Skip to content

Commit 1080f27

Browse files
committed
docs(devlog): close merge train round 3
Five landings, six closures, two rebases, and the flake that held the last PR. Records what the round is evidence of rather than only what it did: three explanations were written for the websocket flake and two were wrong, both plausible enough to justify the same fix -- caught by activation evidence, not review. The '#3128 fixed that flake' citation was repeated across three PRs and taught reviewers to dismiss a red that was real.
1 parent 58be3c5 commit 1080f27

2 files changed

Lines changed: 138 additions & 0 deletions

File tree

Lines changed: 57 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,57 @@
1+
# 061 — wp7 outcome: the flake is fixed, and the work-phase numbering is not
2+
3+
`c8c8dc338` — `test(auth): close the startup-prime window that rotates the credential
4+
mid-fixture (#3139)`. Merged with the roadmap unit in the same PR.
5+
6+
## Result
7+
8+
```
9+
gh pr checks 3139
10+
macos pass 11m40s
11+
ci pass 4s
12+
```
13+
14+
That is the verifier that matters. The same assertion failed on `macos` once for #3133 and
15+
twice for #3137, on heads without the fix. It passed on the **first** run of the fixed head.
16+
17+
Local: `bun test tests/server-auth.test.ts` -> 91 pass / 0 fail / 618 expect().
18+
19+
## Where wp7's work actually happened
20+
21+
In wp6, not wp7. The FSM's active work-phase was wp6 when the fix was written, and wp6 was
22+
the landing phase blocked by exactly this flake — so its plan absorbed the fix rather than
23+
the two units pretending to be independent.
24+
25+
Recording that plainly instead of back-dating an attest: wp7 was registered as a
26+
work-phase, its plan doc (`060`) is real and was written under it, and its implementation
27+
rode wp6's cycle. The ledger shows one cycle, which is what happened.
28+
29+
## What this phase is really a record of
30+
31+
Three explanations, two wrong, one measured — the table is in `060`. Both wrong ones were
32+
plausible, cited real mechanisms, and would have justified the same one-line fix. That is
33+
what made them dangerous rather than harmless: the fix would have worked, the reasoning
34+
would have been wrong, and the next person to touch this fixture would have inherited the
35+
wrong model.
36+
37+
What broke the tie was the runtime's own counter:
38+
39+
```
40+
$ OPENCODEX_DEBUG_QUOTA=1 bun test ... -t "websocket passthrough refreshes pool auth"
41+
[codex-quota] prime done (reason=startup, pool=1, refreshed=1)
42+
```
43+
44+
`refreshed=1` on five runs of the unfixed tree **and** five of the fixed one. Staleness never
45+
varied, so the "cache age crosses the TTL" story was dead — and the surviving explanation is
46+
that the prime always fetches, and what varied was whether it hit the stubbed `fetch` and
47+
pinned clock or the real ones.
48+
49+
`LOOP-MECHANISM-PROOF-01` asks for activation evidence before adopting a mechanism. Here it
50+
did more than confirm: it killed the hypothesis I had already written into a devlog document
51+
and two PR comments.
52+
53+
## Residual
54+
55+
The comments on #3109 and #3112 quote the first wrong explanation. They were left in place —
56+
their operational advice (rerun rather than read a single red as a regression) was correct,
57+
and is now moot because the flake is fixed. `051` carries the pointer to the correction.
Lines changed: 81 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,81 @@
1+
# 070 — outcome: merge train round 3
2+
3+
Terminal outcome: **DONE**. Every item in the round-3 scope reached a terminal state.
4+
5+
## What landed on `dev`
6+
7+
| commit | what | origin |
8+
| --- | --- | --- |
9+
| `abcda8e13` | 2026-08-31 non-priority-70 bug triage record | #3114 |
10+
| `0dc01cdaa` | canonical fake-IP addresses on provider PATCH | #3122 via #3133 |
11+
| `b14b741dc` | Windows cold-start budget + code-page scheduler paths | #3104 via #3134 |
12+
| `c8c8dc338` | startup-prime window fix + this roadmap unit | #3139 |
13+
| `58be3c5bb` | probe for a free pid instead of assuming 4242 is dead | #3042 via #3137 |
