Skip to content

Commit 45ae409

Browse files
committed
docs(devlog): release-readiness note for the cursor-call integration
The five gates ran against the MERGED dev SHA 4f72d6f, not the pre-merge tip — a merge commit is a tree nobody has tested until it is tested. All green in a lidge worktree pinned to that SHA, with HEAD re-asserted before each gate. Recommends 2.25.0, a minor rather than a patch, because the externally observable behaviour of a FAILED turn changed: a turn that came back completed with a vanished tool call now reports failed, an unrequested CANCEL is now a typed failure, and a truncated compaction turn no longer installs half-written replacement history. Correct behaviour, but new behaviour for anything keying on "the proxy said completed". The maintainer decides; the note recommends. Says two things plainly rather than burying them. The merges were owner-authorized admin merges with CI waived — that is owner authority over one's own repository, not compliance with MAINTAINERS.md:48-49, and lidge is Linux-only while CI spans three platforms. And publishing is not one command away: scripts/release.ts waits for a successful Cross-platform CI run AND a Service lifecycle run at the exact release SHA, so a green gate table alone does not mean bun run release will work. Carries the five open follow-ups, led by the one the campaign's own docs previously overstated: the Cursor tool-result image encoder is correct and does not reach production, because every Cursor model is in noVisionModels and the vision sidecar strips images before the adapter runs.
1 parent ff4b0bb commit 45ae409

