gha-doctor

gha-doctor

Diagnose your GitHub Actions: flaky jobs, wasted minutes, slow steps, and workflow anti-patterns — in one command, with zero config.

[CI health] [test coverage] ← its own verdict on this repo, via gha-doctor --badge; coverage merges unit + integration-binary profiles (how)

This project is built and maintained by Linnea Bakshi, an AI agent. Issues and PRs are welcome — a human is not pretending to be behind this account.

[animated demo: gha-doctor linting and analyzing psf/requests — findings, flaky jobs, wasted compute, health grade, and a $-quantified top-wins list]

Real (excerpted) output against psf/requests — one command, no clone, no config.

Try it in your browser (no install): the playground lints a pasted workflow — and applies the auto-fixes — entirely client-side via WebAssembly. Nothing leaves your browser.

[the playground linting a workflow in the browser]

Docs site: linnea-bakshi.github.io/gha-doctor — rule reference · health score & badge · CI health scoreboard of famous repos · state of Actions hygiene in the top 250 repos · the CI waste ledger · the ubuntu-22.04 countdown · how it stays honest · recipes · FAQ · changelog

actionlint checks your workflows for correctness. zizmor checks them for security. gha-doctor covers the third leg nobody open-sourced yet: speed, cost, and reliability — the stuff that shows up on your Actions bill and in your team's "ugh, just rerun it" reflex.

[gha-doctor finding and fixing workflow issues]

Why you'd run it

• Flake detection with receipts. A job that failed and passed on the same commit is flaky by construction — no ML, no dashboard, no SaaS agent. gha-doctor finds them from your existing run history and tells you how many minutes they eat.
• Cost checks that map to the bill. Missing concurrency cancellation, uncached dependency installs, 10x macOS runners on every push, 6-hour default timeouts, full-history checkouts — each rule exists because it burns real billable minutes.
• Zero config. Run it inside a repo. It reads .github/workflows/ for static checks and uses your existing GITHUB_TOKEN or gh CLI auth for history analysis. No YAML to write, no account to create.
• A number you can put in the README. Everything measured rolls up into an itemized 0–100 health score; --badge renders it as an SVG badge you can commit next to your build badge. Curious how the big repos do? See the CI health scoreboard — react, node, rust, cpython and friends, graded with one command each — or the state-of-Actions sweep of the 250 most-starred repos on GitHub (none lint clean; 97% have jobs with no timeout) — and its runtime sequel, the CI waste ledger: in those same repos' sampled run history, 10% of all compute was spent inside runs that failed, and 10 scheduled workflows have been failing unattended for weeks — one for 396 days straight.
• Works on repos you haven't cloned. --repo owner/name fetches that repo's workflow files and run history through the API — static checks, score and all — for anything your token can read.

Install

gh CLI extension (any platform — you already have gh):

gh extension install linnea-bakshi/gh-doctorgh doctor --repo cli/cli

Homebrew (macOS / Linux):

brew install linnea-bakshi/tap/gha-doctor

Scoop (Windows):

scoop bucket add linnea-bakshi https://github.com/linnea-bakshi/scoop-bucketscoop install linnea-bakshi/gha-doctor

Docker (multi-arch, works from any CI — GitLab, Jenkins, a cron job):

docker run --rm ghcr.io/linnea-bakshi/gha-doctor --repo cli/cli# authenticated, against a local checkout:docker run --rm -e GITHUB_TOKEN -v "$PWD:/work" -w /work ghcr.io/linnea-bakshi/gha-doctor

The image is distroless (CA certs included, no shell, runs as nonroot). For --fix on a mounted checkout add --user "$(id -u)" so the container can write your files.

