Reading the live monitor

python3 thoughts.py (or the Open Monitor button on the panel) streams the agents’ real-time activity. Once you can read it, you can tell what the system is doing at a glance.

What you’ll see

Cycle markers 

12:38:42  Cycle 14: 3 agent(s) (0 skipped/cooling) workers=6

A new cycle just started. All three agents have an active goal and will work through one round each in parallel.

Existence loop output 

12:38:50  scout         🧠  existence loop, goal: Execute `baseline_test.py` and inspect...

The agent just decided on a goal for this cycle. The text after goal: is what they’re about to pursue.

Step execution 

12:39:01  scout         ▶  shell_exec
12:39:01                ✓  shell_exec
12:39:42  scout         ▶  ollama_chat
12:41:24                ✓  ollama_chat

Each ▶ is a step starting, each ✓ is success, each ✗ is failure. Time between start and end tells you which capabilities are slow (LLM calls dominate).

Validation 

12:41:30  scout         🔍  artifact ok | file /agentOS/workspace/scout/note.md (substantive)

or

12:41:30  scout         ⚠  artifact MISSING (3/5) | semantic verification rejected: ...

The substrate just checked whether the goal was actually accomplished. The fraction (3/5) is the validation-failure count toward permanent abandonment.

Suffering changes 

12:42:11  scout         💢  suffering: load=0.45 stressors=[stagnation,epistemic_uncertainty]

The agent’s suffering value just changed. The bracketed list is the active stressors. If you see load creeping above 0.55, that agent’s synthesize_capability is locked. Above 0.75, fs_write and fs_edit are too.

Goals abandoned 

12:43:08  scout stalled on 'Verify the persistence hypothesis of the baseline artifact...', goal abandoned

After 5 validation failures or 4 consecutive step failures, a goal is permanently abandoned and any partial artifacts are deleted by _delete_goal_fs_writes.

Goal completions 

12:45:00  analyst → goal=goal-5e2b356d72da progress=1.00 steps=1

progress=1.00 and a → goal= line means a goal completed AND validated. This is the milestone, agent will pick new work on the next cycle.

Anything less than 1.00 is partial progress that carries over to the next cycle (the same goal will be re-attempted).

What “healthy” looks like

  • Cycles complete within 4-15 min on a recommended GPU
  • Multiple progress=1.00 outcomes per hour (goal completions)
  • Suffering loads fluctuating between 0.2 and 0.6 (some pressure, not maxed)
  • Workspace files growing slowly with substantive artifacts

What’s NOT healthy

  • No log activity for 5+ minutes: daemon may be hung. Restart via panel or python3 hollow.py stop; python3 hollow.py.
  • Same goal text appearing in 10+ consecutive cycles: the loop guard is failing. Should be auto-detected and rotated in v5.7.32+.
  • All agents stuck at progress=0.9: completion is unreachable. Should be impossible in v5.7.32+ (delta math fix).
  • Cycle worker timeout (1500s) every cycle: Ollama is too slow or hung. Check curl http://localhost:11434/api/ps for loaded models.

Filtering the stream

The thoughts log is colorized for the terminal. To grep just goal outcomes from logs:

docker logs hollow-api 2>&1 | grep "→ goal=" | tail -20

For validation results:

docker logs hollow-api 2>&1 | grep "VALIDATE_END"

For everything an agent did in a window:

docker logs hollow-api --since 30m 2>&1 | grep "scout"