Skip to content

Official Codebase of 3DCodeVerse

Python

budget: money stops being a ceiling; 3.7-flash becomes the default generator

gaoypeng5be2450d1977
diff --git a/harness/CLAUDE.md b/harness/CLAUDE.md
index 2040295f..db19485e 100644
--- a/harness/CLAUDE.md
+++ b/harness/CLAUDE.md
@@ -63,7 +63,13 @@ push, with `set -o pipefail` so a `| tail` cannot swallow a red exit. The versi
drops workspace MCP servers), `codex` 0.147+ (MCP needs
`default_tools_approval_mode="approve"`), `claude` 2.1, `agy` 1.1 — the last three run
on local subscriptions: test lightly. No Anthropic/OpenAI API keys here.
-- Main test model: `gemini-3.7-flash` (lenient judge; needs the defect checklist for
+- Main test model: `gemini-3.7-flash` — and since 2026-08-28 it is also the DEFAULT
+ generator (`gemini-cli:gemini-3.7-flash`), verified live through the harness on both
+ vendor paths: gemini-cli completes and is not silently substituted, and `agy` 1.1.22
+ lists and serves `gemini-3.7-flash-{high,medium,low}`. A BARE `gemini -m ...` is
+ substituted down to 3.5-flash — the harness's per-session system settings file is
+ what prevents that, so never judge model availability with a raw CLI probe.
+ (lenient judge; needs the defect checklist for
range). Judge default `gemini-3.1-pro-preview` (`Settings.default_judge`; tracks use
n_samples=1); flash + `--n 3` is the cheap fallback, pro for calibration/eval.
- Live runs under `runs/` (`e2e_*`): read-only reference material; never modify.
@@ -79,7 +85,7 @@ cd /home/yipeng/3dcodeverse/harness
3dcodeverse make "..." --profile economy|balanced|quality # one dial: models, judge n, rounds, candidates, texture, ceilings (--profile == CV3D_PROFILE)
3dcodeverse make "..." --track static_object --language threejs --generator gemini-cli:gemini-3.7-flash
3dcodeverse make "..." --track articulated_object --language urdf_blender
-3dcodeverse make "..." --track scene --language scene_threejs --rounds 2 --max-usd 3
+3dcodeverse make "..." --track scene --language scene_threejs --rounds 2 # --max-usd is OFF by default (0 = no ceiling); cost is still recorded
3dcodeverse make "neon rain on a window" --track graphics --language glsl_shader
3dcodeverse make "..." --image ref.png --candidates 3 --rounds 2 --dim height=0.45 --must "three legs" --texture --no-run
3dcv resume <slug> · 3dcv status <slug> · 3dcv render <slug> [--mode wire] · 3dcv judge <slug> [--model ... --n 3]
diff --git a/harness/codeverse/cli/main.py b/harness/codeverse/cli/main.py
index 4e7b700c..38936ddb 100644
--- a/harness/codeverse/cli/main.py
+++ b/harness/codeverse/cli/main.py
@@ -104,7 +104,7 @@ def run_cmd(
judge: Annotated[str | None, typer.Option("--judge", help="fixed judge model for the whole battery")] = None,
parallel: Annotated[int, typer.Option("--parallel", min=1)] = 4,
rounds: Annotated[int, typer.Option("--rounds", min=0)] = 4,
- max_usd: Annotated[float, typer.Option("--max-usd")] = 5.0,
+ max_usd: Annotated[float, typer.Option("--max-usd", help="0 = no ceiling (cost is still recorded)")] = 0.0,
max_minutes: Annotated[float, typer.Option("--max-minutes", help="wall-clock budget per run; size it to the weather "
"(RUNBOOK 7.x: 120 in a 503 storm, else runs burn the hour with no judged round)")] = 60.0,
limit: Annotated[int | None, typer.Option("--limit")] = None,
diff --git a/harness/codeverse/contracts/common.py b/harness/codeverse/contracts/common.py
index fe2a00f9..8dd65a64 100644
--- a/harness/codeverse/contracts/common.py
+++ b/harness/codeverse/contracts/common.py
@@ -133,13 +133,18 @@ class Budget(BaseModel):
The ceilings are non-negative by construction. A negative one is not a small
budget, it is an unrunnable one: ``BudgetGuard.ok()`` is False before a single
- token is spent, the first charge raises ``BudgetExceeded: cost $0.001 exceeds
- max_usd $-2.50``, and ``grant_grace`` cannot lift a hard ceiling back above zero —
+ token is spent and ``grant_grace`` cannot lift a hard ceiling back above zero —
so the workspace and the git-committed spec are created for a run that can only die.
