"""Content smoke tests — gpt-image-2 against real Tartarus assets. Six smokes designed to surface USABLE output for visual evaluation, not just an API handshake: 1. test_character_jade_t2i — character ref from bible (Jade, phase 1) 2. test_character_wren_t2i — character ref from bible (Wren, armored) 3. test_location_corridor_t2i — env ref (int_lower_decks_corridor) 4. test_location_shaft_t2i — env ref (int_lower_decks_maintenance_shaft) 5. test_previs_ep001_sh01_via_build_previs_prompt — exact plan-pass code path with the canonical prompt builder 6. test_image_to_image_edit_existing_previs — i2i endpoint on an existing previz frame Gated by RECOIL_RUN_LIVE_TESTS=1. Per-shot hard ceiling: $1.00 (mirrors model_profiles.gpt-image-2.max_cost_per_shot_usd). Quality: "medium" across the board (better signal than "low", well under the ceiling). Outputs land in /tmp/gpt_image_2_content_smoke/ for review. Each test prints the absolute path + measured cost on success. NOTE: gpt-image-2's prompt-tuning needs are not yet codified. These tests use build_previs_prompt (the generic Flash builder) for #5 — known not to be tuned for gpt-image-2. Tests #1-4 build natural-language prompts inline from bible content, which matches gpt-image-2's documented preference. """ from __future__ import annotations import json import os import shutil from pathlib import Path import pytest LIVE = os.environ.get("RECOIL_RUN_LIVE_TESTS") == "1" _OUT_DIR = Path("/tmp/gpt_image_2_content_smoke") _HARD_BUDGET_CEILING_USD = 1.00 # matches model_profiles.gpt-image-2.max_cost_per_shot_usd _QUALITY = "medium" _REPO_ROOT = Path(__file__).resolve().parents[4] _BIBLE_PATH = _REPO_ROOT / "projects/tartarus/state/visual/global_bible.json" _PLAN_PATH = _REPO_ROOT / "projects/tartarus/state/visual/plans/ep_001_plan.json" def _ensure_out_dir(): _OUT_DIR.mkdir(parents=True, exist_ok=True) def _bible() -> dict: return json.loads(_BIBLE_PATH.read_text()) def _make_ctx(label: str): """Build a DispatchContext + StepRunner for a smoke test.""" from recoil.pipeline.core import dispatch, DispatchContext from recoil.execution.step_runner import StepRunner from recoil.execution.execution_store import ExecutionStore from recoil.execution.step_types import ProjectPaths store = ExecutionStore( project="tartarus", db_path=Path(f"/tmp/gpt_image_2_smoke_store_{label}"), ) paths = ProjectPaths.for_episode("tartarus", 1) sr = StepRunner(store=store, paths=paths) ctx = DispatchContext( caller_id=f"gpt_image_2_smoke_{label}", step_runner=sr, project="tartarus", episode=1, receipts_log_path="DISABLED", ) return dispatch, ctx def _resolve_output_path(raw: str) -> Path: """rr.output_path is project-relative (under projects/{project}/) per StepRunner/ProjectPaths convention. Resolve to absolute against the canonical projects_root. """ p = Path(raw) if p.is_absolute(): return p from recoil.core.paths import projects_root return projects_root() / "tartarus" / raw def _assert_and_save(receipt, dest_name: str) -> tuple[Path, float]: """Common post-call assertions + copy output to /tmp/gpt_image_2_content_smoke/.""" rr = receipt.run_result assert rr.success, f"dispatch failed: {rr.error}" assert rr.output_path is not None, "no output_path on RunResult" src = _resolve_output_path(str(rr.output_path)) assert src.exists() and src.stat().st_size > 0, ( f"output file missing or empty at {src}" ) # gpt-image-2 fal billing format gap: _extract_fal_billing_cost_usd # doesn't read this endpoint's response shape, so cost_usd often lands # as 0.0 with a WARNING. That's a separate substrate fix — don't fail # the content smoke on it. Still enforce the hard ceiling for safety # if the extractor ever does return a value. cost_usd = (rr.metadata or {}).get("cost_usd") or 0.0 if cost_usd > 0: assert cost_usd < _HARD_BUDGET_CEILING_USD, ( f"cost ${cost_usd:.4f} breached " f"${_HARD_BUDGET_CEILING_USD:.2f} per-shot