Hollow

What's new in 5.5.5

Feedback loop integrity

The core insight of this release: agents need an accurate picture of their environment to act coherently. A tool that runs cleanly but returns garbage is worse than a failed deployment — it gives the agent a false map. Several systems now report truth instead of mechanical success.

• Semantic validation in synthesize_capability — after a tool is deployed, the system calls it and asks Ollama whether the return value matches what the description promised. Description is now a contract. Mismatch = deployment rejected with specific error.
• check_claude_status now returns ok: False for rejected and not_found requests. These are failure signals, not informational responses. Agents that check a rejected request will have the goal step fail — triggering eventual goal abandonment instead of infinite loops.
• invoke_claude quality gate — description must be > 40 chars, spec must be > 80 chars. Circular requests (asking Claude to manage the queue itself) are auto-rejected.

Atomic goal execution

• Goals are now wrapped in txn_begin → execute → txn_commit or txn_rollback
• If a goal fails mid-way, all staged fs_write and memory_set operations are rolled back
• Conflict detection: if two agents try to write the same file simultaneously, one gets a conflict error
• Transaction timeout corrected from 60s to 600s (goals take 3–8 minutes; 60s caused silent rollback of everything)
• txn_begin, txn_commit, txn_rollback added as built-in capabilities

Z-score anomaly detection (now wired)

• Every capability execution now logged to memory/audit.log with agent_id, operation, duration, result_code
• Z-score anomaly detection has real per-agent behavioral data to work with
• Baseline builds from first 50 operations per agent per role; anomaly checked every 10 ops after that
• Circuit break at z=5.0 suspends agent and drops rate limits to 10%

Existence prompt redesign

• Removed the priority list entirely ("Highest-value goals: 1. check pending requests, 2. read peers...")
• Replaced with pure environmental context: workspace files with timestamps, peer activity, system state
• Pending invoke_claude requests no longer surfaced every cycle — agents check by choice, not directive
• Result: agents now choose genuinely different goals from a clean start instead of all converging on queue-checking

Stability fixes

• ContextVar propagation through ThreadPoolExecutor — capability profiles and judgment gates now actually track per-agent
• not_found tool calls no longer accumulate in broken_tools.json as ghost entries
• Module-level code in synthesized tools is now caught before import (prevents registry corruption crashes)
• shell_exec and synthesize_capability block Python writes to core system paths
• Agent identity filters: req- IDs and check_claude_status references blocked from worldview, narrative, open_questions
• execution_engine.py, transaction.py, agent_identity.py added to bind mounts

New files

• BUGS.md — live tracker of known issues with status (open/fixed/wontfix)