14+
15+
## Closed
16+
17+
Issues #3009, #3064. Pull requests #3104, #3122, #3042, #3077, and a credit comment on the
18+
already-closed #3067. Every closure names the merged commit and what changed from the
19+
original; none is a bare "superseded".
20+
21+
#3039 was closed by its own author at `2026-09-01T04:17:43Z`, not by this train. The comment
22+
recording which contribution #3104 did **not** carry — the elapsed-time diagnostic, replaced
23+
by the configured budget — landed anyway, so the residual is findable.
24+
25+
## Rebased, not merged
26+
27+
#3109 to `b3b502045` (five commits; `926a8d8c` dropped because it had already landed as
28+
#3128) and #3112 to `f3c4e9f75` (four commits). Both `range-diff`-identical. Neither PR's
29+
review blockers were touched — #3112's three credential-path findings still stand and it
30+
still needs a fresh security review.
31+
32+
## Untouched, deliberately
33+
34+
#3117 reverses a direction `b46164e78` pinned one day earlier and is a policy decision about
35+
#1690, not a mechanical. #3061 has a substantive rebuttal on record. Both were named OUT at
36+
wp0 and stayed out.
37+
38+
## What the round is actually evidence of
39+
40+
**Three wrong explanations, caught by measurement rather than review.** The websocket flake
41+
was explained three times: a 60 s skew margin (wrong — the margin is months), a cache age
42+
crossing a TTL (wrong — `refreshed=1` on every run of both trees), and finally the measured
43+
one. Both wrong versions were plausible, cited real code, and **would have justified the same
44+
fix**. That is what made them worth catching: the fix would have worked and the reasoning
45+
would have been wrong, which is how a fixture acquires folklore.
46+
47+
`LOOP-MECHANISM-PROOF-01` is why it was caught. Asking for activation evidence before
48+
adopting a mechanism killed a hypothesis already written into a devlog document and two PR
49+
comments.
50+
51+
**A citation can be worse than silence.** "#3128 fixed that flake" was repeated across three
52+
PRs and a release-train record. It was false — #3128 is an ancestor of every head that failed
53+
afterwards — and its effect was to teach reviewers to dismiss a red. The correction is now on
54+
#3109, #3112, #3104, and in `051` and `060`.
55+
56+
**A plan audit that returns FAIL is cheap.** Round 1 returned five blockers; three were
57+
folded and changed the train's shape — fork PRs became cherry-pick carries once
58+
`enforce-pr-target.yml:740-746` was read properly, and #3039's closure was withdrawn. Two
59+
were rebutted with evidence. The audit cost one subagent and prevented stranding two
60+
contributor PRs in draft.
61+
62+
**The test suite commits into the developer's checkout.**
63+
`tests/test-runner.test.ts` calls `commitFixture(cwd, ...)`, which makes a real commit in
64+
whatever worktree runs the suite. It rode along on the first push of two carry branches
65+
(author `OpenCodex Test <[email protected]>`, adding `base.txt`) and both had to be
66+
reset and force-pushed. Not fixed here — it is a real trap and belongs to its own unit.
67+
68+
## Verification summary
69+
70+
| check | result |
71+
| --- | --- |
72+
| `bun test tests/service.test.ts` (#3134 carry) | 191 pass / 0 fail |
73+
| `bun test tests/management-provider-validation.test.ts tests/destination-policy-resolved.test.ts` (#3133 carry) | 129 pass / 0 fail |
74+
| `bun test tests/responses-state.test.ts tests/doctor.test.ts tests/cli-status-json.test.ts` (#3137 carry) | 214 pass / 0 fail |
75+
| `bun test tests/server-auth.test.ts` (flake fix) | 91 pass / 0 fail |
76+
| exact-head CI on #3133, #3134, #3137, #3139 | fully green before each merge |
77+
78+
Every merge used `--admin`, because GitHub refuses self-approval and `dev` requires a
79+
reviewed PR. That is a real gap and worth stating rather than burying: what stood in for
80+
review was an independent security lane on #3122, direct maintainer audits on the rest, and
81+
a green exact-head matrix on all four.

0 commit comments

Comments
 (0)