ceiling" ) _ensure_out_dir() dest = _OUT_DIR / dest_name shutil.copy2(src, dest) cost_str = f"${cost_usd:.4f}" if cost_usd > 0 else "$? (extractor gap)" print(f"\n[smoke OK] {dest_name} → {dest} ({cost_str})") return dest, cost_usd # ────────────────────────────────────────────────────────────────────────── # 1. Character — Jade (phase 1 full-mask wardrobe) # ────────────────────────────────────────────────────────────────────────── @pytest.mark.skipif(not LIVE, reason="set RECOIL_RUN_LIVE_TESTS=1 to enable") def test_character_jade_t2i(): bible = _bible() jade = bible["characters"]["JADE"] phase1 = next(p for p in jade["phases"] if p["phase_id"] == "jade_phase_1_full_mask") prompt = ( "Full-body character reference, 9:16 vertical. " f"Subject: {jade['display_name']}, {jade['gender']}. " f"{jade['visual_description']} " f"Wardrobe: {phase1['wardrobe_description']} " "Neutral pose, arms relaxed, facing camera, even soft frontal lighting, " "plain dark gray studio backdrop, full figure visible head-to-toe. " "Photorealistic, cinematic film stock grain, no text, no logos." ) dispatch, ctx = _make_ctx("jade") receipt = dispatch("image_t2i", { "shot_id": "GPTIMG2_CHAR_JADE", "model": "gpt-image-2", "prompt": prompt, "aspect_ratio": "9:16", "quality": _QUALITY, }, context=ctx) _assert_and_save(receipt, "character_jade.png") # ────────────────────────────────────────────────────────────────────────── # 2. Character — Wren (armored figure; tests material rendering) # ────────────────────────────────────────────────────────────────────────── @pytest.mark.skipif(not LIVE, reason="set RECOIL_RUN_LIVE_TESTS=1 to enable") def test_character_wren_t2i(): bible = _bible() wren = bible["characters"]["WREN"] phase1 = wren["phases"][0] prompt = ( "Full-body character reference, 9:16 vertical. " f"Subject: {wren['display_name']}, {wren['gender']}. " f"{wren['visual_description']} " f"Wardrobe: {phase1['wardrobe_description']} " "Standing pose, frontal, even neutral lighting, plain dark gray studio " "backdrop, full figure visible head-to-toe. Photorealistic, gritty " "weathered surfaces, fine metal and rust detail, no text, no logos." ) dispatch, ctx = _make_ctx("wren") receipt = dispatch("image_t2i", { "shot_id": "GPTIMG2_CHAR_WREN", "model": "gpt-image-2", "prompt": prompt, "aspect_ratio": "9:16", "quality": _QUALITY, }, context=ctx) _assert_and_save(receipt, "character_wren.png") # ────────────────────────────────────────────────────────────────────────── # 3. Location — int_lower_decks_corridor # ────────────────────────────────────────────────────────────────────────── @pytest.mark.skipif(not LIVE, reason="set RECOIL_RUN_LIVE_TESTS=1 to enable") def test_location_corridor_t2i(): bible = _bible() loc = bible["locations"]["int_lower_decks_corridor"] palette = ", ".join(loc.get("color_palette") or []) lp = loc.get("lighting_profile") or {} prompt = ( f"Environment reference plate, 9:16 vertical, no characters present. " f"Location: derelict spaceship corridor, {loc['habitat_zone']}. " f"Visual elements: {loc['description']}. " f"Atmosphere: {loc.get('atmosphere', '')[:200]} " f"Lighting: {lp.get('primary_source','')}, {lp.get('direction','')}, " f"{lp.get('quality','')} quality, {lp.get('color_temp','')} color temp. " f"Color palette: {palette}. " "Wide deep-perspective shot down the corridor. Photorealistic, " "cinematic film stock grain, no text, no logos." ) dispatch, ctx = _make_ctx("corridor") receipt = dispatch("image_t2i", { "shot_id": "GPTIMG2_LOC_CORRIDOR", "model": "gpt-image-2", "prompt": prompt, "aspect_ratio": "9:16", "quality": _QUALITY, }, context=ctx) _assert_and_save(receipt, "location_corridor.png") # ────────────────────────────────────────────────────────────────────────── # 4. Location — int_lower_decks_maintenance_shaft # ────────────────────────────────────────────────────────────────────────── @pytest.mark.skipif(not LIVE, reason="set RECOIL_RUN_LIVE_TESTS=1 