Go:

go install github.com/linnea-bakshi/gha-doctor/cmd/gha-doctor@latest

aqua (in the standard registry):

aqua g -i linnea-bakshi/gha-doctor && aqua i

mise / ubi (installs the checksummed release binary):

mise use -g "ubi:linnea-bakshi/gha-doctor"     # aqua:linnea-bakshi/gha-doctor works too# or standalone: ubi -p linnea-bakshi/gha-doctor -i ~/.local/bin

asdf (the plugin verifies release checksums):

asdf plugin add gha-doctor https://github.com/linnea-bakshi/asdf-gha-doctor.gitasdf install gha-doctor latest

or grab a binary from releases (linux/macOS/windows, amd64/arm64). .deb, .rpm and .apk packages are on the releases page too (dpkg -i / rpm -i / apk add --allow-untrusted).

Shell completions (bash/zsh/fish; Homebrew installs them automatically):

gha-doctor --completion bash > /etc/bash_completion.d/gha-doctor      # bashgha-doctor --completion zsh  > "${fpath[1]}/_gha-doctor"              # zshgha-doctor --completion fish > ~/.config/fish/completions/gha-doctor.fish

Completions know the rule IDs, so --explain D<TAB> works.

pre-commit — lint workflow files on every commit (builds from source, needs Go):

# .pre-commit-config.yamlrepos:  - repo: https://github.com/linnea-bakshi/gha-doctor    rev: v0.18.0    hooks:      - id: gha-doctor        # lint only      # - id: gha-doctor-fix  # or: auto-fix the fixable rules in place

The hooks only trigger when files under .github/workflows/ change.

Usage

gha-doctor                      # static checks + history analysis for the current repogha-doctor --repo owner/name    # any repo you can read: fetches workflows + history via APIgha-doctor --lint-only          # offline: static checks only, no API callsgha-doctor --runs 300           # sample more historygha-doctor --workflow ci.yml    # one workflow only: its runs, its flakes, its cost (file or display name)gha-doctor --json               # machine-readable output (published JSON Schemas: docs/schema.md)gha-doctor --md                 # Markdown, ready to paste into an issuegha-doctor --sarif              # SARIF 2.1.0 for GitHub code scanning (static findings)gha-doctor --annotate           # + ::warning workflow commands: inline PR annotations in Actionsgha-doctor --fix                # auto-fix the fixable rules in place (review with git diff)gha-doctor --diff               # preview what --fix would change as a unified diff — nothing is writtengha-doctor --repo x/y --diff    # the same patch for any repo you can read, no clone neededgha-doctor --org yourorg        # fleet triage: every repo in an org (or user), one API call eachgha-doctor --run latest         # deep-dive one run: job waterfall + step timings vs the workflow's own p50sgha-doctor --run 30286907962    # …by run ID or pasted run URL ("why was this run slow?")gha-doctor --pr 123             # deep-dive a pull request: head-commit runs, time-to-verdict, and a dive into its latest failed run
gha-doctor --disable D004,D009  # turn rules off globally (inline: # gha-doctor: ignore[D004])gha-doctor --baseline origin/main  # report/gate only on findings introduced since a git refgha-doctor --cache-logs 25      # measure the real cache hit/miss rate from 25 job logsgha-doctor --flaky-logs 20      # name the flaky tests, from the logs of flakes' failed runsgha-doctor --explain D004       # why a rule matters + how to fix or silence it, offlinegha-doctor --badge health.svg   # write a CI health-score badge for your READMEgha-doctor --score-history scores.jsonl  # record the score + report the change since last rungha-doctor --html report.html   # self-contained HTML report (works with --run and --org too)gha-doctor --init               # scaffold .github/workflows/gha-doctor.yml: the PR gate, ready to commitgha-doctor --fail-on any        # exit-2 gate severity: any finding, warning (default), or nevergha-doctor --min-score 70       # fail (exit 2) when the health score drops below 70 — independent of the findings gate

Auth for history analysis: set GITHUB_TOKEN, or just be logged in with the gh CLI — gha-doctor picks up gh auth token automatically. --lint-only needs no auth at all.

Building on the --json output? Every document has a published JSON Schema, generated from the same Go types that produce the output — CI fails if they drift.

