REPLAYS.md

description	Full-stack live-narration / GM-assistant replay suite.
tags	testing e2e replays
layer	0

Full-stack replay suite

The running backend over HTTP, six surfaces. One entry point (scripts/run_e2e_replays.py), seven sub-scripts, one combined report.

For the in-process loop harness (real character + real scene + real SceneLoop, no HTTP), see HARNESS_FULL_LOOP.md.

Quick start

# 1. Docker stack + UI backend uvicorn at :8000 (usually running already).
docker compose --env-file .env -f infra/docker-compose.yml up -d

# 2. Env.
set -a; source .env; set +a
export MONITOR_PLAYTEST_MODEL=ollama/qwen2.5:latest

# 3. Full suite (~25-30 min wall clock).
python scripts/run_e2e_replays.py

# Subset.
python scripts/run_e2e_replays.py --only forge,copilot

The combined report lands at tests/e2e/logs/replays/replay_run_<UTC>.md.

Suites at a glance

Suite	Script	Wall clock
forge	scripts/forge_replay.py	~25 s
copilot	scripts/live_copilot_observe.py	~60 s
long_form	scripts/long_form_narration.py	~12 min
subsystem	scripts/subsystem_replay.py	~10 min
char_creation	scripts/character_creation_replay.py	~10 min
session_observe	scripts/live_session_observe.py	~12 min

Total full-suite wall clock: ~70 minutes when the player model is fast local ollama.

CLI flags

--api-url      http://localhost:8000/api
--only         forge,copilot
--skip         session_observe
--output-dir   tests/e2e/logs/replays
--report       /path/to/custom_report.md

What "passing" means

• Exit 0 — suite completed with no internal assertion failures.
• Exit 1 — at least one internal assertion failed (e.g., a fallback marker appeared in GM output, the canonical-universe endpoint returned 404).
• Exit 2 — semantic failure the suite can't recover from (e.g., char-creation never reached active_play).

The entry script treats any non-zero as a failure and writes a replay_run_<UTC>.md with a per-suite pass/fail table.

Tag reference

long_form_narration.py (22 tags): char_creation_accept, char_creation_stat_roll, first_in_fiction_action, social_inquiry, lore_recall, social_deal, exploration_intent, first_combat_intent, combat_engaged, combat_finisher, tactical_assessment, deeper_push, puzzle_encounter, puzzle_solve, puzzle_consequence, loot_choice, boss_intro, boss_phase_two, boss_finisher, extraction, extraction_pressure, wrap_up.

subsystem_replay.py (14 tags): intro, trust_seed, lore_recall, faction_probe, social_negotiation, npc_voice, lore_event, relationship_mirror, faction_hook, social_deal, npc_test, lore_blacksite, relationship_close, wrap_up.

character_creation_replay.py (10 tags): intro, ask_species, ask_class, choose_class, roll_stats, choose_background, choose_equipment, ask_sheet, confirm_lock, first_action.