DEV Community: Shachar Solomon

I forgot I installed my own CI tool. It caught me shipping 4 bugs.

Shachar Solomon — Sat, 11 Apr 2026 18:01:02 +0000

Four type errors didn't ship to main last week.

Not because I caught them. I'd already done the ritual — pnpm lint, npx tsc --noEmit -p apps/api/tsconfig.json, pnpm test. Three greens. I was one git push away from landing all four on main.

A pre-push hook I'd forgotten I'd installed stopped the push and said:

typecheck failed — 4 errors

I sat there for a second. I just checked. With my own eyes. Three times.

It was right. I was wrong. Here's exactly what it found, in the order it found them.

Bug 1: coverage enforcement had been silently off for weeks

vitest.config.ts(19,7): error TS2769: No overload matches this call.
  Object literal may only specify known properties,
  and 'lines' does not exist in type 'CoverageOptions'.

Vitest v4 moved coverage thresholds from the top level into a nested thresholds: {} key. My config had lines/functions/branches/statements at the old level:

// before — silently ignored
coverage: {
  provider: 'v8',
  reporter: ['text', 'json', 'html'],
  exclude: [/* ... */],
  lines: 70,
  functions: 70,
  branches: 70,
  statements: 70
}

// after
coverage: {
  provider: 'v8',
  reporter: ['text', 'json', 'html'],
  exclude: [/* ... */],
  thresholds: {
    lines: 70,
    functions: 70,
    branches: 70,
    statements: 70
  }
}

Vitest reads the old key, shrugs, ignores it, enforces nothing. My tests had been "passing coverage" for an indeterminate number of weeks with zero coverage actually checked.

No lint rule catches this. No test catches this. It requires tsc pointed at vitest.config.ts — which nothing in my dev loop ever did, because root-level config files don't live inside any workspace package's tsconfig.json.

Bug 2: a string that forgot to be a union

tests/services/tenant.test.ts(37,41): error TS2345:
  Type 'string' is not assignable to type
  '"us-east-1" | "us-west-2" | "eu-west-1" | "ap-southeast-1"'.

Classic TypeScript widening. config: { region: 'us-east-1' } — TypeScript widens the literal 'us-east-1' to string, which isn't assignable to the Region union my function actually requires. Fix is one as const:

// before
const tenantData = {
  orgId: 'org-123',
  name: 'Acme Corp',
  domain: 'acme.com',
  config: { region: 'us-east-1' }          // widened to string
};

// after
const tenantData = {
  orgId: 'org-123',
  name: 'Acme Corp',
  domain: 'acme.com',
  config: { region: 'us-east-1' as const } // preserves the literal
};

Same structural problem as Bug 1: tests/services/ is a root-level test directory that lives outside any project's tsconfig.json include. My pnpm lint fans out to per-project tsc runs. None of them ever looked at this file.

Bug 3: `.avg` on a `number`

tests/services/metrics.test.ts(167,32): error TS2339: Property 'avg' does not exist on type 'number'.
tests/services/metrics.test.ts(168,32): error TS2339: Property 'min' does not exist on type 'number'.
tests/services/metrics.test.ts(169,32): error TS2339: Property 'max' does not exist on type 'number'.

This one embarrassed me most. I'd refactored aggregateMetrics to return number[] at some point and forgotten to update the test. The test was still asserting aggregated[0].cpu.avg, .min, .max — properties that don't exist on a primitive number.

In TypeScript: three explicit errors. In JavaScript: .avg on a number silently returns undefined, then .cpu on undefined throws at runtime, but only for the specific shape of the test data I happened to pass. On slightly different inputs it might silently return nonsense. On the inputs my test used, it happened to look fine.

Same structural problem as Bugs 1 and 2. Same dev-loop blind spot. Same fix: point tsc at the file.

Bug 4: the meta-bug that hid the other three

apps/api/src/app/types.ts(2,6): error TS2304: Cannot find name 'KVNamespace'.
apps/api/src/app/types.ts(3,6): error TS2304: Cannot find name 'R2Bucket'.
apps/api/src/app/types.ts(4,14): error TS2304: Cannot find name 'Queue'.
apps/api/src/app/types.ts(7,17): error TS2304: Cannot find name 'DurableObjectNamespace'.
apps/api/src/app/types.ts(8,6): error TS2552: Cannot find name 'D1Database'.
... 200+ more errors across 50+ files

My root tsconfig.json was the template every monorepo template ships with. No include, no exclude, no types, no baseUrl, no paths:

