Skip to content

Official Codebase of 3DCodeVerse

Python

review round 2: the residue — a self-mimic's own message, and three prose copies

ziyaof3eef613be2d
diff --git a/harness/codeverse/spatial/joints_model.py b/harness/codeverse/spatial/joints_model.py
index abebb5d4..bd80617d 100644
--- a/harness/codeverse/spatial/joints_model.py
+++ b/harness/codeverse/spatial/joints_model.py
@@ -343,6 +343,8 @@ def _mimic_message(issue: MimicIssue, joints: dict[str, Joint]) -> str:
return f'joint {j}: <mimic joint="{t}"> names no joint; known: {sorted(joints)}'
if kind == "immobile_target":
return f"joint {j}: mimics {t!r}, which is {joints[t].type} and never moves"
+ if kind == "self":
+ return f'joint {j}: <mimic joint="{j}"> names itself'
return f"joint {j}: <mimic> chain is a cycle through {(issue.detail or j)!r}"
diff --git a/harness/docs/ARCHITECTURE.md b/harness/docs/ARCHITECTURE.md
index 9c4b0844..fd4d3126 100644
--- a/harness/docs/ARCHITECTURE.md
+++ b/harness/docs/ARCHITECTURE.md
@@ -357,10 +357,8 @@ server (name `3dcv`) for the vendor CLIs, (c) a native tool schema for any embed
An `Observation` carries two different answers: `ok` is the **verdict** and `failed`
says the tool **could not run** (exception, missing artefact, unusable arguments).
`Observation.error` builds every failure caught at the `ToolDef.call` boundary; three
-tools set `failed` on an observation they compose themselves, where the failure is a
-fact about the result rather than an exception — `build` (runtime success, no readable
-GLB), `scene_probe` (the probe driver died) and `observe.render_observation` (no view AND
-no console error; with one it is a verdict). Those four places are the whole list. MCP `is_error` is `failed` alone: a
+tools set `failed` on an observation they compose themselves; the list of those places
+lives with the field, in `spatial/registry.Observation.error`. MCP `is_error` is `failed` alone: a
negative verdict is a result whose text leads with FAIL, because a vendor CLI retries an
errored call, and over 224 recorded gemini-cli sessions `is_error = not ok` made 62 % of
1 445 `joint_sweep` calls, 23 % of 2 144 `build`s and ~15 % of the connectivity/contract
@@ -553,14 +551,10 @@ errors, cost, fingerprints, `duplicate_of`; `--pack` tars with byte-range locato
`flywheel pairs` emits preference pairs (round i < j by judge Δ ≥ τ) and round-level
repair pairs (in-session trajectory mining died with the api-agent, 2026-08-28 —
vendor CLIs log raw stdout, not structured tool turns); `flywheel refine` emits the loop's OWN transitions — one row per (round i → round
-i+1) where the harness asked for a change: the failing round's gate findings and
-judge complaint, the refine instructions the harness wrote, both code snapshots (git
-`before`/`after` commits, `--with-code` inlines the files), the score before and
-after, and `outcome` (`improved` when the judge gained ≥ 0.05 — the same threshold
-`flywheel/pairs.py` calls a preference — `regressed` when it lost as much,
-`unchanged` in between, `unscored` when either side has no verdict). Rows are
-dropped, with the reason recorded, when a predecessor has no commit, no verdict or
-did not build. That is the supervision the harness produces that a
+i+1) where the harness asked for a change, carrying what condemned the round, the
+instructions written in response, both code snapshots and whether the score moved
+(the row is `flywheel/refine.RefineTransition`; INTERFACES has the fields and the
+drop reasons). That is the supervision the harness produces that a
one-shot corpus cannot: what a failing artefact looked like, what was wrong with it
in the harness's own words, and what the fix changed. `toolkits/llamafactory/`
turns those rows into training files; the harness writes the measurement, not the
