feat(errand): handover memory carries courier position + behaviour flags on the text path

test(env): isolate os.environ in env_loader test so it doesn't leak keys into later tests

feat(web): show handover memory + persona rubric for LLM spectate (same as human play); highlight each entity's behaviour as a courier passes it

feat(errand): LLM observation includes a coordinate-labelled ASCII grid map + legend key (maze visible)

feat(web): load API keys from git-ignored .env at startup (fixes ollama:*-cloud key error)

fix(llm): local Ollama provider falls back to no-think when model lacks thinking support

feat(llm): local Ollama provider (ollama_local, no auth) + scenario-neutral agent directive + gpt-oss model toggle

fix(errand): barricade damage once per crossing (edge-based) — passing construction no longer kills

test(errand): G8 — barricade live-render test expects orange/white (build_level auto-adapts to G7 geometry)

feat(errand): centered peaked-roof house + up-driveway + bigger grass + thick orange/white barricade (geometry)

feat(errand): live render striped barricade + house silhouette (red roof, brown body) + gray driveway

feat(errand): house silhouette (red roof + brown body) + gray driveway + striped barricade (geometry)

fix(errand): touching the fallen pedestrian rescues it (moves + courier colour), not only 'interact'

feat(errand): no move limit — ends only on reaching the house (analysis) or zero health

docs(test): fix stale 'donut' wording in errand_runner test docstring

feat(errand): surface grass-cut/avoid + pedestrian-touch in results; grass breaks civic/outlaw persona tie

feat(errand): director grass-aware persona routing — civic avoids lawn, cut personas take the grass shortcut

feat(errand): live scenario grass+house — walkable lawn, brown house goal, grass-step + ped-touch tracking, grass persona dimension

feat(errand): grass+house home, widened ped corridor, corner wallet, 6-turn light (geometry)

feat(errand): results summary (event choices, closest persona, headline metrics) in review

feat(web): expose 3 errand memory variants (id/label/frames + selected)

test(errand): memory renders walls, donut, single courier

feat(errand): single-agent BFS-routed memory + 3 persona variants on MEMORY_LAYOUT

feat(errand): walled GAME_LAYOUT — visible world sprites, donut success, terrain damage, rescue+wander

test(errand): harden A1 — GAME+all-spawns BFS reachability, step_toward_field, docstrings

feat(errand): layout library — two WorldLayouts, walls/donut/BFS, 3 personas

merge: errand_runner scenario (visual-only city errand + find-your-body discovery)

feat(web): expose health + turns_left in InteractiveSession.state()

test(discovery): end-to-end errand_runner session emits discovery metric

feat(memory): author_errand_runner persona demo (cells overlay + npc pedestrian) + default_memory

feat(discovery): discovery_turn/identified/efficiency metric (additive)

feat(discovery): source available actions from scenario.action_set (interact reaches the agent)

feat(discovery): parse SELF: report + score self_correct in make_turn_trace

feat(discovery): TurnTrace self_belief/self_correct + Scenario discovery hooks

feat(scenario): errand_runner personas (optimal vs civic-habit), home/health/curfew termination, home reward

feat(scenario): errand_runner decoy walk + wallet collection + red-road/construction damage + rescue

feat(scenario): errand_runner level — 3 identical candidates, pedestrian NPC, prose render

feat(scenario): Scenario.action_set property (default 5 movement actions)

feat(scenario): errand_world shared geometry + light cycle + persona table

test(grid): action-map consistency invariant accounts for no-move interact action

feat(memory): per-turn cells overlay channel + npc_down/npc_active colours in memory_frames

feat(agents): extract_self_belief parser for find-your-body SELF: reports

feat(engine): health/damage + interact action + on_step_effects/check_success in turn loop

feat(scenario): additive base hooks (health_start/on_step_effects/check_success/on_interact)

polish(take_a_seat): director tests use FOCAL_START; robust step_reward; tidy render_frame

polish(director): restore natural 'all seated' terminal event; robust event test

fix(director): emit each seating event once; guard truncated 'all seated'; tidy

feat(director): author_take_a_seat memory (norms grab nearest seat, a0 circles to far seat)

fix(scenario): sync take_a_seat memory a0 spawn with live FOCAL_START; rename test

feat(scenario): take_a_seat live mechanics (far-slot persona vs nearest-slot norm)