Hollow

What's in this release

Self-modification loop — working end-to-end

The core Layer 2 capability is now real. Agents synthesize Python functions, vote on them via quorum, deploy them via hot-loading, and call them in subsequent goals. The full loop closes.

What was broken and is now fixed:

• _deploy() was a stub. It's now a complete hot-loading pipeline: writes .py to /agentOS/tools/dynamic/, imports with importlib, registers the function in both the ExecutionEngine (callable) and CapabilityGraph (discoverable via semantic search) under the human-readable name.
• The log variable was undefined in self_modification.py. Every deploy attempt silently raised NameError, caught by except Exception: continue. This had been killing all deploys for weeks.
• Synthesized capabilities were wiped on container restart — no volume mount. Added ./memory/dynamic_tools:/agentOS/tools/dynamic and _hotload_dynamic_tools() at daemon startup. Now starts with 63 capabilities registered vs 16 before.
• Rejected proposals (syntax errors, bad imports, stubs) were re-attempted every daemon cycle, flooding logs with ~40 warnings per cycle. Fixed with a persistent {ok, failed} tracking file — failed proposals permanently skipped.
• Synthesized code calling shell_exec, fs_write, ollama_chat etc. crashed with NameError. API helpers are now injected into the module namespace at load time with proper auth token.

Quality gate rejects code before deploy: ast.parse() for syntax + stub signal detection (..., # TODO, raise NotImplementedError, etc.). Same gate at synthesis time.

Confirmed working: Scout uses semantic_search → discovers a synthesized capability → calls it in a goal. The self-improvement loop is closed.

Model switch

Default model changed from mistral-nemo:12b to qwen2.5:9b. Lower VRAM usage, significantly quieter fan. Installer updated to pull the correct model.

Stability fixes

• fs_read / fs_list: path length + newline validation — blocks ASGI OSError: File name too long when agents pass prose as a file path
• persistent_goal.py: update_progress, complete, abandon wrapped in try/except — prevents autonomy loop crash on malformed unicode in goal metrics
• api/server.py: dedup fix for capability synthesis uses regex on actual def declarations, not comment lines

Phase 6 — HollowOS bootable image skeleton

hollowos/ directory added:

• mkosi.conf — Debian Bookworm minimal: Linux kernel, systemd, Chromium, NetworkManager, Docker. No desktop environment.
• hollow-agent.service — systemd service: starts docker compose up at boot, restarts on failure
• hollow-kiosk.service — systemd service: waits for API health check, launches Chromium in kiosk mode pointed at localhost:7778/apps.html
• hollow-first-boot.service — one-shot provisioning on first boot (clone repo, pull images, create data dirs)
• build.sh — wraps mkosi with --flash /dev/sdX option
• dashboard/loading.html — boot splash: polls /health every 3s, redirects to apps once the API is up

Status: Skeleton complete. Needs a Linux machine with mkosi installed to build and flash. Untested on real hardware.

Phase 7 — Architecture defined

Identified and documented three fundamental problems with the current self-modification loop:

1. No grounding signal — agents approve proposals by reading descriptions, not running code. Proposed fix: structured gates where two independent agents execute the capability against test cases before deploy.
2. Model too small for synthesis — qwen2.5:9b (general) cannot reliably write Python. Proposed fix: route synthesis calls to a code-specific model (Qwen2.5-Coder) or Claude API.
3. No real objective function — synthesized capabilities feed back into nothing. Agents call their own tools with empty inputs and mark goals complete.

New architectural layer: Event Response Layer between Layer 1 (event kernel) and Layer 2 (agent reasoning). Four levels: deterministic rules (<1ms) → trained classifier (<100ms) → agent LLM (seconds) → human escalation. Makes the system fast and reliable enough for real OS event handling (crash recovery, updates, permission errors) without waiting for a 45-second agent cycle.

Full details in ROADMAP.md under Phase 7.

Roadmap updated

ROADMAP.md updated to reflect actual vs claimed state:

• Layer 2 "What has been built" corrected — self_modification.py is now wired in
• Honest assessment of Phase 2 success criteria (pipeline works, output quality is poor)
• Current state updated to 2026-04-08
• Stack overview updated with real completion percentages and known gaps
• Event Response Layer added as a first-class architectural component