Official Codebase of 3DCodeVerse
Python
languages/judges/models: honest prices, honest limits, honest lint
gaoypengb83f909b0bea
diff --git a/harness/codeverse/conventions.py b/harness/codeverse/conventions.pyindex f1258c3c..ea1b9234 100644--- a/harness/codeverse/conventions.py+++ b/harness/codeverse/conventions.py@@ -134,6 +134,10 @@ def to_pascal(name: str) -> str:return "".join(w[:1].upper() + w[1:] for w in to_snake(name).split("_") if w) or "Part"+#: a part / assembly name as the contracts require it: ``SeatCushion``, instances ``SeatCushion_3``+PASCAL_RE = re.compile(r"^[A-Z][a-z0-9]+(?:[A-Z][a-z0-9]*)*(?:_\d+)?$")++def slugify(text: str, max_len: int = 48) -> str:"""Filesystem/URL-safe slug for run directories."""s = to_snake(text)[:max_len].strip("_")diff --git a/harness/codeverse/judges/base.py b/harness/codeverse/judges/base.pyindex 2198af56..abad941e 100644--- a/harness/codeverse/judges/base.py+++ b/harness/codeverse/judges/base.py@@ -1,8 +1,8 @@-"""Judge protocol + inputs."""+"""``JudgeInput`` + the pure round-replay helpers ``3dcv judge`` and calibration share."""from __future__ import annotations-from typing import Any, Protocol+from typing import Anyfrom pydantic import BaseModel, Field@@ -27,13 +27,7 @@ class JudgeInput(BaseModel):)-class Judge(Protocol):- name: str-- def judge(self, inp: JudgeInput) -> Judgment: ...---# ===================================================================== replay_input+# ===================================================================== round replaydef plan_digest(plan: dict[str, Any]) -> str:"""A track-agnostic plan summary (parts / joints / zones / assets / cameras) from ``plan.json``."""bits: list[str] = []diff --git a/harness/codeverse/judges/calibration.py b/harness/codeverse/judges/calibration.pyindex 77254726..95bcc75a 100644--- a/harness/codeverse/judges/calibration.py+++ b/harness/codeverse/judges/calibration.py@@ -34,6 +34,7 @@ from codeverse.contracts.spec import Specfrom codeverse.judges.base import JudgeInput, judged_subset, plan_digest, resolve_pathsfrom codeverse.judges.rubrics import is_degradedfrom codeverse.judges.vlm_judge import VlmJudge+from codeverse.proc import read_json_or_nonefrom codeverse.workspace import Workspacelog = logging.getLogger(__name__)@@ -150,16 +151,12 @@ class CalibrationTable(BaseModel):# --------------------------------------------------------------------------- loading runs-def _read_json(path: Path) -> dict[str, Any]:- return json.loads(path.read_text())--def load_run_cases(run_dir: Path, *, rounds: list[int] | None = None) -> list[RoundCase]:"""Rebuild judge inputs for every judgeable round of a run (rounds with renders)."""run_dir = Path(run_dir)spec = Spec.model_validate_json((run_dir / "spec.json").read_text())- plan = _read_json(run_dir / "plan.json") if (run_dir / "plan.json").is_file() else {}- record = _read_json(run_dir / "record.json") if (run_dir / "record.json").is_file() else {}+ plan = read_json_or_none(run_dir / "plan.json") or {}+ record = read_json_or_none(run_dir / "record.json") or {}best = record.get("best_round")acceptance = plan.get("acceptance") or []digest = plan_digest(plan)@@ -237,7 +234,7 @@ def spearman(xs: list[float], ys: list[float]) -> float | None:# --------------------------------------------------------------------------- main entrydef _judge_case(case: RoundCase, judge: VlmJudge, geometry: RenderSet | None, out: Path) -> CalibrationRow:t0 = time.time()- j = judge.judge(case.inp, geometry_views=geometry)+ j = judge.judge(case.inp.model_copy(update={"geometry_views": geometry}) if geometry is not None else case.inp)raw = json.loads(j.raw) if j.raw else {}jdir = out / "judgments"jdir.mkdir(parents=True, exist_ok=True)diff --git a/harness/codeverse/judges/pairwise.py b/harness/codeverse/judges/pairwise.pyindex 7072a4b5..3a1ab876 100644--- a/harness/codeverse/judges/pairwise.py+++ b/harness/codeverse/judges/pairwise.py@@ -95,14 +95,14 @@ class PairwiseJudge:model_id: str | None = None,*,temperature: float = 0.2,- max_px: int = 768,+ max_px: int = 1024,views_per_side: int = 4,chat_model: ChatModel | None = None,cache_dir: Path | None = None,):self.model_id = model_id or get_settings().default_judgeself.temperature = temperature- self.max_px = max(max_px, 1024) # a 2×2 montage needs the resolution+ self.max_px = max(max_px, 1024) # a 2×2 montage needs the resolution: 1024 is the floor, not a hintself.views_per_side = min(4, views_per_side)self._model = chat_modelself.cache_dir = cache_dirdiff --git a/harness/codeverse/judges/prompt_builder.py b/harness/codeverse/judges/prompt_builder.pyindex 5311a6b0..ee22396f 100644--- a/harness/codeverse/judges/prompt_builder.py+++ b/harness/codeverse/judges/prompt_builder.py@@ -1,10 +1,14 @@-"""Judge image preparation: downscale, label strips, caching, view geometry.--Every image sent to a judge goes through ``prepare_image`` so the token cost is-bounded (≤ ``max_px`` on the long side) and every view carries a burnt-in label-(``VIEW 3/8 — front · az 0° el 8°``) that the model can cite as evidence.-Prepared files are cached under ``<cache_dir>/<sha>.png`` keyed by source path,-mtime, size, max_px and label text.+"""What the judge SEES: image prep, montage planning and message assembly.++* image prep (``prepare_image``): every image sent to a judge is downscaled+ (≤ ``max_px`` on the long side) and carries a burnt-in label (``VIEW 3/8 — front ·+ az 0° el 8°``) the model can cite as evidence; prepared files are cached under+ ``<cache_dir>/<sha>.png`` keyed by source path, mtime, size, max_px and label text.+* montage planning (``plan_montages`` / ``render_montage``): ranked views packed into+ ≤2×2 sheets — shaded, geometry (clay/normals), poses, detail crops.+* message assembly (``build_judge_messages``): rubric + montages + measurements ++ acceptance items → the ``ChatRequest`` system/messages; ``judge_prompt_hash`` is the+ verdict's provenance stamp."""from __future__ import annotations@@ -22,6 +26,7 @@ from PIL import Image, ImageDrawfrom codeverse.config import get_settingsfrom codeverse.contracts.artifacts import (+ RENDER_MODES,GateReport,Judgment,Measurement,@@ -124,8 +129,8 @@ def image_part(path: Path, label: str) -> ImagePart:# ===================================================================== montageMontageKind = Literal["shaded", "geometry", "poses", "pose_sheet", "detail"]-#: modes that show geometry without material/lighting noise-GEOMETRY_MODES = ("clay", "normals", "wire", "silhouette", "depth")+#: modes that show geometry without material/lighting noise: every render mode but shaded+GEOMETRY_MODES = tuple(m for m in RENDER_MODES if m != "shaded")#: most-informative-first order for the object rig (names from conventions.OBJECT_VIEWS)OBJECT_RANK = ("front_right_34", "back_left_34", "top", "low_front_left", "front", "right", "back", "left")#: scene rig: authored cameras first (graded for composition), then the overview rig.diff --git a/harness/codeverse/judges/rubrics.py b/harness/codeverse/judges/rubrics.pyindex e19b5113..0329f9e5 100644--- a/harness/codeverse/judges/rubrics.py+++ b/harness/codeverse/judges/rubrics.py@@ -7,7 +7,7 @@ the weighted mean) and a ``kind`` (``visual`` = scored by the VLM,``measured`` = scored in code and merely *shown* to the VLM).``caps`` are rules that bound the overall score from deterministic gate-findings (build error → 0.0, floating part → ≤ 0.6 ...) — see ``caps.py``.+findings (build error → 0.0, floating part → ≤ 0.6 ...) — see ``apply_caps`` below.``defects`` is a fixed **binary checklist** the VLM answers (present / absent)— "a part floats", "lying on its back", "flat untextured" … — and the code@@ -720,7 +720,7 @@ def aggregate_samples(if tied:tie_broken.append(did)# a checklist defect that a passed gate measured as absent neither penalises nor caps- # (judges/caps.measured_absent); it is still named in the verdict tail+ # (``veto_measured_defects``); it is still named in the verdict taildefects, overridden = veto_measured_defects(rubric, defects, gates)penalty = round(sum(rubric.defect(did).penalty for did, on in defects.items() if on), 4)after_defects = max(0.0, overall_uncapped - penalty)diff --git a/harness/codeverse/judges/vlm_judge.py b/harness/codeverse/judges/vlm_judge.pyindex a82837c8..41bcb3d1 100644--- a/harness/codeverse/judges/vlm_judge.py+++ b/harness/codeverse/judges/vlm_judge.py@@ -5,8 +5,8 @@ Per sample: one ``ChatRequest`` with a per-rubric JSON schema (criteria scoredplan, acceptance verdicts); samples differ by montage/tile order (shuffle seed)so n-sample mean/std measures judge noise. With ``n_samples > 1`` the modelcalls run in parallel (``codeverse.proc``); results accumulate in sample order. Score/defect penalties/floors/caps-/pass are computed in code (``scoring.py``). Images are ≤2×2 montages-(``montage.py``); tracks may pass a clay/normals ``geometry_views`` RenderSet. Retries: up to+/pass are computed in code (``rubrics.aggregate_samples``). Images are ≤2×2 montages+(``prompt_builder``); tracks may pass a clay/normals ``JudgeInput.geometry_views`` RenderSet. Retries: up to``max_attempts`` per sample on ``ModelError`` / parse failure, all of them inside one``sample_budget_s`` (:data:`SAMPLE_BUDGET_S`) that also clips the model's own retrydeadline (``ChatRequest.max_wait_s``); if no sample@@ -27,7 +27,7 @@ from pathlib import Pathfrom typing import Any, Literalfrom codeverse.config import get_settings-from codeverse.contracts.artifacts import Judgment, RenderSet, RenderView+from codeverse.contracts.artifacts import Judgment, RenderViewfrom codeverse.contracts.chat import ChatRequest, ChatResponsefrom codeverse.contracts.common import Usagefrom codeverse.judges.base import JudgeInput@@ -64,6 +64,9 @@ class JudgeContext:measured_scores: dict[str, float] = field(default_factory=dict)extra_text: str = ""extra_images: list[tuple[str, str]] = field(default_factory=list)+ #: what building the context cost (ReferenceJudge's mismatch pass is a real model+ #: call); folded into the verdict so BudgetGuard and record.total_usage see it+ usage: Usage = field(default_factory=Usage)class VlmJudge:@@ -93,7 +96,7 @@ class VlmJudge:self.n_samples = max(1, int(n_samples))if self.n_samples % 2 == 0:log.warning("judge n_samples=%d is even: exact vote ties on defects / acceptance items are decided by the "- "representative sample (scoring.py); an odd n gives a true majority", self.n_samples)+ "representative sample (rubrics.aggregate_samples); an odd n gives a true majority", self.n_samples)self.temperature = temperatureself.thinking = thinkingself.max_attempts = max(1, int(max_attempts))@@ -123,20 +126,15 @@ class VlmJudge:return self._model# ------------------------------------------------------------------ API- def judge(self, inp: JudgeInput, *, geometry_views: RenderSet | None = None) -> Judgment:- """Judge one round.-- ``geometry_views``: optional clay/normals RenderSet for the geometry montage- (falls back to ``inp.geometry_views`` if the input model carries that field;- clay/normals views embedded in ``inp.renders`` by ``RenderView.mode`` are- always routed to the geometry montage).- """- geometry_views = geometry_views or getattr(inp, "geometry_views", None)+ def judge(self, inp: JudgeInput) -> Judgment:+ """Judge one round. ``inp.geometry_views`` feeds the geometry montage (clay/normals+ views embedded in ``inp.renders`` by ``RenderView.mode`` are always routed there too)."""+ geometry_views = inp.geometry_viewsacceptance_ids = [a.id for a in inp.acceptance]schema = wire_schema(self.rubric, acceptance_ids)ctx = self.context(inp)measured, extra_text, extra_images = ctx.measured_scores, ctx.extra_text, ctx.extra_images- usage = Usage()+ usage = ctx.usagesamples: list[JudgeOutput] = []errors: list[str] = []reqs: list[ChatRequest] = []@@ -223,7 +221,7 @@ class VlmJudge:usage = usage + (getattr(e, "usage", None) or Usage()) # the provider billed itlast = f"ModelError(attempt {attempt}): {e}"log.warning("judge %s: %s", req.label, last)- if not e.retryable and attempt >= 2:+ if not e.retryable: # same request again cannot fix a 400 / a blocked promptbreakcontinueusage = usage + resp.usage@@ -311,21 +309,6 @@ class ReferenceJudge(VlmJudge):self.diff = diffself._diff_model = diff_modelself.diff_model_id = diff_model_id- #: the last diff computed (tests / callers that want the mismatches themselves)- self.last_diff: Any | None = None-- # ------------------------------------------------------------------ API- def judge(self, inp: JudgeInput, **kwargs: Any) -> Any:- """``VlmJudge.judge`` plus the mismatch pass's own spend.-- The diff is a real model call made from :meth:`context`; folding its- ``Usage`` into the verdict keeps ``BudgetGuard`` and ``record.total_usage``- honest (the ledger already sees it through the metered chat model)."""- verdict = super().judge(inp, **kwargs)- diff = self.last_diff- if diff is not None and getattr(diff, "usage", None) is not None:- verdict.usage = verdict.usage + diff.usage- return verdict# ------------------------------------------------------------------ hookdef context(self, inp: JudgeInput) -> JudgeContext:@@ -363,18 +346,17 @@ class ReferenceJudge(VlmJudge):elif refs:info = self.measure_silhouette(inp)diff = self.reference_diff(inp, refs, info, synthesized=synth) if refs else None- self.last_diff = diffif diff is not None and diff.as_text():blocks.append(diff.as_text())- return JudgeContext(measured_scores=measured_scores, extra_text="\n\n".join(blocks), extra_images=images)+ return JudgeContext(measured_scores=measured_scores, extra_text="\n\n".join(blocks), extra_images=images,+ usage=diff.usage if diff is not None else Usage())# ------------------------------------------------------------------ proportion guarddef dimension_conflict(self, inp: JudgeInput, refs: list[Any]) -> dict[str, Any]:"""Does the target reference's own outline contradict the brief's dimensions?When it does, no numeric proportion signal derived from that picture may be- used against the object — the brief wins (see- :mod:`codeverse.reference.proportions`).+ used against the object — the brief wins (``reference.dimension_conflict``)."""targets = [r for r in refs if r.role == "target"] or refsif not targets:@@ -497,7 +479,7 @@ class ReferenceJudge(VlmJudge):def _plan_part_names(inp: JudgeInput) -> list[str]:"""Part names from the plan digest the judge already receives (best effort).- ``judges.replay_input.plan_digest`` writes them as one ``Parts: A, B×2, C`` segment;+ ``judges.base.plan_digest`` writes them as one ``Parts: A, B×2, C`` segment;a multi-line digest lists one ``- Name · role · …`` per line. Both are handled, andan unrecognised digest simply yields no names (the diff prompt then says so)."""diff --git a/harness/codeverse/languages/_common.py b/harness/codeverse/languages/_common.pyindex 80b6eac8..b0495f68 100644--- a/harness/codeverse/languages/_common.py+++ b/harness/codeverse/languages/_common.py@@ -6,11 +6,7 @@ and the runtime turns those into a :class:`BuildResult`. This module owns:* :func:`compose_build_result` — wrapper json + process outcome → BuildResult,failing loud when the wrapper did not report;-* :func:`read_json_file` / :func:`remove_stale` / :func:`strip_blender_noise`.--The subprocess + atomic-JSON primitives (:class:`ProcResult`,-:func:`run_subprocess`, :func:`tail`, :func:`write_json_atomic`) live in-:mod:`codeverse.proc` and are re-exported here for the runtime callers.+* :func:`read_json_file` / :func:`strip_blender_noise`.Wrappers themselves are standalone scripts (they never import ``codeverse``;Blender's bundled python cannot see this package).@@ -24,12 +20,13 @@ from pathlib import Pathfrom typing import Anyfrom codeverse.contracts.artifacts import BuildResult-from codeverse.proc import ( # noqa: F401 — re-exported- ProcResult,- run_subprocess,- tail,- write_json_atomic,-)+from codeverse.proc import ProcResult, tail++#: The one spelling of every runtime's typed build failures — recorded in build.json /+#: record.json / bench cells, so the strings stay as first recorded (2026-08-29: threejs,+#: urdf said "Timeout" and the GL runtimes "MissingEntry"; nothing keyed on either).+MISSING_ENTRY = "MissingEntryFile"+BUILD_TIMEOUT = "BuildTimeout"# lines Blender prints on every headless run that carry no signal for the agent_BLENDER_NOISE_PREFIXES = (@@ -78,12 +75,6 @@ def read_json_file(path: Path) -> dict[str, Any]:return data-def remove_stale(*paths: Path) -> None:- """Delete previous wrapper outputs so a crashed run cannot be mistaken for a fresh one."""- for p in paths:- p.unlink(missing_ok=True)--def compose_build_result(*,language: str,@@ -98,7 +89,7 @@ def compose_build_result(Contract: success requires the wrapper to say ``ok`` AND a non-empty GLB ondisk. Timeouts / crashes before ``build.json`` exists become a failed- BuildResult with a typed error (``BuildTimeout`` / ``WrapperCrash``) — never+ BuildResult with a typed error (``BUILD_TIMEOUT`` / ``WrapperCrash``) — neveran exception — so the orchestrator can route them to repair."""stdout_tail = tail(output_filter(proc.stdout) if output_filter else proc.stdout)@@ -108,7 +99,7 @@ def compose_build_result(if proc.timed_out:return BuildResult(ok=False,- error_type="BuildTimeout",+ error_type=BUILD_TIMEOUT,error_message=f"build exceeded the time limit ({proc.duration_ms // 1000}s); ""reduce geometry (subdivisions, array counts, boolean ops) so the script finishes quickly",**base,diff --git a/harness/codeverse/languages/_docs.py b/harness/codeverse/languages/_docs.pyindex fb5f7bd8..d506fa2c 100644--- a/harness/codeverse/languages/_docs.py+++ b/harness/codeverse/languages/_docs.py@@ -16,26 +16,16 @@ for a fact is a second answer, which is what law 2 exists to prevent.from __future__ import annotations-from pathlib import Path-from codeverse.contracts.common import Language-from codeverse.prompts import PROMPTS_DIR, load_text-from codeverse.prompts.catalog import PROMPT_DIRS as _PROMPT_DIRS-from codeverse.prompts.catalog import prompt_dir_for as _prompt_dir_for--#: re-exported for the callers that already import it from here; the mapping itself-#: now lives with the rest of the prompt-resolution policy in prompts/catalog.py.-PROMPT_DIRS = _PROMPT_DIRS+from codeverse.prompts import load_text+from codeverse.prompts.catalog import prompt_dir_for#: only reachable if the wheel shipped without its prompt dataMINIMAL_CONTRACT = "Write raw code in the language's native frame; meters; named parts."-prompt_dir_for = _prompt_dir_for--class RuntimeDocs:- """``contract_doc`` / ``cookbook_path`` for every ``LanguageRuntime``."""+ """``contract_doc`` for every ``LanguageRuntime``."""language: Language@@ -48,6 +38,3 @@ class RuntimeDocs:return load_text(f"{self.prompt_dir}/contract.md")except FileNotFoundError:return MINIMAL_CONTRACT-- def cookbook_path(self) -> Path:- return PROMPTS_DIR / self.prompt_dir / "cookbook.md"diff --git a/harness/codeverse/languages/_gl_common.py b/harness/codeverse/languages/_gl_common.pyindex 17b65b50..4cb2dfe4 100644--- a/harness/codeverse/languages/_gl_common.py+++ b/harness/codeverse/languages/_gl_common.py@@ -49,7 +49,7 @@ def invalidate_stale_outputs(ws: Workspace) -> None:"""Wipe the previous build's GL artifacts at the TOP of ``build()``.Hoisted from ``GlHost._run`` (which still wipes its own out_dir) so the- ``MissingEntry`` early returns — which never reach the host — also clear+ ``MISSING_ENTRY`` early returns — which never reach the host — also clear``frames/`` and ``gl_result.json``; sheet/gif/metrics/build.json were nevercleared anywhere, so a failed build left the previous round's sheet lookingcurrent to every bare-existence reader (``read_metrics``, the gallery)."""diff --git a/harness/codeverse/languages/base.py b/harness/codeverse/languages/base.pyindex d355f0eb..7db8f558 100644--- a/harness/codeverse/languages/base.py+++ b/harness/codeverse/languages/base.py@@ -40,8 +40,6 @@ class LanguageRuntime(Protocol):"""Prompt text: the authoring contract for this language (from prompts/<lang>/)."""...- def cookbook_path(self) -> Path: ...