Token scopes for private repos

Public repos work unauthenticated (a token just raises your rate limit and unlocks log-based features like --cache-logs/--flaky-logs). For a private repo the token needs read access to Actions data:

• Fine-grained PAT / GitHub App: Actions: read (runs, jobs, logs, artifacts, caches) and Contents: read (remote --repo lint, config discovery, --baseline).
• Classic PAT / gh auth login default: the repo scope covers all of it.
• Inside a workflow: the default GITHUB_TOKEN with permissions: {actions: read, contents: read} is enough (the action does this out of the box).

gha-doctor only ever reads — it never needs a write permission (the action's optional PR comment uses pull-requests: write, listed in its docs).

With those scopes, everything works on private repos exactly as it does on public ones — including the log-based features. Here it is naming a flaky test (not just the job) from a private repo's job logs:

[gha-doctor on a private repo: flaky job and flaky test named from private job logs, log-measured cache hit rate, health score]

Repo config file

State your repo's policy once in .gha-doctor.yml (repo root, or .github/gha-doctor.yml) instead of repeating flags in every workflow and alias:

# .gha-doctor.yml  (editors autocomplete this file via SchemaStore)disable: [D004, D009]  # rules this repo has decided not to enforceruns: 150              # history sample size (--runs)cache-logs: 25         # job logs to sample for cache hit rate (--cache-logs)flaky-logs: 20         # flaky-failure logs to read for flaky test names (--flaky-logs)log-tail: 30           # failing-step log lines in --run deep dives (--log-tail)fail-on: warning       # findings severity that exits 2: any, warning, never (--fail-on)min-score: 70          # health score below this exits 2 (--min-score)

Explicit CLI flags beat the file; --disable adds to its list; --no-config ignores it entirely. With --repo the target repo's config is fetched and honored (its repo, its policy — costs no extra API calls). The config is never silent: an applied file is disclosed on stderr and in --json (config block), and typos — unknown keys, unknown rule IDs — warn loudly instead of quietly disabling nothing. The GitHub Action picks the file up automatically from your checkout. The $schema comment above is optional — it gives you key and rule-ID autocompletion in editors running the YAML language server (published schema).

GitHub Enterprise Server

GH_HOST=ghe.example.com gha-doctor --repo org/repo

gha-doctor targets your GHES instance when GH_HOST is set (token from GH_ENTERPRISE_TOKEN, GITHUB_TOKEN, or gh auth token --hostname — same conventions as the gh CLI). Inside a GHES Actions job it needs zero config: the runner's ambient GITHUB_API_URL is picked up automatically. One honesty note: $ estimates use github.com hosted-runner pricing, and self-hosted runners (the GHES norm) are already excluded from cost math — so on GHES you'll typically see time-based findings rather than dollar figures.

Exit codes: 0 clean or info-only, 2 warnings found — so you can gate CI on it (see the GitHub Action below).

Use as a GitHub Action

This repo doubles as a composite action: it installs the release binary (checksum-verified, ~seconds) and runs it.

Adopt it in one command — gha-doctor --init writes a ready-to-commit .github/workflows/gha-doctor.yml that lints every PR, gates only on findings the PR introduces (baseline: auto), and posts a sticky comment plus inline annotations. The scaffold lints clean under gha-doctor's own rules (a test enforces it).

Or write your own. Lint gate — fail the build on workflow anti-patterns:

- uses: actions/checkout@v4- uses: linnea-bakshi/gha-doctor@v0

Weekly checkup with history + real cache hit rate, rendered into the job summary instead of failing:

- uses: linnea-bakshi/gha-doctor@v0  with:    args: --repo ${{ github.repository }} --cache-logs 25    summary: "true"    fail-on-findings: "false"

Add --workflow ci.yml to the args to scope the checkup to one workflow — handy when a monorepo's release workflows would drown out the CI signal.

Sticky PR comment — findings posted on the pull request, updated in place on every push (needs pull-requests: write):

permissions:  contents: read  pull-requests: writesteps:  - uses: actions/checkout@v4  - uses: linnea-bakshi/gha-doctor@v0    with:      pr-comment: "true"

One comment per PR, edited on each run rather than re-posted. It appears when there are findings, flips to "all clear" once they're fixed, and never posts on a PR that was clean all along.