{
  "compilerOptions": {
    "target": "ES2022",
    "module": "ES2022",
    "moduleResolution": "bundler",
    "strict": true,
    "esModuleInterop": true,
    "skipLibCheck": true,
    "forceConsistentCasingInFileNames": true,
    "resolveJsonModule": true,
    "isolatedModules": true,
    "declaration": true,
    "declarationMap": true,
    "sourceMap": true
  }
}

Because it had no include, running tsc --noEmit from the repo root pulled in every .ts file in the project, tried to compile them under the root's type set, and exploded with 200+ errors from Cloudflare Worker code that depends on @cloudflare/workers-types — which the root tsconfig doesn't load.

Total noise. Unrunnable. My response, months ago, was to pin my dev-loop tsc command to -p apps/api/tsconfig.json — the one project that worked cleanly. That gave me fast, clean typechecks for apps/api, and a silent blind spot everywhere else.

Bug 4 didn't cause Bugs 1–3. Bug 4 was the condition in which Bugs 1–3 could live forever.

Fixing Bug 4 was the last step. After adding include, exclude, types: [], and workspace paths, bare tsc from the repo root now exits 0. It took four minutes. It could have taken four minutes any time in the last six months. I just never had a reason to do it, because I'd already built a workaround I didn't have to think about.

The line I want you to remember

It's not that my CI catches things my local environment doesn't. It's that my CI runs the commands I stopped running.

That's the whole thesis. Every one of those four bugs was catchable by tsc. The same tsc that ships with TypeScript. The same binary I already have installed. I just couldn't — or wouldn't — type the command that would have caught them, because I'd spent months engineering my dev loop around a broken root config I didn't want to fix.

A pre-push hook that runs your pipeline automatically eliminates that drift by removing the human who "just skips this one this time." That's it. That's the product.

What the tool did, exactly

Here's the part I want to be specific about, because the specifics are the value:

pushci init walked my repo and detected two Node workspace projects — apps/api and apps/web. It read each project's tsconfig.json independently instead of trying to compile the whole tree under one root config. No KVNamespace noise, no avoidance pattern required.
It read my pnpm-lock.yaml and chose pnpm run build over bare npx vite build, so the auto-detected commands actually run in a pnpm workspace. (That one was a recent fix. I was dogfooding the fix when it caught me. Recursion everywhere.)
It generated a pushci.yml pipeline I didn't have to write, with install, build, test, lint, typecheck stages. I edited it later. I didn't need to.
It installed a git pre-push hook I forgot about in about ten minutes, because the best developer tools are the ones you stop noticing. The next push I did — the one with the four bugs in it — the hook fired, ran my pipeline, caught them, blocked the push. None of the four reached main. The commit that did reach main had all four already fixed.

Scoreboard

Metric	Value
Bugs about to ship	4
Bugs that shipped	0
Install time	90 seconds
Cost	$0
SaaS accounts created	0
YAML files written by hand	0
GitHub webhooks installed	0
Pieces of local state I had to maintain	0

What this is, in one breath

One binary. npm install -g pushci. Auto-detects your stack across 33 languages. Generates pushci.yml from your repo structure. Installs a git pre-push hook that runs the pipeline locally before every push. No SaaS, no dashboard, no account, no credit card, no GitHub App, no .github/workflows/ to maintain. Runs inside Claude Code, Cursor, and Windsurf sandboxes where GitHub Actions can't reach — which is, for what it's worth, what I was originally dogfooding when it caught me.

I wrote it. I forgot I installed it. It caught me anyway.

Ninety seconds to install it and find your own seam:

npm install -g pushci
pushci init

Other ways that work today:

brew install finsavvyai/tap/pushci                # macOS + Linux
curl -fsSL https://pushci.dev/install.sh | sh     # any POSIX shell
npx pushci init                                    # no install

Tomorrow you'll either find zero bugs and forget about it, or you'll find one bug and keep it forever. Those are the only two outcomes.

— Shahar

pushci.dev

PushCI v1.3.0: Your CI Tool Supports Three Languages and You Are Fine With That?

Shachar Solomon — Thu, 09 Apr 2026 21:39:59 +0000

33 language stacks. Performance tracing. Flaky test detection. Still $0.

Let Me Get This Straight

Your CI platform — the one you're paying for — supports Go, Node, and Python. That's it. That's the list. Meanwhile your actual team has a Terraform module, a Rust CLI, a Kotlin microservice, and an intern who insists on writing everything in Elixir.