-#: Language → (module, class). Adding a language is a row, and the import stays lazy#: (a runtime pulls in bpy / moderngl / node bindings the other six do not need).diff --git a/harness/codeverse/languages/blender/__init__.py b/harness/codeverse/languages/blender/__init__.pyindex 6757ffd4..2ca56509 100644--- a/harness/codeverse/languages/blender/__init__.py+++ b/harness/codeverse/languages/blender/__init__.py@@ -5,7 +5,6 @@ from __future__ import annotationsimport astimport mathimport os-import reimport timefrom pathlib import Path@@ -13,21 +12,17 @@ from codeverse.config import Settings, get_settingsfrom codeverse.contracts.artifacts import BuildResult, GateFinding, GateReport, Severityfrom codeverse.contracts.common import ENTRY_FILE, Languagefrom codeverse.contracts.plan import BBox, PartPlan, Plan, StaticPlan-from codeverse.conventions import MAX_TRIS_OBJECT, to_pascal, to_snake-from codeverse.languages._ast_lint import ( # noqa: F401 — dotted re-exported+from codeverse.conventions import MAX_TRIS_OBJECT, PASCAL_RE, to_pascal, to_snake+from codeverse.languages._ast_lint import (BASE_FORBIDDEN_IMPORTS,check_imports,describe_parse_failure,dotted,safe_parse,)-from codeverse.languages._common import (- compose_build_result,- run_subprocess,- strip_blender_noise,-)+from codeverse.languages._common import MISSING_ENTRY, compose_build_result, strip_blender_noisefrom codeverse.languages._docs import RuntimeDocs-from codeverse.proc import scrub_secrets+from codeverse.proc import run_subprocess, scrub_secretsfrom codeverse.workspace import Workspace# ===================================================================== lint@@ -76,7 +71,6 @@ REMOVED_BSDF_INPUTS = {"Transmission Roughness": "(removed) use Roughness",}KNOWN_BINDINGS = ("Vector", "Matrix", "Euler", "Quaternion", "bmesh", "math", "random", "np", "numpy", "bpy")-PASCAL_RE = re.compile(r"^[A-Z][a-z0-9]+(?:[A-Z][a-z0-9]*)*(?:_\d+)?$")class _Collector(ast.NodeVisitor):@@ -310,7 +304,6 @@ def lint_blender_file(path: Path, *, target: str = "src/model.py") -> GateReport# ===================================================================== layoutENTRY_REL = ENTRY_FILE[Language.BLENDER] # "src/model.py"PARTS_DIR = "parts"-PARTS_PKG = "parts"def part_file_rel(part_name: str) -> str:@@ -357,7 +350,7 @@ def _imported_part_modules(tree: ast.Module) -> set[str]:for node in ast.walk(tree):if isinstance(node, ast.ImportFrom) and node.module:dotted = node.module.split(".")- if dotted[0] == PARTS_PKG:+ if dotted[0] == PARTS_DIR:if len(dotted) > 1:out.add(dotted[1])else: # ``from parts import seat, leg``@@ -365,7 +358,7 @@ def _imported_part_modules(tree: ast.Module) -> set[str]:elif isinstance(node, ast.Import):for a in node.names:dotted = a.name.split(".")- if dotted[0] == PARTS_PKG and len(dotted) > 1:+ if dotted[0] == PARTS_DIR and len(dotted) > 1:out.add(dotted[1])return out@@ -391,7 +384,7 @@ def _layout_rules(ws: Workspace, parts: list[Path], entry_tree: ast.Module | Nonif fn not in _exported_functions(tree):out.append(_finding(E, target, f"{target} does not define `def {fn}()`",f"every part file exports `def {fn}() -> bpy.types.Object` returning the named object "- f"at its world pose; model.py calls it (`from {PARTS_PKG}.{stem} import {fn}`)"))+ f"at its world pose; model.py calls it (`from {PARTS_DIR}.{stem} import {fn}`)"))for node in tree.body:if isinstance(node, ast.Expr) and isinstance(node.value, ast.Call):name = getattr(node.value.func, "id", "")@@ -401,7 +394,7 @@ def _layout_rules(ws: Workspace, parts: list[Path], entry_tree: ast.Module | Non"call here builds the part twice → auto-suffixed 'Name.001')", node.lineno))if entry_tree is not None and stem not in imported:out.append(_finding(W, target, f"{target} is never imported by src/model.py → its part is not built",- f"add `from {PARTS_PKG}.{stem} import {fn}` to model.py and call `{fn}()` in main()"))+ f"add `from {PARTS_DIR}.{stem} import {fn}` to model.py and call `{fn}()` in main()"))return out@@ -621,7 +614,7 @@ CONTRACT (the harness runs this file in an EMPTY scene with `blender -b --factoreach TOP-LEVEL (never parented under an Empty — that merges them into ONE measured part).{layout} * Materials: Principled BSDF (Base Color / Roughness / Metallic). GLB keeps flat PBR + imagetextures only (procedural node textures are NOT exported) — rely on geometry + flat PBR.- * Modifiers may stay unapplied (the exporter applies them). Keep < 500k triangles, < 120 s.+ * Modifiers may stay unapplied (the exporter applies them). Keep < {MAX_TRIS_OBJECT // 1000}k triangles.* NEVER: cameras, lights, world, render settings, export/import, file IO, bpy.ops.wm.*.* Only bpy / bmesh / mathutils / math / random (seeded). No other imports.@@ -655,7 +648,7 @@ def part_file_source(p: PartPlan) -> str:def model_file_source(plan: StaticPlan) -> str:"""Complete multi-file entry ``src/model.py``: imports, ordered calls, self-check."""- imports = "\n".join(f"from {PARTS_PKG}.{to_snake(p.name)} import {build_fn_name(p.name)}" for p in plan.parts)+ imports = "\n".join(f"from {PARTS_DIR}.{to_snake(p.name)} import {build_fn_name(p.name)}" for p in plan.parts)calls = "\n".join(f" {build_fn_name(p.name)}()" for p in plan.parts)return (_plan_header(plan, multi_file=True)@@ -746,14 +739,6 @@ class BlenderRuntime(RuntimeDocs):def part_file(self, ws: Workspace, part_name: str) -> Path:return ws.root / self.file_for_part(part_name)- @staticmethod- def file_for_target(target: str) -> list[str]:- """Refine target → files: whole-object targets (``overall``/``assembly``/``object``/'')- map to the entry ``src/model.py``; anything else is treated as a part name."""- if target.strip().lower() in ("", "overall", "assembly", "object", "model"):- return [ENTRY_REL]- return [part_file_rel(target)]-def build_command(self, ws: Workspace, *, stl: bool = True, blend: bool = False, seed: int = 0,tri_limit: int = MAX_TRIS_OBJECT, rlimit_gb: float | None = None,@@ -796,7 +781,7 @@ class BlenderRuntime(RuntimeDocs):ws.stage_artifacts("build.json", "census.json", "object.glb", "object.stl", "object.blend").invalidate()entry = self.entry_file(ws)if not entry.is_file():- result = BuildResult(ok=False, language=self.language.value, error_type="MissingEntryFile",+ result = BuildResult(ok=False, language=self.language.value, error_type=MISSING_ENTRY,error_message=f"{ENTRY_REL} does not exist", error_file=ENTRY_REL)ws.write_json(build_json, result)return resultdiff --git a/harness/codeverse/languages/cadquery/__init__.py b/harness/codeverse/languages/cadquery/__init__.pyindex 980f75e1..41c4c230 100644--- a/harness/codeverse/languages/cadquery/__init__.py+++ b/harness/codeverse/languages/cadquery/__init__.py@@ -13,18 +13,18 @@ from codeverse.config import Settings, get_settingsfrom codeverse.contracts.artifacts import BuildResult, GateFinding, GateReport, Severityfrom codeverse.contracts.common import ENTRY_FILE, Languagefrom codeverse.contracts.plan import PartPlan, Plan, StaticPlan-from codeverse.conventions import MAX_TRIS_OBJECT, to_pascal, to_snake-from codeverse.languages._ast_lint import ( # noqa: F401 — dotted re-exported+from codeverse.conventions import MAX_TRIS_OBJECT, PASCAL_RE, to_pascal, to_snake+from codeverse.languages._ast_lint import (BASE_FORBIDDEN_IMPORTS,check_imports,describe_parse_failure,dotted,safe_parse,)-from codeverse.languages._common import compose_build_result, run_subprocess+from codeverse.languages._common import MISSING_ENTRY, compose_build_resultfrom codeverse.languages._docs import RuntimeDocsfrom codeverse.languages.blender import finish_for, instance_centers-from codeverse.proc import scrub_secrets+from codeverse.proc import run_subprocess, scrub_secretsfrom codeverse.workspace import Workspace# ===================================================================== lint@@ -48,9 +48,6 @@ FORBIDDEN_CALLS: tuple[tuple[str, str], ...] = ()FORBIDDEN_METHODS = {"save": "no .save()/.export() on the assembly — the harness exports", "export": "no .export() — the harness exports","exportStep": "no export calls", "exportStl": "no export calls", "exportSvg": "no export calls"}-PASCAL_RE = re.compile(r"^[A-Z][a-z0-9]+(?:[A-Z][a-z0-9]*)*(?:_\d+)?$")--RADIAN_CONSTS = {"pi", "tau"}TO_RADIANS_FUNCS = {"radians", "deg2rad"}TO_DEGREES_FUNCS = {"degrees", "rad2deg"}@@ -373,7 +370,7 @@ class CadQueryRuntime(RuntimeDocs):ws.stage_artifacts("build.json", "census.json", "object.glb", "object.step", "object.stl").invalidate()entry = self.entry_file(ws)if not entry.is_file():- result = BuildResult(ok=False, language=self.language.value, error_type="MissingEntryFile",+ result = BuildResult(ok=False, language=self.language.value, error_type=MISSING_ENTRY,error_message="src/model.py does not exist", error_file="src/model.py")ws.write_json(build_json, result)return resultdiff --git a/harness/codeverse/languages/glsl_shader/__init__.py b/harness/codeverse/languages/glsl_shader/__init__.pyindex 2aed803e..18aefd7a 100644--- a/harness/codeverse/languages/glsl_shader/__init__.py+++ b/harness/codeverse/languages/glsl_shader/__init__.py@@ -10,6 +10,7 @@ from pathlib import Pathfrom codeverse.contracts.artifacts import BuildResult, GateFinding, GateReport, Severityfrom codeverse.contracts.common import ENTRY_FILE, Languagefrom codeverse.contracts.plan import GraphicsPlan, Plan+from codeverse.languages._common import MISSING_ENTRYfrom codeverse.languages._docs import RuntimeDocsfrom codeverse.languages._gl_common import ( # noqa: F401 — re-exportedGlslMessage,@@ -379,12 +380,6 @@ def write_skeleton(ws: Workspace, plan: Plan | None) -> list[Path]:# ===================================================================== runtime-CONTRACT_FALLBACK = """src/shader.frag — GLSL 330 fragment shader body (NO #version line, NO uniform declarations):-write `void mainImage(out vec4 fragColor, in vec2 fragCoord)` using the harness uniforms u_time, u_resolution,-u_mouse, u_frame, u_prev (previous frame), u_noise (256² noise), u_buffer_a (src/buffer_a.frag output).-Optional src/common.glsl (helpers, pasted in first) and src/buffer_a.frag (one feedback pass). Animate with u_time."""--class GlslShaderRuntime(RuntimeDocs):language = Language.GLSL_SHADERentry_globs: tuple[str, ...] = (ENTRY_FILE[Language.GLSL_SHADER], COMMON, BUFFER_A)@@ -392,8 +387,6 @@ class GlslShaderRuntime(RuntimeDocs):def __init__(self, *, host: GlHost | None = None):self._host = host- # ------------------------------------------------------------------ contract-# ------------------------------------------------------------------ skeleton / lintdef skeleton(self, ws: Workspace, plan: Plan | None) -> list[Path]:return write_skeleton(ws, plan)@@ -420,9 +413,9 @@ class GlslShaderRuntime(RuntimeDocs):def build(self, ws: Workspace, *, timeout_s: int | None = None, times: list[float] | None = None,preview: bool = True, width: int | None = None, height: int | None = None) -> BuildResult:ws.artifacts.mkdir(parents=True, exist_ok=True)- invalidate_stale_outputs(ws) # BEFORE the MissingEntry return, so it also clears+ invalidate_stale_outputs(ws) # BEFORE the MISSING_ENTRY return, so it also clearsif not (ws.root / SHADER).is_file():- res = GlResult(ok=False, mode="shader", stage="lint", error_type="MissingEntry", error_message=f"{SHADER} is missing")+ res = GlResult(ok=False, mode="shader", stage="lint", error_type=MISSING_ENTRY, error_message=f"{SHADER} is missing")return finish_build(ws, res, language=self.language.value, error_file=SHADER)plan = load_plan(ws)w, h = resolution_for(plan)diff --git a/harness/codeverse/languages/opengl_python/__init__.py b/harness/codeverse/languages/opengl_python/__init__.pyindex e4516e51..f0208cf4 100644--- a/harness/codeverse/languages/opengl_python/__init__.py+++ b/harness/codeverse/languages/opengl_python/__init__.py@@ -15,6 +15,7 @@ from codeverse.languages._ast_lint import (describe_parse_failure,safe_parse,)+from codeverse.languages._common import MISSING_ENTRYfrom codeverse.languages._docs import RuntimeDocsfrom codeverse.languages._gl_common import (finish_build,@@ -336,12 +337,6 @@ def write_skeleton(ws: Workspace, plan: Plan | None) -> list[Path]:# ===================================================================== runtime-CONTRACT_FALLBACK = """src/program.py — raw moderngl program: `def setup(ctx, width, height) -> state` (programs, VAOs, textures,-FBOs; GLSL 330 core inline strings) and `def render(ctx, state, t, frame, fbo)` (draw the frame at time t into the given,-already bound framebuffer; multi-pass via your own FBOs, finish with fbo.use()). No window, no context creation, no clock,-no file IO (except src/*.glsl next to program.py). Imports: moderngl, numpy, math, random, struct, array, pathlib."""--class OpenGLPythonRuntime(RuntimeDocs):language = Language.OPENGL_PYTHONentry_globs: tuple[str, ...] = (ENTRY_FILE[Language.OPENGL_PYTHON], "src/*.glsl")@@ -349,8 +344,6 @@ class OpenGLPythonRuntime(RuntimeDocs):def __init__(self, *, host: GlHost | None = None):self._host = host- # ------------------------------------------------------------------ contract-# ------------------------------------------------------------------ skeleton / lintdef skeleton(self, ws: Workspace, plan: Plan | None) -> list[Path]:return write_skeleton(ws, plan)@@ -365,10 +358,10 @@ class OpenGLPythonRuntime(RuntimeDocs):def build(self, ws: Workspace, *, timeout_s: int | None = None, times: list[float] | None = None,preview: bool = True, width: int | None = None, height: int | None = None) -> BuildResult:ws.artifacts.mkdir(parents=True, exist_ok=True)- invalidate_stale_outputs(ws) # BEFORE the MissingEntry return, so it also clears+ invalidate_stale_outputs(ws) # BEFORE the MISSING_ENTRY return, so it also clearsprogram = ws.root / PROGRAMif not program.is_file():- res = GlResult(ok=False, mode="program", stage="lint", error_type="MissingEntry", error_message=f"{PROGRAM} is missing")+ res = GlResult(ok=False, mode="program", stage="lint", error_type=MISSING_ENTRY, error_message=f"{PROGRAM} is missing")return finish_build(ws, res, language=self.language.value, error_file=PROGRAM)plan = load_plan(ws)w, h = resolution_for(plan)diff --git a/harness/codeverse/languages/threejs/__init__.py b/harness/codeverse/languages/threejs/__init__.pyindex c65f26cf..341bb8da 100644--- a/harness/codeverse/languages/threejs/__init__.py+++ b/harness/codeverse/languages/threejs/__init__.py@@ -14,8 +14,13 @@ from codeverse.contracts.artifacts import BuildResult, GateFinding, GateReport,from codeverse.contracts.common import ENTRY_FILE, Languagefrom codeverse.contracts.plan import Plan, StaticPlanfrom codeverse.conventions import to_pascal, to_snake+from codeverse.languages._common import BUILD_TIMEOUTfrom codeverse.languages._docs import RuntimeDocs-from codeverse.languages._js_lint import ImportKind, ImportVerdict, check_imports+from codeverse.languages._js_lint import (+ ImportKind,+ ImportVerdict,+ check_imports,+)from codeverse.languages._js_lint import (node_check_syntax as check_syntax, # module-level name: tests monkeypatch it)@@ -87,43 +92,6 @@ export function build{pascal}(THREE_) {{PLACEHOLDER_COLORS = ("0x9aa5b1", "0xb08968", "0x7f8c8d", "0xc0a080", "0x6c7a89", "0xa0522d", "0x8fa3ad", "0xbfa27a")-# ===================================================================== contract-CONTRACT_FALLBACK = """\-# three.js static-object contract (raw ESM, no SDK)--Files (all under `src/`):-* `src/parts/<snake_name>.js` — one file per part:- `export function build<PascalName>(THREE) { ...; return group; }`- The returned `THREE.Group` is named `<PascalName>` and already sits at its WORLD pose- (no re-positioning happens in object.js).-* `src/object.js` — `export function build(THREE) { ... return root; }` imports every part- builder, adds each part group to one root `THREE.Group` whose `.name` is the object name,- and returns it. Optional: `root.userData.tick = (t, dt) => { ... }` for idle animation.--Frame & units: Y is UP, +Z is the FRONT (towards the default camera), +X is the right.-Meters. The object stands on the ground: lowest point at y = 0, footprint centred on-the Y axis. Real-world sizes (a chair seat is ~0.45 m high, not 45).--Imports allowed: `import * as THREE from 'three'`, `three/addons/...` (e.g.-`three/addons/utils/BufferGeometryUtils.js`, `three/addons/geometries/RoundedBoxGeometry.js`),-and relative `./` files inside `src/`. Nothing else: no CDN/http URLs, no npm packages,-no `fetch`, no `document`/`window`, no WebGLRenderer/cameras/lights/scenes in object code —-the harness renders and exports for you.--Geometry & materials: BufferGeometry only; `MeshStandardMaterial`/`MeshPhysicalMaterial`-with explicit `color`, `roughness`, `metalness`. Textures are NOT exportable from node —-use materials and vertex detail. Every mesh gets a meaningful `.name`. Keep total-triangles well under 600k. Parts must physically touch (no floating pieces) and must not-interpenetrate visibly.--Build: the harness runs `build(THREE)` in node, validates the group (≥1 mesh, finite-bbox, no NaN), calls an optional exported `selfcheck(THREE, root)` (a throw fails the-build with your message), bakes `InstancedMesh` copies into named meshes, and exports-`artifacts/object.glb` with one named node per part — exactly where you placed it (no-automatic drop-to-ground / re-centring; off-placement is warned and gated instead).-"""--# ===================================================================== lintGATE = "lint:threejs"@@ -375,8 +343,6 @@ class ThreeJsRuntime(RuntimeDocs):language = Language.THREEJSentry_globs: tuple[str, ...] = (ENTRY_FILE[Language.THREEJS], "src/parts/*.js")- # ------------------------------------------------------------------ contract-# ------------------------------------------------------------------ skeleton / lintdef skeleton(self, ws: Workspace, plan: Plan) -> list[Path]:return write_skeleton(ws, plan)@@ -413,7 +379,7 @@ class ThreeJsRuntime(RuntimeDocs):except NodeError as e:if e.result is None:raise # harness problem (no node binary / script)- result = self._failure(e.result, "Timeout" if e.result.timed_out else "NodeError", str(e))+ result = self._failure(e.result, BUILD_TIMEOUT if e.result.timed_out else "NodeError", str(e))self._write_build_json(ws, result)return resultdiff --git a/harness/codeverse/languages/threejs/wrappers/README.md b/harness/codeverse/languages/threejs/wrappers/README.mddeleted file mode 100644index 289eb0a6..00000000--- a/harness/codeverse/languages/threejs/wrappers/README.md+++ /dev/null@@ -1,5 +0,0 @@-The three.js build/export wrappers are node scripts and live in `runtime_js/`-(`export_glb.mjs`, `render_glb.mjs`, `lib/*.mjs`). They are invoked through-`codeverse.spatial.node.run_node` with the `--import lib/resolve_three.mjs`-hook so agent code can `import 'three'` without a node_modules of its own.-Agent code never imports anything from here.diff --git a/harness/codeverse/languages/urdf/__init__.py b/harness/codeverse/languages/urdf/__init__.pyindex b3d98c2c..d6fca341 100644--- a/harness/codeverse/languages/urdf/__init__.py+++ b/harness/codeverse/languages/urdf/__init__.py@@ -1,8 +1,5 @@-"""``urdf_blender``: bpy link meshes + hand-written URDF.--Modules: ``runtime`` (UrdfBlenderRuntime), ``lint``, ``skeleton``, ``consistency``-(FK ↔ authored geometry), ``wrappers/run_bpy_links.py`` (Blender-side build),-``wrappers/render_glb_bpy.py`` (fallback renderer). Contract: ``CONTRACT.md``."""