Official Codebase of 3DCodeVerse
Python
prompts: correct the medians the lean-prompt docstrings ship
ziyao3e94ec230210
diff --git a/harness/codeverse/config.py b/harness/codeverse/config.pyindex 05e76947..9a026356 100644--- a/harness/codeverse/config.py+++ b/harness/codeverse/config.py@@ -116,7 +116,7 @@ class Limits(BaseModel):"AGENTS.md/GEMINI.md; the tool list arrives three times (prompt 4 088 + GEMINI.md + ""the MCP declarations); the whole 24 k cookbook is inlined though a brief-driven ""chapter selector exists; and a refine prompt (median 18 604 chars over the 48 "- "recorded there) carries all 20 acceptance items when 6 failed, and every part row "+ "recorded there) carries all 21 acceptance items when 6 failed, and every part row ""for a median of 4 targets. OFF until the A/B reads out; ""`CV3D_LEAN_PROMPT=1` (read at call time by `lean_prompt_enabled`) is the variant arm.",)diff --git a/harness/codeverse/tracks/prompting.py b/harness/codeverse/tracks/prompting.pyindex a9d554d0..3b89613f 100644--- a/harness/codeverse/tracks/prompting.py+++ b/harness/codeverse/tracks/prompting.py@@ -364,8 +364,8 @@ def parts_table_for_targets(plan: Plan | None, targets: Sequence[str]) -> str:aa_articulated). Narrowing is only honest when the round is ABOUT those parts: a groupthat also carries the whole-artifact target (``overall`` — where a gate error with nopart of its own lands) is about the assembly, and an assembly is judged against rows- this would have dropped. That is most rounds — 39 of those 48 prompts carry ``overall``- and only 4 narrow — which is the price of not lying about what the table is. Instance+ this would have dropped. That is most rounds — 40 of those 48 prompts carry ``overall``+ and only 3 narrow — which is the price of not lying about what the table is. Instancesuffixes come off first: ``_canon_target`` hands on the plan's spelling but keeps the``Leg_1``, and the plan has no such row."""keys = {to_snake(instance_base(t)) for t in targets}diff --git a/harness/tests/orchestrator_tracks/test_lean_prompt.py b/harness/tests/orchestrator_tracks/test_lean_prompt.pyindex be798dc9..e7abb5c5 100644--- a/harness/tests/orchestrator_tracks/test_lean_prompt.py+++ b/harness/tests/orchestrator_tracks/test_lean_prompt.py@@ -4,7 +4,7 @@ Measured on aa_articulated (2026-09-02, prompts dated 2026-09-01): the baseline53 562 chars and the CLI additionally loads GEMINI.md (18 444 B) plus 6 618 chars of MCP tooldeclarations on EVERY turn — with prompts/urdf/contract.md in both GEMINI.md and the prompt,the tool list in all three, the whole 24 k cookbook inlined, and a refine prompt (median-18 604 chars over the 48 recorded there) carrying all 20 acceptance items when 6 had failed,+18 604 chars over the 48 recorded there) carrying all 21 acceptance items when 6 had failed,and every part row for a median of 4 targets.The switch is an A/B arm, so the control must be untouched: every test here renders the SAME