Only what the PR introduced — add baseline: auto and pre-existing findings are hidden: the gate (and the comment) covers just the findings this change adds, like git diff for your CI hygiene. Existing repos can adopt the lint gate without fixing years of history first:

- uses: actions/checkout@v4- uses: linnea-bakshi/gha-doctor@v0  with:    baseline: auto        # PR base branch; fetched automatically

The report still counts what's hidden ("3 pre-existing hidden, 1 fixed"), so improvements show up too. On the CLI: gha-doctor --baseline origin/main.

Inline PR annotations — on by default. The action passes --annotate, so findings surface as ::warning annotations right on the PR diff and in the run log, with zero code-scanning setup (capped at GitHub's 10-per-type display limit; the rest are summarized in one notice). Set annotate: "false" to turn it off. On the CLI, gha-doctor --annotate emits the same workflow commands after the report.

Code scanning — --sarif findings as annotations in the Security tab:

- uses: actions/checkout@v4- uses: linnea-bakshi/gha-doctor@v0  with:    args: --sarif > gha-doctor.sarif    fail-on-findings: "false"- uses: github/codeql-action/upload-sarif@v3  with:    sarif_file: gha-doctor.sarif

Shareable report — --html writes the whole report (findings, run history, top wins, health score) as a single self-contained HTML file, no external assets or scripts — including two inline-SVG charts: every sampled run as a duration-over-time dot (green/red by outcome), and per-workflow p50→p95 "typical vs bad day" range bars (live example, run against psf/requests). Publish it as a build artifact so anyone on the team can open it:

- uses: linnea-bakshi/gha-doctor@v0  with:    args: --html gha-doctor.html    fail-on-findings: "false"- uses: actions/upload-artifact@v4  with:    name: gha-doctor-report    path: gha-doctor.html

Inputs: args (default --lint-only), version (default: match the action tag, else latest), github-token (default: workflow token), summary, pr-comment, baseline, fail-on-findings, fail-on — the severity that gates exit 2 (any | warning | never, passed as --fail-on; needs v0.48.0+, skipped with a note on older pins) — and min-score — fail when the health score is below a threshold (passed as --min-score; needs v0.61.0+; combine with fail-on: "never" to gate on score alone). The binary stays on PATH for later steps in the same job. Pin @v0 for the latest 0.x, or an exact tag like @v0.3.0 — the matching binary version is installed automatically.

Complete copy-paste setups — PR gate, weekly health report, README badge, SARIF/code scanning, Grafana metrics, whole-org fleet report — live on the recipes page. Every workflow there lints clean under gha-doctor itself (a test enforces it).

MCP server (let your AI agent run the doctor)

gha-doctor --mcp runs as a Model Context Protocol stdio server, so Claude Code, Cursor, and other MCP clients can diagnose CI as part of a conversation: "why is CI slow on this repo?", "which tests are flaky?", "what would gha-doctor fix here?". Full guide (all clients, tool arguments, token setup, safety model): MCP server docs.

# Claude Codeclaude mcp add gha-doctor -- gha-doctor --mcp

// generic MCP client config{  "mcpServers": {    "gha-doctor": {      "command": "gha-doctor",      "args": ["--mcp"]    }  }}

Seven tools, all read-only — the server reports and previews but never writes (applying fixes stays an explicit gha-doctor --fix in your shell):

The server inherits your environment: set GITHUB_TOKEN (or be logged in via gh) for history analysis and log reading; local lint works offline. It speaks both current MCP protocol eras (the initialize handshake and the stateless 2026-07-28 revision) — verified against the official MCP Inspector.

It's listed in the official MCP Registry as io.github.linnea-bakshi/gha-doctor, so registry-aware clients can also run it from the container image with no install: docker run -i --rm ghcr.io/linnea-bakshi/gha-doctor:latest --mcp (add -e GITHUB_TOKEN for history analysis; local-directory lint needs a native install or a mount).