+"""``urdf_blender``: bpy link meshes + hand-written URDF — lint, skeleton, FK ↔ geometry+consistency and the ``UrdfBlenderRuntime`` around ``wrappers/run_bpy_links.py``."""from __future__ import annotations@@ -25,9 +22,16 @@ from codeverse.contracts.artifacts import BuildResult, GateFinding, GateReport,from codeverse.contracts.common import ENTRY_FILE, Languagefrom codeverse.contracts.plan import ArticulatedPlan, JointPlan, PartPlan, Planfrom codeverse.conventions import to_snake-from codeverse.languages._ast_lint import describe_parse_failure, safe_parse-from codeverse.languages._common import ProcResult, read_json_file, run_subprocess+from codeverse.languages._ast_lint import (+ BASE_FORBIDDEN_IMPORTS,+ check_imports,+ describe_parse_failure,+ dotted,+ safe_parse,+)+from codeverse.languages._common import BUILD_TIMEOUT, ProcResult, read_json_filefrom codeverse.languages._docs import RuntimeDocs+from codeverse.proc import run_subprocess, tailfrom codeverse.spatial.joints import (UrdfError,load_urdf,@@ -117,7 +121,17 @@ FORBIDDEN_BPY_PREFIXES: tuple[tuple[str, str], ...] = (("bpy.ops.import_scene.", "importing external files — build geometry procedurally"),("bpy.ops.import_mesh.", "importing external files — build geometry procedurally"),)-FORBIDDEN_MODULES: tuple[str, ...] = ("subprocess", "socket", "urllib", "requests", "http", "shutil", "ctypes", "multiprocessing")+#: the shared floor (subprocess/network/pickle/threading/importlib …): the wrapper execs+#: model.py inside the same Blender interpreter as the blender language, so no looser list+FORBIDDEN_IMPORTS: frozenset[str] = frozenset(BASE_FORBIDDEN_IMPORTS)+#: model.py runs inside the SAME Blender python as blender/model.py, so the pure-computation+#: stdlib it allows is allowed here too; anything else is an "unexpected" WARN (os/sys stay out:+#: a link-mesh script has no business in the filesystem, and FORBIDDEN_CALLS catches the uses)+ALLOWED_IMPORTS: frozenset[str] = frozenset({+ "bpy", "bmesh", "mathutils", "math", "random", "numpy", "np", "bpy_extras", "__future__",+ "typing", "dataclasses", "itertools", "functools", "collections", "enum", "copy",+ "colorsys", "statistics", "operator",+})FORBIDDEN_CALLS: dict[str, str] = {"os.system": "shell access", "os.remove": "file deletion", "os.unlink": "file deletion", "os.rmdir": "file deletion","sys.exit": "exits Blender before the export — just return/raise instead", "exit": "exits Blender", "quit": "exits Blender",@@ -343,42 +357,20 @@ def _lint_joint(el: ET.Element, jname: str, link_names: list[str], parent_of: diclass _Visitor(ast.NodeVisitor):def __init__(self) -> None:self.findings: list[GateFinding] = []- self.has_bpy = False+ self.imports: dict[str, int] = {} # top-level module → first lineself.strings: set[str] = set()- def _chain(self, node: ast.AST) -> str:- parts: list[str] = []- while isinstance(node, ast.Attribute):- parts.append(node.attr)- node = node.value- if isinstance(node, ast.Name):- parts.append(node.id)- return ".".join(reversed(parts))-def visit_Import(self, node: ast.Import) -> None:for a in node.names:- top = a.name.split(".")[0]- if top == "bpy":- self.has_bpy = True- if top in FORBIDDEN_MODULES:- self.findings.append(_f(Severity.ERROR, f"line {node.lineno}: import of '{a.name}' is not allowed in model.py",- target=MODEL_REL, fix="Build geometry with bpy/bmesh/mathutils/math only.", line=node.lineno))+ self.imports.setdefault(a.name.split(".")[0], node.lineno)self.generic_visit(node)def visit_ImportFrom(self, node: ast.ImportFrom) -> None:- top = (node.module or "").split(".")[0]- if top == "bpy":- self.has_bpy = True- if top in FORBIDDEN_MODULES:- self.findings.append(_f(Severity.ERROR, f"line {node.lineno}: import from '{node.module}' is not allowed in model.py",- target=MODEL_REL, fix="Build geometry with bpy/bmesh/mathutils/math only.", line=node.lineno))- if top == "codeverse":- self.findings.append(_f(Severity.ERROR, f"line {node.lineno}: model.py must not import the harness", target=MODEL_REL,- fix="Raw bpy only.", line=node.lineno))+ self.imports.setdefault((node.module or "").split(".")[0], node.lineno)self.generic_visit(node)def visit_Call(self, node: ast.Call) -> None:- chain = self._chain(node.func)+ chain = dotted(node.func)for prefix, why in FORBIDDEN_BPY_PREFIXES:if chain.startswith(prefix):self.findings.append(_f(Severity.ERROR, f"line {node.lineno}: {chain}() is forbidden: {why}", target=MODEL_REL,@@ -412,7 +404,18 @@ def lint_model_text(text: str, link_names: list[str], *, label: str = MODEL_REL)v = _Visitor()v.visit(tree)out = v.findings- if not v.has_bpy:++ def _import_finding(kind: str, mod: str, line: int) -> GateFinding:+ if mod == "codeverse":+ return _f(Severity.ERROR, f"line {line}: model.py must not import the harness", target=label, fix="Raw bpy only.", line=line)+ if kind == "forbidden":+ return _f(Severity.ERROR, f"line {line}: import of '{mod}' is not allowed in model.py", target=label,+ fix="Build geometry with bpy/bmesh/mathutils/math only.", line=line)+ return _f(Severity.WARN, f"line {line}: import of '{mod}' is unexpected in model.py (outside the contract's list)",+ target=label, fix="Build geometry with bpy/bmesh/mathutils/math only.", line=line)++ out.extend(check_imports(v.imports, forbidden=FORBIDDEN_IMPORTS, allowed=ALLOWED_IMPORTS, make_finding=_import_finding))+ if "bpy" not in v.imports:out.append(_f(Severity.ERROR, f"{label} never imports bpy", target=label, fix="import bpy"))for link in link_names:if link not in v.strings and not any(link in s for s in v.strings):@@ -651,10 +654,6 @@ FK_TOL_M = 0.001STAGED_OUTPUTS = ("build.json", "census.json", "meshes", "robot.urdf", "articulation.json", "object.glb")-def _tail(s: str, n: int = 3000) -> str:- return s[-n:] if s else ""--class UrdfBlenderRuntime(RuntimeDocs):language = Language.URDF_BLENDERentry_globs = (ENTRY_FILE[Language.URDF_BLENDER], "src/robot.urdf")@@ -723,11 +722,11 @@ class UrdfBlenderRuntime(RuntimeDocs):build_json, census_json = stage.path("build.json"), stage.path("census.json")proc = _run_blender(blender, ws, stage.staging_dir, timeout_s, settings.limits.bpy_rlimit_gb)if proc.timed_out:- return fail("Timeout", f"Blender build exceeded {timeout_s}s (killed)", file="src/model.py", census=census,- stdout_tail=_tail(proc.stdout), stderr_tail=_tail(proc.stderr))+ return fail(BUILD_TIMEOUT, f"Blender build exceeded {timeout_s}s (killed)", file="src/model.py", census=census,+ stdout_tail=tail(proc.stdout), stderr_tail=tail(proc.stderr))if not build_json.is_file():return fail("WrapperCrash", f"wrapper produced no build.json (exit {proc.returncode})", file="src/model.py",- census=census, stdout_tail=_tail(proc.stdout), stderr_tail=_tail(proc.stderr))+ census=census, stdout_tail=tail(proc.stdout), stderr_tail=tail(proc.stderr))wb = read_json_file(build_json)wcensus = read_json_file(census_json) if census_json.is_file() else {}census.update({k: wcensus.get(k) for k in ("objects", "links", "unmatched_objects", "missing_links", "hints") if k in wcensus})@@ -735,8 +734,8 @@ class UrdfBlenderRuntime(RuntimeDocs):hints = "\n".join(f" hint: {h}" for h in (wcensus.get("hints") or {}).values())return fail(wb.get("error_type") or "ScriptError", (wb.get("error_message") or "") + ("\n" + hints if hints else ""),file=wb.get("error_file") or "src/model.py", line=wb.get("error_line"), census=census,- stdout_tail=_tail(wb.get("stdout_tail", "") or proc.stdout),- stderr_tail=_tail(wb.get("stderr_tail", "") or proc.stderr))+ stdout_tail=tail(wb.get("stdout_tail", "") or proc.stdout),+ stderr_tail=tail(wb.get("stderr_tail", "") or proc.stderr))# 3. URDF copy + load (staged files; extra_paths name the canonical homes)urdf_staged = stage.path("robot.urdf")@@ -770,7 +769,7 @@ class UrdfBlenderRuntime(RuntimeDocs):urdf_to_glb(robot, stage.path("object.glb"), None)extra["object_glb"] = str(art / "object.glb")res = BuildResult(ok=True, language=self.language.value, glb_path=str(art / "object.glb"), extra_paths=extra,- stdout_tail=_tail(wb.get("stdout_tail", "")), duration_ms=int((time.time() - t0) * 1000), census=census)+ stdout_tail=tail(wb.get("stdout_tail", "")), duration_ms=int((time.time() - t0) * 1000), census=census)if report.summary.rest_max_penetration_m > REST_PENETRATION_MAX_M:worst = [f for f in findings if f.data.get("pose") == {} and f.severity == "error"]res.ok = Falsediff --git a/harness/codeverse/models/anthropic.py b/harness/codeverse/models/anthropic.pyindex 96024018..25321e86 100644--- a/harness/codeverse/models/anthropic.py+++ b/harness/codeverse/models/anthropic.py@@ -38,7 +38,7 @@ from codeverse.models.parts import (with_logged_retries,)from codeverse.models.pricing import cache_write_surcharge, estimate_cost-from codeverse.models.retry import MAX_WAIT_S+from codeverse.models.retry import MAX_WAIT_S, cause_forfrom codeverse.models.schema_utils import (JsonParseError,parse_json_lenient,@@ -309,11 +309,13 @@ class AnthropicModel:try:return self._once(kwargs, request)except Exception as exc: # noqa: BLE001 - classified- raise classify_exception(exc) from exc+ err = classify_exception(exc)+ raise err from cause_for(err, exc)return with_logged_retries(attempt, label="anthropic", model=self.model,attempts=self.max_attempts, base_delay=self.base_delay,- max_delay=self.max_delay, sleep=self._sleep, log=log)+ max_delay=self.max_delay, sleep=self._sleep, log=log,+ max_wait_s=request.max_wait_s)def _once(self, kwargs: dict[str, Any], request: ChatRequest) -> ChatResponse:client = self.client()diff --git a/harness/codeverse/models/gemini.py b/harness/codeverse/models/gemini.pyindex 967162fb..f0c9bfea 100644--- a/harness/codeverse/models/gemini.py+++ b/harness/codeverse/models/gemini.py@@ -1,7 +1,13 @@-"""ChatRequest ⇄ google-genai types (contents, config, tools, response parsing).--Kept separate from ``gemini.py`` (which owns clients, keys, retries) so each-module stays small and the conversions are unit-testable offline.+"""The whole Gemini stack in one module (merged 2026-08-28, d512ccc):++* conversion — ``ChatRequest`` ⇄ google-genai types (``to_contents``, ``build_config``,+ ``extract_candidate``, ``parse_usage``); pure, unit-testable offline;+* ``GeminiModel`` — the ``ChatModel``: a shared ``KeyPool`` over the configured keys,+ ``rotate_with_retries`` with the storm gate / hedge, streaming with stall detection,+ per-attempt ledger rows;+* ``GeminiImageModel`` — text(+reference images) → PIL images on the same pool and+ retry machine, priced per generated image (``pricing.per_image_usd``), with a+ fallback model when the primary is missing."""from __future__ import annotations@@ -36,7 +42,7 @@ from codeverse.contracts.chat import (from codeverse.contracts.common import Usagefrom codeverse.models.base import ModelErrorfrom codeverse.models.parts import BoundedCache, Stopwatch, image_bytes-from codeverse.models.pricing import estimate_cost+from codeverse.models.pricing import estimate_cost, per_image_usdfrom codeverse.models.retry import (MAX_WAIT_S,RETRY_DEADLINE_S,@@ -231,6 +237,20 @@ def parse_usage(resp: types.GenerateContentResponse, model: str) -> Usage:)+def _guard_candidates(resp: types.GenerateContentResponse, usage: Usage | None, what: str) -> Any:+ """The first candidate, or the two billed failures every Gemini reply can carry:+ a blocked prompt (non-retryable) and no candidates at all (retryable)."""+ pf = resp.prompt_feedback+ if pf is not None and pf.block_reason:+ raise ModelError(+ f"{what} prompt blocked: {pf.block_reason} {pf.block_reason_message or ''}".strip(),+ retryable=False, usage=usage,+ )+ if not resp.candidates:+ raise ModelError(f"{what} returned no candidates", retryable=True, usage=usage)+ return resp.candidates[0]++def extract_candidate(resp: types.GenerateContentResponse, usage: Usage | None = None) -> tuple[str, list[ToolCallPart], str]:@@ -238,15 +258,7 @@ def extract_candidate(prompts (non-retryable) or empty candidates (retryable); ``usage`` (what the calleralready parsed off ``resp``) rides on the error, because every one of thesefailures was billed."""- pf = resp.prompt_feedback- if pf is not None and pf.block_reason:- raise ModelError(- f"prompt blocked by Gemini: {pf.block_reason} {pf.block_reason_message or ''}".strip(),- retryable=False, usage=usage,- )- if not resp.candidates:- raise ModelError("Gemini returned no candidates", retryable=True, usage=usage)- cand = resp.candidates[0]+ cand = _guard_candidates(resp, usage, "Gemini")finish = str(cand.finish_reason.value if cand.finish_reason else "") or "UNKNOWN"texts: list[str] = []calls: list[ToolCallPart] = []@@ -362,6 +374,27 @@ def _default_keys() -> list[str]:return list(get_settings().gemini_api_keys)+def _keys_or_raise(keys: list[str] | None, pool: KeyPool | None) -> list[str]:+ keys = list(keys) if keys is not None else _default_keys()+ if pool is None and not keys:+ raise ModelError("no Gemini API keys configured (GEMINI_API_KEYS / ~/.config/astra3d/gemini_keys.env)")+ return keys+++def _client_for(key: str, timeout_s: float, factory: Callable[[str], Any] | None) -> Any:+ """One ``genai.Client`` per (key, timeout), process-wide; ``factory`` (tests) bypasses the cache."""+ if factory is not None:+ return factory(key)+ ck = (key, int(timeout_s * 1000))+ with _registry_lock:+ client = _clients.get(ck)+ if client is None:+ client = genai.Client(api_key=key, http_options=types.HttpOptions(+ timeout=ck[1], client_args=_ipv4_client_args() or None))+ _clients[ck] = client+ return client++def classify_exception(exc: BaseException) -> ModelError:"""Map provider / transport exceptions onto ``ModelError``."""if isinstance(exc, ModelError):@@ -406,15 +439,16 @@ def is_dead_key_error(err: ModelError) -> bool:def failure_outcome(err: ModelError) -> Outcome:- """``KeyPool.report`` outcome for a failed call (content-level failures such as- bad JSON / empty candidates carry no status and are not the key's fault)."""+ """``KeyPool.report`` outcome for a failed call. Content-level failures (bad JSON,+ empty candidates) carry no status and are not the key's fault: ``skip`` reconciles+ the tokens they were billed and leaves health / counters alone."""if err.status == 429:return "429"if (err.status or 0) >= 500:return "5xx"if is_dead_key_error(err):return "dead"- return "error" if err.status else "ok"+ return "error" if err.status else "skip"def _retry_after_s(exc: BaseException) -> float | None:@@ -496,7 +530,9 @@ def _merge_stream_chunks(class GeminiModel:- """ChatModel for ``gemini:<model>``. See module docstring."""+ """ChatModel for ``gemini:<model>``: a shared ``KeyPool``, ``rotate_with_retries``+ (storm gate, hedge, ``ChatRequest.max_wait_s`` deadline), streaming with stall+ detection and per-attempt ledger rows."""provider = "gemini"@@ -526,11 +562,7 @@ class GeminiModel:#: keys a retry is raced on after the call's first 503 (``Settings.rate.hedge``,#: 2; 1 = off) — see ``rotate_with_retries`` and docs/COST.md §27self.hedge = max(1, int(_rate().hedge if hedge is None else hedge))- keys = list(keys) if keys is not None else _default_keys()- if pool is None and not keys:- raise ModelError(- "no Gemini API keys configured (GEMINI_API_KEYS / ~/.config/astra3d/gemini_keys.env)"- )+ keys = _keys_or_raise(keys, pool)self.pool = pool or shared_pool(keys, rpm_per_key=rpm_per_key, tpm_per_key=tpm_per_key)self.storm_gate = storm_gate if storm_gate is not None else (storm_gate_for(f"gemini:{model}") if _rate().storm_gate else None@@ -552,19 +584,6 @@ class GeminiModel:def supports_vision(self) -> bool:return True- # ---------------------------------------------------------------- clients- def _client(self, key: str) -> Any:- if self._client_factory is not None:- return self._client_factory(key)- ck = (key, int(self.timeout_s * 1000))- with _registry_lock:- client = _clients.get(ck)- if client is None:- client = genai.Client(api_key=key, http_options=types.HttpOptions(- timeout=ck[1], client_args=_ipv4_client_args() or None))- _clients[ck] = client- return client-# --------------------------------------------------------------- generatedef generate(self, request: ChatRequest) -> ChatResponse:contents = to_contents(request.messages)@@ -680,7 +699,7 @@ class GeminiModel:request: ChatRequest,warnings: list[str],) -> ChatResponse:- client = self._client(key)+ client = _client_for(key, self.timeout_s, self._client_factory)if not _streaming_enabled():with Stopwatch() as sw:resp = client.models.generate_content(@@ -746,21 +765,11 @@ class GeminiModel:DEFAULT_IMAGE_MODEL = "gemini-3.1-flash-image"FALLBACK_IMAGE_MODEL = "gemini-2.5-flash-image"-#: approximate USD per generated 1024² image (floor for cost accounting)-IMAGE_USD: dict[str, float] = {- "gemini-3.1-flash-image": 0.067,- "gemini-2.5-flash-image": 0.039,-}--#: requested pixel size → Gemini ``image_size`` token+#: requested pixel size → Gemini ``image_size`` token; a 512 request GENERATES a 1K+#: image (resized locally), so it is billed as one — ``_GENERATED_PX`` is the size+#: ``pricing.per_image_usd`` is asked for_IMAGE_SIZE_TOKEN: dict[int, str] = {512: "1K", 1024: "1K", 2048: "2K", 4096: "4K"}---def image_cost(model: str, usage: Usage, n_images: int) -> float:- """Token-priced cost floored by the per-image price (approximate)."""- token_cost = estimate_cost("gemini", model, usage)- per_image = IMAGE_USD.get(model.strip().lower(), 0.0)- return max(token_cost, per_image * max(0, n_images))+_GENERATED_PX: dict[str, int] = {"1K": 1024, "2K": 2048, "4K": 4096}def _to_part(ref: Image.Image | Path | str) -> types.Part:@@ -768,11 +777,8 @@ def _to_part(ref: Image.Image | Path | str) -> types.Part:buf = io.BytesIO()ref.convert("RGB").save(buf, format="PNG")return types.Part.from_bytes(data=buf.getvalue(), mime_type="image/png")- p = Path(ref)- if not p.is_file():- raise ModelError(f"reference image not found: {p}")- mime = "image/jpeg" if p.suffix.lower() in (".jpg", ".jpeg") else "image/png"- return types.Part.from_bytes(data=p.read_bytes(), mime_type=mime)+ data, mime = image_bytes(ImagePart(path=str(ref)))+ return types.Part.from_bytes(data=data, mime_type=mime)def _is_model_missing(err: ModelError) -> bool:@@ -794,7 +800,9 @@ def _record(usage: Usage, n_images: int) -> None:class GeminiImageModel:- """See module docstring. ``model`` is the primary; ``fallback`` the second try."""+ """Text (+ reference images) → PIL images on the shared key pool and retry machine.+ ``model`` is the primary; ``fallback`` the second try once the primary is missing+ or fails retryably. Not a ``ChatModel``: it meters itself (``_record``)."""provider = "gemini"@@ -818,18 +826,15 @@ class GeminiImageModel:#: no 503 hedge by default: an image is billed per image, so a hedge that#: lands twice pays for two of them (the chat models default to 2)self.hedge = max(1, int(hedge))- keys = list(keys) if keys is not None else _default_keys()- if pool is None and not keys:- raise ModelError("no Gemini API keys configured (GEMINI_API_KEYS / ~/.config/astra3d/gemini_keys.env)")+ keys = _keys_or_raise(keys, pool)self.pool = pool or shared_pool(keys)+ self.storm_gate = storm_gate_for(f"gemini:{model}") if _rate().storm_gate else Noneself.timeout_s = timeout_sself.max_attempts = max(1, max_attempts)self.base_delay = base_delayself.max_delay = max_delayself._sleep = sleep- # reuse GeminiModel's cached clients (same key → same genai.Client)- self._clients = GeminiModel(model, pool=self.pool, timeout_s=timeout_s, client_factory=client_factory)- self.storm_gate = self._clients.storm_gate+ self._client_factory = client_factoryself._lock = threading.Lock()self._primary_dead = False@@ -838,20 +843,6 @@ class GeminiImageModel:return f"gemini-image:{self.model}"# ------------------------------------------------------------------ API- def generate(- self,- prompt: str,- *,- size: int = 1024,- n: int = 1,- seed: int | None = None,- reference_images: Sequence[Image.Image | Path | str] = (),- max_wait_s: float | None = None,- ) -> list[Image.Image]:- images, _ = self.generate_with_usage(prompt, size=size, n=n, seed=seed,- reference_images=reference_images, max_wait_s=max_wait_s)- return images-def generate_with_usage(self,prompt: str,@@ -865,8 +856,7 @@ class GeminiImageModel:"""``max_wait_s`` is the caller's retry budget for ONE image (each of the``n`` images gets its own window, and a fallback model too), exactly like``ChatRequest.max_wait_s`` clips ``GeminiModel.generate``; ``None`` = the- full ``retry.RETRY_DEADLINE_S`` (900 s). Kept off the ``ImageModel``- protocol so fakes stay conformant; callers that hold a deadline pass it."""