to enable") def test_location_shaft_t2i(): bible = _bible() loc = bible["locations"]["int_lower_decks_maintenance_shaft"] palette = ", ".join(loc.get("color_palette") or []) lp = loc.get("lighting_profile") or {} prompt = ( f"Environment reference plate, 9:16 vertical, no characters present. " f"Location: derelict spaceship vertical maintenance shaft, {loc['habitat_zone']}. " f"Visual elements: {loc['description']}. " f"Atmosphere: {loc.get('atmosphere', '')[:200]} " f"Lighting: {lp.get('primary_source','')}, {lp.get('direction','')}, " f"{lp.get('quality','')} quality, {lp.get('color_temp','')} color temp. " f"Color palette: {palette}. " "Looking down into the shaft, sense of depth and vertigo. " "Photorealistic, cinematic film stock grain, no text, no logos." ) dispatch, ctx = _make_ctx("shaft") receipt = dispatch("image_t2i", { "shot_id": "GPTIMG2_LOC_SHAFT", "model": "gpt-image-2", "prompt": prompt, "aspect_ratio": "9:16", "quality": _QUALITY, }, context=ctx) _assert_and_save(receipt, "location_shaft.png") # ────────────────────────────────────────────────────────────────────────── # 5. Previs via build_previs_prompt — exact plan-pass code path # ────────────────────────────────────────────────────────────────────────── @pytest.mark.skipif(not LIVE, reason="set RECOIL_RUN_LIVE_TESTS=1 to enable") def test_previs_ep001_sh01_via_build_previs_prompt(): """Real plan-pass smoke. Loads EP001_SH01 from the actual plan, runs it through build_previs_prompt (the canonical builder dispatched by the orchestrator's previz step), fires through gpt-image-2. """ from recoil.pipeline._lib.prompt_engine import build_previs_prompt bible = _bible() plan = json.loads(_PLAN_PATH.read_text()) shots = plan.get("shots") or [ s for seq in (plan.get("sequences") or {}).values() for s in seq ] sh01 = next(s for s in shots if s["shot_id"] == "EP001_SH01") config = {"film_stock": "Kodak Vision3 500T"} prompt = build_previs_prompt(sh01, bible, config) assert prompt and len(prompt) > 20, "build_previs_prompt produced empty/trivial output" print(f"\n[plan-pass previs prompt ({len(prompt.split())} words)]:\n {prompt[:240]}...\n") dispatch, ctx = _make_ctx("previs_sh01") receipt = dispatch("image_t2i", { "shot_id": "GPTIMG2_PREVIS_EP001_SH01", "model": "gpt-image-2", "prompt": prompt, "aspect_ratio": sh01.get("aspect_ratio") or "9:16", "quality": _QUALITY, }, context=ctx) _assert_and_save(receipt, "previs_ep001_sh01.png") # ────────────────────────────────────────────────────────────────────────── # 6. Image-to-image edit — gpt-image-2's edit endpoint on existing previz # ────────────────────────────────────────────────────────────────────────── @pytest.mark.skipif(not LIVE, reason="set RECOIL_RUN_LIVE_TESTS=1 to enable") def test_image_to_image_edit_existing_previs(): """Smoke for fal-ai/gpt-image-2/image-to-image. Takes the existing SH01 hero previs and asks gpt-image-2 to push the lighting darker / add visible rust + steam. Verifies the edit endpoint is wired (not just t2i) and that single-ref edit produces a usable result. Per model_profiles: max_reference_images=1, multi_ref_supported=false. Pass exactly one reference_image. """ src = _REPO_ROOT / "projects/tartarus/output/previs/ep_001/shot_001_take018.png" if not src.exists(): pytest.skip(f"expected previs frame not present at {src}") prompt = ( "Edit this frame: push lighting moodier and darker, add visible " "steam wisps from the cracked pipes, intensify the rust streaks " "on the metal grating, preserve composition and subject placement." ) dispatch, ctx = _make_ctx("edit_sh01") receipt = dispatch("image_t2i", { "shot_id": "GPTIMG2_EDIT_EP001_SH01", "model": "gpt-image-2", "prompt": prompt, "aspect_ratio": "9:16", "quality": _QUALITY, # Trigger the i2i endpoint per _infer_gpt_image_2_action in fal.py:77 "reference_images": [str(src)], }, context=ctx) _assert_and_save(receipt, "edit_ep001_sh01.png")