diff --git a/harness/docs/COST.md b/harness/docs/COST.md
index 49da7772..ae08c5fe 100644
--- a/harness/docs/COST.md
+++ b/harness/docs/COST.md
@@ -1447,12 +1447,9 @@ A second flag, not a threshold. `Observation.failed` is what the MCP server rep
`is_error`; `ok` stays the verdict. `Observation.error` builds every failure caught at the
`ToolDef.call` boundary (exception, `ToolUnavailable`, `ToolUsageError`, missing or
unreadable artefact), and three tools set `failed` on a result they compose themselves,
-where the failure is a fact about the result rather than an exception: `build` (the runtime
-reported success and left no readable GLB), `scene_probe` (the probe driver died) and
-`observe.render_observation` (no view AND nothing logged — a render with no picture and no
-reason is not a verdict; console errors ARE one, empty RenderSet or not, because a scene
-that fails to boot returns exactly that and the errors say what to fix). Those four places
-are the whole list.
+where the failure is a fact about the result rather than an exception. Which places those
+are is a property of the code, so the list lives there (`spatial/registry.Observation.error`)
+and not in three documents.
Every negative verdict now LEADS its text with `… FAIL` (`gate_observation` for the gates,
`JOINT SWEEP: FAIL — penetration …`, `FRAME GATE: FAIL — n error(s)`, `RENDER: FAIL — n
diff --git a/harness/docs/RUNBOOK.md b/harness/docs/RUNBOOK.md
index 7d1d7454..1dc3fc42 100644
--- a/harness/docs/RUNBOOK.md
+++ b/harness/docs/RUNBOOK.md
@@ -159,13 +159,10 @@ unless you type `--host` yourself; it never serves a path outside the declared r
[--captions-dir caps/] [--drop-duplicates]
3dcv flywheel pairs runs/ pairs.jsonl [--min-delta 0.05]
3dcv flywheel refine runs/ refine.jsonl [--with-code]
- # one row per round i -> i+1 the harness asked to change:
- # gate findings + judge complaint + the refine instructions,
- # before/after commits (--with-code inlines the files),
- # score_before/score_after, outcome improved|regressed|
- # unchanged|unscored, and dropped rows with the reason
- # (no_predecessor / no_commit / predecessor_unjudged / ...).
- # Training formats live in toolkits/llamafactory/, not here.
+ # one row per round the harness asked to change; the row
+ # schema is codeverse/flywheel/refine.RefineTransition and
+ # INTERFACES has the call signatures. Training formats live
+ # in toolkits/llamafactory/, not here.
3dcv flywheel caption <slug> [--model …] [--out caps/] # --out = side-car mode, run untouched
3dcv flywheel gallery runs/ gallery.html [--title …] # alias of `3dcv gallery build --embed` (§4)
3dcv flywheel index runs/ runs_index.sqlite · 3dcv flywheel dedupe dataset/
diff --git a/harness/tests/urdf_joints/test_mimic.py b/harness/tests/urdf_joints/test_mimic.py
index a024c662..31c0e777 100644
--- a/harness/tests/urdf_joints/test_mimic.py
+++ b/harness/tests/urdf_joints/test_mimic.py
@@ -316,3 +316,13 @@ def test_the_lint_says_when_only_the_first_instance_of_a_driver_is_followed(tmp_
warns = [f for f in lint_workspace(ws).findings if "stay independent inputs" in f.message]
assert len(warns) == 1 and warns[0].target == "rib_a_hinge"
assert "runner_slide_2" in warns[0].message and warns[0].severity.value == "warn"
+
+
+def test_a_self_mimic_says_it_names_itself(tmp_path):
+ """The loader reported a self-coupling as "chain is a cycle", which is true of the
+ graph and useless to whoever wrote ``<mimic joint="itself">``. The shared rules have a
+ ``self`` kind; the URDF renderer now has the sentence for it."""
+ text = RIB.replace('<mimic joint="runner_slide" multiplier="4" offset="0"/>',
+ '<mimic joint="rib_a_hinge" multiplier="1" offset="0"/>')
+ with pytest.raises(UrdfError, match="names itself"):
+ _robot(tmp_path, text)