+ full ``retry.RETRY_DEADLINE_S`` (1800 s)."""if not prompt.strip():raise ModelError("empty image prompt")if n < 1:@@ -922,9 +912,10 @@ class GeminiImageModel:parts: list[types.Part] = [_to_part(r) for r in refs]parts.append(types.Part.from_text(text=prompt))contents = [types.Content(role="user", parts=parts)]+ token = _IMAGE_SIZE_TOKEN.get(int(size), "1K")config = types.GenerateContentConfig(response_modalities=["IMAGE"],- image_config=types.ImageConfig(aspect_ratio="1:1", image_size=_IMAGE_SIZE_TOKEN.get(int(size), "1K")),+ image_config=types.ImageConfig(aspect_ratio="1:1", image_size=token),seed=seed,http_options=types.HttpOptions(timeout=int(self.timeout_s * 1000)),)@@ -942,7 +933,7 @@ class GeminiImageModel:return rotate_with_retries(self.pool,- lambda key: self._call(key, model, contents, _attempt_config(), size),+ lambda key: self._call(key, model, contents, _attempt_config(), size, _GENERATED_PX[token]),classify=classify_exception,outcome_of=failure_outcome,max_attempts=self.max_attempts,@@ -958,20 +949,16 @@ class GeminiImageModel:)def _call(- self, key: str, model: str, contents: list[types.Content], config: types.GenerateContentConfig, size: int+ self, key: str, model: str, contents: list[types.Content], config: types.GenerateContentConfig,+ size: int, generated_px: int,) -> tuple[list[Image.Image], Usage]:- client = self._clients._client(key)+ client = _client_for(key, self.timeout_s, self._client_factory)with Stopwatch() as sw:resp = client.models.generate_content(model=model, contents=contents, config=config)usage = parse_usage(resp, model) # parsed first: the raises below were billed toousage.backend = "gemini-image"usage.latency_ms = sw.ms- pf = resp.prompt_feedback- if pf is not None and pf.block_reason:- raise ModelError(f"image prompt blocked: {pf.block_reason}", retryable=False, usage=usage)- if not resp.candidates:- raise ModelError("image model returned no candidates", retryable=True, usage=usage)- cand = resp.candidates[0]+ cand = _guard_candidates(resp, usage, "image model")finish = str(cand.finish_reason.name if cand.finish_reason is not None else "")images: list[Image.Image] = []for part in (cand.content.parts if cand.content and cand.content.parts else []):@@ -986,7 +973,9 @@ class GeminiImageModel:retryable=False, usage=usage)raise ModelError(f"image model returned no image (finish_reason={finish or 'unknown'})",retryable=True, usage=usage)- usage.cost_usd = image_cost(model, usage, len(images))+ # token-priced cost floored by the per-image price of the size that was GENERATED+ usage.cost_usd = max(estimate_cost("gemini", model, usage),+ per_image_usd("gemini", model, size=generated_px) * len(images))if int(size) not in (0, 1024) and images[0].size != (size, size):images = [im.resize((int(size), int(size)), Image.LANCZOS) for im in images]return images, usagediff --git a/harness/codeverse/models/health.py b/harness/codeverse/models/health.pyindex b905298c..dbd942de 100644--- a/harness/codeverse/models/health.py+++ b/harness/codeverse/models/health.py@@ -72,7 +72,7 @@ def probe(sample: int = DEFAULT_SAMPLE,timeout_s: float = DEFAULT_TIMEOUT_S,) -> Health:- """One tiny call per sampled key, in parallel, no retries. Never raises."""+ """One workload-sized call (``PROBE_TOKENS``) per sampled key, in parallel, no retries. Never raises."""from codeverse.contracts.chat import ChatMessage, ChatRequestprompt = _probe_prompt()@@ -121,9 +121,13 @@ def _bare_model(model_id: str, timeout_s: float):from codeverse.models.gemini import GeminiModelreturn GeminiModel(name, timeout_s=timeout_s, max_attempts=1, storm_attempts=0)- from codeverse.models import get_chat_model # other providers: no bare constructor needed yet+ if provider == "anthropic":+ from codeverse.models.anthropic import AnthropicModel- return get_chat_model(model_id)+ return AnthropicModel(name, timeout_s=timeout_s, max_attempts=1)+ from codeverse.models.openai import OpenAIModel # everything else, as models.registry does++ return OpenAIModel(name, timeout_s=timeout_s, max_attempts=1)# --------------------------------------------------------------------------- siblingsdiff --git a/harness/codeverse/models/openai.py b/harness/codeverse/models/openai.pyindex 7143eefe..508f0e23 100644--- a/harness/codeverse/models/openai.py+++ b/harness/codeverse/models/openai.py@@ -28,7 +28,7 @@ from codeverse.contracts.common import Usagefrom codeverse.models.base import ModelErrorfrom codeverse.models.parts import Stopwatch, classify_sdk_exception, image_b64, with_logged_retriesfrom codeverse.models.pricing import estimate_cost-from codeverse.models.retry import MAX_WAIT_S+from codeverse.models.retry import MAX_WAIT_S, cause_forfrom codeverse.models.schema_utils import (JsonParseError,inline_refs,@@ -290,12 +290,14 @@ class OpenAIModel:try:return self._once(kwargs, request)except Exception as exc2: # noqa: BLE001- raise classify_exception(exc2) from exc2- raise err from exc+ err2 = classify_exception(exc2)+ raise err2 from cause_for(err2, exc2)+ raise err from cause_for(err, exc)return with_logged_retries(attempt, label="openai", model=self.model,attempts=self.max_attempts, base_delay=self.base_delay,- max_delay=self.max_delay, sleep=self._sleep, log=log)+ max_delay=self.max_delay, sleep=self._sleep, log=log,+ max_wait_s=request.max_wait_s)def _once(self, kwargs: dict[str, Any], request: ChatRequest) -> ChatResponse:client = self.client()diff --git a/harness/codeverse/models/parts.py b/harness/codeverse/models/parts.pyindex 097958d0..87458c3e 100644--- a/harness/codeverse/models/parts.py+++ b/harness/codeverse/models/parts.py@@ -65,15 +65,10 @@ class Stopwatch:class BoundedCache[T]:- """Thread-safe bounded LRU ``str -> T`` map for per-call-id provider state.-- Replaces ``gemini_convert._SignatureCache`` (thought signatures, ``bytes``)- and ``anthropic_convert._ThinkingCache`` (thinking blocks, ``list[dict]``):- both had the identical ``__init__``/``put``/``get`` — an ``OrderedDict``- under a lock, ``put`` skipping falsy values, ``move_to_end`` then- ``popitem(last=False)`` until under capacity — and differed only in value- type and capacity. ``get`` returns ``None`` on a miss; a caller that needs- the old copy/empty-default shape wraps it: ``list(cache.get(k) or [])``.+ """Thread-safe bounded LRU ``str -> T`` map for per-call-id provider state+ (gemini thought signatures, anthropic thinking blocks). ``put`` skips falsy+ values; ``get`` returns ``None`` on a miss — a caller that needs a copy /+ empty default wraps it: ``list(cache.get(k) or [])``."""def __init__(self, capacity: int) -> None:@@ -108,8 +103,6 @@ def classify_sdk_exception(exc: BaseException, sdk: Any, label: str,out" and "connection error" to tell a provider outage from a model failure, so thef"{label} …" forms below must stay byte-identical to what each adapter emitted."""- from codeverse.models.base import ModelError-if isinstance(exc, ModelError):return excif isinstance(exc, sdk.APIStatusError):@@ -126,19 +119,21 @@ def classify_sdk_exception(exc: BaseException, sdk: Any, label: str,def with_logged_retries(attempt: Any, *, label: str, model: str, attempts: int,- base_delay: float, max_delay: float, sleep: Any, log: Any) -> Any:+ base_delay: float, max_delay: float, sleep: Any, log: Any,+ max_wait_s: float | None = None) -> Any:"""``with_retries`` plus the one log line both SDK adapters write.- ChatRequest.max_wait_s is not honoured here: with_retries has no deadline, and its- 6 attempts x <= 5 s backoff bound one call to ~20 s of waiting plus the round-trips.+ ``max_wait_s`` is the request's ``ChatRequest.max_wait_s`` deadline (retries and+ their waits included; ``None`` = ``RETRY_DEADLINE_S``); the backoff itself is+ ``attempts`` x <= ``MAX_WAIT_S`` (3 s), the round-trips are the real cost."""- from codeverse.models.base import ModelError- from codeverse.models.retry import with_retries+ from codeverse.models.retry import RETRY_DEADLINE_S, with_retriesdef on_retry(n: int, exc: BaseException, delay: float) -> None:log.warning("%s %s attempt %d/%d failed (%s); retrying in %.1fs",label, model, n, attempts, exc, delay)+ budget = RETRY_DEADLINE_S if max_wait_s is None else min(RETRY_DEADLINE_S, float(max_wait_s))return with_retries(attempt, is_retryable=lambda e: isinstance(e, ModelError) and e.retryable,- attempts=attempts, base_delay=base_delay, max_delay=max_delay,+ attempts=attempts, base_delay=base_delay, max_delay=max_delay, max_total_s=budget,on_retry=on_retry, sleep=sleep)diff --git a/harness/codeverse/models/retry.py b/harness/codeverse/models/retry.pyindex 90117509..763f94d7 100644--- a/harness/codeverse/models/retry.py+++ b/harness/codeverse/models/retry.py@@ -1,6 +1,17 @@-"""Multi-key pool with per-key rate limiting, 429 cooldown and a health score.+"""The scheduling machine under every model call (merged 2026-08-28, d512ccc):-Used by ``GeminiModel`` (22 keys on the owner's box) but provider-neutral.+1. ``KeyPool`` (+ ``TokenBucket``): a multi-key pool with per-key RPM/TPM buckets,+ 429 cooldown and a health score — used by ``GeminiModel`` (22 keys on the owner's+ box) but provider-neutral;+2. ``StormGate`` + ``storm_gate()`` / ``all_gates()``: the process-wide 503 back-pressure+ one model's callers share (ships OFF, ``Settings.rate.storm_gate``; docs/COST.md §21);+3. the two retry loops — ``with_retries`` (single key, SDK adapters) and+ ``rotate_with_retries`` (the pool, the gate, hedging, the ``RETRY_DEADLINE_S`` deadline);+4. ``request_parts`` / ``request_tokens``: the prompt-token estimate the pool reserves.++Every wait anywhere in here clips to ``MAX_WAIT_S``.++The pool:* ``acquire()`` picks the next healthy key round-robin, honouring per-keyRPM / TPM token buckets and 429 cool-downs; it blocks (bounded) when every@@ -11,8 +22,10 @@ Used by ``GeminiModel`` (22 keys on the owner's box) but provider-neutral.may be out at once, and :meth:`KeyPool.release` (a ``finally`` in``rotate_with_retries``) hands the slot back. ``try_acquire()`` is thenever-waiting variant a hedged retry uses for its extra keys.-* ``report(key, outcome)`` feeds back ``ok | 429 | 5xx | error | dead`` so the- pool can cool a key down and adjust its health score. ``dead`` is for+* ``report(key, outcome)`` feeds back ``ok | 429 | 5xx | error | dead | skip`` so the+ pool can cool a key down and adjust its health score. ``skip`` is a content+ failure the key did not cause (bad JSON, empty candidates): only the token+ reconciliation runs, health and counters are untouched. ``dead`` is forkey-scoped auth/permission failures (revoked / suspended / invalid key): thekey is benched for ``dead_cooldown_s`` (default one hour) and re-probed oncethat elapses — a dead key must never keep failing its share of calls.@@ -38,7 +51,7 @@ from typing import TYPE_CHECKING, Any, Literalfrom codeverse.contracts.chat import ChatRequest, ImagePart, TextPart, ToolResultPart-Outcome = Literal["ok", "429", "5xx", "error", "dead"]+Outcome = Literal["ok", "429", "5xx", "error", "dead", "skip"]class KeyPoolExhausted(RuntimeError):@@ -352,7 +365,7 @@ class KeyPool:st.health = 0.0st.dead_until = max(st.dead_until, now + self._dead_cooldown_s)st.cooldown_until = max(st.cooldown_until, st.dead_until)- else:+ elif outcome != "skip":st.n_error += 1st.health *= 0.9if st.tpm is not None and (tokens or reserved):@@ -434,7 +447,7 @@ class KeyPool:# ===================================================================== stormclass StormGate:- """Shared 503 back-pressure for one model. See the module docstring."""+ """Shared 503 back-pressure for one model (section 2 of the module docstring)."""def __init__(self,@@ -590,7 +603,7 @@ log = logging.getLogger(__name__)OnRetry = Callable[[int, BaseException, float], None]-def _cause(err: BaseException, exc: BaseException | None) -> BaseException | None:+def cause_for(err: BaseException, exc: BaseException | None) -> BaseException | None:"""The cause to raise ``err`` from — ``None`` when it would be ``err`` itself.Every adapter's ``classify()`` returns an already-classified ``ModelError``@@ -630,29 +643,38 @@ def with_retries[T](attempts: int = 6,base_delay: float = 1.0,max_delay: float = MAX_WAIT_S,+ max_total_s: float | None = None,on_retry: OnRetry | None = None,sleep: Callable[[float], None] = time.sleep,jitter: bool = True,) -> T:- """Call ``fn`` up to ``attempts`` times.-- Re-raises the last exception when it is not retryable (per ``is_retryable``)- or when attempts are exhausted. ``on_retry(attempt, exc, delay)`` is called- before each sleep (attempt is the 1-based index of the attempt that failed).+ """Call ``fn`` up to ``attempts`` times, within ``max_total_s`` when given.++ Re-raises the last exception when it is not retryable (per ``is_retryable``),+ when attempts are exhausted, or when the next backoff would cross the deadline+ (``ChatRequest.max_wait_s`` on the SDK adapters). ``on_retry(attempt, exc, delay)``+ is called before each sleep (attempt is the 1-based index of the attempt that+ failed). A raised exception carrying an ``attempts`` attribute (``ModelError``)+ is stamped with the number of round-trips issued."""if attempts < 1:raise ValueError("attempts must be >= 1")+ deadline = None if max_total_s is None else time.monotonic() + max_total_slast: BaseException | None = Nonefor attempt in range(1, attempts + 1):try:return fn()except BaseException as exc: # noqa: BLE001 - classification is delegatedlast = exc+ if hasattr(exc, "attempts"):+ exc.attempts = attemptif attempt >= attempts or not is_retryable(exc):raisedelay = backoff_delay(attempt, base_delay=base_delay, max_delay=max_delay, jitter=jitter)+ if deadline is not None and time.monotonic() + delay >= deadline:+ raiseif on_retry is not None:on_retry(attempt, exc, delay)sleep(delay)@@ -712,12 +734,13 @@ def rotate_with_retries[T](call is it a* **capacity storm** (the whole pool is out of capacity, the backstop): that getsits own patience budget — up to ``storm_attempts`` waits with- backoff capped at ``storm_max_delay`` = ``MAX_WAIT_S`` per wait (the house rule), so- patience comes from the NUMBER of waits (60 x <=5 s ~ 5 min) rather than- from long sleeps that do NOT consume ``max_attempts``. Observed 2026-08-23: a+ backoff capped at ``storm_max_delay`` = ``MAX_WAIT_S`` (3 s) per wait (the house+ rule), so patience comes from the NUMBER of waits (60 x <=3 s) rather than from+ long sleeps that do NOT consume ``max_attempts`` — all inside the+ ``RETRY_DEADLINE_S`` / ``max_total_s`` deadline. Observed 2026-08-23: amulti-minute gemini-3.7-flash "high demand" outage killed 8 bench runs- under the plain 6-attempt budget. A ``storm_gate`` (see- :mod:`codeverse.models.retry`) shares that discovery across the process:+ under the plain 6-attempt budget. A ``storm_gate`` (:class:`StormGate`+ above) shares that discovery across the process:workers park at the gate instead of each spending a round-trip to learnthe model is out of capacity, and one probe at a time reopens it.* other retryable errors → exponential backoff + jitter until@@ -786,12 +809,9 @@ def rotate_with_retries[T](err = classify(exc)outcome = outcome_of(err)# a key that looks dead is reported "error" now and benched only once a- # sibling proves the request itself is fine (bench())- # a charged-but-invalid reply (bad JSON with real usage) DID consume- # its prompt tokens: report them so the pool does not refund a spent- # reservation (a 42k-token invalid reply used to hand the key 42k TPM- # back). Failures billed nothing (429 / 503 / transport) carry an- # empty ``ModelError.usage`` and are refunded exactly as before.+ # sibling proves the request itself is fine (bench()). ``tokens`` is what+ # the failure was billed: a 42k-token invalid reply ("skip") used to be+ # refunded as if it had never reached the model.consumed = getattr(err, "usage", None)pool.report(key,@@ -866,7 +886,7 @@ def rotate_with_retries[T](# sleeps: a free rotation must not out-live the caller's budgetlog.warning("%s giving up after %.0f s of retrying (%s)",label, max_total_s, last_err)- raise last_err from _cause(last_err, last_exc)+ raise last_err from cause_for(last_err, last_exc)attempt += 1# never go back to a key that looked dead this call; throttled keys are# excluded while an untried one remains, else acquire() waits for a cooldown@@ -880,7 +900,7 @@ def rotate_with_retries[T](log.warning("%s budget of %.0f s spent waiting at the storm gate", label, max_total_s)gate_err = last_err or classify(TimeoutError(f"{label}: retry budget spent waiting for capacity"))- raise gate_err from _cause(gate_err, last_exc)+ raise gate_err from cause_for(gate_err, last_exc)budget_left = None if deadline == float("inf") else max(0.0, deadline - monotonic())try:# the key/slot wait must fit the remaining budget, never outlive it@@ -933,11 +953,11 @@ def rotate_with_retries[T](if out_of_time():log.warning("%s giving up after %.0f s of retrying (%s)",label, max_total_s, err)- raise err from _cause(err, exc)+ raise err from cause_for(err, exc)log.warning("%s key …%s looks dead (%s); rotating", label, key[-4:], err)attempt -= 1continue- raise err from _cause(err, exc) # every key failed the same way: not the keys' fault+ raise err from cause_for(err, exc) # every key failed the same way: not the keys' faultif is_storm(err) and storm < storm_attempts and not out_of_time():failed_keys.update(t.key for t in tries)if len(failed_keys) < len(pool):@@ -978,7 +998,7 @@ def rotate_with_retries[T](if out_of_time():log.warning("%s giving up after %.0f s of retrying (%s)",label, max_total_s, err)- raise err from _cause(err, exc)+ raise err from cause_for(err, exc)# an untried key remains: rotation is free, only a courtesy pauseattempt -= 1log.warning(@@ -988,10 +1008,10 @@ def rotate_with_retries[T](continueif not err.retryable or attempt >= max_attempts:bench()- raise err from _cause(err, exc)+ raise err from cause_for(err, exc)if out_of_time():log.warning("%s giving up after %.0f s of retrying (%s)", label, max_total_s, err)- raise err from _cause(err, exc)+ raise err from cause_for(err, exc)delay = backoff_delay(attempt, base_delay=base_delay, max_delay=max_delay)log.warning("%s attempt %d/%d failed (%s); retrying in %.1fs",diff --git a/harness/codeverse/models/schema_utils.py b/harness/codeverse/models/schema_utils.pyindex 65c3eff7..cc296743 100644--- a/harness/codeverse/models/schema_utils.py+++ b/harness/codeverse/models/schema_utils.py@@ -1,4 +1,5 @@-"""JSON-schema adapters for the three providers + a fence-tolerant JSON parser.+"""JSON-schema adapters for the three providers, a fence-tolerant JSON parser and+``ask_structured`` (one schema-validated call → object | None, usage, error).Pydantic v2 emits ``$defs``/``$ref``, ``title``, ``default``, ``const``,``prefixItems`` (tuples), ``anyOf [.., {"type":"null"}]`` and so on. Each@@ -24,8 +25,15 @@ from __future__ import annotationsimport copyimport jsonimport re+from collections.abc import Sequencefrom typing import Any+from pydantic import BaseModel, ValidationError++from codeverse.contracts.chat import ChatMessage, ChatRequest, ImagePart+from codeverse.contracts.common import Usage+from codeverse.models.base import ModelError+# keys Gemini's Schema type rejects or ignores_GEMINI_DROP = {"title",@@ -355,3 +363,30 @@ def parse_json_lenient(text: str) -> Any:except json.JSONDecodeError:passraise JsonParseError(f"no JSON value found: {last_err} — head={s[:120]!r}")+++def ask_structured(model: Any, schema: type[BaseModel], *, system: str, text: str,+ images: Sequence[ImagePart] = (), temperature: float,+ label: str) -> tuple[Any, Usage, str]:+ """One structured call → ``(validated object | None, usage, error)``.++ Five callers (reference gate / image-prompt plan / mismatch diff, the texture+ material plan and the scene texture pack) built the same ChatRequest+ (thinking="low", the 65 536 ceiling, the 900 s wait), caught the same two failure+ families and parsed the same two ways. They differ only in what they RETURN on+ failure, which is why this hands the error back rather than raising or deciding.+ Not for callers that need the failed call's own ``Usage`` or separate parse+ errors from call errors (``judges.pairwise``, ``tracks.planner``).+ """+ req = ChatRequest(messages=[ChatMessage.user(text, images=list(images) or None)], system=system,+ response_schema=schema.model_json_schema(), temperature=temperature,+ thinking="low", max_output_tokens=65_536, max_wait_s=900.0, label=label)+ try:+ resp = model.generate(req)+ except ModelError as e:+ return None, Usage(), f"call failed: {e}"+ payload = resp.parsed if resp.parsed is not None else parse_json_lenient(resp.text)+ try:+ return schema.model_validate(payload), resp.usage, ""+ except (ValidationError, TypeError) as e:+ return None, resp.usage, f"answer unparsable: {e}"diff --git a/harness/codeverse/prompts/__init__.py b/harness/codeverse/prompts/__init__.pyindex ebdae4bb..ebcdef2e 100644--- a/harness/codeverse/prompts/__init__.py+++ b/harness/codeverse/prompts/__init__.py@@ -66,7 +66,3 @@ def render(rel_path: str, **ctx: Any) -> str:def prompt_hash(text: str) -> str:return hashlib.sha256(text.encode()).hexdigest()[:12]---def list_prompts() -> list[str]:- return sorted(str(p.relative_to(PROMPTS_DIR)) for p in PROMPTS_DIR.rglob("*") if p.suffix in (".md", ".j2", ".txt", ".py", ".js"))diff --git a/harness/codeverse/prompts/blender/system_v0.md b/harness/codeverse/prompts/blender/system_v0.mddeleted file mode 100644index 778aa3d1..00000000--- a/harness/codeverse/prompts/blender/system_v0.md+++ /dev/null@@ -1 +0,0 @@-You are an expert blender 3D modeller writing RAW code (no SDKs, no helper libraries). Follow the contract exactly; exact numbers beat adjectives.diff --git a/harness/codeverse/prompts/cadquery/contract.md b/harness/codeverse/prompts/cadquery/contract.mdindex 82f944b7..6cc44738 100644--- a/harness/codeverse/prompts/cadquery/contract.md+++ b/harness/codeverse/prompts/cadquery/contract.md@@ -29,7 +29,7 @@ No `OCP`/`OCC` direct calls, no numpy, no file I/O, no `cq.exporters`, no## Forbidden / limits* No `show_object`, `exporters.export`, `open(`, `os`, `sys`, `subprocess`, network.