+
+ ``max_usd = 0`` means NO MONEY CEILING, and that is the default (owner, 2026-08-28:
+ "先不计成本,就是run就好了"). Cost is still measured on every call — the ledger,
+ `3dcv cost`, the run record and the bench economics are untouched; money simply
+ stops being a thing that can end a run. Pass ``--max-usd`` to put the ceiling back.
+ Time still bounds a run: that is ``max_minutes``, and without it nothing ends.
"""
max_rounds: int = Field(default=4, ge=0)
- max_usd: float = Field(default=5.0, ge=0)
+ max_usd: float = Field(default=0.0, ge=0, description="0 = no ceiling (cost is still recorded)")
max_minutes: float = Field(default=60.0, ge=0)
max_repair_attempts: int = Field(default=3, ge=0) # per build failure before escalating
@@ -150,7 +155,7 @@ class Backends(BaseModel):
* API chat models: ``gemini:gemini-3.7-flash`` · ``anthropic:claude-sonnet-5``
· ``openai:gpt-5.6-sol``
* Coding agents: ``gemini-cli:gemini-3.7-flash`` · ``claude-code:sonnet``
- · ``codex:gpt-5.6-sol`` · ``agy:gemini-3.6-flash-high``
+ · ``codex:gpt-5.6-sol`` · ``agy:gemini-3.7-flash-high``
The coding agent is always a VENDOR agent: the harness supplies the workspace, the
prompt and its 3D tools (over MCP) and reads the result. The in-process
@@ -159,7 +164,7 @@ class Backends(BaseModel):
"""
planner: str = "gemini:gemini-3.7-flash"
- generator: str = "gemini-cli:gemini-3.6-flash"
+ generator: str = "gemini-cli:gemini-3.7-flash"
# The judge drives the refine loop: the pro tier has ~3x lower sample noise than
# flash (calibration 2026-08-23: std 0.03 vs 0.08-0.12) for ~$0.07 per verdict.
judge: str = "gemini:gemini-3.1-pro-preview"
diff --git a/harness/codeverse/cost/profiles.py b/harness/codeverse/cost/profiles.py
index 977ce17c..be2e699a 100644
--- a/harness/codeverse/cost/profiles.py
+++ b/harness/codeverse/cost/profiles.py
@@ -66,7 +66,7 @@ class Profile:
