[jahands]

Result: about 80–160s off each release (−13% to −23%)

Since Windows arm64 stopped being the last desktop job, about a third of a release is the jobs after the desktop builds, which run one after another: Publish GitHub Release, then npm → hosted web → Discord, with AUR alongside. On the last two cached nightlies these took 204–273s. Most of that was checkout and Setup Vite+ installing the workspace to run a 1–6s step: npm publish spent 49–62s on setup plus 9–20s in post-step cleanup.

Old and new jobs, side by side in three runs on a fork (same runners, publish steps dry-run or pointed at a local webhook):

Job	Before	After
Publish GitHub Release, setup + manifest merge	22–39s	8–16s
Publish CLI to npm	37–38s	9–12s
Deploy hosted web app	23–27s	16–22s
Announce on Discord	35–41s	13–14s
AUR, checkout step	12–15s	2s

Estimate per release, applying those savings to the last two cached nightlies: 639s → ~556s and 713s → ~552s. After this, AUR (59–86s, mostly building the package) is about as long as the npm → web → Discord chain, so trimming either one further gains little.

Pipeline, before and after

Before: every job after the desktop builds checks out the repo and installs the workspace before doing a few seconds of work (times from upstream nightlies 39543980476 and 39505921112):

flowchart LR
  D["Desktop builds"] --> R
  R["Publish GitHub Release · 68–97s<br/>checkout + install 24–46s<br/>merge manifests, upload"] --> N
  N["Publish CLI to npm · 67–91s<br/>checkout + install 49–62s<br/>build t3.tgz, npm publish"] --> W
  W["Deploy hosted web · 25–31s<br/>checkout + setup 14–16s<br/>vercel alias"] --> X
  X["Announce on Discord · 30–37s<br/>checkout + install 26–33s<br/>post webhook"]
  R --> A["AUR · 59–86s<br/>full checkout 15–22s<br/>makepkg, push"]

Loading

After: a new job builds those inputs while the desktop builds run, and the jobs after them only download them (times from the fork bench):

flowchart LR
  P["Prepare publish tools · ~41s<br/>build t3.tgz, bundle<br/>merge + Discord scripts"] -. "publish-tools artifact" .-> R
  P -.-> N
  P -.-> X
  D["Desktop builds · ~250s"] --> R
  R["Publish GitHub Release<br/>download 2s<br/>merge manifests, upload"] --> N
  N["Publish CLI to npm · 9–12s<br/>package.json + setup-node<br/>npm publish"] --> W
  W["Deploy hosted web · 16–22s<br/>package.json + setup<br/>vercel alias"] --> X
  X["Announce on Discord · 13–14s<br/>download 2s<br/>post webhook"]
  R --> A["AUR<br/>packaging/ checkout 2s<br/>makepkg, push"]

Loading

Prepare publish tools starts right after preflight and runs alongside the ~250s desktop builds, so nothing waits on it.

Change

• New prepare_publish job, alongside the desktop builds and off the critical path: builds t3.tgz and esbuild-bundles merge-update-manifests and notify-discord-release (same approach as the Windows verify scripts), plus a copy of mirror-play-apk.cjs, into one artifact. release now needs it.
• release and announce_discord download that artifact instead of checking out and installing. The scripts run on the runner's own Node (v22 on Blacksmith images); import.meta.main is defined at bundle time so they run on Node versions that lack it.
• publish_cli runs npm publish --provenance ./t3.tgz after a package.json-only checkout and actions/setup-node, so npm still comes from the repo's Node 24 (npm 11.19; trusted publishing needs 11.5.1+). It stays on the GitHub-hosted runner. cli.ts publish had no other callers and is removed.
• deploy_web checks out only package.json for Setup Vite+. It keeps vp dlx vercel; npx vercel took 23s against 3s.
• publish-aur.yml checks out only packaging/, which is all packaging/aur/scripts/release.sh reads.
• Job ordering is unchanged: npm after the release, web after npm, Discord after web.

Parity, from the fork runs: merged updater manifests are byte-identical, the npm tarball has identical file contents (npm publish --dry-run output matches), and the Discord webhook payload is identical apart from its timestamp.

Not tested on the fork

The fork has no npm trusted publisher, Vercel or Discord secrets, and the AUR job only runs on pingdotgg/t3code. The real npm publish --provenance, the Vercel alias, the Discord post and makepkg from the sparse checkout are first exercised on the next upstream nightly.

Built with Claude Opus 5.5 via Claude Code.

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	5.0 KiB	5.0 KiB	−7 B (−0.1%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.2 KiB	1.2 KiB	−7 B (−0.6%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	20.9 KiB	20.9 KiB	0 B (0.0%)	29.3 KiB	✅
Codex	Live turn messages	2	2	0 (0.0%)	8	✅
Claude	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	21.2 KiB	21.2 KiB	0 B (0.0%)	29.3 KiB	✅
Claude	Live turn messages	2	2	0 (0.0%)	8	✅

Baseline: 21da306 · PR result: 0d892b5 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.