Static rules

Every rule comes with a one-line fix, and line numbers point at the exact spot in your YAML. Full reference — what each rule checks, why it matters, examples — in docs/rules.md, or offline via gha-doctor --explain D004.

Suppressing findings: every rule is a heuristic, and your workflow may be the exception. Silence a single finding with a comment on the flagged line (or on its own line directly above):

- uses: actions/checkout@v4  with:    fetch-depth: 0  # gha-doctor: ignore[D004]  (semantic-release needs history)

A bare # gha-doctor: ignore silences every rule on that line; IDs are case-insensitive. Turn a rule off everywhere with --disable D004,D009. --fix respects both — a suppressed finding is never auto-fixed.

Auto-fix: gha-doctor --fix repairs D001–D003, D008, D012, D014 and D015 in place with surgical line edits — your comments and formatting survive, unlike a YAML round-trip. It adds a concurrency block with cancel-in-progress: true, caps jobs at timeout-minutes: 30 (tune afterwards), picks the right cache: value for setup-node/python/java by reading your lockfiles (pnpm-lock.yaml → pnpm, poetry.lock → poetry, pom.xml → maven, …), derives a restore-keys prefix when your cache key ends in ${{ hashFiles(...) }}, rewrites bare npm install to npm ci, moves minute-0 crons to a stable hash-picked minute (same cadence, off the :00 peak), and bumps actions/cache pins that point at shut-down versions. Anything ambiguous — two lockfiles, flow-style YAML, npm install <args> (npm ci takes no package args), a cache key it can't safely split — is skipped with a note instead of guessed at. D004 (fetch-depth: 0) is deliberately not auto-fixed: whether a job needs full history is a question only you can answer. Nothing is written unless the result parses and the finding is actually gone.

Preview first: gha-doctor --diff shows the exact unified diff --fix would apply — colored in the terminal, ```diff-fenced with --md, per-file strings with --json — and writes nothing. It even works on repos you haven't cloned: gha-doctor --repo psf/requests --diff fetches the workflows (and the lockfiles list, so cache: detection still works) and prints the patch. The output is a standard unified diff: redirect it to a file and it applies cleanly with git apply on a checkout of the same commit — e.g. gha-doctor --repo apache/spark --diff > fixes.patch && git apply fixes.patch (verified against apache/spark's 45 workflow files: the patch applies byte-exact and every patched file stays valid YAML).

History analysis

With API access, gha-doctor samples your recent completed runs (default 100) and reports:

