latest-prerelease lookup strands users on 1.0.81-9
Summary
copilot update prerelease refuses to advance from 1.0.81-9 to 1.0.81-10, reporting the older build as the latest available release:
> copilot --no-auto-update update prerelease
Checking for updates...
Checking GitHub for the latest release...
No update needed, current version is 1.0.81-9, fetched latest release is v1.0.81-9
At the time of that run, v1.0.81-10 had been published for ~1.5 hours and was not a draft:
> gh api repos/github/copilot-cli/releases/tags/v1.0.81-10 \
--jq '"tag=\(.tag_name) draft=\(.draft) prerelease=\(.prerelease) published=\(.published_at) assets=\(.assets|length)"'
tag=v1.0.81-10 draft=false prerelease=true published=2026-08-25T21:15:43Z assets=20
Root cause
Every 1.0.81-N release shares an identical created_at, so GitHub's list endpoint cannot order them by creation time and falls back to a lexicographic tiebreak. "10" sorts below "2", so -10 lands in the middle of the list rather than at the top:
> gh api "repos/github/copilot-cli/releases?per_page=12" \
--jq '.[] | "\(.tag_name) created=\(.created_at) published=\(.published_at)"'
v1.0.81-9 created=2026-08-14T20:20:00Z published=2026-08-24T17:42:32Z
v1.0.81-8 created=2026-08-14T20:20:00Z published=2026-08-23T14:46:46Z
v1.0.81-7 created=2026-08-14T20:20:00Z published=2026-08-21T18:39:24Z
v1.0.81-6 created=2026-08-14T20:20:00Z published=2026-08-20T17:59:54Z
v1.0.81-5 created=2026-08-14T20:20:00Z published=2026-08-19T23:16:31Z
v1.0.81-4 created=2026-08-14T20:20:00Z published=2026-08-19T18:23:39Z
v1.0.81-3 created=2026-08-14T20:20:00Z published=2026-08-19T08:46:57Z
v1.0.81-2 created=2026-08-14T20:20:00Z published=2026-08-19T04:28:01Z
v1.0.81-10 created=2026-08-14T20:20:00Z published=2026-08-25T21:15:43Z <-- newest, ranked 9th
v1.0.81-1 created=2026-08-14T20:20:00Z published=2026-08-18T18:30:50Z
v1.0.81-0 created=2026-08-14T20:20:00Z published=2026-08-14T23:47:01Z
v1.0.80 created=2026-08-10T16:19:17Z published=2026-08-14T02:28:39Z
The selection appears to take the first prerelease in that default ordering. In app.js, the update routine logs the message above from:
if (nR.lte(s.tag_name, Kh())) {
let u = `No update needed, current version is ${Kh()}, fetched latest release is ${s.tag_name}`;
...
}
where s originates from lx() → uht("latest-prerelease", ...) → native githubLookupRelease(...).
Worth emphasising: the comparison is not the bug. nR.lte is semver-aware and ranks 1.0.81-9 < 1.0.81-10 correctly (numeric prerelease identifiers compare numerically). The wrong value is the release that gets fetched. Since GitHub exposes no "latest prerelease" endpoint (/releases/latest returns v1.0.80, the newest stable), the native lookup must enumerate — and it trusts list order.
Impact
Any prerelease line that reaches -10 strands every user on -9 for the remainder of that line. This is not self-correcting: it persists until a new stable or minor version is published. It affects both explicit copilot update prerelease and the automatic startup update check, on all platforms.
Steps to reproduce
- Be on
1.0.81-9 (or install it).
- Run
copilot --no-auto-update update prerelease.
- Observe
No update needed ... fetched latest release is v1.0.81-9, despite v1.0.81-10 being published.
The --no-auto-update prefix matters for a clean repro: without it, the startup auto-update check inflates the reported running version, so a plain copilot update prerelease compares against an already-elevated value and no-ops for a second, unrelated reason.
Suggested fix
Don't rely on list ordering. After enumerating non-draft releases, select the maximum by semver (or by published_at) before comparing against the running version. A one-line jq equivalent that resolves correctly today:
[.[] | select(.draft == false)] | sort_by(.published_at) | reverse | .[0].tag_name
Separately, giving each release a distinct created_at at publish time would remove the ambiguity at the source, though the client-side sort is the durable fix.
Workaround
Download the release asset for the desired tag directly and replace the binary, bypassing the update check entirely.
Environment
- Windows 11,
win32-x64
- Copilot CLI
1.0.81-9 (standalone exe), attempting to reach 1.0.81-10
latest-prereleaselookup strands users on1.0.81-9Summary
copilot update prereleaserefuses to advance from1.0.81-9to1.0.81-10, reporting the older build as the latest available release:At the time of that run,
v1.0.81-10had been published for ~1.5 hours and was not a draft:Root cause
Every
1.0.81-Nrelease shares an identicalcreated_at, so GitHub's list endpoint cannot order them by creation time and falls back to a lexicographic tiebreak."10"sorts below"2", so-10lands in the middle of the list rather than at the top:The selection appears to take the first prerelease in that default ordering. In
app.js, the update routine logs the message above from:where
soriginates fromlx()→uht("latest-prerelease", ...)→ nativegithubLookupRelease(...).Worth emphasising: the comparison is not the bug.
nR.lteis semver-aware and ranks1.0.81-9 < 1.0.81-10correctly (numeric prerelease identifiers compare numerically). The wrong value is the release that gets fetched. Since GitHub exposes no "latest prerelease" endpoint (/releases/latestreturnsv1.0.80, the newest stable), the native lookup must enumerate — and it trusts list order.Impact
Any prerelease line that reaches
-10strands every user on-9for the remainder of that line. This is not self-correcting: it persists until a new stable or minor version is published. It affects both explicitcopilot update prereleaseand the automatic startup update check, on all platforms.Steps to reproduce
1.0.81-9(or install it).copilot --no-auto-update update prerelease.No update needed ... fetched latest release is v1.0.81-9, despitev1.0.81-10being published.The
--no-auto-updateprefix matters for a clean repro: without it, the startup auto-update check inflates the reported running version, so a plaincopilot update prereleasecompares against an already-elevated value and no-ops for a second, unrelated reason.Suggested fix
Don't rely on list ordering. After enumerating non-draft releases, select the maximum by semver (or by
published_at) before comparing against the running version. A one-line jq equivalent that resolves correctly today:Separately, giving each release a distinct
created_atat publish time would remove the ambiguity at the source, though the client-side sort is the durable fix.Workaround
Download the release asset for the desired tag directly and replace the binary, bypassing the update check entirely.
Environment
win32-x641.0.81-9(standalone exe), attempting to reach1.0.81-10