Deploy prod

Skill seldonframe/seldonframe/.claude/skills/deploy-prod

Deploy main to Vercel production the safe way (clean worktree, env baked in) and smoke it with Reelier at Level 0 — zero tokens. Use after any merge to main; Vercel git integration does NOT reliably auto-deploy main.From its SKILL.md

Install

npx -y skills add seldonframe/seldonframe --skill deploy-prod

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

runs commandsInstructs the agent to run 8 commands, including `git fetch origin main` and 7 more.

SKILL.md

2.4 KB, 564 tokens by cl100k_base, as published. Nobody here has run it

deploy-prod — clean-worktree deploy + Reelier smoke

Vercel's git integration does not reliably auto-deploy main (verified 2026-07-18), and deploying from the dirty local checkout is forbidden. This is the canonical sequence.

Steps

Get the merge sha: git fetch origin main and note the commit to deploy.

Clean worktree deploy (env vars are baked at build, so this also picks up flag changes):

git worktree add "$env:TEMP\sf-main-deploy" <sha>
Copy-Item .vercel "$env:TEMP\sf-main-deploy\.vercel" -Recurse -Force
Set-Location "$env:TEMP\sf-main-deploy"; npx vercel deploy --prod

Verify: npx vercel inspect <deployment-url> → target production + status Ready.
Smoke with Reelier (default — NOT the smoke-runner agent):
```
Set-Location C:\Users\maxim\CascadeProjects\reelier
node dist/cli.js run skills/sf-post-deploy-smoke.skill.md
```
Level-0 deterministic replay: ~1-3s, 0 LLM tokens, per-step receipts in .reelier/runs/. Exit 0 = deploy smoked. This replaces a ~35-60k-token smoke-runner agent dispatch.
Escalation only on divergence: if a step fails, do NOT immediately re-run — a failed sentinel usually means the page copy changed (update the skill's assert) or the deploy is genuinely broken (check npx vercel logs). Dispatch the smoke-runner agent ONLY when the reelier failure needs judgment (e.g. deciding whether copy drift is intentional).
If the deploy touched migrations: verify on Neon that the new table/column exists (information_schema) — the build runs db:migrate, but confirm, don't assume.
Clean up: git worktree remove "$env:TEMP\sf-main-deploy" --force; git worktree prune.

Why Reelier-first

Deterministic replay with assertions IS the smoke test: same checks every time, receipts written, zero spend. The agent is the escalation path, not the default — the same L0-first/escalate-on-divergence ladder the replay product itself uses. When the smoke skill's sentinels drift because marketing copy changed, updating the assert line in reelier/skills/sf-post-deploy-smoke.skill.md is the fix (and gets a changelog line).

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Gives 0 of the 12 instructions most ship operate skills give in 564 tokens

Counted across 1,077 of the 1,713 authors here whose files we hold, read 2026-09-06

Create GitHub releasein 44 of 1077, across 43 files
Run the test suitein 30 of 1077, across 25 files
Create and push git tagin 27 of 1077, across 26 files
Push commits and tagsin 27 of 1077
Create annotated tagin 25 of 1077, across 22 files
Ensure working tree is cleanin 24 of 1077
Check for product marketing context firstin 23 of 1077, across 6 files
Commit version bump changesin 22 of 1077, across 21 files
Update CHANGELOG.mdin 21 of 1077, across 20 files
Structure launch marketing across three channel typesin 20 of 1077, across 5 files
Commit and tag the releasein 20 of 1077, across 18 files
Update the CHANGELOG for new releasesin 19 of 1077

Said here and by no other author read

Copy vercel configuration
Inspect the Vercel deployment
Smoke test with Reelier
Verify migrations on Neon

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.