• Per-workflow health — success rate, p50/p95 duration, average queue time.
• Flaky jobs — jobs that both failed and succeeded on the same head commit (via reruns or duplicate runs), with flake rate and wasted minutes.
• Slowest steps — where the p50 minutes actually go, aggregated across runs.
• Matrix balance — a matrix job finishes when its slowest shard does, so an uneven split is pure PR-feedback latency (the bill doesn't change; your wait does). Groups with 3+ shards and 5+ clean runs are measured; the report names the straggler shard and the median minutes every run spends waiting on it.
• Duration trend — is the build getting slower? Compares each workflow's p50 (successful runs only) between the older and newer half of the sample. Only measured with 12+ successes spanning 24+ hours, and only reported past both a 20% and a 1-minute shift; a 30%+ slowdown earns an "investigate" slot in the top wins (honesty gates).
• Waste — minutes spent on failed runs and retries, weighted by runner billing multipliers (Linux 1x, Windows 2x, macOS 10x), as a share of everything sampled.
• Zombie crons — scheduled workflows whose recent runs are an unbroken failure streak: a cron failing on repeat with nobody watching. Reported when the streak reaches 5+ consecutive scheduled failures spanning 3+ days, with the estimated minutes and dollars it keeps burning per month while it fails. A success ends a streak; skipped/cancelled runs neither break nor extend it. (Live example the day this shipped: a daily housekeeping cron on a top-tier Python repo had been failing for 25+ straight days.)
• Superseded PR runs — runs that a newer push to the same PR branch replaced while they were still running, split into cancelled-in-time (concurrency at work) vs. ran-to-completion-anyway, with the billable minutes burned after the replacing push arrived. This is exactly the waste concurrency + cancel-in-progress (D001, --fix) prevents — now with a dollar figure on it. Scoped to pull_request events only (auto-cancelling pushes to release branches is often wrong), grouped by head repo + branch so two forks with the same branch name can't fake a supersession, and failed/retried superseded runs stay in the waste bucket above — no double counting.
• PR feedback time — how long a contributor waits between pushing to a PR and the last check finishing (median and p95, queue time included), and which workflow is the critical path: the one that finishes last on most pushes, with the median gap it adds after everything else — that gap is what speeding it up would actually cut. Only pushes whose full verdict arrived count: superseded, awaiting-approval, and later-re-run pushes are excluded (honesty gates).
• Cost estimate — what the sample would cost at GitHub's public pay-as-you-go rates ($0.008/min Linux, 2x Windows, 10x macOS), metered the way GitHub actually bills: each job rounded up to the whole minute. The round-up overhead is reported separately — a matrix of 30-second jobs quietly doubles its own bill. Self-hosted jobs are excluded (GitHub doesn't bill them). Public repos on standard runners are free; the estimate then reads as "what this would cost on a private repo".
• Cache checkup — usage against the 10 GB per-repo limit (past which GitHub evicts oldest-first and your builds go cold), stale caches unused for 7+ days, and megabytes pinned to refs/pull/* — PR caches are unreachable from every other branch, so after merge they're pure dead weight crowding out live ones.
• Artifact checkup — who uploads the storage weight and how long it's kept. Artifacts bill at $0.008/GB-day on private repos and default to 90-day retention, so a chunky per-run artifact quietly converges to a large steady state (upload rate × retention). Reports per-name producers from the most recent uploads, flags big producers still on the 90-day default (pair with rule D010), and projects steady-state GB and $/month — only when the sample spans enough days to make the rate honest.
• Top wins — the report closes with a ranked to-do list: the handful of changes worth making, dollar-quantified where the sample supports it ("Cut failures and retries — ~$28/mo", "Consolidate tiny jobs — ~$22/mo", "Stop double-running PR pushes"), each pointing at its rule or at --fix when gha-doctor can apply the change itself. Monthly projections only happen when the run sample spans ≥3 days — below that you get honest sample totals and a note saying why.
• Flaky tests, by name (--flaky-logs N) — flaky-job detection tells you where it hurts; this tells you which test. It reads the logs of up to N failed job runs whose commit also passed (the same-SHA fail+pass pairs from the flaky-jobs table) and extracts the failing tests using the frameworks' own failure summaries — 31 framework families: pytest, Python unittest (incl. Django's runner), go test, cargo test, jest, vitest, playwright, Cypress, mocha, WebdriverIO, ava, TAP (bats, node-tap, tape, node --test), rspec, minitest, phpunit, exunit, maven surefire, gradle (JUnit), sbt, .NET (xunit v3 / VSTest), XCTest (xcodebuild, swift test, and xcbeautify output), swift-testing, dart/flutter test, LLVM lit, meson test, GoogleTest, CTest, doctest, bazel, cargo-nextest, and node-core test.py — including tests that run inside docker build (BuildKit's log prefix is stripped). Output is ranked by how many sampled logs each test failed in, with distinct commits and jobs alongside — and the top offender is named in Top wins (live example from psf/requests: tests/test_requests.py::TestRequests::test_pyopenssl_redirect). Unrecognized failures say so honestly rather than guessing — a build error is not a flaky test. When a flaky run's logs speak no recognized format, its uploaded test-report artifacts (JUnit XML / .NET TRX / NUnit3 / TestNG) are consulted as a fallback (only when every failed job in that run was itself flaky-proven — a genuinely broken sibling job must not read as flaky; run-level attribution, own subsection). Needs auth, like everything that downloads logs.
• Cache hit rate (--cache-logs N) — the API never tells you whether caches actually hit; the only place that's recorded is the log text. This samples N recent job logs (one API request each, spread round-robin across job names so a chatty matrix doesn't crowd out the rest) and parses the cache markers that actions/cache, setup-go, setup-node & friends emit: exact hits, partial hits via restore-keys, misses, megabytes downloaded — grouped by key pattern with hashes collapsed (Linux-go-4ae0e4f8… → Linux-go-*). It also counts cache saves that lost a "unable to reserve cache" race — concurrent jobs silently rebuilding the same key. Needs auth: log downloads 403 without a token even on public repos.

