All characters on nemotron-4b; one cold-load per fork

@GPU (duration=120) + DECISIONS_PER_DRAIN=2 for tighter drains

Timestamped logs via game.log.tlog; explains visible drain pauses

Add Cyra hero (saboteur); sharpen saboteur shell-as-weapon line

Fix Pause to halt drain; Spawn click feedback; Info modal with The Dot

Drop Flux 2; 4 starting chars; pin Ada=engineer Bram=diplomat

interpreter: per-fork interpreter so shell output reaches the UI (pid-mismatch detection)

interpreter: stop spurious 'mass death' flag (is_alive() unreliable across @spaces.GPU fork)

Named-place movement: atomic move_to, drop coord reasoning, 10/gather, per-archetype reward, default llama-3b, smart-mock updated

actions: strip surrounding quotes from move_to target (fixes "'cave'" loop)

actions: mine_coal/draw_water yield 10 units per action (was 1); tool descriptions updated so agent knows

actions: move_to short-circuits when already at destination so agent can mine/draw/post in the same decision

stream: strip <think> blocks + broaden meta-leak filter for Nemotron reasoning output

scheduler: idle drift skips characters at named places (fixes mine_coal not at cave) + strip <think> blocks from agent thought

stream: stop SoC meta-leak (few-shot exemplars + leak filter)

Recording prep: JS mutex, mixed models (Ada/Bram=Nemotron, rest=SmolLM2), drain=3, init=3, duration=90s, idle-drift, mock latency profile

Sync latest local snapshot

Default model: always Nemotron 3 Nano 4B, regardless of env

Trace: strip smolagents </code> marker from displayed thought

SoC: include in character's own next-decision perception prompt

SoC: hook into _decide + load/save with state.json

SoC: new game/stream.py — refresh every 10 decisions per char

SoC: stream_of_consciousness + decisions_since_soc on Character

Trace: record each _decide via tool monkey-patch + smolagents memory

Trace: TraceEntry + ring buffer (1000) on Character

Fix: shell lock release no longer gated on exact shell_queue tile

Roster: add SmolLM2 360M for local-friendly testing

The Dot: embed full game mechanics (resources, locker, shell semaphore)

Filesystem state persistence with flock to survive @spaces.GPU fork

Log PID + TID at endpoint entry to diagnose multi-worker isolation

Architectural fix: kill bg tick thread, tick synchronously inside @spaces.GPU and snapshot calls

Diagnostic: log every _advance_move call

Diagnostics: heartbeat with Townlet id; tick try/except

Diagnostics: log pending size + action params

MAX_AGENT_STEPS 1 -> 2 for parse-error retry

Fix: ensure smol_adapter.py is the GBNF-off version

Fix: restore scheduler.py contents (was overwritten by smol_adapter on previous push)

Drop GBNF default; log per-decision outcomes

Townlet: initial private upload with Zero-GPU fixes