So what happens when someone pushes the Terraform module? Nothing. It goes through. Unvalidated. Like a passport control agent who only speaks three languages working at an international airport.

I got tired of pretending this was acceptable.

PushCI v1.3.0 supports 33 language stacks. Because that's how many languages people actually use.

The Numbers

33 stacks: Go, Node/TS, Python, Rust, Java, C#, Ruby, PHP, Swift, Dart, Elixir, Zig, Kotlin, Lua, Perl, R, Julia, OCaml, Nim, Crystal, Erlang, V, Terraform, Helm, Solidity/Foundry, Bun, Fortran — plus framework pipelines for Expo, Electron, Angular, Vue, CRA, and T3
22 CLI commands
69 marketplace skills
57 integration tests + 11 E2E tests

Every stack tested. Every pipeline validated. Not "community maintained." Not "beta support." Tested.

What's New

Performance Tracing with Perfetto

I asked my DevOps guy where the bottleneck was. He said "somewhere in the build." Somewhere. In the build. That's like a doctor saying the problem is "somewhere in your body."

pushci run --trace

Generates Chrome Trace Event JSON. Drag it into ui.perfetto.dev and see exactly which step took 47 seconds while the rest finished in 3. No other CI tool does this. Zero. I checked.

Flaky Test Detection

The test passed on your machine. It passed in CI. It passed 9 out of 10 times. But on that tenth time, the whole team stops what they're doing for a meeting about "test reliability."

pushci run --stress 10

Runs each check 10 times. Color-coded flake rate report. Now you have proof that test_payment_webhook fails 30% of the time instead of everyone pretending it doesn't.

Parallel Agents with Git Worktree Isolation

Build, test, security scan, and deploy — all running simultaneously on isolated git worktrees. Race strategy (first agent wins) or consensus (all must agree).

agents:
  strategy: race
  parallel:
    - build
    - test
    - security
    - deploy

Your CI pipeline just became a parallel processing system instead of a sad, sequential queue.

5 Notification Channels

Slack, Discord, Email, Telegram, Webhook. All of them. On every run.

notifications:
  slack:
    webhook: $SLACK_WEBHOOK
  discord:
    webhook: $DISCORD_WEBHOOK
  email:
    to: team@company.com

No marketplace plugin. No third-party integration. Just add it to pushci.yml and it works.

Interactive Dashboard Charts

Victory.js charts replacing the hand-rolled SVG situation we had before. Build time trends, flaky test rates, cost savings over time. Looks like a real product now because it is one.

Tailscale Fleet Discovery

Got a fleet of runners across machines? PushCI discovers peers via Tailscale MagicDNS with tag-based filtering.

pushci fleet discover --tag ci-runner

Distributed CI across your actual infrastructure. Not rented infrastructure. Yours.

Llamafile Lifecycle

Run AI diagnosis without sending your code to anyone's cloud:

pushci llamafile download mistral-7b
pushci llamafile start
pushci diagnose --local

Download, start, stop LLM models locally. AI-powered pipeline diagnosis at $0 with zero data leaving your machine.

The Free Tier Promise

Let me be very clear about what "free" means here:

Unlimited repos
Unlimited local runs
All 33 language stacks
Performance tracing
Flaky test detection
All install methods
$0. Forever. Not a trial.

Your machine runs the tests. Your machine pays the electricity bill. We don't get involved.

Pro ($9/mo) adds AI diagnosis, cloud runners, and priority support for when you want someone else's computer involved.

Team ($29/seat) adds SSO, audit logs, 25 members, and an SLA for when your company requires a piece of paper that says someone is responsible.

Try It

# Install
npm install -g pushci

# Or Homebrew
brew install finsavvyai/tap/pushci

# Or just run it
npx pushci init

# Set up CI in 30 seconds
pushci init

# Run with performance tracing
pushci run --trace

# Stress test for flaky tests
pushci run --stress 10

Links

Website: pushci.dev
GitHub: github.com/finsavvyai/push-ci.dev
Homebrew: brew install finsavvyai/tap/pushci
npm: npm install -g pushci
Dashboard: app.pushci.dev

So you're paying per minute for a CI tool that supports three languages, can't tell you where the bottleneck is, has no idea which tests are flaky, and runs everything sequentially. And when you asked about it, someone said "that's just how CI works." No. That's just how bad CI works.

pushci.dev

I Told My CI/CD to Configure Itself. It Did. In 30 Seconds.

Shachar Solomon — Tue, 07 Apr 2026 11:59:30 +0000

Let me get this straight.