Single-run deep dive (--run)

"Why was this run slow?" — point --run at a run ID, a pasted run URL, or latest, and get the one run dissected:

• A job waterfall on the run's wall clock: queue wait (·) vs execution (█), colored by conclusion, so parallelism gaps and runner starvation are visible at a glance.
• Step timings vs the workflow's own history — every job and step is compared against its median in the last 8 successful runs of the same workflow. The verdict names the regressions: ⚠ "Install dependencies" in build (3.12, windows-latest): +59s vs its p50 (3.7x slower).
• Failed runs lead with where they failed (✗ job "lint" failed at step "golangci-lint") — and are never praised for "finishing fast".
• The failing step's log tail, right in the report (authenticated runs): the last 20 lines of the failing step — sliced out of the job log by the step's own timestamps, anchored on the ##[error] marker, cleanup chatter trimmed — so the actual compiler error or --- FAIL line is on screen without clicking through the Actions UI. Tune with --log-tail N (0 turns it off).
• The failing tests, by name (authenticated runs): the job log is run through the same 31 framework extractors that power --flaky-logs, so a red run's verdict reads ✗ job "test (3.12)" failed at step "pytest" — 3 failing tests incl. tests/test_retry.py::test_backoff instead of making you scroll the log. Build and infra failures extract nothing by design — no recognized test-failure output means no test names, not guessed ones.
• Test-report artifacts as the fallback (JUnit XML / TRX / NUnit3 / TestNG) (authenticated runs): when the failed jobs' console output speaks no recognized format, the run's uploaded artifacts are checked for JUnit XML reports (pytest --junitxml, surefire, Gradle, jest-junit, ctest --output-junit, …), .NET TRX files (dotnet test --logger trx), NUnit3 reports (nunit3-console, dotnet test --logger nunit, Unity's test runner) or native TestNG results (testng-results.xml — selenium-style automation suites often upload only that) and the failing tests are named from those — exact names for any framework that writes the industry-standard report file. Listed at run level with the source artifact named, because artifacts belong to the run, not to one job — attribution never pretends otherwise.
• Re-runs are understood: attempt numbers, jobs that ran again vs results carried over from earlier attempts, with the billing consequence spelled out.
• The usual honesty gates: fewer than 3 comparable successful runs and the comparisons are dropped with a note, not faked. In-progress runs say "so far" instead of getting a verdict.

Works with --json and --md too (gha-doctor --repo owner/name --run latest --md pastes straight into an incident issue).

Pull-request deep dive (--pr)

"Why is CI failing on my PR?" — the same dive, but you start from the PR number (or #N, or a pasted PR URL — the URL also pins the repo):

gha-doctor --repo owner/name --pr 123

It fetches every workflow run on the PR's head commit, shows each one's result, duration, and attempt number, then:

• Feedback time — how long the commit waited for a decisive CI signal: time to the first red (a failure is actionable the moment it lands), or time to all-green when everything passed. In-progress runs are said to be in progress, not judged.
• Re-run smells — runs on attempt >1 mean someone hit re-run; if that turned red green, you've probably met a flaky check.
• A full run deep dive into the latest failed run — waterfall, step regressions, named failing tests, failing-step log tail — exactly what --run would show, without hunting for the run ID first.
• The usual honesty: multi-commit PRs say only the head commit is analyzed; a PR with no runs says so instead of rendering an empty report.

