Skip to content

Commit 43e470a

Browse files
committed
devlog: close the v2.32.1 train — GO/NO-GO, and two deferrals
The readiness report freezes dev at faaa78d with the verdict GO, and records the audit that got it there: the first freeze at 02c302a was rejected, correctly, on three counts. Two of the three unresolved review threads it found were live defects that had been opened minutes before their PRs merged and so were never addressed — a malformed namespace authorizing a tool alias, and the snapshot fast path never restoring broadened permissions on a file that holds request and response bodies. Both are fixed and both needed a second pass, because the first fix for the namespace case was itself incomplete against a pre-flattened wire name. The third finding was the fairest: the full-suite gate was red and the report argued an exception for it. Arguing an exception in the document that reports the result is gate-weakening after the fact. The gate is now decomposed the way CI actually partitions the suite, and passes in that form: 14565 pass across the general batches, storage-policy green in its own job, with the single api-usage failure proven identical on the untouched pre-train baseline. Also records the two units that closed without merging. #2472 is NOT_REPRODUCED: a regression was written, passed, and then deleted once review showed it pinned the pre-execution announcement path rather than the post-execution loss the issue describes. #2427 is deferred on five runs of data — four different tests flaked, each green in isolation — because a runner that fails one run in three would make the freeze gate itself unfalsifiable.
1 parent faaa78d commit 43e470a

3 files changed

Lines changed: 247 additions & 0 deletions

File tree

‎devlog/_plan/260824_v2_32_1_hotfix_train/070_wp2_pr2427_parallel_test_runner.md‎

Lines changed: 65 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -73,3 +73,68 @@ IN: `scripts/test.ts`, `bunfig.toml`, `tests/test-runner.test.ts`.
7373
OUT: #2429 (`test:changed`), which is stacked on this PR and belongs to the next
7474
minor.
7575

76+
---
77+
78+
## Outcome (wp2 close, 2026-08-25): DEFERRED
79+
80+
Ran last, exactly as the roadmap audit required, and the deferral is this
81+
document's own rule applied rather than a new judgement: *"If it does not, it is
82+
deferred and the train proceeds on the existing runner. It is a convenience,
83+
never a blocker."*
84+
85+
### The measurement
86+
87+
Five full-suite runs on an idle machine, across two heads. The PR head moved
88+
mid-phase — the author pushed `cdeda10c` bounding the default to
89+
`--parallel=4` while the first runs were in flight, so the first two rows are
90+
stale and are kept only to show the bound's effect.
91+
92+
| head | workers | result | wall |
93+
|------|---------|--------|------|
94+
| `e03b9fca` | 15x | 14601 pass / 0 fail | 47s |
95+
| `e03b9fca` | 15x | 14600 pass / **1 fail** (`codex-shim`) | 47s |
96+
| `cdeda10c` | 4x | 14601 pass / 0 fail | 130s |
97+
| `cdeda10c` | 4x | 14601 pass / 0 fail | 125s |
98+
| `cdeda10c` | 4x | 14600 pass / **1 fail** (`issue-452`) | 123s |
99+
100+
Serial baseline on the same machine: ~560s. The speedup is real and the four-worker
101+
bound measurably reduces the failure rate. Neither fact was the deciding one.
102+
103+
### Why it was deferred
104+
105+
Four **different** tests failed intermittently across those runs —
106+
`cursor-native-exec-shell`, `openai-provider-option-e2e`, `codex-shim`,
107+
`issue-452-empty-503` — and every one passes in isolation (17/17 and 88/88
108+
respectively). An independent reviewer additionally had one run stop emitting
109+
output for ten minutes without a terminal summary.
110+
111+
These are **pre-existing latent order dependencies that parallelism exposes**, not
112+
defects the PR introduces. The reviewer named a concrete mechanism worth chasing:
113+
`scripts/test.ts` supplies one common startup `HOME`, and `homedir()` is fixed at
114+
process start, so the `.claude` sentinel in `openai-provider-option-e2e` can observe
115+
a path shared across workers even after preload rewrites the environment.
116+
117+
The blocking argument is specific to this train's position: wp8's freeze gate **is**
118+
a full-suite run, and every remaining criterion depends on it meaning something. A
119+
runner that fails roughly one run in three for unrelated reasons makes a red result
120+
indistinguishable from noise — the same attribution problem that moved this PR from
121+
first to last, arriving one step later.
122+
123+
### What would land it
124+
125+
Fix or quarantine the order-dependent tests (the shared-`HOME` sentinel first), then
126+
three consecutive green full-suite runs at one head plus Linux and Windows CI. It
127+
belongs early in the next cycle: 2 minutes versus 9 changes how often the suite gets
128+
run at all.
129+
130+
Criterion c-7's #2427 half is met by this recorded deferral. Nothing in the v2.32.1
131+
train depends on it; wp8 freezes on the existing serial runner. Posted to the PR at
132+
https://github.com/lidge-jun/opencodex/pull/2427#issuecomment-5402835810.
133+
134+
**LOOP-PESSIMIST-01.** What died here is my own P-phase recommendation: I reached A
135+
saying MERGE on one green run at what turned out to be a stale head. Two lessons,
136+
both cheap to state and easy to skip: re-read the PR head immediately before
137+
claiming exact-head evidence, because a contributor can push mid-verification; and
138+
one green run of a flaky-capable suite is not evidence of stability — the third run
139+
is what produced the finding.
140+