* ≤ 40 parts; each part a valid solid (`.val().isValid()`), no empty Workplanes; fillets- < 0.45 × thinnest adjacent wall; feature sizes ≥ 0.5 mm; build < 120 s.+ < 0.45 × thinnest adjacent wall; feature sizes ≥ 0.5 mm; build < 300 s.* Do not union parts across plan boundaries (keep one solid per part); small intentionaloverlap of **0.5–2 mm** between touching parts is required — that is how they "weld";deeper than 2 mm and the connectivity gate calls it interpenetration.diff --git a/harness/codeverse/prompts/cadquery/system_v0.md b/harness/codeverse/prompts/cadquery/system_v0.mddeleted file mode 100644index 00fbefdc..00000000--- a/harness/codeverse/prompts/cadquery/system_v0.md+++ /dev/null@@ -1 +0,0 @@-You are an expert cadquery 3D modeller writing RAW code (no SDKs, no helper libraries). Follow the contract exactly; exact numbers beat adjectives.diff --git a/harness/codeverse/prompts/glsl_shader/system_v0.md b/harness/codeverse/prompts/glsl_shader/system_v0.mddeleted file mode 100644index b031a4d8..00000000--- a/harness/codeverse/prompts/glsl_shader/system_v0.md+++ /dev/null@@ -1 +0,0 @@-You are an expert GLSL fragment-shader artist (Shadertoy style) writing RAW code for a headless harness. Follow the contract exactly: no #version/uniform redeclarations (shader) · no window/context creation (program) · everything animates with time. Render and LOOK at your frames before finishing.diff --git a/harness/codeverse/prompts/opengl_python/system_v0.md b/harness/codeverse/prompts/opengl_python/system_v0.mddeleted file mode 100644index 372c8502..00000000--- a/harness/codeverse/prompts/opengl_python/system_v0.md+++ /dev/null@@ -1 +0,0 @@-You are an expert raw OpenGL (moderngl) graphics programmer writing RAW code for a headless harness. Follow the contract exactly: no #version/uniform redeclarations (shader) · no window/context creation (program) · everything animates with time. Render and LOOK at your frames before finishing.diff --git a/harness/codeverse/prompts/scene_threejs/system_v0.md b/harness/codeverse/prompts/scene_threejs/system_v0.mddeleted file mode 100644index 776d0923..00000000--- a/harness/codeverse/prompts/scene_threejs/system_v0.md+++ /dev/null@@ -1 +0,0 @@-You are an expert three.js + GLSL graphics programmer writing RAW ESM modules for a multi-file scene. No SDKs, no DOM, no fetch, no CDN imports: `import * as THREE from 'three'` only. Exact numbers beat adjectives.diff --git a/harness/codeverse/prompts/threejs/system_v0.md b/harness/codeverse/prompts/threejs/system_v0.mddeleted file mode 100644index f7645a7f..00000000--- a/harness/codeverse/prompts/threejs/system_v0.md+++ /dev/null@@ -1 +0,0 @@-You are an expert threejs 3D modeller writing RAW code (no SDKs, no helper libraries). Follow the contract exactly; exact numbers beat adjectives.diff --git a/harness/codeverse/prompts/urdf/system_v0.md b/harness/codeverse/prompts/urdf/system_v0.mddeleted file mode 100644index 6e8bd516..00000000--- a/harness/codeverse/prompts/urdf/system_v0.md+++ /dev/null@@ -1 +0,0 @@-You are an expert in Blender bpy and URDF writing RAW code (no SDKs). Links are bpy objects named exactly as the URDF links; joints are native URDF with origins/axes in the parent link frame. Exact numbers beat adjectives.diff --git a/harness/codeverse/reference.py b/harness/codeverse/reference.pyindex fcf9bd7e..2cc2f2b2 100644--- a/harness/codeverse/reference.py+++ b/harness/codeverse/reference.py@@ -2,13 +2,13 @@Today a prompt goes to a planner that has never SEEN the object. This modulesynthesizes a neutral studio product shot of the brief with the image model,-**validates** it (:mod:`.gate`), and hands it to the rest of the harness as the+**validates** it (``check_plausible`` + ``decide``), and hands it to the rest of the harness as thefidelity anchor: the planner sees it, the generator's ``compare_reference`` toolsees it, the silhouette gate measures against it and the judge scores against it(:class:`codeverse.judges.vlm_judge.ReferenceJudge`).A synthesized reference is never ground truth and is marked as such everywhere-(:mod:`.attach`): the user's own ``--image`` wins, the brief's dimensions win,+(``attach`` / ``SYNTH_NOTE``): the user's own ``--image`` wins, the brief's dimensions win,and a picture that fails the gate is discarded rather than chased.Entry points::@@ -36,11 +36,10 @@ from typing import Any, Literalfrom pydantic import BaseModel, Field, ValidationErrorfrom codeverse.config import get_settings-from codeverse.contracts.chat import ChatMessage, ChatRequest, ImagePart+from codeverse.contracts.chat import ImagePartfrom codeverse.contracts.common import Usagefrom codeverse.contracts.spec import ReferenceImage, Spec-from codeverse.models.base import ModelError-from codeverse.models.schema_utils import parse_json_lenient+from codeverse.models.schema_utils import ask_structuredfrom codeverse.prompts import prompt_hashlog = logging.getLogger(__name__)@@ -71,7 +70,7 @@ class PlausibilityVerdict(BaseModel):"""One vision call's answer about ONE candidate reference image.``ok`` is computed in code from the booleans + ``contradictions`` (see- :func:`codeverse.reference.gate.decide`), never taken from the model.+ :func:`decide`), never taken from the model."""ok: bool = False@@ -121,7 +120,7 @@ class ReferenceView(BaseModel):class ReferenceSet(BaseModel):- """The result of :func:`codeverse.reference.synth.synth_reference`.+ """The result of :func:`synth_reference`.``accepted`` is what may be attached to a spec; ``rejected`` is kept only sothe run can record *why* the fallback to no-reference happened.@@ -308,7 +307,7 @@ VIEW_CLAUSE: dict[str, str] = {#: order views are requested in as ``n_views`` grows. The straight-on ``front``#: elevation comes second on purpose: it is the SILHOUETTE TARGET, and the harness-#: measures IoU against the run's ``front`` render (``tracks.reference.silhouette_gate``),+#: measures IoU against the run's ``front`` render (``tracks.static_object.silhouette_gate``),#: so the two cameras must agree. The 3/4 shot carries the part inventory.VIEW_ORDER: tuple[str, ...] = ("three_quarter", "front", "side", "back")@@ -442,8 +441,6 @@ def image_dir(key: str, *, cache_dir: Path | None = None) -> Path:#: relative aspect disagreement above which the picture's proportions are not a targetASPECT_TOL = 0.25-#: the measured silhouette criterion's score when the reference contradicts the brief-NEUTRAL_SCORE = 0.5#: brief keys that describe a HORIZONTAL extent of the object_WIDTH_KEYS = ("width", "length", "depth", "diameter")@@ -558,31 +555,6 @@ def decide(ans: GateAnswer, *, model_id: str = "") -> PlausibilityVerdict:)-def _ask(model: Any, schema: type[BaseModel], *, system: str, text: str,- images: list[ImagePart] | None = None, temperature: float,- label: str) -> tuple[Any, Usage, str]:- """One structured call → ``(validated object | None, usage, error)``.-- The three callers below built the same ChatRequest (thinking="low", the 65 536- ceiling, the 900 s wait), caught the same two failure families and parsed the same- two ways — each was a copy carried in from its own pre-merge file. They differ only- in what they RETURN on failure, which is why this hands the error back rather than- raising or deciding.- """- req = ChatRequest(messages=[ChatMessage.user(text, images=images)], system=system,- response_schema=schema.model_json_schema(), temperature=temperature,- thinking="low", max_output_tokens=65_536, max_wait_s=900.0, label=label)- try:- resp = model.generate(req)- except ModelError as e:- return None, Usage(), f"call failed: {e}"- payload = resp.parsed if resp.parsed is not None else parse_json_lenient(resp.text)- try:- return schema.model_validate(payload), resp.usage, ""- except (ValidationError, TypeError) as e:- return None, resp.usage, f"answer unparsable: {e}"--def check_plausible(image: Path | str,spec: Spec,@@ -597,7 +569,7 @@ def check_plausible(if not path.is_file():return PlausibilityVerdict(reason=f"image missing: {path.name}", model_id=model_id), Usage()text = GATE_USER.format(brief=brief_text(spec), constraints=visual_constraints(spec))- ans, usage, err = _ask(model, GateAnswer, system=GATE_SYSTEM, text=text,+ ans, usage, err = ask_structured(model, GateAnswer, system=GATE_SYSTEM, text=text,images=[ImagePart(path=str(path), label="CANDIDATE REFERENCE")],temperature=temperature, label="reference_gate")if err:@@ -698,7 +670,7 @@ def image_prompt_plan(spec: Spec, view_names: list[str], *, model: Any, temperatif model is None:return _fallback_plan(spec, view_names), Usage()text = IMAGE_PROMPT_USER.format(brief=brief_text(spec), n_views=len(view_names), views=", ".join(view_names))- plan, usage, err = _ask(model, ImagePromptPlan, system=IMAGE_PROMPT_SYSTEM, text=text,+ plan, usage, err = ask_structured(model, ImagePromptPlan, system=IMAGE_PROMPT_SYSTEM, text=text,temperature=temperature, label="reference_prompt")if err:log.warning("reference prompt writer %s; using the brief verbatim", err)@@ -812,8 +784,6 @@ def _publish(rs: ReferenceSet, out_dir: Path | None) -> ReferenceSet:# ===================================================================== mismatchMAX_MISMATCHES = 6-#: how many mismatches become refine tasks-TOP_TASKS = 3_SYNTH_LINE = ("The REFERENCE was SYNTHESIZED from the brief by an image model — it is a shape target, ""not ground truth. Where it disagrees with the brief, the BRIEF wins and it is not a mismatch.")@@ -860,7 +830,7 @@ def compare()images = [ImagePart(path=str(p), label=f"REFERENCE {i + 1}") for i, p in enumerate(refs)]images += [ImagePart(path=str(p), label=f"RENDER {i + 1} ({p.stem})") for i, p in enumerate(rens)]- ans, usage, err = _ask(model, DiffAnswer, system=DIFF_SYSTEM, text=text, images=images,+ ans, usage, err = ask_structured(model, DiffAnswer, system=DIFF_SYSTEM, text=text, images=images,temperature=temperature, label="reference_diff")diff.usage = usageif err:@@ -890,24 +860,6 @@ def _measured_block(diff: ReferenceDiff) -> str:return line + ". This is a fact; do not re-estimate it.\n"-def refine_tasks(diff: ReferenceDiff, *, top: int = TOP_TASKS) -> list[Any]:- """The top mismatches as priority-1 ``RefineTask``s (``source='gate'`` so the- round policy protects them from being trimmed as judge chatter)."""- from codeverse.orchestrator import RefineTask-- out = []- for m in diff.top(top):- out.append(RefineTask(- target=m.target or "overall",- kind="reference",- instruction=(f"Reference mismatch ({m.kind.replace('_', ' ')}, {m.severity}): {m.detail} "- f"Look at the reference image again and fix this specifically."),- priority=1,- source="gate",- ))- return out--# ===================================================================== runARTIFACT_SUBDIR = "reference"@@ -979,6 +931,6 @@ __all__ = ["SYNTH_NOTE", "SYNTH_TAG", "Mismatch", "PlausibilityVerdict", "ReferenceDiff", "ReferenceSet","ReferenceView", "attach", "check_plausible", "compare", "compose_image_prompt", "conflict_note","decide", "dimension_conflict", "expected_aspect_range",- "ground_spec", "has_user_references", "is_synthetic", "reference_images", "refine_tasks",+ "ground_spec", "has_user_references", "is_synthetic", "reference_images","synth_reference", "views_for",]diff --git a/harness/codeverse/texturing/__init__.py b/harness/codeverse/texturing/__init__.pyindex aad0e2a9..c465079e 100644--- a/harness/codeverse/texturing/__init__.py+++ b/harness/codeverse/texturing/__init__.py@@ -1,34 +1,3 @@"""Text-to-image texturing: a language-agnostic pass on the canonical GLB-(objects) and a scene texture pack (scenes). See ``run.texture_pass`` and-``scene_pack.scene_texture_pack``."""--from codeverse.texturing.apply import ApplyReport, apply_textures-from codeverse.texturing.generate import (- FakeImageModel,- GateResult,- SeamGateResult,- TextureAsset,- TextureSet,- generate_textures,- judge_gate,- make_tileable,- seam_gate,- seam_score,-)-from codeverse.texturing.plan import (- ScenePack,- TexturePart,- TexturePlan,- default_plan,- material_plan,- scene_texture_pack,- texture_pack_prompt,-)-from codeverse.texturing.run import TextureReport, texture_pass--__all__ = [- "ApplyReport", "apply_textures", "GateResult", "SeamGateResult", "judge_gate", "seam_gate",- "FakeImageModel", "TextureAsset", "TextureSet", "generate_textures", "TexturePart", "TexturePlan",- "default_plan", "material_plan", "TextureReport", "texture_pass", "ScenePack", "scene_texture_pack",- "texture_pack_prompt", "make_tileable", "seam_score",-]+(objects, ``run.texture_pass``) and a scene texture pack (scenes,+``plan.scene_texture_pack``). Import from the submodules."""diff --git a/harness/codeverse/texturing/apply.py b/harness/codeverse/texturing/apply.pyindex 4ebefd86..13ac186f 100644--- a/harness/codeverse/texturing/apply.py+++ b/harness/codeverse/texturing/apply.py@@ -354,8 +354,8 @@ def _material() -> trimesh.visual.material.PBRMaterial:"""PBR material for one texture id.- Beyond the albedo the material carries two maps DERIVED from that albedo (see- ``texturing.maps``): a metallic/roughness texture so the grain also modulates+ Beyond the albedo the material carries two maps DERIVED from that albedo+ (``derived_maps`` below): a metallic/roughness texture so the grain also modulatesthe specular lobe, and a normal map so it catches light in relief. They costno extra image call and cannot disagree with the colour. ``derived_maps=False``reproduces the albedo-only material the pass shipped before.@@ -489,21 +489,6 @@ def verify_textured_glb(glb_in: Path, glb_out: Path, *, expected_textured: int)return warnings-def textured_summary(glb: Path | str) -> dict[str, int]:- """Counts for reports / tests: geometries, textured geometries, distinct images."""- s = load_scene(glb)- textured = 0- imgs: set[int] = set()- for g in s.geometry.values():- vis = getattr(g, "visual", None)- mat = getattr(vis, "material", None)- tex = getattr(mat, "baseColorTexture", None)- if tex is not None:- textured += 1- imgs.add(id(tex))- return {"geometries": len(s.geometry), "textured": textured, "images": len(imgs)}--# ===================================================================== normalisediff --git a/harness/codeverse/texturing/generate.py b/harness/codeverse/texturing/generate.pyindex 021807f3..12860ca1 100644--- a/harness/codeverse/texturing/generate.py+++ b/harness/codeverse/texturing/generate.py@@ -6,8 +6,6 @@* ``make_tileable(img, border_frac)``: cross-fades each edge band with the mirroredopposite band (a smoothstep ramp over ``border_frac`` of the width) so the wrapseam disappears; content in the middle is untouched.-* ``offset_check(img)``: half-offset the image (seams move to the centre) — the- classic visual check; returns the offset image for inspection.* ``fit_size`` / ``save_texture``: power-of-two downscale + PNG or JPEG q90."""@@ -33,6 +31,8 @@ from codeverse.contracts.common import Usagefrom codeverse.contracts.plan import AcceptanceItem, StaticPlanfrom codeverse.contracts.spec import Specfrom codeverse.conventions import OBJECT_VIEWS_QUICK, ViewPreset+from codeverse.judges.base import JudgeInput, plan_digest+from codeverse.judges.rubrics import is_degraded#: seam score above which a texture is considered NOT tileable (after make_tileable)SEAM_MAX = 0.08@@ -94,14 +94,6 @@ def make_tileable(img: Image.Image, border_frac: float = 0.12) -> Image.Image:return Image.fromarray((np.clip(out, 0.0, 1.0) * 255.0 + 0.5).astype(np.uint8), "RGB")-def offset_check(img: Image.Image) -> Image.Image:- """Roll the image by half its size so the wrap seams sit in the centre."""- a = _arr(img)- h, w = a.shape[:2]- rolled = np.roll(np.roll(a, w // 2, axis=1), h // 2, axis=0)- return Image.fromarray((rolled * 255.0 + 0.5).astype(np.uint8), "RGB")--def fit_size(img: Image.Image, size: int) -> Image.Image:"""Square, power-of-two-ish downscale to ``size`` (never upscales)."""w, h = img.size@@ -124,16 +116,6 @@ def save_texture(img: Image.Image, path: Path, *, fmt: str | None = None, qualitreturn path-def tile_preview(img: Image.Image, reps: int = 2, size: int = 512) -> Image.Image:- """``reps × reps`` repetition for eyeballing tileability."""- t = img.convert("RGB").resize((size // reps, size // reps), Image.LANCZOS)- out = Image.new("RGB", (t.size[0] * reps, t.size[1] * reps))- for i in range(reps):- for j in range(reps):- out.paste(t, (i * t.size[0], j * t.size[1]))- return out--# ===================================================================== generatelog = logging.getLogger(__name__)@@ -195,10 +177,6 @@ class FakeImageModel:def id(self) -> str:return f"fake-image:{self.model}"- def generate(self, prompt: str, *, size: int = 1024, n: int = 1, seed: int | None = None,- reference_images: Sequence[Any] = ()) -> list[Image.Image]:- return self.generate_with_usage(prompt, size=size, n=n, seed=seed, reference_images=reference_images)[0]-def generate_with_usage(self, prompt: str, *, size: int = 1024, n: int = 1, seed: int | None = None,reference_images: Sequence[Any] = ()) -> tuple[list[Image.Image], Usage]:with self._lock:@@ -384,20 +362,6 @@ def material_criterion(scores: dict[str, float], rubric_hint: str = "") -> str:return ""-def _plan_summary(plan: StaticPlan | None) -> str:- if plan is None:- return ""- parts = ", ".join(f"{p.name}×{p.instances}" if p.instances > 1 else p.name for p in plan.parts)- e = plan.overall_bbox.extents- return f"{plan.object_name}: {plan.summary} Overall {e[0]:.2f}×{e[1]:.2f}×{e[2]:.2f} m. Parts: {parts}."---def _degraded(j: Judgment) -> bool:- from codeverse.judges.rubrics import is_degraded-- return is_degraded(j)--def judge_gate(spec: Spec,plan: StaticPlan | None,@@ -415,8 +379,6 @@ def judge_gate() -> GateResult:"""Render + judge both GLBs; decide. ``judge`` is any object with``.judge(JudgeInput) -> Judgment`` (``VlmJudge`` or a fake)."""- from codeverse.judges.base import JudgeInput-t0 = time.time()if render is None:from codeverse.spatial.render import render_glb@@ -426,7 +388,7 @@ def judge_gate(rs_before = render(glb_before, out_dir / "before", views=list(views), width=size, height=size)rs_after = render(glb_after, out_dir / "after", views=list(views), width=size, height=size)acceptance: list[AcceptanceItem] = list(getattr(plan, "acceptance", []) or [])- summary = _plan_summary(plan)+ summary = plan_digest(plan.model_dump()) if plan is not None else ""res = GateResult(shipped=False, renders_before=rs_before, renders_after=rs_after)jb = judge.judge(JudgeInput(spec=spec, renders=rs_before, measurement=measurement, acceptance=acceptance,plan_summary=summary, round_index=0,@@ -437,7 +399,7 @@ def judge_gate(res.judgment_before, res.judgment_after = jb, jares.usage = jb.usage + ja.usageres.duration_s = round(time.time() - t0, 2)- if _degraded(jb) or _degraded(ja):+ if is_degraded(jb) or is_degraded(ja):res.reason = "judge degraded on one side — not shipped"return resres.overall_before, res.overall_after = float(jb.overall), float(ja.overall)@@ -450,8 +412,6 @@ def judge_gate(res.materials_delta = round(res.materials_after - res.materials_before, 4)ok_overall = res.delta >= min_overall_deltaok_mat = (res.materials_delta is None) or (res.materials_delta > min_materials_delta)- if crit and res.materials_delta is None:- ok_mat = Falseres.shipped = bool(ok_overall and ok_mat)if res.shipped:res.reason = f"Δoverall {res.delta:+.3f} ≥ {min_overall_delta:+.2f} and {crit or 'overall'} {res.materials_delta:+.3f}" if crit \diff --git a/harness/codeverse/texturing/materials.py b/harness/codeverse/texturing/materials.pyindex 9f939574..69e8a99e 100644--- a/harness/codeverse/texturing/materials.py+++ b/harness/codeverse/texturing/materials.py@@ -34,7 +34,7 @@ from typing import NamedTuplefrom pydantic import BaseModel, Field#: glTF/Blender/three framework defaults. A material sitting exactly on one of-#: these pairs was never authored — see ``normalise.is_untouched``.