Skip to content

Official Codebase of 3DCodeVerse

Python

lifecycle: the last round is re-judged before the run ends, so `--rounds 0` with a failed judge says judge_unavailable

gaoypeng76f2560cdf88
diff --git a/harness/codeverse3d/tracks/lifecycle.py b/harness/codeverse3d/tracks/lifecycle.py
index 63398839..e5c1d2c6 100644
--- a/harness/codeverse3d/tracks/lifecycle.py
+++ b/harness/codeverse3d/tracks/lifecycle.py
@@ -549,36 +549,39 @@ class BaseTrack:
ctx.events.emit("stop", reason=reason.value, rounds=len(rounds))
return reason
+ def _unjudged(rec: RoundRecord | None) -> bool:
+ return rec is not None and rec.judgment is None and rec.build is not None and rec.build.ok
+
while True:
index = len(rounds)
+ last = rounds[-1] if rounds else None
+ if _unjudged(last) and last.index not in rejudged and ctx.budget.ok():
+ # judge outage/degraded verdict on a clean build: re-judge the SAME commit once
+ # (no regeneration) — before planning refinements from nothing, and before the
+ # last round ends the run unscored (`--rounds 0` used to stop MAX_ROUNDS with an
+ # unjudged baseline: the re-judge lived in the refine branch only)
+ rejudged.add(last.index)
+ prev_j = rounds[-2].judgment if len(rounds) > 1 else None
+ if rejudge_round(ctx, pipeline, last, previous=prev_j):
+ self._save_budget(ctx)
if index > ctx.policy.max_rounds: # the baseline + max_rounds refine rounds all ran
- return _stop(RunStatus.MAX_ROUNDS)
+ return _stop(RunStatus.JUDGE_UNAVAILABLE if _unjudged(last) else RunStatus.MAX_ROUNDS)
if not ctx.budget.ok():
if not rounds: # the stages spent the clock without raising: the ONE budget path (salvage)
ctx.budget.check()
return _stop(RunStatus.BUDGET)
- previous = rounds[-1].judgment if rounds else None
- if index == 0:
+ previous = last.judgment if last else None
+ if last is None:
ctx.state.status = RunStatus.GENERATING
tasks, instructions, kind = self.baseline_tasks(ctx), [], "baseline"
else:
ctx.state.status = RunStatus.REFINING
- last = rounds[-1]
- if last.judgment is None and last.build is not None and last.build.ok and last.index not in rejudged:
- # judge outage/degraded verdict on a clean build: re-judge the SAME
- # commit once (no regeneration) before planning refinements from nothing.
- rejudged.add(last.index)
- prev_j = rounds[-2].judgment if len(rounds) > 1 else None
- if rejudge_round(ctx, pipeline, last, previous=prev_j):
- self._save_budget(ctx)
- continue # plan the next round from the recovered verdict
tasks, instructions = self.refine_tasks(ctx, last)
kind = "refine"
if not tasks:
# no gate error, no failed must-item, no judge plan: nothing to ask for. When
# that is only because the judge never scored the round, say so.
- unjudged = last.judgment is None and last.build is not None and last.build.ok
- return _stop(RunStatus.JUDGE_UNAVAILABLE if unjudged else RunStatus.NO_REFINE_TASKS)
+ return _stop(RunStatus.JUDGE_UNAVAILABLE if _unjudged(last) else RunStatus.NO_REFINE_TASKS)
ctx.state.save(ctx.ws)
try:
if index == 0 and tasks and ctx.policy.n_candidates > 1:
diff --git a/harness/docs/ARCHITECTURE.md b/harness/docs/ARCHITECTURE.md
index 7f8ef147..449c1310 100644
--- a/harness/docs/ARCHITECTURE.md
+++ b/harness/docs/ARCHITECTURE.md
@@ -487,7 +487,8 @@ plateau, regression or diminishing-returns stop, no rewrite or surface-detail ro
escalates on identical signatures) → gates → … → judge (SKIPPED only where the verdict is never
bought at all: no judge or no renders — a round that finishes past the clock is still judged)
→ commit src/ (the round's commit) and copy its build to artifacts/rNN/
- a round left without a verdict (judge outage / degraded) is re-judged once before the next is planned
+ a round left without a verdict (judge outage / degraded) is re-judged once before the next is planned,
+ and so is the LAST round before the run ends (still none → judge_unavailable, `--rounds 0` included)
a round's cost is the ledger rows it booked (a cost.tally) and its steps are timed (RoundRecord.steps); a round
that raises mid-way still reports what it burned (rounds/aborted_rNN.json + record.extra["aborted_rounds"])
a round whose every task failed raises RoundFailed, TYPED (2026-09-22 — the loop never reads its message):
diff --git a/harness/tests/orchestrator_tracks/test_fix_batch2.py b/harness/tests/orchestrator_tracks/test_fix_batch2.py
index c4c6bdb2..300e7013 100644
--- a/harness/tests/orchestrator_tracks/test_fix_batch2.py
+++ b/harness/tests/orchestrator_tracks/test_fix_batch2.py
@@ -116,6 +116,24 @@ def test_degraded_verdict_never_scores_and_run_stops_as_judge_unavailable(tmp_pa
assert "judge degraded" in saved["notes"] and saved["usage"]["cost_usd"] == pytest.approx(0.016, abs=1e-6)
[email protected](("verdicts", "status", "score"), [
+ (["degraded", 0.7], RunStatus.MAX_ROUNDS, 0.7),
+ (["degraded", "degraded"], RunStatus.JUDGE_UNAVAILABLE, None),
+])
+def test_the_last_round_is_rejudged_before_the_run_ends(tmp_path, chair_plan, settings, verdicts, status, score):
+ """`--rounds 0`: the baseline IS the last round. Its failed verdict was never re-judged (the
+ re-judge lived in the refine branch) and the run stopped MAX_ROUNDS unscored; now it gets the
+ same one re-judge as any round, and a second failure says judge_unavailable (D80)."""
+ ws = Workspace(tmp_path / "runs" / "r")
+ judge = ScriptedJudge(verdicts)
+ track = StaticObjectTrack(services=FakeServices(), judge=judge, agent=FakeAgent(_writer),
+ planner_model=_planner(chair_plan.model_dump(mode="json")), settings=settings,
+ runtime=FakeRuntime(Language.THREEJS))
+ rec = track.run(make_spec(max_rounds=0), ws)
+ assert rec.status is status and len(rec.rounds) == 1 and len(judge.all_calls) == 2
+ assert rec.rounds[0].score == (pytest.approx(score) if score is not None else None)
+
+
def test_degraded_verdict_recovers_via_rejudge_of_same_commit(tmp_path, chair_plan, settings):
spec = make_spec(max_rounds=3)
ws = Workspace(tmp_path / "runs" / "r")