‎devlog/_plan/260824_v2_32_1_hotfix_train/090_wp9_issue2472_mixed_sequence_regression.md‎

Lines changed: 64 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -84,3 +84,67 @@ where the fix goes.
8484
OUT: implementing that fix inside this phase; restarting or reconfiguring the
8585
user's running proxy; any live provider call.
8686

87+
---
88+
89+
## Outcome (wp9 close, 2026-08-25)
90+
91+
**Terminal outcome: cannot be driven deterministically → #2472 deregistered as a GO
92+
criterion and recorded as a deferred known defect.** This is the third of the three
93+
outcomes this document allowed, and it is the honest one.
94+
95+
### What was attempted
96+
97+
The planned regression was written: `tests/cursor-zero-output-turn.test.ts`, seven
98+
tests driving the native/host call-id dedupe, including three routes that each
99+
produce a turn whose only event is the terminal `done`. It passed. It was then
100+
**deleted**, because an independent review showed it pins the wrong mechanism.
101+
102+
### Why it was wrong
103+
104+
The dedupe lives on the **pre-execution announcement** side of the tool boundary.
105+
`planMcpArgsHandling` deliberately ends turn 1 as `done` and cancels the Cursor run
106+
without a result — `live-transport.ts:220-225` states outright that the real tool
107+
result arrives on the NEXT `/v1/responses` request as structured history. #2472
108+
reports output lost **after** the calling agent already produced non-empty text,
109+
which is downstream of that boundary. A test that reproduced an empty-looking turn
110+
on the announcement side would have looked like evidence while proving nothing.
111+
112+
The other two routes were equally unreachable: the empty-argument case only goes
113+
silent under `allowEmptyArgs: false`, and the live bridge passes `true`
114+
(`live-transport.ts:258`), where malformed shell arguments raise an explicit error.
115+
116+
### What is settled
117+
118+
- The bridge-version theory from the issue's own point 3 is closed: `88b7cc057`
119+
(zero-output combo failover) is an ancestor of `dev`, and the regression the issue
120+
asked for exists and passes — `tests/combo-stream-preflight.test.ts`, *"converts a
121+
zero-output failed terminal into a retryable HTTP failure"*, 4 pass / 0 fail.
122+
- The dedupe is correct and stays. Without it every repeated `tool_call_start`
123+
becomes another Responses `function_call` item, i.e. a duplicate execution request.
124+
125+
### What remains open
126+
127+
There is no turn-wide semantic-output ledger. `finalizeTurnEvents` reports an error
128+
for a call left OPEN at turn end but is silent for a turn that closed with zero
129+
output, and the bridge emits `response.completed` with an empty snapshot. The
130+
`empty-completion-guard` would catch it but is opt-in and defaults to false. The gap
131+
is real **if a reachable producer exists**; none was constructible on current `dev`.
132+
133+
Settling it needs a reproduction at the `function_call_output`/next-request boundary,
134+
not another adapter-level probe. The microsecond `wall_time_seconds` in the report is
135+
the strongest remaining lead and deserves its own telemetry issue.
136+
137+
### Consequence for this train
138+
139+
Per 000's canary section, #2472 **stops being a GO criterion**. Criterion c-9 is met
140+
by this recorded disposition rather than by a passing canary. Nothing about the six
141+
merged runtime fixes depends on it, and the issue stays open with the investigation
142+
posted at https://github.com/lidge-jun/opencodex/issues/2472#issuecomment-5402463174.
143+
144+
**LOOP-PESSIMIST-01.** The hypothesis that died is mine: that the zero-output symptom
145+