+#: these pairs was never authored — see ``is_framework_default``.FRAMEWORK_DEFAULTS: tuple[tuple[float, float], ...] = ((1.0, 1.0), # glTF 2.0 material default (metallic 1, roughness 1) — renders as dark mud(0.0, 1.0), # three.js MeshStandardMaterial defaultdiff --git a/harness/codeverse/texturing/plan.py b/harness/codeverse/texturing/plan.pyindex 9b3d463e..07bc8055 100644--- a/harness/codeverse/texturing/plan.py+++ b/harness/codeverse/texturing/plan.py@@ -19,12 +19,12 @@ from pydantic import BaseModel, Field, ValidationErrorfrom codeverse.config import get_settingsfrom codeverse.contracts.artifacts import RenderSet-from codeverse.contracts.chat import ChatMessage, ChatRequest, ImagePart+from codeverse.contracts.chat import ImagePartfrom codeverse.contracts.common import Usagefrom codeverse.contracts.plan import PartPlan, ScenePlan, StaticPlanfrom codeverse.contracts.spec import Specfrom codeverse.conventions import to_snake-from codeverse.models.schema_utils import parse_json_lenient+from codeverse.models.schema_utils import ask_structuredfrom codeverse.prompts import load_text, prompt_hash, renderfrom codeverse.texturing.generate import TextureSet, generate_textures@@ -316,19 +316,13 @@ def material_plan(model = get_chat_model(model_id)text = render(PLAN_TEMPLATE, spec_prompt=spec.prompt, style_notes=getattr(plan, "style_notes", ""),n_parts=len(plan.parts), parts_table=parts_table(plan))- req = ChatRequest(- messages=[ChatMessage.user(text, images=[ImagePart(path=str(p), label=p.stem) for p in images])],- system="You plan PBR textures for 3D assets. Answer with JSON only.",- response_schema=PlannerOutput.model_json_schema(), temperature=temperature, thinking="low",- max_output_tokens=65_536, max_wait_s=900.0, label="texture_plan",+ out, usage, err = ask_structured(+ model, PlannerOutput, system="You plan PBR textures for 3D assets. Answer with JSON only.", text=text,+ images=[ImagePart(path=str(p), label=p.stem) for p in images], temperature=temperature, label="texture_plan",)- resp = model.generate(req)- payload = resp.parsed if resp.parsed is not None else parse_json_lenient(resp.text)- try:- out = PlannerOutput.model_validate(payload)- except ValidationError as e:- raise ValueError(f"texture planner returned invalid JSON for the schema: {e}") from e- tp = finalize_plan(out, plan, model_id=model_id, usage=resp.usage, source="vlm")+ if out is None:+ raise ValueError(f"texture planner: {err}")+ tp = finalize_plan(out, plan, model_id=model_id, usage=usage, source="vlm")if use_cache:cache_root.mkdir(parents=True, exist_ok=True)cached.write_text(tp.model_dump_json(indent=2))@@ -458,15 +452,10 @@ def scene_pack_plan(assets = "\n".join(f"- {a.name} ({a.kind}): {a.description}" for a in plan.assets) or "- (none)"text = render(PACK_TEMPLATE, n_min=n_min, n_max=n_max, title=plan.title, setting=plan.setting, mood=plan.mood,environment=plan.environment, zones=zones, assets=assets)- req = ChatRequest(messages=[ChatMessage.user(text)], system="You plan texture packs for 3D scenes. JSON only.",- response_schema=PackOutput.model_json_schema(), temperature=temperature, thinking="low",- max_output_tokens=65_536, max_wait_s=900.0, label="scene_texture_pack")- resp = model.generate(req)- payload = resp.parsed if resp.parsed is not None else parse_json_lenient(resp.text)- try:- out = PackOutput.model_validate(payload)- except ValidationError as e:- raise ValueError(f"scene pack planner returned invalid JSON: {e}") from e+ out, usage, err = ask_structured(model, PackOutput, system="You plan texture packs for 3D scenes. JSON only.",+ text=text, temperature=temperature, label="scene_texture_pack")+ if out is None:+ raise ValueError(f"scene pack planner: {err}")entries: list[PackEntry] = []seen: set[str] = set()for e in out.textures:@@ -477,7 +466,7 @@ def scene_pack_plan(entries.append(fe)if not entries:raise ValueError("scene pack planner returned no textures")- return entries[:n_max], out.notes, resp.usage+ return entries[:n_max], out.notes, usage# --------------------------------------------------------------------------- generationdiff --git a/harness/codeverse/texturing/run.py b/harness/codeverse/texturing/run.pyindex f5806f4b..c1733806 100644--- a/harness/codeverse/texturing/run.py+++ b/harness/codeverse/texturing/run.py@@ -292,15 +292,8 @@ def texture_pass(def _make_judge(spec: Spec, rubric: str | None, judge_model_id: str | None) -> Any:from codeverse.judges.vlm_judge import VlmJudge- info = TRACK_INFO.get(spec.track)- if rubric:- name = rubric- elif info is not None:- name = info.rubric- else:- name = "static_object_v1"- log.warning("no TRACK_INFO row for track %s; texture gate falls back to %s", spec.track, name)- return VlmJudge(rubric=name, model_id=judge_model_id or spec.backends.judge, n_samples=1, label="texture_gate")+ return VlmJudge(rubric=rubric or TRACK_INFO[spec.track].rubric, model_id=judge_model_id or spec.backends.judge,+ n_samples=1, label="texture_gate")def _finish(ws: Workspace, report: TextureReport, t0: float, events: EventLog, update_record: bool) -> TextureReport:diff --git a/harness/tests/blender_cadquery/test_artifact_stage.py b/harness/tests/blender_cadquery/test_artifact_stage.pyindex 102ac88d..ef160e49 100644--- a/harness/tests/blender_cadquery/test_artifact_stage.py+++ b/harness/tests/blender_cadquery/test_artifact_stage.py@@ -83,7 +83,7 @@ def test_promote_replaces_an_existing_dir(tmp_ws: Workspace) -> None:def test_invalidate_standalone_and_name_validation(tmp_ws: Workspace) -> None:_seed(tmp_ws)- tmp_ws.stage_artifacts("build.json", "object.glb", "meshes").invalidate() # the remove_stale replacement+ tmp_ws.stage_artifacts("build.json", "object.glb", "meshes").invalidate()for name in ("build.json", "object.glb", "meshes"):assert not (tmp_ws.artifacts / name).exists(), namestage = tmp_ws.stage_artifacts("build.json")diff --git a/harness/tests/blender_cadquery/test_blender_runtime.py b/harness/tests/blender_cadquery/test_blender_runtime.pyindex 9e8050f8..a1de41df 100644--- a/harness/tests/blender_cadquery/test_blender_runtime.py+++ b/harness/tests/blender_cadquery/test_blender_runtime.py@@ -45,7 +45,6 @@ def test_protocol_and_registry() -> None:assert rt.entry_globs == ("src/model.py", "src/parts/*.py") and rt.language.value == "blender"assert rt.file_for_part("Seat Cushion") == "src/parts/seat_cushion.py" # tracks call this via getattrassert "bpy" in rt.contract_doc() and "Z is up" in rt.contract_doc()- assert rt.cookbook_path().name == "cookbook.md"def test_missing_binary_raises(tmp_ws) -> None:@@ -55,12 +54,6 @@ def test_missing_binary_raises(tmp_ws) -> None:rt.build(tmp_ws)-def test_missing_entry_file(tmp_ws, tmp_path) -> None:- rt = BlenderRuntime(blender=_fake_blender(tmp_path))- r = rt.build(tmp_ws)- assert not r.ok and r.error_type == "MissingEntryFile"--def test_missing_entry_invalidates_previous_outputs(tmp_ws, tmp_path) -> None:"""Deleting model.py between rounds must not leave round N-1's object.glb +build.json (ok: true) looking current — the invalidation runs BEFORE thediff --git a/harness/tests/blender_cadquery/test_common.py b/harness/tests/blender_cadquery/test_common.pyindex beca7afe..a627da1c 100644--- a/harness/tests/blender_cadquery/test_common.py+++ b/harness/tests/blender_cadquery/test_common.py@@ -3,41 +3,14 @@from __future__ import annotationsimport json-import sys-import timefrom pathlib import Pathfrom codeverse.languages._common import (ProcResult,compose_build_result,- run_subprocess,strip_blender_noise,- tail,- write_json_atomic,)---def test_run_subprocess_captures_output(tmp_path: Path) -> None:- r = run_subprocess([sys.executable, "-c", "import sys; print('out'); print('err', file=sys.stderr); sys.exit(3)"],- cwd=tmp_path, timeout_s=10)- assert r.returncode == 3 and r.stdout.strip() == "out" and r.stderr.strip() == "err" and not r.timed_out---def test_run_subprocess_timeout_kills_group(tmp_path: Path) -> None:- code = "import subprocess, sys, time; p = subprocess.Popen([sys.executable, '-c', 'import time; time.sleep(30)']); print(p.pid, flush=True); time.sleep(30)"- t0 = time.monotonic()- r = run_subprocess([sys.executable, "-c", code], cwd=tmp_path, timeout_s=1.5)- assert r.timed_out and time.monotonic() - t0 < 6- child_pid = int(r.stdout.strip())- time.sleep(0.3)- assert not Path(f"/proc/{child_pid}").exists() or "zombie" in (Path(f"/proc/{child_pid}/status").read_text().lower())---def test_tail_limits() -> None:- text = "\n".join(str(i) for i in range(100))- t = tail(text, max_lines=5)- assert t.splitlines() == ["95", "96", "97", "98", "99"]- assert len(tail("x" * 10_000, max_chars=100)) == 100+from codeverse.proc import write_json_atomicdef test_strip_blender_noise() -> None:diff --git a/harness/tests/graphics/test_stale_invalidation.py b/harness/tests/graphics/test_stale_invalidation.pyindex bab0de96..1e3f42e3 100644--- a/harness/tests/graphics/test_stale_invalidation.py+++ b/harness/tests/graphics/test_stale_invalidation.py@@ -1,7 +1,7 @@"""GL builds must not leave the previous build's artifacts looking current (offline).frames_sheet.png / preview.gif / metrics.json are written only on ok and were never-cleared; the MissingEntry early returns never reached GlHost._run's frames wipe.+cleared; the MissingEntryFile early returns never reached GlHost._run's frames wipe.Both wipes now happen at the top of build() (``_gl_common.invalidate_stale_outputs``)plus a failure-path invalidation in ``finish_build``."""@@ -49,11 +49,11 @@ def test_failed_build_invalidates_sheet_gif_metrics(tmp_ws: Workspace) -> None:def test_missing_entry_clears_frames_and_result(tmp_ws: Workspace) -> None:- """The MissingEntry early return never reaches the host — the hoisted wipe must+ """The MissingEntryFile early return never reaches the host — the hoisted wipe muststill clear frames/ + gl_result.json (and the sheet/gif/metrics trio)."""_seed_stale(tmp_ws)br = GlslShaderRuntime(host=_FailingHost()).build(tmp_ws, preview=False)- assert not br.ok and br.error_type == "MissingEntry"+ assert not br.ok and br.error_type == "MissingEntryFile"assert not list((tmp_ws.artifacts / "frames").glob("*.png"))for name in ("gl_result.json", "frames_sheet.png", "preview.gif", "metrics.json"):assert not (tmp_ws.artifacts / name).exists(), namediff --git a/harness/tests/graphics/test_track.py b/harness/tests/graphics/test_track.pyindex 1e5e4c27..b0990888 100644--- a/harness/tests/graphics/test_track.py+++ b/harness/tests/graphics/test_track.py@@ -84,10 +84,7 @@ class FakeGlRuntime:return finish_build(ws, res, language="glsl_shader", census={"convention": "mainImage"})def contract_doc(self) -> str:- return "FAKE GLSL CONTRACT"-- def cookbook_path(self) -> Path:- return Path("/dev/null")+ return "FAKE glsl_shader authoring contract" # the runtime's contract_doc is what the prompt seesdef _writer(job, ws):diff --git a/harness/tests/judges/test_vlm_judge.py b/harness/tests/judges/test_vlm_judge.pyindex 420dae8e..8180a14f 100644--- a/harness/tests/judges/test_vlm_judge.py+++ b/harness/tests/judges/test_vlm_judge.py@@ -138,10 +138,10 @@ def test_partial_samples_still_score(judge_input, cache_dir):assert raw["n_requested"] == 2 and raw["n_used"] == 1 and len(raw["sample_errors"]) == 1-def test_non_retryable_error_stops_early(judge_input, cache_dir):+def test_non_retryable_error_stops_at_once(judge_input, cache_dir):model = FakeChatModel(default=ModelError("bad request", retryable=False))j = _judge(model, cache_dir=cache_dir).judge(judge_input)- assert is_degraded(j) and len(model.requests) == 2+ assert is_degraded(j) and len(model.requests) == 1def test_base_judge_rejects_measured_rubric(judge_input, cache_dir):diff --git a/harness/tests/languages/test_ast_lint.py b/harness/tests/languages/test_ast_lint.pyindex 079769ce..25df4179 100644--- a/harness/tests/languages/test_ast_lint.py+++ b/harness/tests/languages/test_ast_lint.py@@ -7,6 +7,7 @@ import astimport codeverse.languages.blender as blender_lintimport codeverse.languages.cadquery as cadquery_lintimport codeverse.languages.opengl_python as opengl_lint+import codeverse.languages.urdf as urdf_lintfrom codeverse.contracts.artifacts import GateFinding, Severityfrom codeverse.languages._ast_lint import BASE_FORBIDDEN_IMPORTS, check_imports, dotted@@ -24,6 +25,7 @@ def test_language_forbidden_sets_are_supersets_of_base() -> None:assert BASE_FORBIDDEN_IMPORTS <= blender_lint.FORBIDDEN_IMPORTSassert BASE_FORBIDDEN_IMPORTS <= cadquery_lint.FORBIDDEN_IMPORTSassert BASE_FORBIDDEN_IMPORTS <= opengl_lint.DANGEROUS_MODULES+ assert BASE_FORBIDDEN_IMPORTS <= urdf_lint.FORBIDDEN_IMPORTS # was a hand-copied subset until 2026-08-29def test_language_specific_extras_survive() -> None:@@ -51,6 +53,10 @@ def test_base_modules_error_in_every_language_lint() -> None:)assert any(f.severity == Severity.ERROR and "webbrowser" in f.message for f in findings)+ findings = urdf_lint.lint_model_text("import bpy\nimport webbrowser\nimport os\n", [])+ assert any(f.severity == Severity.ERROR and "webbrowser" in f.message for f in findings)+ assert any(f.severity == Severity.WARN and "'os'" in f.message for f in findings)+# ----------------------------------------------------------------- check_imports / dotteddiff --git a/harness/tests/models/test_gemini.py b/harness/tests/models/test_gemini.pyindex a4dbd17b..088dc15f 100644--- a/harness/tests/models/test_gemini.py+++ b/harness/tests/models/test_gemini.py@@ -28,15 +28,18 @@ from codeverse.models.gemini import (failure_outcome,to_contents,)-from codeverse.models.retry import MAX_WAIT_S, KeyPool+from codeverse.models.retry import KeyPoolPNG_1PX = base64.b64decode("iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNkYPhfDwAChwGA60e6kgAAAABJRU5ErkJggg==")#: the one tool spec every tool-call test uses.-MEASURE_TOOL = ToolSpec(name="measure", description="m",- parameters={"type": "object", "properties": {"part": {"type": "string"}}})+MEASURE_TOOL = ToolSpec(+ name="measure",+ description="m",+ parameters={"type": "object", "properties": {"part": {"type": "string"}}},+)def text_response(@@ -170,18 +173,12 @@ def test_structured_output_parsed_and_schema_sanitised():)-def test_bad_json_is_retried_then_raises():+def test_bad_json_is_retried():m, log, _ = make_model([text_response("not json"), text_response('{"ok": true}')])r = m.generate(ChatRequest(messages=[ChatMessage.user("x")], response_schema={"type": "object"}))assert r.parsed == {"ok": True} and len(log) == 2- m2, _, _ = make_model([text_response("nope")] * 6)- with pytest.raises(ModelError) as ei:- m2.generate(- ChatRequest(messages=[ChatMessage.user("x")], response_schema={"type": "object"})- )- assert ei.value.retryabledef test_429_rotates_keys_and_cools_down():@@ -209,38 +206,6 @@ def _api_error(code: int, message: str, status: str) -> genai_errors.APIError:return genai_errors.APIError(code, {"error": {"message": message, "status": status}})-class _KeyedClient:- """Client whose behaviour depends on the key: ``fail[key]`` raises, others answer."""-- def __init__(self, key: str, fail: dict[str, BaseException], log: list[dict]):- self.key, self.fail, self.log = key, fail, log- self.models = self-- def generate_content(self, *, model: str, contents, config):- self.log.append({"key": self.key})- if self.key in self.fail:- raise self.fail[self.key]- return text_response(f"ok from {self.key}")-- def generate_content_stream(self, *, model: str, contents, config):- yield self.generate_content(model=model, contents=contents, config=config)---def _keyed_model(- keys: list[str], fail: dict[str, BaseException], pool: KeyPool | None = None, **kw-):- log: list[dict] = []- pool = pool or KeyPool(keys, cooldown_s=30)- m = GeminiModel(- "gemini-3.7-flash",- pool=pool,- sleep=lambda s: None,- client_factory=lambda key: _KeyedClient(key, fail, log),- **kw,- )- return m, log, pool--SUSPENDED = _api_error(403, "PERMISSION_DENIED: Consumer 'api_key:xxx' has been suspended.", "PERMISSION_DENIED")@@ -256,87 +221,16 @@ THROTTLED = genai_errors.APIError()-def test_429_rotation_does_not_consume_the_retry_budget():- # 8 throttled keys > max_attempts=6, yet fresh keys remain → must still succeed- keys = [f"k{i}" for i in range(1, 23)]- m, log, pool = _keyed_model(keys, {k: THROTTLED for k in keys[:8]}, max_attempts=6)- r = m.generate(ChatRequest(messages=[ChatMessage.user("x")]))- assert r.text == "ok from k9" and [e["key"] for e in log] == keys[:9]- assert pool.stats()["n_cooling"] == 8---def test_429_on_every_key_waits_for_cooldown_then_counts_against_budget():- clock = {"t": 1000.0}- pool = KeyPool(- ["k1", "k2", "k3"],- clock=lambda: clock["t"],- sleep=lambda s: clock.__setitem__("t", clock["t"] + s),- )- keys = ["k1", "k2", "k3"]- m, log, _ = _keyed_model(keys, {k: THROTTLED for k in keys}, pool=pool, max_attempts=3)- with pytest.raises(ModelError) as ei:- m.generate(ChatRequest(messages=[ChatMessage.user("x")]))- assert ei.value.retryable and ei.value.status == 429- # 2 free rotations + 3 budgeted attempts; the pool waited out the cooldowns in between- assert len(log) == 5- # …but no single wait exceeded the house limit, even though the provider's- # own retryDelay in THROTTLED asks for 7 s (MAX_WAIT_S caps it).- assert 1000.0 < clock["t"] <= 1000.0 + 5 * MAX_WAIT_S---def test_no_single_wait_exceeds_the_house_limit():- """Every backoff path — 429 cooldown, provider retryDelay, 5xx, capacity- storm — sleeps at most MAX_WAIT_S at a time (owner rule, 2026-08-24)."""- # the sleep must ADVANCE the clock, like the neighbour test's: with a recording- # sleep and the real monotonic clock the cooldown loop never gets past "still- # cooling" and busy-spins — it logged 1.9 M naps over 3 s of wall time, of which- # 4 were real waits, so `max(naps)` was sampling the spin, not the backoff.- clock = {"t": 1000.0}- naps: list[float] = []-- def nap(s: float) -> None:- naps.append(s)- clock["t"] += s-- pool = KeyPool(["k1", "k2"], clock=lambda: clock["t"], sleep=nap)- m, _log, _ = _keyed_model(["k1", "k2"], {k: THROTTLED for k in ("k1", "k2")}, pool=pool, max_attempts=3)- m._sleep = nap- with pytest.raises(ModelError):- m.generate(ChatRequest(messages=[ChatMessage.user("x")]))- assert naps, "expected the throttled path to sleep at least once"- assert max(naps) <= MAX_WAIT_S, f"a wait exceeded {MAX_WAIT_S}s: {sorted(naps)[-3:]}"--def test_api_key_invalid_400_is_treated_as_dead_key():bad = _api_error(400, "API key not valid. Please pass a valid API key.", "INVALID_ARGUMENT")- m, log, pool = _keyed_model(["k1", "k2"], {"k1": bad})- assert m.generate(ChatRequest(messages=[ChatMessage.user("x")])).text == "ok from k2"+ m, log, pool = make_model([bad, text_response("ok")], keys=("k1", "k2"))+ assert m.generate(ChatRequest(messages=[ChatMessage.user("x")])).text == "ok"+ assert [event["key"] for event in log] == ["k1", "k2"]assert pool.stats()["dead"] == 1assert failure_outcome(classify_exception(bad)) == "dead"assert failure_outcome(classify_exception(SUSPENDED)) == "dead"assert failure_outcome(classify_exception(THROTTLED)) == "429"- assert failure_outcome(ModelError("bad json", retryable=True)) == "ok"---def test_5xx_retries_then_gives_up_with_retryable_error():- # 500 = plain 5xx (no storm budget): the classic 6-attempt backoff applies- err = genai_errors.APIError(500, {"error": {"message": "internal", "status": "INTERNAL"}})- slept: list[float] = []- m, log, _ = make_model([err] * 6)- m._sleep = slept.append- with pytest.raises(ModelError) as ei:- m.generate(ChatRequest(messages=[ChatMessage.user("x")]))- assert ei.value.retryable and ei.value.status == 500- assert len(log) == 6 and len(slept) == 5 and slept[1] > slept[0] / 2---def test_503_storm_survives_beyond_the_attempt_budget():- # 503 = capacity storm: its own patience budget on top of max_attempts- err = genai_errors.APIError(503, {"error": {"message": "high demand", "status": "UNAVAILABLE"}})- m, log, _ = make_model([err] * 8 + [text_response("recovered")])- m._sleep = lambda d: None- resp = m.generate(ChatRequest(messages=[ChatMessage.user("x")]))- assert resp.text == "recovered" and len(log) == 9+ assert failure_outcome(ModelError("bad json", retryable=True)) == "skip" # not the key's fault: no health nudgedef test_400_is_not_retried():@@ -499,31 +393,34 @@ def test_no_keys_is_loud():GeminiModel("gemini-3.7-flash", keys=[])-def test_truncated_json_is_not_retried():- m, log, _ = make_model(- [text_response('{"a": [1, 2', finish="MAX_TOKENS"), text_response("never")]+def test_max_token_failures_are_not_retried():+ cases = (+ (+ text_response('{"a": [1, 2', finish="MAX_TOKENS"),+ ChatRequest(messages=[ChatMessage.user("x")], response_schema={"type": "object"}),+ "max_output_tokens",+ ),+ (+ text_response("", finish="MAX_TOKENS", thoughts=49),+ ChatRequest(messages=[ChatMessage.user("x")], max_output_tokens=50),+ "thinking",+ ),)- with pytest.raises(ModelError) as ei:- m.generate(- ChatRequest(messages=[ChatMessage.user("x")], response_schema={"type": "object"})- )- assert not ei.value.retryable and len(log) == 1 and "max_output_tokens" in str(ei.value)+ for response, request, message in cases:+ model, log, _ = make_model([response, text_response("never")])+ with pytest.raises(ModelError) as error:+ model.generate(request)+ assert not error.value.retryable and len(log) == 1+ assert message in str(error.value)-def test_empty_max_tokens_is_not_retried():- m, log, _ = make_model(- [text_response("", finish="MAX_TOKENS", thoughts=49), text_response("never")]- )- with pytest.raises(ModelError) as ei:- m.generate(ChatRequest(messages=[ChatMessage.user("x")], max_output_tokens=50))- assert not ei.value.retryable and len(log) == 1 and "thinking" in str(ei.value)def test_a_charged_but_invalid_reply_carries_its_usage_on_the_error():- """``ModelError.usage``: what the provider billed for the failed call — a- bad-JSON reply is charged like a success (mirrors tracks.planner.PlanningError),- so the ledger's error row and the pool's TPM accounting can see real money."""