Official Codebase of 3DCodeVerse
Python
planner: a truncated plan is retried with low thinking and a compact-answer note, not only a bigger budget
ziyao7e90ca09621b
diff --git a/harness/codeverse/tracks/planner.py b/harness/codeverse/tracks/planner.pyindex c8e293e7..ea46489f 100644--- a/harness/codeverse/tracks/planner.py+++ b/harness/codeverse/tracks/planner.py@@ -626,6 +626,17 @@ def plan_tokens(budget: PlanBudget, floor: int) -> int:return min(PLAN_TOKENS_MAX, max(floor, want))+TRUNCATION_NOTE = ("\n\nCOMPACT ANSWER REQUIRED: your previous answer did not fit in the output budget "+ "({tokens} tokens). Return the same plan with every `description` at most 25 words, at most 6 "+ "`children` per part and nothing outside the JSON. Answer directly without long deliberation.")+++def _with_note(msg: ChatMessage, note: str) -> ChatMessage:+ """The same user message with ``note`` appended to its text (images kept)."""+ images = [p for p in msg.parts if isinstance(p, ImagePart)] or None+ return ChatMessage.user(msg.text + note, images=images)++def _truncated(exc: Exception) -> bool:"""Did this model error mean 'the answer did not fit'?"""return "MAX_TOKENS" in str(exc)@@ -731,10 +742,11 @@ def plan_with_usage[P: BaseModel](spec: Spec, model_id: str, plan_model: type[P]last_error = ""invalid = requeried = grown = 0tokens = plan_tokens(budget, max_output_tokens)+ thinking = "medium"geo_reasked = 0for attempt in range(2 + MAX_VALIDATION_REASKS + MAX_QUALITY_REASKS + MAX_GEOMETRY_REASKS):req = ChatRequest(messages=messages, system=system, response_schema=schema, temperature=temperature,- thinking="medium", max_output_tokens=tokens, label=f"planner{'-retry' if attempt else ''}",+ thinking=thinking, max_output_tokens=tokens, label=f"planner{'-retry' if attempt else ''}",max_wait_s=plan_wait_s(tokens, guard))try:resp = model.generate(req)@@ -743,8 +755,14 @@ def plan_with_usage[P: BaseModel](spec: Spec, model_id: str, plan_model: type[P]raise # not a truncation, or already at the model ceiling: nothing to growgrown += 1tokens = min(PLAN_TOKENS_MAX, int(tokens * TRUNCATION_GROWTH))+ # A bigger budget alone does not help when thinking ate the last one: with the pro+ # planner a 55k-token retry ran ~16 min into the provider's deadline (504) and the+ # cell was lost (compare_art_v4_pp, 2026-08-28). Retry with low thinking and a+ # compact-answer note on the last user turn.+ thinking = "low"+ messages = messages[:-1] + [_with_note(messages[-1], TRUNCATION_NOTE.format(tokens=tokens))]if events is not None:- events.emit("plan.truncated", attempt=attempt, max_output_tokens=tokens)+ events.emit("plan.truncated", attempt=attempt, max_output_tokens=tokens, thinking=thinking)continueusage = usage + resp.usageif guard is not None:diff --git a/harness/tests/orchestrator_tracks/test_planner_truncation_reask.py b/harness/tests/orchestrator_tracks/test_planner_truncation_reask.pynew file mode 100644index 00000000..3c52448a--- /dev/null+++ b/harness/tests/orchestrator_tracks/test_planner_truncation_reask.py@@ -0,0 +1,35 @@+"""A truncated plan is retried with low thinking and a compact-answer note, not just more tokens."""++from __future__ import annotations++from codeverse.contracts.common import Language+from codeverse.contracts.plan import StaticPlan+from codeverse.proc import EventLog+from codeverse.tracks.planner import TRUNCATION_NOTE, plan+from tests.orchestrator_tracks.conftest import make_spec+from tests.orchestrator_tracks.fakes import FakeChatModel, FakeRuntime+from tests.orchestrator_tracks.test_generation_planner_repair import _valid_plan_dict+++def test_truncation_retries_compact_with_low_thinking(tmp_ws):+ calls = {"n": 0}++ def respond(req):+ calls["n"] += 1+ if calls["n"] == 1:+ raise RuntimeError("model stopped: finish_reason=MAX_TOKENS")+ return _valid_plan_dict()++ model = FakeChatModel(respond)+ events = EventLog(tmp_ws.events_path)+ p = plan(make_spec(), "fake:planner", StaticPlan, tmp_ws, model=model, events=events,+ runtime=FakeRuntime(Language.THREEJS))+ assert p.object_name == "DiningChair" and len(model.requests) == 2+ first, second = model.requests+ assert first.thinking == "medium" and second.thinking == "low"+ assert second.max_output_tokens > first.max_output_tokens+ assert second.messages[-1].text.endswith(TRUNCATION_NOTE.format(tokens=second.max_output_tokens))+ assert second.messages[-1].text.startswith(first.messages[-1].text) # the original request is kept+ assert len(second.messages) == len(first.messages) # no dangling user turn+ kinds = [e["event"] for e in events.read()]+ assert "plan.truncated" in kinds and "plan.done" in kinds