could be reproduced from the adapter's event mapper. Two cycles in a row (wp7's
146+
config-dir guard, wp9's dedupe theory) I built a plausible mechanism and had to
147+
discard it against evidence. The pattern worth carrying: a reproduction that only
148+
exercises code I chose to call is not a reproduction — it has to start from the
149+
reported observable and work backwards to a path the runtime actually takes.
150+
Lines changed: 118 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,118 @@
1+
# 900 — v2.32.1 release-candidate readiness: GO/NO-GO
2+
3+
Frozen `dev` SHA: **`faaa78dc05489625e5c9bf450050a46a7fa91d1f`**
4+
Train started from: `origin/dev` `c44e43f00`, `origin/main` `96e2f67c3` (v2.32.0)
5+
Report written: 2026-08-25. Supersedes an earlier draft frozen at `02c302a54`,
6+
which a freeze audit rejected — see "What the audit changed" below.
7+
8+
## Verdict
9+
10+
**GO** for promoting `dev` → `main` and publishing **v2.32.1** as a bugfix-only
11+
release. Promotion, tagging, and publishing were deliberately not performed; they
12+
are human decisions and this unit ends before them.
13+
14+
## What landed
15+
16+
| # | PR | Merge SHA | What review changed |
17+
|---|-----|-----------|---------------------|
18+
| wp1 | #2487 | `73a11a8f1` | Baseline was misread as divergence; `dev` was an *ancestor* of `main` |
19+
| wp3 | #2483 | `3e3a028fe` | The fix regressed `claude-opus-4-8.1`; tail corrected to `(?!\d)` |
20+
| wp4 | #2481 | `a60d51748` | **My** replacement was disproved and reverted |
21+
| wp5 | #2473 | `84ade0f15` | Clean — the only unit needing no code correction |
22+
| wp6 | #2477 | `1d4a92a32` | Two defects: the flagged `allowed_tools` hole and a cross-kind residual |
23+
| wp7 | #2476 | `02c302a54` | Skip-on-digest was a persistence regression |
24+
| — | #2500 | `43227ac07` | Post-merge: malformed `namespace`; unrestored file permissions |
25+
| — | #2501 | `faaa78dc0` | Post-merge: malformed selector using a pre-flattened wire name |
26+
27+
Two units closed without a merge, both pre-registered outcomes: **#2472**
28+
NOT_REPRODUCED (deregistered as a GO criterion), **#2427** DEFERRED (three runs
29+
gave 0/0/1 failures; four different tests flaked across five runs).
30+
31+
## What the audit changed
32+
33+
The first freeze at `02c302a54` was audited and **failed**, correctly, on three
34+
counts. All three are now closed:
35+
36+
1. **Three unresolved review threads on merged PRs**, which the GO criteria forbid.
37+
Two were live defects that had been opened minutes before their PRs merged: a
38+
malformed `namespace` authorizing an alias, and the snapshot fast path never
39+
restoring broadened file permissions on a file holding request/response bodies.
40+
Fixed in #2500, then #2501 after review found #2500 was itself incomplete — a
41+
malformed selector carrying an already-flattened wire name still matched the
42+
alias map exactly. **All threads across all seven PRs are now resolved: 0.**
43+
2. **The full-suite gate was red** and the first draft argued an exception in the
44+
report itself. That is retroactive gate-weakening and the audit was right to
45+
reject it. The gate is now decomposed the way CI actually partitions it, below.
46+
3. **Missing frozen-head receipts.** Recorded below.
47+
48+
## Gate results at the frozen SHA
49+
50+
CI partitions the suite because three files are known to be load-sensitive:
51+
`scripts/ci/run-bun-test-batches.sh:50` excludes `api-storage-policy*`,
52+
`api-storage`, and `api-usage` from the general batches, and `ci.yml` runs each in
53+
its own job. Running `bun run test` as one process is therefore *not* the same
54+
gate CI applies. Both forms are recorded:
55+
56+
| Gate | Command | Result |
57+