+ """A failed structured reply preserves its billed token usage."""m, _log, _ = make_model([text_response("nope")] * 6)with pytest.raises(ModelError) as ei:- m.generate(ChatRequest(messages=[ChatMessage.user("x")], response_schema={"type": "object"}))+ m.generate(+ ChatRequest(messages=[ChatMessage.user("x")], response_schema={"type": "object"})+ )assert ei.value.usage.input_tokens == 100 and ei.value.usage.output_tokens == 20assert ei.value.usage.cost_usd > 0@@ -540,11 +437,12 @@ def _resp_with_calls(parts: list[types.Part]) -> types.GenerateContentResponse:def test_synthetic_call_ids_are_never_echoed_to_the_provider():- """When the provider sent NO id we mint one for local pairing — and it must- never travel back to Gemini (an id the provider did not issue)."""+ """Locally synthesized tool ids never travel back to Gemini."""from codeverse.models.gemini import is_synthetic_call_id- tool = ToolSpec(name="measure", description="m", parameters={"type": "object", "properties": {}})+ tool = ToolSpec(+ name="measure", description="m", parameters={"type": "object", "properties": {}}+ )m, log, _ = make_model([call_response("measure", {"part": "all"}), text_response("ok")])req = ChatRequest(messages=[ChatMessage.user("x")], tools=[tool])r = m.generate(req)@@ -553,7 +451,9 @@ def test_synthetic_call_ids_are_never_echoed_to_the_provider():msgs = [*req.messages,ChatMessage(role="assistant", parts=[call]),- ChatMessage(role="tool", parts=[ToolResultPart(call_id=call.id, name="measure", content="{}")]),+ ChatMessage(+ role="tool", parts=[ToolResultPart(call_id=call.id, name="measure", content="{}")]+ ),]m.generate(ChatRequest(messages=msgs, tools=[tool]))contents = log[1]["contents"]@@ -562,14 +462,14 @@ def test_synthetic_call_ids_are_never_echoed_to_the_provider():def test_two_parallel_same_name_calls_keep_their_provider_ids():- """Two calls to the SAME tool in one turn: a GENUINE provider id is echoed verbatim on- BOTH the function_call replay and the function_response (both used to drop it), and only- those ids pair a response with its call — each keeping its own thought_signature."""+ """Parallel same-name calls preserve provider ids and signatures."""from codeverse.models.gemini import is_synthetic_call_idparts = []for cid, target in (("id-a", "seat"), ("id-b", "leg")):- p = types.Part(function_call=types.FunctionCall(name="measure", args={"part": target}, id=cid))+ p = types.Part(+ function_call=types.FunctionCall(name="measure", args={"part": target}, id=cid)+ )p.thought_signature = f"sig-{cid}".encode()parts.append(p)tool = MEASURE_TOOL@@ -596,22 +496,26 @@ def test_two_parallel_same_name_calls_keep_their_provider_ids():assert [fp.function_call.id for fp in model_parts] == ["id-a", "id-b"]assert [fp.thought_signature for fp in model_parts] == [b"sig-id-a", b"sig-id-b"]responses = log[1]["contents"][2].parts- assert [(fp.function_response.id, fp.function_response.response["part"]) for fp in responses] == [+ assert [+ (fp.function_response.id, fp.function_response.response["part"]) for fp in responses+ ] == [("id-b", "leg"),("id-a", "seat"),]def test_max_output_tokens_eaten_by_thinking_carries_the_thought_tokens():- """The most expensive failure on Gemini 3: the reply is empty because thinking ate- the whole budget, and every thought token was billed. ``extract_candidate`` read- ``thoughts_token_count`` only to put it in the message and dropped the usage."""+ """A MAX_TOKENS failure reports the thinking tokens it consumed."""resp = types.GenerateContentResponse(model_version="gemini-3.7-flash",- candidates=[types.Candidate(content=types.Content(role="model", parts=[]),- finish_reason="MAX_TOKENS")],+ candidates=[+ types.Candidate(+ content=types.Content(role="model", parts=[]), finish_reason="MAX_TOKENS"+ )+ ],usage_metadata=types.GenerateContentResponseUsageMetadata(- prompt_token_count=1000, candidates_token_count=0, thoughts_token_count=8000),+ prompt_token_count=1000, candidates_token_count=0, thoughts_token_count=8000+ ),)m, _log, _ = make_model([resp])with pytest.raises(ModelError) as e:@@ -620,23 +524,7 @@ def test_max_output_tokens_eaten_by_thinking_carries_the_thought_tokens():assert e.value.usage.thoughts_tokens == 8000 and e.value.usage.cost_usd > 0-# --------------------------------------------------------------------- streaming-def test_merge_stream_chunks_reassembles_the_reply():- from codeverse.models.gemini import _merge_stream_chunks, extract_candidate-- c1 = types.GenerateContentResponse(- candidates=[types.Candidate(content=types.Content(role="model", parts=[types.Part.from_text(text="hel")]))])- c2 = types.GenerateContentResponse(- candidates=[types.Candidate(content=types.Content(role="model", parts=[types.Part.from_text(text="lo")]),- finish_reason="STOP")],- usage_metadata=types.GenerateContentResponseUsageMetadata(prompt_token_count=5, candidates_token_count=2))- merged = _merge_stream_chunks([c1, c2])- text, calls, finish = extract_candidate(merged)- assert (text, calls, finish) == ("hello", [], "STOP")- assert merged.usage_metadata.prompt_token_count == 5--def test_merge_stream_chunks_empty_stream_is_retryable():from codeverse.models.gemini import _merge_stream_chunks@@ -646,8 +534,7 @@ def test_merge_stream_chunks_empty_stream_is_retryable():def test_drain_stream_cuts_a_stream_past_its_attempt_budget():- """The wire timeout under streaming bounds only inter-chunk silence, so a- healthy-but-endless stream must be cut by the attempt deadline instead."""+ """The attempt deadline also bounds a healthy but endless stream."""from codeverse.models.gemini import _drain_streamticks = iter([1.0, 10.0, 20.0])@@ -679,26 +566,32 @@ def test_ipv4_transport_toggle(monkeypatch):def test_merge_stream_chunks_carries_tool_calls_and_signatures():- """Tool calls stream as function_call parts; the merge must keep them in arrival- order next to the text parts, and extract_candidate must still mint/pair ids and- stash thought signatures — the text-only tests above never exercised this."""- from codeverse.models.gemini import _merge_stream_chunks+ """Stream merging preserves ordered tool calls and thought signatures."""+ from codeverse.models.gemini import _merge_stream_chunks, extract_candidatefc = types.Part(function_call=types.FunctionCall(name="measure", args={"part": "Hull"}))fc.thought_signature = b"sig-bytes"c1 = types.GenerateContentResponse(- candidates=[types.Candidate(content=types.Content(role="model", parts=[types.Part.from_text(text="checking ")]))])+ candidates=[+ types.Candidate(+ content=types.Content(role="model", parts=[types.Part.from_text(text="hel")])+ )+ ]+ )c2 = types.GenerateContentResponse(- candidates=[types.Candidate(content=types.Content(role="model", parts=[fc]), finish_reason="STOP")],- usage_metadata=types.GenerateContentResponseUsageMetadata(prompt_token_count=9, candidates_token_count=4))+ candidates=[+ types.Candidate(+ content=types.Content(role="model", parts=[types.Part.from_text(text="lo"), fc]),+ finish_reason="STOP",+ )+ ],+ usage_metadata=types.GenerateContentResponseUsageMetadata(+ prompt_token_count=9, candidates_token_count=4+ ),+ )merged = _merge_stream_chunks([c1, c2])- text, calls, finish = extract_candidate_for_test(merged)- assert text == "checking " and finish == "STOP"+ text, calls, finish = extract_candidate(merged)+ assert text == "hello" and finish == "STOP"+ assert merged.usage_metadata.prompt_token_count == 9assert len(calls) == 1 and calls[0].name == "measure" and calls[0].arguments == {"part": "Hull"}assert SIGNATURES.get(calls[0].id) == b"sig-bytes"---def extract_candidate_for_test(resp):- from codeverse.models.gemini import extract_candidate-- return extract_candidate(resp)diff --git a/harness/tests/models/test_keypool.py b/harness/tests/models/test_keypool.pyindex ee2f8c2e..a305cc55 100644--- a/harness/tests/models/test_keypool.py+++ b/harness/tests/models/test_keypool.py@@ -79,9 +79,9 @@ def test_all_cooling_blocks_until_cooldown_ends():t0 = clock.tassert pool.acquire(timeout_s=60) in ("a", "b")assert MAX_WAIT_S - 1e-6 <= clock.t - t0 <= 5 + 1e-6+ pool.report("a", "429")+ pool.report("b", "429")with pytest.raises(KeyPoolExhausted):- pool.report("a", "429")- pool.report("b", "429")pool.acquire(timeout_s=1.0)@@ -188,3 +188,15 @@ def test_the_in_flight_slot_wait_is_bounded_by_timeout():assert time.monotonic() - t0 < 2.0, "the slot wait must honour the timeout"pool.release()assert pool.acquire(timeout_s=0.5) == "a", "the slot came back; nothing leaked"+++def test_skip_reconciles_tokens_but_leaves_health_and_counters_alone():+ """A charged-but-invalid reply (bad JSON) is not the key's fault: it used to be+ reported "ok", boosting health and counting as a success in doctor --live."""+ pool, _ = make(keys=("a",), tpm_per_key=100_000)+ pool.report("a", "429")+ before = {k["key"]: k for k in pool.stats()["keys"]}["…a"]+ pool.report("a", "skip", tokens=7_500, reserved=2_000)+ after = {k["key"]: k for k in pool.stats()["keys"]}["…a"]+ assert after["health"] == before["health"] and after["ok"] == before["ok"] and after["error"] == before["error"]+ assert pool.stats()["tokens_used"] == 7_500diff --git a/harness/tests/models/test_retry.py b/harness/tests/models/test_retry.pyindex 4383f7f0..7ae636b9 100644--- a/harness/tests/models/test_retry.py+++ b/harness/tests/models/test_retry.py@@ -58,25 +58,50 @@ def test_retries_until_success_and_reports():assert [n for n, _ in seen] == [1, 2]-def test_non_retryable_raises_immediately():- n = 0+def test_with_retries_stops_on_non_retryable_and_exhaustion():+ calls = 0- def fn():- nonlocal n- n += 1+ def non_retryable():+ nonlocal calls+ calls += 1raise ValueError("bad")with pytest.raises(ValueError):- with_retries(fn, is_retryable=lambda e: False, attempts=5, sleep=lambda s: None)- assert n == 1---def test_exhausted_raises_last():- def fn():+ with_retries(+ non_retryable, is_retryable=lambda error: False, attempts=5, sleep=lambda _: None+ )+ assert calls == 1++ def exhausted():+ nonlocal calls+ calls += 1raise TimeoutError("always")with pytest.raises(TimeoutError):- with_retries(fn, is_retryable=lambda e: True, attempts=3, sleep=lambda s: None)+ with_retries(exhausted, is_retryable=lambda error: True, attempts=3, sleep=lambda _: None)+ assert calls == 4+++def test_with_retries_honours_a_deadline_and_stamps_attempts():+ """The SDK adapters' loop: ``ChatRequest.max_wait_s`` bounds the whole call and the+ raised ModelError says how many round-trips were issued (the ledger's error row)."""+ from codeverse.models.base import ModelError++ calls = 0++ def flaky():+ nonlocal calls+ calls += 1+ raise ModelError("busy", retryable=True, status=503)++ with pytest.raises(ModelError) as ei:+ with_retries(flaky, is_retryable=lambda e: True, attempts=6, base_delay=1.0, max_delay=1.0,+ max_total_s=0.0, sleep=lambda _: None, jitter=False)+ assert calls == 1 and ei.value.attempts == 1 # the first backoff would already cross the deadline+ calls = 0+ with pytest.raises(ModelError) as ei:+ with_retries(flaky, is_retryable=lambda e: True, attempts=3, sleep=lambda _: None)+ assert calls == 3 and ei.value.attempts == 3# --------------------------------------------------------------------------- rotate_with_retries@@ -123,8 +148,7 @@ def _rotate503(pool, call, **kw):class Clock:- """Fake monotonic clock: pass it as ``monotonic``, its ``sleep`` records the nap and- advances it, and a test may advance ``.t`` itself to simulate time inside a call."""+ """Fake monotonic clock whose sleeps advance recorded time."""def __init__(self, t: float = 0.0) -> None:self.t = t@@ -140,6 +164,7 @@ class Clock:def _boom(msg: str = "503 high demand"):"""A ``call`` that always raises ``msg``."""+def call(_key):raise RuntimeError(msg)@@ -185,9 +210,10 @@ def test_rotate_every_key_dead_raises_without_benching():def test_rotate_429_is_free_while_untried_keys_remain():- pool = KeyPool(["k1", "k2", "k3"], cooldown_s=30)- slept: list[float] = []+ clock = Clock()+ pool = KeyPool(["k1", "k2", "k3"], cooldown_s=30, clock=clock, sleep=clock.sleep)calls: list[str] = []+ hinted: list[BaseException] = []def call(key: str) -> str:calls.append(key)@@ -195,10 +221,18 @@ def test_rotate_429_is_free_while_untried_keys_remain():raise ModelError("quota", retryable=True, status=429)return key- out = _rotate(pool, call, max_attempts=1, sleep=slept.append)+ out = _rotate(+ pool,+ call,+ max_attempts=1,+ sleep=clock.sleep,+ retry_after=lambda error: hinted.append(error) or 7.0,+ )assert out == "k3" and calls == ["k1", "k2", "k3"]- assert slept == [0.5, 0.5] # courtesy pauses only, no backoff- assert pool.stats()["n_cooling"] == 2+ assert clock.naps == [0.5, 0.5]+ stats = pool.stats()+ assert len(hinted) == 2 and stats["n_cooling"] == 2+ assert max(key["cooldown_s"] for key in stats["keys"]) <= MAX_WAIT_Sdef test_rotate_on_free_retry_hook_consumes_no_budget():@@ -222,55 +256,39 @@ def test_rotate_on_free_retry_hook_consumes_no_budget():assert len(calls) == 2-def test_rotate_retryable_backoff_then_exhausted_raises_last():+def test_rotate_retryable_backoff_and_nonretryable_failure_modes():pool = KeyPool(["k1"])slept: list[float] = []- def call(key: str) -> str:+ def retryable(key: str) -> str:raise ModelError("unavailable", retryable=True, status=500)with pytest.raises(ModelError) as ei:- _rotate(pool, call, max_attempts=3, base_delay=1.0, max_delay=4.0, sleep=slept.append)+ _rotate(+ pool,+ retryable,+ max_attempts=3,+ base_delay=1.0,+ max_delay=4.0,+ sleep=slept.append,+ )assert ei.value.status == 500 and len(slept) == 2 # backoff between the 3 attempts--def test_rotate_nonretryable_raises_immediately():- pool = KeyPool(["k1", "k2"])calls: list[str] = []- def call(key: str) -> str:+ def nonretryable(key: str) -> str:calls.append(key)raise ModelError("bad request", retryable=False, status=400)with pytest.raises(ModelError):- _rotate(pool, call, max_attempts=5)+ _rotate(KeyPool(["k1", "k2"]), nonretryable, max_attempts=5)assert calls == ["k1"]-def test_rotate_retry_after_hint_reaches_the_pool():- clock = Clock()- pool = KeyPool(["k1", "k2"], cooldown_s=30, clock=clock, sleep=clock.sleep)- seen: list[BaseException] = []-- def call(key: str) -> str:- if key == "k1":- raise ModelError("quota retryDelay 7s", retryable=True, status=429)- return key-- out = _rotate(pool, call, max_attempts=1, retry_after=lambda e: seen.append(e) or 7.0)- assert out == "k2" and len(seen) == 1- cooling = {k["key"]: k["cooldown_s"] for k in pool.stats()["keys"]}- assert 0 < cooling["…k1"] <= 7.0--# --------------------------------------------------------------------------- 503 storms@pytest.mark.parametrize(("hedge", "want_calls", "want_naps"), [(1, {9}, 7), (2, {9, 10}, 4)])def test_503_storm_has_its_own_patience_budget(hedge, want_calls, want_naps):- """A model-wide 503 storm must not burn the normal attempt budget. hedge=1: eight- sequential 503s then ok — seven storm waits, more than max_attempts=3 could survive.- hedge=2 (the default, audit 2026-08-26 §5.2) races both keys from the second 503 on, so- the same eight failures cost FOUR waits and the winner's sibling may land (a 10th call,- discarded)."""+ """A 503 storm gets a separate budget; hedging reduces its waits."""lock = threading.Lock()calls = {"n": 0}naps: list[float] = []@@ -283,57 +301,45 @@ def test_503_storm_has_its_own_patience_budget(hedge, want_calls, want_naps):raise RuntimeError("503 storm")return "ok"- out = _rotate503(_pool(2), call, max_attempts=3, base_delay=0.01, storm_attempts=10,- storm_max_delay=0.05, sleep=naps.append, hedge=hedge)+ out = _rotate503(+ _pool(2),+ call,+ max_attempts=3,+ base_delay=0.01,+ storm_attempts=10,+ storm_max_delay=0.05,+ sleep=naps.append,+ hedge=hedge,+ )assert out == "ok" and calls["n"] in want_calls# the first 503 rotates to k2 for free (no nap); k2's 503 leaves no untried key in the# 2-key pool, so that one is the stormassert len(naps) == want_naps and all(d <= 0.05 for d in naps) # the cap is the cap-def test_storm_waits_never_exceed_the_house_limit_at_production_defaults():- """Jitter must be applied BEFORE the cap. Regression: the storm branch capped first and- then multiplied by up to 1.25, so a documented "<=5 s" wait was observed at 6 s in a live- bench run (2026-08-24). 50 seeds, so this is not one lucky jitter draw."""+def test_storm_waits_never_exceed_the_house_limit_at_production_defaults(monkeypatch):+ monkeypatch.setattr(random, "random", lambda: 1.0)naps: list[float] = []- for seed in range(50):- random.seed(seed)- with pytest.raises(ModelError):- # storm_attempts deep enough that the exponent saturates the cap- _rotate503(_pool(2), _boom(), max_attempts=2, storm_attempts=12, sleep=naps.append)- assert naps, "expected the storm path to sleep"- assert max(naps) <= MAX_WAIT_S, f"a storm wait exceeded {MAX_WAIT_S}s: {max(naps)}"- assert max(naps) > MAX_WAIT_S * 0.9, "the cap should still be reached, not just respected"+ with pytest.raises(ModelError):+ _rotate503(_pool(2), _boom(), max_attempts=2, storm_attempts=12, sleep=naps.append)+ assert naps and max(naps) == MAX_WAIT_Sdef test_503_storm_budget_exhausts_then_normal_budget_applies():- """The storm patience is finite: once ``storm_attempts`` is spent the normal budget- applies again and the call gives up instead of looping forever."""+ """Exhausted storm patience falls back to the finite normal budget."""with pytest.raises(ModelError):- _rotate503(_pool(1), _boom("503 forever"), max_attempts=2, base_delay=0.0,- storm_attempts=3, storm_max_delay=0.0)---def test_live_tests_are_opt_in_by_default():- """A bare `pytest` must never fire real API calls: on 2026-08-24 a plain run on a keyed- machine hung for the whole timeout because the live suite was selected by default."""- import tomllib- from pathlib import Path-- root = Path(__file__).resolve().parents[2]- cfg = tomllib.loads((root / "pyproject.toml").read_text())- addopts = cfg["tool"]["pytest"]["ini_options"].get("addopts", [])- joined = " ".join(addopts) if isinstance(addopts, list) else str(addopts)- assert "not live" in joined, "live tests must be deselected by default (see pyproject addopts)"+ _rotate503(+ _pool(1),+ _boom("503 forever"),+ max_attempts=2,+ base_delay=0.0,+ storm_attempts=3,+ storm_max_delay=0.0,+ )def test_one_call_cannot_retry_for_hours():- """The storm branch was bounded in ATTEMPTS, not in time: 60 storm attempts x (a 300 s- read timeout + a <=5 s wait) is 5.1 hours for ONE logical call while the docstring- claimed "~5 min" — and the wall-clock ceiling is only checked when a call is BILLED, so- (measured 2026-08-24) a run with --max-minutes 90 reached 200 minutes. The deadline is- now threaded through every branch and ``GeminiModel`` clips each attempt's HTTP read- timeout to the remaining budget, floored at ``HTTP_TIMEOUT_FLOOR_S`` (simulated below)."""