Skip to content

Commit b5563d4

Browse files
authored
devlog: record the community bug sweep (#2636)
The plan's DONE criterion (zero lidge-jun issues, every bug PR terminal) was met partway through; the request behind it was broader, so the loop continued into the bug-labelled community backlog. Sixteen open, three left. Records what the sweep found on its own (#2458's real cause one layer off the report, #2509's unnoticed second half, #2459's lying health check), the six that were already fixed by earlier work in this same run and how each was verified, and why the remaining three are open rather than skipped.
1 parent 87179e8 commit b5563d4

1 file changed

Lines changed: 69 additions & 0 deletions

File tree

Lines changed: 69 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,69 @@
1+
# 150 — community bug sweep
2+
3+
The plan's DONE criterion was zero `lidge-jun` issues and every bug PR terminal. Both were met
4+
partway through. The request behind the plan was broader — no outstanding bugs — so the loop
5+
continued into the `bug`-labelled community backlog rather than stopping at its own inventory.
6+
7+
Sixteen open, three left. Final `dev` at `87179e86a`: 15037 pass / 16 skip / 0 fail, typecheck
8+
and `privacy:scan` green.
9+
10+
## Six were already fixed by this same run
11+
12+
| Issue | Fixed by | Verified |
13+
|---|---|---|
14+
| #2499 Windows catalog-state latency | #2580 (`342911fec7`) | ancestry + 51 pass |
15+
| #2545 Antigravity thought_signature | #2577 (`4d3d2716e`) | ancestry + 150 pass |
16+
| #2210 Cursor stall timeout | `994e5ba87` | ancestry + 252 pass |
17+
| #2156 Muse Spark mid-tool-call | `1a15d6292` + `08cc2ac89` | ancestry + 54 pass |
18+
| #2300 Cursor slowness | #2307 | ancestry + falsification |
19+
| #2509 stale pool account (core) | #2515 | ancestry + falsification |
20+
21+
Each was checked with `git merge-base --is-ancestor` **and** a covering test run before closing.
22+
Closing on a report that something was fixed is how a backlog acquires issues that were never
23+
actually resolved.
24+
25+
## What the sweep found on its own
26+
27+
**#2458 — the reported cause was one layer off.** A 502 on an undeclared `get_video_duration`
28+
tool call looked like a bridge problem. The real defect was that Chat Completions silently
29+
discarded `video_url` (`src/chat/inbound.ts:56`), so the model reasoned about a video it had
30+
never received, and the bridge refused the resulting call. OpenCodex was rejecting a symptom of
31+
its own dropped input.
32+
33+
**#2509 — a second half nobody had noticed.** #2515 fixed selection to preview per candidate
34+
quota scope and #2623 added the entitlement filter, but the encrypted-recovery path got the
35+
scope and not the filter. Same stale-selection class, one layer over.
36+
37+
**#2459 — the health check was the worst part.** A bare npm reinstall under a live proxy leaves
38+
`/healthz` answering 200 while every `/v1/responses` fails. A lying health check is worse than
39+
an honest failure, because every supervisor believes it.
40+
41+
**A load flake, not a defect.** One full-suite failure at exactly 5003 ms — Bun's default 5s
42+
timeout — on a case that starts a real server while its neighbours declare `SERVER_BUDGET_MS`.
43+
It passed 3/3 in isolation on two machines. Fixed as a budget, not chased as a bug.
44+
45+
## Three left open, each for a stated reason
46+
47+
**#2221 — native main token refresh.** The safe-looking half is not separable. A refresh rotates
48+
the grant, `atomicWriteFile` has no compare-and-swap against an external writer, and the Codex
49+
CLI writes the same `auth.json`. Refreshing without a crash-safe publisher can strand the login:
50+
the current behavior fails one request, the naive fix can cost the credential. #2497 proposes a
51+
publisher and is itself held under security review with a rename-before-link crash window. This
52+
needs a designed protocol, not a smaller patch.
53+
54+
**#1527 — Cursor large-context.** Both mechanisms behind the reported symptoms changed: rate
55+
limits are excluded from transport retry, and checkpoint continuation replaced full-history
56+
replay — which was the credible differentiator in the reporter's own proxy-vs-direct comparison.
57+
The residual claim needs a matched live probe, which only the reporter can run.
58+
59+
**#1419 — macOS SIGTRAP.** The reported offsets symbolicate to Bun's own crash handler and
60+
nothing past it; the frames that would name the faulting subsystem are absent, and no upstream
61+
release establishes that 1.4.0 fixes that signature. The runtime is pinned so it cannot regress;
62+
the crash is unproven. Recorded alongside it: `ocx gui` starts an unref'd detached process, so a
63+
native crash there strands the proxy with no supervisor, unlike an installed service.
64+
65+
## The pattern worth keeping
66+
67+
Four of these closed as "already fixed" and three as "not fixable here, here is why". Neither is
68+
a failure to do work. A backlog that only ever accepts code changes as outcomes accumulates
69+
issues nobody can close, and closes issues nobody actually verified.

0 commit comments

Comments
 (0)