# NOT reduced by any profile — the score effect is inside the
# judge's own noise in both directions (docs/COST.md §14)
#: budget ceilings a run of this shape should not need to exceed
- max_usd: float = 5.0
+ max_usd: float = 0.0 # 0 = no ceiling; cost is still recorded
max_minutes: float = 60.0
#: measured expectation (docs/COST.md) — reported by ``3dcv cost profiles``
expected_usd: float = 0.0
@@ -109,7 +109,7 @@ PROFILES: dict[str, Profile] = {
# the instrument's noise do not license a payload cut worth a fifth of a cent, so the
# experiment is recorded as INCONCLUSIVE in docs/COST.md §14 and the budget is unchanged.
judge_max_px=1024, judge_montages=3, judge_detail_crops=2,
- max_usd=1.50, max_minutes=30.0,
+ max_usd=0.0, max_minutes=30.0,
expected_usd=0.30,
expected_score="graphics 0.81 median (5/6 passed); objects clear the gates less often "
"— a one-shot flash generator scored 0.14 on compare_v1's hard cells",
@@ -118,26 +118,26 @@ PROFILES: dict[str, Profile] = {
),
"balanced": Profile(
name="balanced",
- generator="gemini-cli:gemini-3.6-flash",
+ generator="gemini-cli:gemini-3.7-flash",
planner="gemini:gemini-3.7-flash",
judge="gemini:gemini-3.1-pro-preview",
captioner="gemini:gemini-3.7-flash",
rounds=4, candidates=1, judge_samples=1, max_turns=0, texture=False,
judge_max_px=1024, judge_montages=3, judge_detail_crops=2,
- max_usd=5.0, max_minutes=60.0,
+ max_usd=0.0, max_minutes=60.0,
expected_usd=1.47,
expected_score="0.835 mean on compare_v1, 36/61 runs passed ($2.50 per passing artifact)",
note="today's defaults — the arm every number in docs/COST.md was measured on",
),
"quality": Profile(
name="quality",
- generator="gemini-cli:gemini-3.6-flash",
+ generator="gemini-cli:gemini-3.7-flash",
planner="gemini:gemini-3.7-flash",
judge="gemini:gemini-3.1-pro-preview",
captioner="gemini:gemini-3.7-flash",
rounds=4, candidates=2, judge_samples=3, max_turns=0, texture=True,
judge_max_px=1024, judge_montages=3, judge_detail_crops=2,
- max_usd=8.0, max_minutes=90.0,
+ max_usd=0.0, max_minutes=90.0,
expected_usd=3.20,
expected_score="best-of-2 lifted the stool baseline 0.563 → 0.612 and the texture pass "
"0.686 → 0.701; judge σ 0.017 at n=3",
diff --git a/harness/codeverse/orchestrator.py b/harness/codeverse/orchestrator.py
index b32287c4..bd3e57b0 100644
--- a/harness/codeverse/orchestrator.py
+++ b/harness/codeverse/orchestrator.py
@@ -1008,7 +1008,9 @@ class BudgetGuard:
"""Raise ``BudgetExceeded`` if any hard ceiling has been crossed."""
spent = self.billed_usd
elapsed = self.elapsed_minutes()
- if spent > self.hard_usd:
+ # max_usd == 0 means no money ceiling (the default since 2026-08-28); cost is
+ # still accumulated above, so every report and the ledger still know what it cost.
+ if self.budget.max_usd > 0 and spent > self.hard_usd:
raise BudgetExceeded(
f"cost ${spent:.3f} exceeds max_usd ${self.hard_usd:.2f}",
spent_usd=spent,
@@ -1033,7 +1035,7 @@ class BudgetGuard:
"""Reason string when the soft sub-budget is used up, else ``""``."""
usd, minutes = self.soft_limits()
spent, elapsed = self.billed_usd, self.elapsed_minutes()
- if spent > usd:
+ if self.budget.max_usd > 0 and spent > usd:
return f"cost ${spent:.3f} exceeds soft cap ${usd:.2f} ({self.soft_fraction:.0%} of ${self.budget.max_usd:.2f})"
if elapsed > minutes:
return f"elapsed {elapsed:.1f} min exceeds soft cap {minutes:.1f} min ({self.soft_fraction:.0%} of {self.budget.max_minutes:.1f})"
diff --git a/harness/codeverse/tracks/lifecycle.py b/harness/codeverse/tracks/lifecycle.py
index 6f03598d..e20248b3 100644
--- a/harness/codeverse/tracks/lifecycle.py
+++ b/harness/codeverse/tracks/lifecycle.py
@@ -709,9 +709,10 @@ class BaseTrack:
already stopping on budget."""
if rounds or ctx.state.completed_rounds or ctx.plan is None:
return # nothing was built yet (the PLAN itself blew the budget) → nothing to salvage
- if not ctx.budget.ok() and ctx.budget.spent.cost_usd > ctx.budget.hard_usd * 2:
- return # runaway spend: do not throw good money after bad
- grace_usd = max(0.15, ctx.spec.budget.max_usd * 0.10)
+ cap = ctx.spec.budget.max_usd
+ if cap > 0 and not ctx.budget.ok() and ctx.budget.spent.cost_usd > ctx.budget.hard_usd * 2:
+ return # runaway spend against a real ceiling: do not throw good money after bad
+ grace_usd = max(0.15, cap * 0.10)
grace_min = max(5.0, ctx.spec.budget.max_minutes * 0.15)
ctx.budget.grant_grace(usd=grace_usd, minutes=grace_min)
ctx.events.emit("budget.salvage", reason="no round completed before the budget stop",
diff --git a/harness/docs/COST.md b/harness/docs/COST.md
index b051202a..caf61306 100644
--- a/harness/docs/COST.md
+++ b/harness/docs/COST.md
@@ -1159,6 +1159,13 @@ burned earlier the same morning.
## 25. `max_usd` guards money, and a subscription costs none
+> **Since 2026-08-28 the money guard is OFF by default** (`max_usd = 0` in `Budget` and
+> in every profile, owner's call: run first, count later). Nothing below stops being
+> true — every call is still priced into `telemetry/cost.jsonl`, `3dcv cost` and the
+> run record — but money can no longer END a run. `--max-usd <n>` puts the ceiling
+> back, and everything in this section then applies exactly as written. What still
+> bounds a run is `max_minutes`.
