation no longer hangs. L2 semantic and L5 codebase fact-check Ollama calls are now wrapped in concurrent.futures.ThreadPoolExecutor with hard wall-clock deadlines. On Windows hosts, Ollama can evict the model under VRAM pressure despite keep_alive=-1, and httpx requests then sit forever below the exception layer. Bounded execution guarantees return. Validations now complete in 9-22 seconds; previously they could hang 14-40 minutes.
Other:
- Mechanical validator prefers real artifacts (
fs_write,memory_set) over evidence steps (shell_execstdout). Stops semantic verification from comparing goals against directory listings. - Fallback goal generator rotates away from recently-used stressor types. Stops agents picking "Address the repeated_failure stressor" 25 times in a row.
- Watchdog plumbed correctly.
entrypoint.shnow backgrounds both daemon and uvicorn; whichever dies brings the container down for Docker to restart. The previousos.kill(1)watchdog call was a no-op (PID 1 was uvicorn, not the daemon). - Setup wizard surfaces compose-failure errors with platform-specific fixes for mounts-denied (#17) and missing
nvidia-container-toolkit(#18). - Web dashboard removed (#21). The operator panel (
panel.bat/panel.py) is the canonical UI. - Wiki shipped at https://ninjahawk.github.io/hollow-wiki/.
python hollow.py statusno longer crashes on Windows cp1252 console.
Tested on the recommended config: qwen3.6:35b-a3b on a 12 GB GPU with partial offload, Windows 11. Smaller models in the model table have not been verified at this version. Please report issues.
Closes #14, #16, #17, #18, #20, #21, #22.
Breaking Changes
- Web dashboard removed (#21); operator panel is now the canonical UI.
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
Beta — feedback welcome: [email protected]