You WROTE a config file. To TELL a computer. To run a TEST. That YOUR computer could run. For FREE.

And you're paying $0.008 per minute for that privilege.

I'm looking at my GitHub Actions bill right now. We're paying... to LINT. We're paying someone to run eslint. I could run eslint. My toaster could run eslint.

The Vibe-Coding Era Has a CI/CD Problem

We're in 2026. Nobody writes code anymore. You tell Claude to build your app, Cursor autocompletes your dreams, and Copilot finishes your sentences before you think them.

But CI/CD? Still stuck in 2019.

Still writing 50-line YAML files by hand. Still Googling "github actions cache node_modules" for the 47th time. Still debugging indentation errors in a config file that runs npm test.

The AI writes your code. But YOU have to write the YAML that tests the code the AI wrote.

Let that sink in.

So I Fixed It

npx pushci init

That's the whole setup. 30 seconds. AI scans your repo, detects your stack, generates the pipeline. You push code, it runs. Done.

No YAML. No config file. No 45-minute detour into GitHub Actions documentation.

What Just Happened

PushCI used AI to:

Look at your repo and figure out you're running Next.js with TypeScript
Generate build, test, lint, and deploy steps
Install a git hook so it runs before every push
Run everything on YOUR machine. For $0. Forever.

Supports 19 languages, 40+ frameworks, 16 deploy targets. Works with GitHub, GitLab, AND Bitbucket.

The Numbers That Make DevOps Guys Go Quiet

	PushCI	GitHub Actions	GitLab CI	Jenkins
Setup	30 sec	30+ min	30+ min	Hours
Config	AI auto	YAML	YAML	Groovy
Cost	$0	$0.008/min	$0.008/min	Server
AI	Built-in	No	No	No
Local runs	Yes	No	Limited	No
Works with	GH+GL+BB	GitHub only	GitLab only	All

1,000 builds/month at 8 minutes each?

GitHub Actions: $768/year
PushCI: $0/year

Forever.

It Also Fixes Itself

Your pipeline breaks at 2am? PushCI auto-diagnoses it:

pushci heal

26+ failure patterns. Missing dependencies, flaky tests, format errors, timeouts. It figures out the root cause and patches it. While you sleep.

No other CI tool does this. Not GitHub Actions. Not GitLab. Not Jenkins. Not the $15/user one.

Your AI Agent Can Set It Up

Claude Code, Cursor, Windsurf — they can all use PushCI directly:

{
  "mcpServers": {
    "pushci": {
      "command": "pushci",
      "args": ["mcp"]
    }
  }
}

Tell your AI: "set up CI for this project." It does the rest. The AI writes the code AND the CI. Full circle.

The Awkward Part

I showed PushCI to our DevOps engineer. Ran npx pushci init. Pipeline generated. Tests passing. 30 seconds.

Long silence.

"Well... this is embarrassing."

"What is?"

"I spent three days on our GitHub Actions workflow."

"Three DAYS?"

"There were caching issues."

Try It

npx pushci init               # AI detects your stack
pushci run                     # Run CI locally, $0
pushci heal                    # Auto-fix broken pipelines
pushci ask "deploy to staging" # Natural language CI

pushci.dev | vs GitHub Actions | Cost Calculator | GitHub

Stop writing YAML. It's 2026.

DEV Community: Shachar Solomon

I forgot I installed my own CI tool. It caught me shipping 4 bugs.

Bug 1: coverage enforcement had been silently off for weeks

Bug 2: a string that forgot to be a union

Bug 3: .avg on a number

Bug 4: the meta-bug that hid the other three

The line I want you to remember

What the tool did, exactly

Scoreboard

What this is, in one breath

PushCI v1.3.0: Your CI Tool Supports Three Languages and You Are Fine With That?

Let Me Get This Straight

The Numbers

What's New

Performance Tracing with Perfetto

Flaky Test Detection

Parallel Agents with Git Worktree Isolation

5 Notification Channels

Interactive Dashboard Charts

Tailscale Fleet Discovery

Llamafile Lifecycle

The Free Tier Promise

Try It

Links

I Told My CI/CD to Configure Itself. It Did. In 30 Seconds.

The Vibe-Coding Era Has a CI/CD Problem

So I Fixed It

What Just Happened

The Numbers That Make DevOps Guys Go Quiet

It Also Fixes Itself

Your AI Agent Can Set It Up

The Awkward Part

Try It

Bug 3: `.avg` on a `number`