+
`Usage.cost_usd` answers *"what would these tokens cost at list price?"*. That is the
right number for a report, a $/complexity point, or a flywheel record — it is comparable
across backends and independent of who is paying. It is the wrong number to hand a
diff --git a/harness/tests/cost/test_profiles.py b/harness/tests/cost/test_profiles.py
index 53d3b13a..1b466fec 100644
--- a/harness/tests/cost/test_profiles.py
+++ b/harness/tests/cost/test_profiles.py
@@ -46,7 +46,8 @@ def test_balanced_is_todays_defaults():
s = Settings()
assert (p.generator, p.planner, p.judge) == (s.default_generator, s.default_planner, s.default_judge)
assert p.rounds == 4 and p.candidates == s.default_candidates and p.judge_samples == 1
- assert p.max_usd == 5.0 and p.max_minutes == 60.0 and p.judge_max_px == 1024
+ # max_usd 0 = no money ceiling (owner, 2026-08-28); cost is still recorded on every call
+ assert p.max_usd == 0.0 and p.max_minutes == 60.0 and p.judge_max_px == 1024
def test_economy_is_single_shot_flash_and_two_rounds_with_no_turn_cap_of_its_own():
@@ -156,7 +157,7 @@ def test_make_profile_writes_the_whole_shape_onto_the_spec(tmp_path: Path):
assert r.exit_code == 0, r.output
spec = json.loads((runs / "pot" / "spec.json").read_text())
assert spec["options"] == {"candidates": 2, "texture": True, "profile": "quality"}
- assert spec["budget"]["max_rounds"] == 4 and spec["budget"]["max_usd"] == 8.0
+ assert spec["budget"]["max_rounds"] == 4 and spec["budget"]["max_usd"] == 0.0
assert spec["backends"]["judge"] == "gemini:gemini-3.1-pro-preview"
@@ -180,7 +181,7 @@ def test_no_profile_flag_leaves_the_defaults_alone(tmp_path: Path):
# the resolved dial is recorded whichever way it was named, so `3dcv resume`
# reproduces it; with no flag and no env that dial is the default, balanced
assert spec["options"]["profile"] == "balanced" and spec["options"]["texture"] is False
- assert spec["budget"]["max_rounds"] == 4 and spec["budget"]["max_usd"] == 5.0
+ assert spec["budget"]["max_rounds"] == 4 and spec["budget"]["max_usd"] == 0.0
def test_judge_samples_reach_the_round_policy_only_when_a_profile_asks(tmp_path: Path):
@@ -234,17 +235,17 @@ def test_the_flag_and_the_env_var_resolve_to_the_same_dial(name, monkeypatch):
"judge": "gemini:gemini-3.7-flash", "judge_samples": 2,
"rounds": 2, "candidates": 1, "texture": False,
"judge_max_px": 1024, "judge_montages": 3, "judge_detail_crops": 2,
- "agent_max_turns": 0, "max_usd": 1.50, "max_minutes": 30.0}),
- ("balanced", {"generator": "gemini-cli:gemini-3.6-flash",
+ "agent_max_turns": 0, "max_usd": 0.0, "max_minutes": 30.0}),
+ ("balanced", {"generator": "gemini-cli:gemini-3.7-flash",
"judge": "gemini:gemini-3.1-pro-preview", "judge_samples": 1,
"rounds": 4, "candidates": 1, "texture": False,
"judge_max_px": 1024, "judge_montages": 3, "judge_detail_crops": 2,
- "agent_max_turns": 0, "max_usd": 5.0, "max_minutes": 60.0}),
- ("quality", {"generator": "gemini-cli:gemini-3.6-flash",
+ "agent_max_turns": 0, "max_usd": 0.0, "max_minutes": 60.0}),
+ ("quality", {"generator": "gemini-cli:gemini-3.7-flash",
"judge": "gemini:gemini-3.1-pro-preview", "judge_samples": 3,
"rounds": 4, "candidates": 2, "texture": True,
"judge_max_px": 1024, "judge_montages": 3, "judge_detail_crops": 2,
- "agent_max_turns": 0, "max_usd": 8.0, "max_minutes": 90.0}),
+ "agent_max_turns": 0, "max_usd": 0.0, "max_minutes": 90.0}),
],
)
def test_each_profile_resolves_to_its_documented_dial(name, expected, monkeypatch):