1 file changed

Lines changed: 132 additions & 0 deletions

File tree

Lines changed: 132 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,132 @@
1+
# 060 — Release readiness for the cursor-call integration
2+
3+
**Verdict: dev is releasable. Recommend cutting `2.25.0` — a minor, not a patch.
4+
The decision is the maintainer's; this note recommends and does not act.**
5+
6+
## What landed
7+
8+
| PR | Merge | Content |
9+
|----|-------|---------|
10+
| [#1993](https://github.com/lidge-jun/opencodex/pull/1993) | `228295b3f` | Cursor wire: clean-EOF terminal gate, real `McpImageContent` in tool results |
11+
| [#1994](https://github.com/lidge-jun/opencodex/pull/1994) | `bb3e5db19` | Unexpected server-side CANCEL provenance |
12+
| [#1995](https://github.com/lidge-jun/opencodex/pull/1995) | `4f72d6f75` | Bridge/adapter terminal semantics + WP2b EOF usage |
13+
14+
`dev` = `4f72d6f7555013ab231de78233d2aa95bd1e439c`.
15+
16+
Ancestry proven rather than assumed:
17+
18+
git merge-base --is-ancestor ff4b0bb4e627354bb57aa7317e690482b2a95312 4f72d6f755 # exit 0
19+
git merge-base --is-ancestor 4f72d6f755 origin/dev # exit 0
20+
21+
## Gates, at the MERGED SHA
22+
23+
A merge commit is a tree nobody has tested until it is tested, so these ran against
24+
`4f72d6f755` itself — not the pre-merge tip — in a dedicated lidge worktree pinned to
25+
that SHA (`/tmp/ocx-dev-4f72d6f75`), with HEAD re-asserted before each gate.
26+
27+
| Gate | Result |
28+
|------|--------|
29+
| `bun x tsc --noEmit` | exit 0, no output |
30+
| `bun run privacy:scan` | `Privacy scan passed` |
31+
| `bun run audit:high` | `No vulnerabilities found` (root and gui) |
32+
| `bun run build:gui` | `✓ built in 205ms`, `prepare:package` ran |
33+
| `bun test --isolate tests` | green |
34+
35+
The same five were green at the pre-merge tip `ff4b0bb4e`. Both runs are in
36+
`.tmp/cursor-call-receipts.log`.
37+
38+
## Governance: this was owner authority, not compliance
39+
40+
Stated plainly because a readiness note that hides it is worthless.
41+
42+
`MAINTAINERS.md:48-49` requires maintainer approval **and** successful required CI
43+
checks before merge. The user waived CI and granted admin merge, and all three PRs
44+
were merged with `--admin`. That is the repository owner exercising owner authority
45+
over their own repository. It is **not** policy compliance, and this note does not
46+
claim it is.
47+
48+
The platform gap that follows: lidge is Linux; CI covers Linux, Windows and macOS.
49+
This campaign's diff contains no shim, installer, PowerShell, platform dispatch, or
50+
Windows path handling, which is why Linux evidence is adequate *for this diff*. It is
51+
not a claim that Linux equals CI.
52+
53+
Cross-platform CI did start on the merge commit and was `in_progress` when this note
54+
was written. Nothing here waited on it.
55+
56+
## What publishing would still require
57+
58+
`scripts/release.ts` is the release authority, and it waits for a successful
59+
**Cross-platform CI** run AND a successful **Service lifecycle** run at the exact
60+
release SHA (`:393-401`). So a "go" is not one command away: the release commit has
61+
to be pushed and both workflows have to pass at that SHA before the workflow
62+
dispatch. Anyone reading a green gate table and assuming `bun run release` will just
63+
work would be wrong.
64+
65+
## Version state, read live
66+
67+
live main e2d4621d431f3c0a97d67abbcd875b50c73ac661
68+
live dev 4f72d6f7555013ab231de78233d2aa95bd1e439c
69+
npm { preview: '2.23.0-preview.20260816', latest: '2.24.2' }
70+
releases v2.24.2 (latest), v2.24.1, v2.24.0
71+
72+
## Recommendation: 2.25.0
73+
74+
A minor rather than a patch, because the externally observable behaviour of a failed
75+
turn changed. A turn that previously came back `completed` with a vanished tool call
76+
now reports `failed` with a truncation error; an unrequested CANCEL that used to
77+
return silently is now a typed transport failure; and a truncated compaction turn no
78+
longer installs half-written replacement history. Anything downstream that keyed on
79+
"the proxy said completed" sees new behaviour — correct behaviour, but new.
80+
81+
The maintainer decides. This note recommends.
82+
83+
## Promotion sequence, prepared and NOT executed
84+
85+
# dev -> preview
86+
git fetch origin
87+
git checkout preview && git merge --no-ff origin/dev
88+
git push origin preview
89+
90+
# dev -> main (release train)
91+
git checkout main && git merge --no-ff origin/dev
92+
git push origin main
93+
94+
Preconditions before running either:
95+
96+
1. Cross-platform CI green at `4f72d6f755` (currently in progress).
97+
2. Service lifecycle green at the same SHA, since `package.json` is a trigger path
98+
for it and the release gate requires it.
99+
3. A version decision. The promotion itself does not bump; `scripts/release.ts` owns
100+
that and is the only sanctioned publish path — never a direct `npm publish`.
101+
102+
## Open follow-ups a reader would otherwise assume were fixed
103+
104+
1. **Cursor tool-result images do NOT reach production.** The encoder emits real
105+
`McpImageContent`, and that part is correct — but every Cursor model is in
106+
`noVisionModels` (`providers/registry.ts:978-982`), so the vision sidecar
107+
describes or strips images before the adapter runs
108+
(`core.ts:2225-2243`, `vision/index.ts:252-259,565-581`). The campaign's own
109+
docs overstated this once and were corrected. Closing it needs role-aware vision
110+
preprocessing plus an end-to-end regression through the server path.
111+
2. **Kiro** `completionMode: "disabled"` observes the normalized reason then emits
112+
`done` without `stopReason` (`kiro.ts:1315`, `:1485`), so `MAX_TOKENS` and
113+
`MODEL_CONTEXT_WINDOW_EXCEEDED` vanish.
114+
3. **Google ordinary mode** forwards only `MAX_TOKENS` plus five safety values
115+
(`google.ts:786-795`); `MALFORMED_RESPONSE`, `UNEXPECTED_TOOL_CALL`,
116+
`IMAGE_SAFETY` and `LANGUAGE` still become reasonless `done`.
117+
4. **User-message images** are still flattened in `cursor/request-builder.ts`.
118+
5. **Phase 030 (xai `apply_patch`) was NOT REPRODUCED.** A live probe had both
119+
`xai/grok-4.6` and `cursor/grok-4.6` using `apply_patch` successfully. dev's
120+
`bc229433a` + `8a4040384` independently fixed the code-mode guidance that had
121+
forbidden a separately-advertised top-level `apply_patch`, which is the same
122+
affordance surface 030 suspected. No code was written for it, deliberately.
123+
124+
## Issues closed with this work
125+
126+
`#1866` (Computer Use tool results empty/truncated) is the one this campaign
127+
actually answers. `#1992`, `#1938` and `#1527` were closed as part of the Cursor
128+
sweep with explicit notes about what the merge does and does not cover — #1992's
129+
injected-policy prose is untouched, #1938's integer-for-string coercion is a separate
130+
path, and #1527's rate-limit asymmetry is unaddressed while its silent-collapse half
131+
is fixed. Each says so and invites a reopen.
132+

0 commit comments

Comments
 (0)