|
| 1 | +# 090 — Merge-loop outcome |
| 2 | + |
| 3 | +HOTL loop, session `01a01949`, ten work-phases. **14 merged, 8 held.** |
| 4 | + |
| 5 | +## Merged |
| 6 | + |
| 7 | +| Work-phase | PRs | |
| 8 | +|---|---| |
| 9 | +| wp2 | #2085, #2086 | |
| 10 | +| wp3 | #2035, #2031, #1878 | |
| 11 | +| wp4 | #2103 | |
| 12 | +| wp5 | #1876 (closes #1852) | |
| 13 | +| wp6 | #2112, #1934, #2080 | |
| 14 | +| wp7-9 | #2019, #2023, #2036 — the split stack, bottom-up | |
| 15 | +| wp10 | #2119 — this loop's own two fixes | |
| 16 | + |
| 17 | +## Held, with posted blockers |
| 18 | + |
| 19 | +#2102, #2105, #2053, #2100, #2077, #2056, #2062, #2063 — all verified still |
| 20 | +OPEN at closeout. Nothing was swept in by a batch merge. |
| 21 | + |
| 22 | +## What the batching rule actually bought |
| 23 | + |
| 24 | +The user's instruction was to merge in small batches rather than sweeping |
| 25 | +stage 1 at once. Three defects were caught that a sweep would have shipped |
| 26 | +behind green CI: |
| 27 | + |
| 28 | +**#2105** — `reconcileShellHook(false)` unconditionally removes the hook, but |
| 29 | +`false` also means "`claude` is not on **this process's** `PATH`". A |
| 30 | +service-started proxy does not inherit the interactive login shell's `PATH`, so |
| 31 | +a user with Claude Code installed would have had their working `.zshrc` hook |
| 32 | +deleted. The change made the safe direction conditional and left the destructive |
| 33 | +one unconditional. |
| 34 | + |
| 35 | +**#1876** — the async catalog fix returned a pre-write `fresh` to a request that |
| 36 | +was already waiting when an invalidating write landed. `fresh` is the one state |
| 37 | +authorizing positive model guidance, so it would advertise a catalog the |
| 38 | +app-server no longer had. **A slow answer was the bug; a wrong answer is worse |
| 39 | +than the bug.** Fixed on our branch, and the existing regression had been |
| 40 | +asserting the defective `fresh` — the test was pinning the bug. |
| 41 | + |
| 42 | +**#2080** — a paid priority tier enabled from an assertion rather than evidence. |
| 43 | +Being wrong charges the user. |
| 44 | + |
| 45 | +## Three things this loop taught that outlive it |
| 46 | + |
| 47 | +**1. A verdict is bound to a head.** Every work-phase re-reviewed because heads |
| 48 | +had moved, and it mattered twice: #2102's author had changed the exact code the |
| 49 | +earlier verdict covered, and #2086's moved head added a `noVisionModels` |
| 50 | +precedence fix that was not in the reviewed diff. |
| 51 | + |
| 52 | +**2. Stale base is a claim about the base, not the change — four times over.** |
| 53 | +#2031 (60 behind), #1876 (67), #2019/#2023/#2036 (29). Each had red CI that went |
| 54 | +green on rebase with no source change. The converse stayed honest: a rebase |
| 55 | +removes the base as an explanation, it does not prove the change is good. |
| 56 | + |
| 57 | +**3. "No test is possible" was a claim, not a fact.** This document asserted |
| 58 | +across three campaigns that a barrel extraction's oracle is `tsc` plus its |
| 59 | +importers, and that a barrel test "restates the compiler". A review lane |
| 60 | +disproved it: forking `MODEL_ADAPTER_OVERRIDE_ALLOWED` into a second `Set` |
| 61 | +inside the barrel leaves `tsc --noEmit` at exit 0, and no test imported the |
| 62 | +leaves directly, so barrel and leaf were never compared. That test now exists |
| 63 | +(`tests/types-barrel-identity.test.ts`) and drives red against exactly that |
| 64 | +fork. |
| 65 | + |
| 66 | +The pattern across all three: **the conclusions held and the reasons did not.** |
| 67 | +Same finding as the previous campaign's closeout, arrived at independently. |
| 68 | + |
| 69 | +## A process failure worth recording |
| 70 | + |
| 71 | +wp6 held three PRs with three blockers. Two got follow-up fixes before merging; |
| 72 | +#2112's did not, and it landed with `codexToolMode` still absent from |
| 73 | +`providerConfigSchema`. Because that schema ends in `.passthrough()`, nothing |
| 74 | +failed — a misspelled value was accepted and silently resolved to the default. |
| 75 | + |
| 76 | +Caught at closeout and fixed (`d697e2553`), but the honest reading is that the |
| 77 | +hold did not hold. **A hold is worth exactly what the re-check before merge is |
| 78 | +worth**, and a passthrough schema leaves no trace at the merge boundary. |
| 79 | + |
| 80 | +## Still open |
| 81 | + |
| 82 | +- The 8 held PRs need author responses. |
| 83 | +- The preview/soak gate from `070` has not run. The split is on `dev`; it has |
| 84 | + not been exercised as a published build. |
| 85 | +- `dev` CI on the final head was still in flight at closeout; no failing leg was |
| 86 | + observed on any run in this loop. |
0 commit comments