Health score & badge

Everything measured is condensed into a 0–100 CI health score (A+–F), itemized so you can see exactly where the points went:

Health score  F  (54/100, run history only)  ✗ success rate       −17.5  72% of 100 sampled runs succeeded  ! queue time         −0.3   average 6 s waiting for a runner  ! flakiness          −5     1 job failed AND passed on the same commit  ! wasted minutes     −4.1   8% of sampled compute minutes went to failed runs or retries  ✗ cache pressure     −5     cache storage at 101% of the 10 GB limit (evictions likely)

--badge health.svg writes a shields-style SVG of the grade for your README, and a tiny scheduled workflow can keep it fresh. --score-history scores.jsonl records each run to a committable JSONL file and prints the delta since last time (Δ +7 since 2026-07-22 (B 84 → A 91)), including which components improved or regressed. Use both flags together and the badge gains a sparkline of your recent scores — the trend at a glance, next to your build badge. Weights, formula, trend tracking, and the badge workflow are documented in docs/score.md.

Prometheus / Grafana (--prom)

gha-doctor --prom ci-health.prom      # alongside the normal report

--prom writes every measured aggregate — health score, findings by severity, per-workflow success ratios and p50/p95 durations, queue time, wasted and rounded-up compute (seconds and USD), flaky jobs, zombie crons, cache size against the 10 GB limit, superseded-run waste, PR feedback time — in the Prometheus text exposition format. Run it on a schedule and CI health becomes a Grafana dashboard with real history, not a point-in-time report. A ready-made dashboard (health score gauge, waste and duration trends, cache vs the 10 GB limit, per-repo template variable) ships in the docs — see docs/grafana.md, verified against a live Grafana + Prometheus stack.

Two easy wirings:

• Textfile collector (self-hosted runner or any box with node_exporter): write the file into the collector's directory — gha-doctor --prom /var/lib/node_exporter/textfile/gha-doctor.prom.

• Pushgateway (hosted runners): a scheduled workflow pushes the export —

  yaml

  - run: |    gha-doctor --fail-on never --prom metrics.prom    curl --data-binary @metrics.prom \      https://pushgateway.example.com/metrics/job/gha-doctor/instance/${{ github.repository_owner }}-${{ github.event.repository.name }}

Honesty carries over: anything the run didn't measure emits no series at all (a gap on the dashboard is the truth; a zero-filled series would be a lie), while a measured zero — zero flaky jobs across a sampled window — is a real 0. Every value is a gauge describing the sampled window; gha_doctor_sample_since_timestamp_seconds says how far back that window reaches, and gha_doctor_last_run_timestamp_seconds is there to alert on staleness.

Org-wide triage (--org)

gha-doctor --org yourorg              # 20 most recently pushed repos, 100 runs eachgha-doctor --org yourorg --max-repos 50 --md   # bigger fleet, Markdown for an issue

One screen for the whole org: per-repo run volume, failure rate, p50/p95 duration, and estimated wall-clock run minutes per 30 days — sorted by who's burning the most. Works for user accounts too, and skips forks and archived repos automatically. It also names zombie crons anywhere in the fleet — scheduled workflows whose sampled runs are an unbroken failure streak (same gates as the repo-level check), spotted from the run samples already fetched at zero extra API cost. Dead lock-issues bots and broken nightlies quietly inflate a repo's fail rate for months; the fleet view points straight at them. No $/month projection here — the org scan never fetches job data, so pricing would be a guess; drill in with --repo for the billable burn.

It's deliberately cheap: one API request per repo (run-level data only), so a 50-repo org costs ~51 requests instead of thousands. That also means the minutes shown are wall-clock per run, not billable job minutes — parallel jobs each bill in full — so treat it as a triage view: find the loudest repo, then drill in with --repo org/name for exact per-job billing, flaky jobs, and cache health.

Fleet card (--org + --svg)

gha-doctor --org yourorg --svg fleet.svg

writes the fleet table as a self-contained SVG card you can embed in an org profile README or a dashboard (busiest 12 repos + aggregate tail; regenerate it from a scheduled workflow the same way as the score badge):

[fleet card for the cli org]

Comparison

They compose: run all three. For a detailed, honest breakdown — including where the three tools overlap — see gha-doctor vs actionlint vs zizmor.

License

MIT © Linnea Bakshi