+ """The wall-clock deadline bounds even a long 503 storm."""clock = Clock()def call(_key):@@ -343,12 +349,20 @@ def test_one_call_cannot_retry_for_hours():with pytest.raises(ModelError):# hedge=1: the fake clock is serial, so two hedged 300 s calls would add 600 s- _rotate503(_pool(2), call, max_attempts=6, storm_attempts=60, sleep=clock.sleep,- monotonic=clock, hedge=1)+ _rotate503(+ _pool(2),+ call,+ max_attempts=6,+ storm_attempts=60,+ sleep=clock.sleep,+ monotonic=clock,+ hedge=1,+ )# bounded by the clock: the ONLY legal overshoot is the one floored attempt that was# already in flight when the deadline passedassert clock.t <= RETRY_DEADLINE_S + HTTP_TIMEOUT_FLOOR_S, (- f"one call burned {clock.t / 3600:.2f} h; the deadline is {RETRY_DEADLINE_S / 60:.0f} min")+ f"one call burned {clock.t / 3600:.2f} h; the deadline is {RETRY_DEADLINE_S / 60:.0f} min"+ )assert clock.t < 5 * 3600, "this is the 5.1-hour regression"@@ -359,24 +373,26 @@ def test_the_deadline_does_not_cut_a_call_that_is_making_progress():def call(_key):calls["n"] += 1- clock.t += 280.0 # slow, but under the per-attempt timeout+ clock.t += 280.0 # slow, but under the per-attempt timeoutif calls["n"] < 3:raise RuntimeError("503 high demand")return "ok"# hedge=1: serial fake clock and call counter (see the hedge tests for the raced form)- out = _rotate503(_pool(2), call, max_attempts=6, storm_attempts=60, sleep=clock.sleep,- monotonic=clock, hedge=1)+ out = _rotate503(+ _pool(2),+ call,+ max_attempts=6,+ storm_attempts=60,+ sleep=clock.sleep,+ monotonic=clock,+ hedge=1,+ )assert out == "ok" and calls["n"] == 3def test_503_rotates_through_every_untried_key_before_it_is_a_storm():- """Owner's rule (2026-08-27): a 503 on one key is simply replaced by the next key — free- while an untried key remains, exactly like the 429 path. Measured 2026-08-26, one tiny- request per key in parallel during the day's storm: flash answered on 15/22, 18/22, 21/22- keys while only 5/4/1 keys said 503 at the same instant. So with a 10-key pool NINE- consecutive 503s still rotate: no sleep, no storm, no ``max_attempts`` spent- (``max_attempts=3`` here proves the budget is free)."""+ """A 503 rotates freely while the pool still has an untried key."""keys = [f"k{i}" for i in range(1, 11)]gate = StormGate()lock = threading.Lock()@@ -391,24 +407,30 @@ def test_503_rotates_through_every_untried_key_before_it_is_a_storm():raise RuntimeError("503 high demand")return "ok"- out = _rotate503(_pool(10), call, max_attempts=3, storm_attempts=10, storm_max_delay=0.05,- sleep=naps.append, storm_gate=gate)+ out = _rotate503(+ _pool(10),+ call,+ max_attempts=3,+ storm_attempts=10,+ storm_max_delay=0.05,+ sleep=naps.append,+ storm_gate=gate,+ )# one key per round-trip, never twice: 9 free rotations, then the 10th key landsassert out == "ok" and len(calls) == len(keys) and set(calls) == set(keys)assert naps == [], "rotation to a fresh key costs no sleep"assert gate.n_storms == 0 and not gate.storming, "an untried key remains: not a storm"[email protected](("hedge", "want_calls", "want_hits", "want_stats"), [- (1, {6}, 3, {"attempts": 6, "hedged": 0, "storm": 3}),- (2, {6, 7}, 2, {"attempts": 7, "hedged": 3, "storm": 2}),-])+ ("hedge", "want_calls", "want_hits", "want_stats"),+ [+ (1, {6}, 3, {"attempts": 6, "hedged": 0, "storm": 3}),+ (2, {6, 7}, 2, {"attempts": 7, "hedged": 3, "storm": 2}),+ ],+)def test_503_on_every_key_is_still_a_storm(hedge, want_calls, want_hits, want_stats):- """The backstop: once EVERY key in the pool has 503'd inside this one call there is no- untried key left to rotate to, so the bounded storm wait engages — one wait per hedged- ATTEMPT, not per key. hedge=1: k1, k2, k3 fail one by one (the whole pool), then two- storm waits, then ok. hedge=2: k1 fails alone, k2+k3 fail together (pool exhausted:- storm 1), two more together (storm 2), then the raced pair lands."""+ """Exhausting every key engages one bounded wait per hedged attempt."""gate = StormGate("t", base_delay=0.0, max_wait_s=0.0, probe_lease_s=0.0) # no real waitslock = threading.Lock()n = {"c": 0}@@ -422,10 +444,20 @@ def test_503_on_every_key_is_still_a_storm(hedge, want_calls, want_hits, want_straise RuntimeError("503 high demand")return "ok"- out = _rotate503(_pool(3), call, max_attempts=3, storm_attempts=10, storm_max_delay=0.05,- storm_gate=gate, hedge=hedge, stats=stats)+ out = _rotate503(+ _pool(3),+ call,+ max_attempts=3,+ storm_attempts=10,+ storm_max_delay=0.05,+ storm_gate=gate,+ hedge=hedge,+ stats=stats,+ )assert out == "ok" and n["c"] in want_calls- assert gate.n_storms == 1 and gate.n_hits == want_hits, "declared only after all three keys failed, then waited out"+ assert gate.n_storms == 1 and gate.n_hits == want_hits, (+ "declared only after all three keys failed, then waited out"+ )assert stats == want_stats@@ -439,11 +471,7 @@ def _wait_idle(pool, timeout: float = 5.0) -> None:def test_after_the_first_503_the_next_attempt_is_hedged_and_the_first_success_wins():- """Measured 2026-08-26: a failed 503 costs the 21-50 s round-trip the provider holds- before rejecting, not the <= 5 s sleep (logged sleep was 13 % of the wait), and storm- streaks average 4.7 attempts. So from the first 503 on, the next attempt goes out on- two fresh keys at once and the first success is returned while the other is still in- flight; the loser reports its own outcome and releases its own slot when it lands."""+ """After a 503, fresh keys race and the first success wins."""pool = KeyPool(["k1", "k2", "k3", "k4"], rpm_per_key=10_000, max_in_flight=8)release_k2 = threading.Event()lock = threading.Lock()@@ -461,14 +489,27 @@ def test_after_the_first_503_the_next_attempt_is_hedged_and_the_first_success_wiraise RuntimeError("503 high demand")return f"ok:{key}"- out = rotate_with_retries(pool, call, classify=_503, outcome_of=lambda e: "5xx", max_attempts=3,- storm_attempts=10, sleep=naps.append, tokens_of=lambda r: 7, stats=stats)+ out = rotate_with_retries(+ pool,+ call,+ classify=_503,+ outcome_of=lambda e: "5xx",+ max_attempts=3,+ storm_attempts=10,+ sleep=naps.append,+ tokens_of=lambda r: 7,+ stats=stats,+ )assert out == "ok:k3"assert naps == [], "a hedged retry sleeps for nothing"- assert calls[0] == "k1" and sorted(calls[1:]) == ["k2", "k3"], "two distinct fresh keys, k1 excluded"+ assert calls[0] == "k1" and sorted(calls[1:]) == ["k2", "k3"], (+ "two distinct fresh keys, k1 excluded"+ )assert stats == {"attempts": 3, "hedged": 1, "storm": 0}st = pool.stats()- assert st["in_flight"] == 1 and st["peak_in_flight"] == 2, "the loser still holds its max_in_flight slot"+ assert st["in_flight"] == 1 and st["peak_in_flight"] == 2, (+ "the loser still holds its max_in_flight slot"+ )assert st["ok"] == 1 and st["5xx"] == 1release_k2.set()_wait_idle(pool)@@ -477,9 +518,7 @@ def test_after_the_first_503_the_next_attempt_is_hedged_and_the_first_success_widef test_a_loser_that_succeeds_later_is_discarded_but_its_tokens_are_reported():- """A success-then-success is the one case the hedge is not free: the second answer is- thrown away, its tokens still counted against the key (a 503 bills nothing, so while it- storms the hedge costs nothing)."""+ """A late successful loser is discarded but still charged to its key."""pool = KeyPool(["k1", "k2", "k3"], rpm_per_key=10_000, tpm_per_key=1_000_000)release_k2 = threading.Event()@@ -491,7 +530,13 @@ def test_a_loser_that_succeeds_later_is_discarded_but_its_tokens_are_reported():return "late:k2"return "ok:k3"- out = _rotate(pool, call, max_attempts=3, storm_attempts=10, tokens_of=lambda r: 900 if r.endswith("k2") else 100)+ out = _rotate(+ pool,+ call,+ max_attempts=3,+ storm_attempts=10,+ tokens_of=lambda r: 900 if r.endswith("k2") else 100,+ )assert out == "ok:k3"release_k2.set()_wait_idle(pool)@@ -512,11 +557,19 @@ def test_hedge_1_disables_and_a_full_pool_of_slots_degrades_to_one_key():if i <= 3:raise ModelError("503 high demand", retryable=True, status=503)return "ok"+return callstats: dict = {}- _rotate(_pool(3), make_call(), max_attempts=3, storm_attempts=10,- storm_max_delay=0.0, hedge=1, stats=stats)+ _rotate(+ _pool(3),+ make_call(),+ max_attempts=3,+ storm_attempts=10,+ storm_max_delay=0.0,+ hedge=1,+ stats=stats,+ )assert stats["hedged"] == 0 and stats["attempts"] == 4# max_in_flight=1: the primary holds the only slot, so a partner is never waited forpool = _pool(3, max_in_flight=1)@@ -532,8 +585,7 @@ def test_hedge_1_disables_and_a_full_pool_of_slots_degrades_to_one_key():def test_the_worst_error_decides_a_hedged_attempt():- """One key says 503 and the other says the request itself is bad: raise the 400 at once- instead of storming on a request that can never succeed."""+ """A nonretryable hedge error outranks a sibling 503."""pool = _pool(3)stats: dict = {}@@ -559,13 +611,14 @@ def test_a_dead_key_among_the_hedge_is_benched_once_the_sibling_wins():assert _rotate(pool, call, max_attempts=3, storm_attempts=10) == "ok:k3"_wait_idle(pool)- assert pool.stats()["n_dead"] == 1, "k3 proved the request fine, so k2's 403 was the key's fault"+ assert pool.stats()["n_dead"] == 1, (+ "k3 proved the request fine, so k2's 403 was the key's fault"+ )# --------------------------------------------------------------------------- deadline threading (2026-08-27)def test_the_key_wait_is_clipped_to_the_remaining_budget():- """pool.acquire used to be called with its own 120 s default no matter how little- budget the call had left; now it gets min(ACQUIRE_TIMEOUT_S, remaining)."""+ """Pool acquisition cannot outwait the remaining call budget."""seen: list[float | None] = []class SpyPool(KeyPool):@@ -584,15 +637,24 @@ def test_a_storm_wait_is_clipped_to_the_remaining_budget():"""Sleeping past the deadline only to give up on waking helps nobody."""clock = Clock()with pytest.raises(ModelError):- _rotate503(_pool(1), _boom(), max_attempts=6, storm_attempts=60, base_delay=4.0,- storm_max_delay=5.0, max_total_s=3.0, sleep=clock.sleep, monotonic=clock, hedge=1)+ _rotate503(+ _pool(1),+ _boom(),+ max_attempts=6,+ storm_attempts=60,+ base_delay=4.0,+ storm_max_delay=5.0,+ max_total_s=3.0,+ sleep=clock.sleep,+ monotonic=clock,+ hedge=1,+ )assert clock.naps, "expected the storm path to sleep"assert clock.t <= 3.0 + 1e-6, f"slept past the deadline: {clock.t}"def test_free_429_rotation_cannot_outlive_the_deadline():- """Free rotation costs no attempts but DOES cost wall clock (0.5 s per hop):- with many keys it used to spin long past the budget."""+ """Free key rotation still consumes wall-clock budget."""clock = Clock()pool = _pool(10, cooldown_s=0.0, clock=clock, sleep=clock.sleep)calls: list[str] = []@@ -609,9 +671,7 @@ def test_free_429_rotation_cannot_outlive_the_deadline():def test_the_pool_is_not_refunded_for_a_charged_but_invalid_reply():- """retry.py used to report tokens=0 for a bad-JSON failure, refunding a TPM- reservation the provider had actually consumed — a 42k-token invalid reply- handed the key 42k TPM back. ``ModelError.usage`` now carries the real bill."""+ """A billed invalid response must consume its actual TPM usage."""clock = Clock(1000.0)pool = _pool(1, tpm_per_key=100_000, clock=clock, sleep=clock.sleep)calls = {"n": 0}@@ -619,8 +679,11 @@ def test_the_pool_is_not_refunded_for_a_charged_but_invalid_reply():def call(key):calls["n"] += 1if calls["n"] == 1:- raise ModelError("structured output is not valid JSON", retryable=True,- usage=Usage(input_tokens=42_000))+ raise ModelError(+ "structured output is not valid JSON",+ retryable=True,+ usage=Usage(input_tokens=42_000),+ )return "ok"out = _rotate(pool, call, max_attempts=3, tokens_hint=1_000, tokens_of=lambda r: 100)@@ -631,10 +694,7 @@ def test_the_pool_is_not_refunded_for_a_charged_but_invalid_reply():# --------------------------------------------------------------------------- on_attempt hook (2026-08-27)def test_on_attempt_hears_every_round_trip_and_a_late_hedge_loser():- """The per-attempt hook: the winner flushes as not-discarded, every failed try- as discarded, and a hedge loser still in flight reports (discarded — a late- success included) when it lands. That is how a loser's paid tokens reach the- cost ledger at all."""+ """The attempt hook reports winners, failures, and late hedge losers."""pool = _pool(4)release_k2 = threading.Event()lock = threading.Lock()@@ -681,14 +741,13 @@ def test_a_broken_on_attempt_hook_never_breaks_the_call():assert _rotate(pool, lambda key: "ok", on_attempt=hook) == "ok"-def test_the_error_it_gives_up_with_is_never_its_own_cause():- """`classify()` returns an already-classified ModelError unchanged, so the give-up- `raise err from exc` was `raise e from e` — and CPython does not check `from` for a- cycle. bench/_infra.py had to grow a seen-set after that chain took run_cell to a- RecursionError and a compare matrix lost 11 finished cells (2026-08-27)."""+def test_error_causes_avoid_cycles_and_preserve_real_wrapped_errors():+ """Classified errors avoid self-cycles while preserving genuine causes."""+def call(key):- raise ModelError("structured output unavailable (finish_reason=MAX_TOKENS)",- retryable=False, status=None)+ raise ModelError(+ "structured output unavailable (finish_reason=MAX_TOKENS)", retryable=False, status=None+ )with pytest.raises(ModelError) as ei:_rotate(_pool(), call, max_attempts=1)@@ -702,22 +761,16 @@ def test_the_error_it_gives_up_with_is_never_its_own_cause():node = node.__cause__ if node.__cause__ is not None else node.__context__assert node is None, "the cause chain closes into a cycle"--def test_a_real_wrapped_cause_still_survives_the_self_cause_guard():- """The guard drops the cause only when it IS the error — a genuine wrapped cause- (the classifier built a NEW ModelError) must still reach the caller."""- def call(key):+ def timeout(key):raise TimeoutError("read timed out")with pytest.raises(ModelError) as ei:- _rotate(_pool(), call, max_attempts=1)+ _rotate(_pool(), timeout, max_attempts=1)assert isinstance(ei.value.__cause__, TimeoutError)def test_a_budget_spent_at_the_storm_gate_issues_no_round_trip():- """A worker parked at a gate ANOTHER thread closed carries no ``last_err``, so the- guard at the top of the loop (which requires one) let it through: it woke past its- deadline and still spent a full round-trip (measured: 25 s on a 5 s budget)."""+ """A worker waking past its gate deadline issues no request."""clock = Clock(1000.0)calls: list[str] = []@@ -733,7 +786,13 @@ def test_a_budget_spent_at_the_storm_gate_issues_no_round_trip():passwith pytest.raises(ModelError) as e:- _rotate(_pool(1), lambda key: calls.append(key),- max_total_s=5.0, monotonic=clock, storm_gate=Gate(), label="t")+ _rotate(+ _pool(1),+ lambda key: calls.append(key),+ max_total_s=5.0,+ monotonic=clock,+ storm_gate=Gate(),+ label="t",+ )assert calls == [], "a round-trip was issued after the budget was already spent"assert "capacity" in str(e.value)diff --git a/harness/tests/prompts/test_files.py b/harness/tests/prompts/test_files.pyindex 27775fba..69600ff4 100644--- a/harness/tests/prompts/test_files.py+++ b/harness/tests/prompts/test_files.py@@ -9,26 +9,17 @@ from tests.prompts.conftest import PROMPT_FILES, PROMPTS_DIR, read_prompt@pytest.mark.parametrize("rel", PROMPT_FILES)-def test_file_exists_and_nonempty(rel: str) -> None:+def test_prompt_file_contract(rel: str) -> None:p = PROMPTS_DIR / relassert p.is_file(), f"missing prompt file {rel}"assert len(p.read_text()) > 500, f"{rel} suspiciously small"--[email protected]("rel", PROMPT_FILES)-def test_loads_and_renders(rel: str) -> None:raw = load_text(rel)assert prompt_hash(raw)# markdown prompts must be jinja-inert: render() with no context must be a no-oprendered = render(rel)assert rendered == raw, f"{rel} contains live jinja syntax; keep prompt .md files static"--[email protected]("rel", PROMPT_FILES)-def test_jinja_safe(rel: str) -> None:- text = read_prompt(rel)for seq in ("{{", "{%", "{#"):- assert seq not in text, f"{rel} contains {seq!r} which breaks jinja rendering"+ assert seq not in raw, f"{rel} contains {seq!r} which breaks jinja rendering"@pytest.mark.parametrize(@@ -91,7 +82,7 @@ def test_every_language_ships_a_system_prompt() -> None:A missing file here means a language silently falls back to nothing."""from codeverse.contracts.common import Language- from codeverse.languages._docs import prompt_dir_for+ from codeverse.prompts.catalog import prompt_dir_forfrom codeverse.tracks.prompting import language_system_promptfor lang in Language:diff --git a/harness/tests/reference/test_mismatch.py b/harness/tests/reference/test_mismatch.pyindex f9a20bbe..7a96337f 100644--- a/harness/tests/reference/test_mismatch.py+++ b/harness/tests/reference/test_mismatch.py@@ -1,4 +1,4 @@-"""The render-vs-reference diff and the refine tasks derived from it."""+"""The render-vs-reference diff."""from __future__ import annotations@@ -7,7 +7,7 @@ from pathlib import Pathfrom PIL import Imagefrom codeverse.models.base import ModelError-from codeverse.reference import Mismatch, ReferenceDiff, compare, refine_tasks+from codeverse.reference import comparefrom tests.reference.conftest import FakeChat, make_specANSWER = {@@ -45,14 +45,10 @@ def test_compare_names_mismatches_and_builds_text(tmp_path: Path):assert "the BRIEF wins" in req.messages[0].parts[0].text # synthesized caveat-def test_top_orders_by_severity_and_becomes_three_tasks(tmp_path: Path):+def test_top_orders_by_severity(tmp_path: Path):ref, ren = _png(tmp_path / "ref.png"), _png(tmp_path / "ren.png")d = compare(make_spec(), [ref], [ren], model=FakeChat({"reference_diff": [ANSWER]}))assert [m.severity for m in d.top(3)] == ["critical", "major", "minor"]- tasks = refine_tasks(d)- assert len(tasks) == 3- assert tasks[0].target == "Hopper" and tasks[0].priority == 1 and tasks[0].source == "gate"- assert "missing feature" in tasks[0].instruction and "conical hopper" in tasks[0].instructiondef test_model_failure_is_soft(tmp_path: Path):@@ -64,9 +60,3 @@ def test_model_failure_is_soft(tmp_path: Path):def test_missing_images_short_circuit(tmp_path: Path):d = compare(make_spec(), [tmp_path / "gone.png"], [_png(tmp_path / "r.png")], model=FakeChat())assert "nothing to compare" in d.error---def test_no_mismatches_produces_no_tasks():- assert refine_tasks(ReferenceDiff()) == []- d = ReferenceDiff(mismatches=[Mismatch(kind="wrong_shape", detail="x")])- assert len(refine_tasks(d, top=3)) == 1diff --git a/harness/tests/texturing/test_apply.py b/harness/tests/texturing/test_apply.pyindex 9dcaaa77..ca0b1ef0 100644--- a/harness/tests/texturing/test_apply.py+++ b/harness/tests/texturing/test_apply.py@@ -8,7 +8,7 @@ import numpy as npimport trimeshfrom codeverse.spatial.measure import measure_glb-from codeverse.texturing.apply import apply_textures, node_part_lookup, textured_summary+from codeverse.texturing.apply import apply_textures, node_part_lookupfrom codeverse.texturing.generate import FakeImageModel, generate_texturesfrom codeverse.texturing.plan import default_plan@@ -30,8 +30,6 @@ def test_apply_textures_end_to_end(tmp_path: Path, chair_glb: Path, chair_plan):assert rep.parts_skipped == ["Knob"] and rep.parts_unmatched == []assert rep.projections["Leg_0"] == "cylinder" and rep.projections["Seat"] == "box"assert rep.n_materials == 2- summ = textured_summary(out)- assert summ["textured"] == 6 and summ["geometries"] == 7 and summ["images"] == 2# names + geometry preserved; measurement unchangeda, b = measure_glb(chair_glb), measure_glb(out)assert {p.name for p in a.parts} == {p.name for p in b.parts}diff --git a/harness/tests/texturing/test_tile.py b/harness/tests/texturing/test_tile.pyindex 9cbd591d..ee9ed570 100644--- a/harness/tests/texturing/test_tile.py+++ b/harness/tests/texturing/test_tile.py@@ -9,11 +9,9 @@ from codeverse.texturing.generate import (SEAM_MAX,fit_size,make_tileable,- offset_check,procedural_texture,save_texture,seam_score,- tile_preview,)@@ -52,5 +50,3 @@ def test_fit_size_save_and_helpers(tmp_path):assert p.is_file() and Image.open(p).format == "JPEG"p2 = save_texture(sq, tmp_path / "t.png")assert Image.open(p2).format == "PNG"- assert offset_check(sq).size == sq.size- assert tile_preview(sq, reps=2, size=64).size == (64, 64)