|------|---------|--------|
58+
| Typecheck | `bun x tsc --noEmit` | exit 0 |
59+
| Privacy | `bun run privacy:scan` | `Privacy scan passed`, exit 0 |
60+
| **General suite** (CI's partition) | `bun test --isolate` over 1787 files, excluding the three segregated | **14565 pass, 0 fail, exit 0** |
61+
| Storage-policy job | `bun test --isolate` over the six files `ci.yml` names | **9 pass, 0 fail, exit 0** |
62+
| api-usage job | `bun test --isolate ./tests/api-usage.test.ts` | 31 pass, **1 fail** — see below |
63+
| Whole suite in one process | `bun run test` | 14604 pass, 3 fail — the segregated storage-policy family |
64+
65+
**The one `api-usage` failure is pre-existing and environmental.** The same test
66+
fails identically on the untouched pre-train baseline `c44e43f00`, no merged unit
67+
touches usage or overlay code, and CI's own `api usage` job is green at this SHA.
68+
It is a local-environment artifact, not a candidate defect.
69+
70+
## Push-event CI at the frozen SHA
71+
72+
Run **`32793104507`**, event `push`, head `faaa78dc0`, conclusion **success**.
73+
Every job green: `test 1/4`–`4/4`, `storage policy`, `api usage`, `gates`,
74+
`macos`, `keyring` and `npm-global` on ubuntu/windows/macOS, and the `ci`
75+
aggregate. This is the artifact `release.yml` requires.
76+
77+
## GO conditions
78+
79+
| Condition | Evidence |
80+
|-----------|----------|
81+
| `main` lineage in `dev` | `git merge-base --is-ancestor origin/main origin/dev` → 0 |
82+
| Version line synced | `origin/dev:package.json` → `2.32.0` |
83+
| Every PR merged post-wp1 | eight merge SHAs, each `--is-ancestor` verified |
84+
| Maintainer approval | `reviewDecision=APPROVED` on all merged PRs |
85+
| **Zero unresolved review threads** | **0 across #2483, #2481, #2473, #2477, #2476, #2500, #2501** |
86+
| #2477 security review | independent lane recorded in wp6; both follow-ups landed |
87+
| Push-event CI green at frozen SHA | run `32793104507` success |
88+
| Typecheck / privacy / general suite | all exit 0 at `faaa78dc0` |
89+
| Scope clean | `origin/main..origin/dev` is devlog + the runtime units only; no package/lockfile/workflow delta, no tag |
90+
91+
## NO-GO conditions, each checked
92+
93+
| Condition | Status |
94+
|-----------|--------|
95+
| Foreign tool-type selector authorizes a namespace alias | **Closed** (wp6) |
96+
| Malformed `namespace` authorizes an alias | **Closed** (#2500, #2501) |
97+
| Oversized turn opens a socket before falling back | **Closed** (wp5) |
98+
| Snapshot skip loses state or permissions | **Closed** (wp7, #2500) |
99+
| #2476 changed the 24 MiB cap, TTL, or eviction order | **Not changed** |
100+
| A hygiene-blocked PR reached the train | **None** |
101+
| A merge justified by remembered results | **None** — every close carries a bound receipt |
102+
| New runtime feature work after freeze | **None** |
103+
104+
## Known defects shipping in v2.32.1
105+
106+
**#2407** (Kiro `tool_search`), **#2458** (Gemini video 502 — deferred because its
107+
fix touches the guard wp6 hardened), **#2459** (Windows reinstall module graph),
108+
**#2472** (zero-output, not reproduced), **#2491** (four divergent slug-equivalence
109+
relations, filed during this train). Queue at freeze: ~50 open PRs, 20 open `bug`
110+
issues. The train deliberately took six.
111+
112+
## What promotion still requires
113+
114+
Left to a human, per `MAINTAINERS.md`: the `dev` → `main` promotion and its
115+
version bump to `2.32.1`, fresh Cross-platform CI **and** Service lifecycle runs at
116+
the promoted `main` SHA (the `package.json` bump activates that gate), the tag, and
117+
`npm publish` with dry-run and install verification.
118+

0 commit comments

Comments
 (0)