Code complexity stats pr

Code complexity / cost stats on every PR

A GitHub Actions workflow that runs scc (Sloc, Cloc and Code) over the repo on each pull request and posts the results — including a COCOMO estimated cost/effort to build the codebase — as a single sticky comment that updates in place on every push. Drop-in workflow: templates/code-complexity.yml.

What you get

scc prints a per-language table (files, lines, code, comments, blanks, complexity) plus, with --avg-wage, a COCOMO block: estimated person-months, schedule, people, and a dollar cost to develop. The workflow wraps that in a fenced block and posts it as a PR comment titled 📊 Code Complexity Analysis.

It reports the whole codebase at the PR's head, not the diff — it's a "what's this project worth / how big it is" signal, not a per-PR delta.

Install

Copy templates/code-complexity.yml to .github/workflows/code-complexity.yml. It works as-is on ubuntu-latest; the only thing you'll likely change is the AVG_WAGE / WAGE_LABEL workflow env (see Customise).

How it works (the parts that matter)

• Trigger: pull_request: [opened, synchronize, reopened] — runs on open and every push. Kept as its own workflow so it always runs even when other workflows honour a [skip ci]-style marker; these stats are informational.
• Permissions: a PR comment is posted via the Issues API, so pull-requests: write is what actually authorises the create/update; issues: write is included belt-and-suspenders. (Both are in the template.)
• Sticky comment: the body starts with a hidden marker <!-- code-complexity -->. The script lists existing comments, finds the one with that marker, and updates it (else creates one) — so a PR keeps exactly one stats comment that refreshes instead of stacking new ones each push.
• Concurrency: cancel-in-progress: true keyed on the PR number (github.event.pull_request.number) — a new push kills the in-flight run so you don't race two comment updates. Keyed on the number, not head_ref, so two forks sharing a branch name can't cancel each other.
• COCOMO cost: scc --avg-wage <annual> drives the cost figure. The number is unitless to scc — it's read in your currency. The template defines AVG_WAGE (360000, HKD 30,000/month) and WAGE_LABEL once as workflow env, so the number scc uses and the label in the comment can't drift — change the salary in one place.

Customise

Want to…	Change
Use a different currency / salary	the AVG_WAGE and WAGE_LABEL workflow env (one place — the comment label reads WAGE_LABEL)
Exclude more build/generated dirs	--exclude-dir list (comma-separated). scc honours .gitignore by default; this is for output that isn't ignored
Pin / upgrade scc	SCC_VERSION in the install step
Run on ARM runners	swap the tarball to scc_Linux_arm64.tar.gz and use an ARM runner (e.g. ubuntu-24.04-arm)
Count only part of a monorepo	add a path arg to scc (e.g. scc apps/web)

Gotchas

• Fork PRs can't post the comment. For PRs from forks, pull_request runs with a read-only token and no repo secrets, so the comment API call would 403. Declaring pull-requests: write does not override this — GitHub enforces read-only on the token itself for fork-triggered runs. The template guards the whole job with if: github.event.pull_request.head.repo.full_name == github.repository, so on fork PRs the job skips cleanly (green check) instead of failing red — and no compute is wasted running scc for a comment that can't post. Options: (a) accept that stats only post for same-repo branches (fine for solo/team repos — the common case; this is the template default); or (b) split into two workflows — compute on pull_request (no secrets, untrusted code), upload scc-output.txt as an artifact, then comment from a separate workflow_run job that has write access. Do not just switch this workflow to pull_request_target: that runs with a write token in the base repo context, and checking out + running fork code there is a token-exfiltration footgun.
• Long output is truncated, not failed. GitHub caps a comment body at 65,536 characters. Default scc output (a per-language table) is tiny, but --by-file on a big repo could blow past it, which would 422 the step — so the template trims to ~64k with a "truncated" note and the comment still posts.
• Pin scc; don't track latest. A latest download can change scc's output format and silently reshape every PR comment. The template pins SCC_VERSION and verifies the tarball against a pinned SCC_SHA256 (sha256sum -c) before running it, so a swapped/compromised release asset fails the run instead of executing in CI. When you bump the version — or switch to the arm64 asset — update the checksum from that release's checksums.txt.
• Match the tarball to the runner arch. The release asset is arch-specific (scc_Linux_x86_64.tar.gz vs _arm64). The wrong one fails to run.
• Paginate when finding the sticky comment. A busy PR can have more comments than one API page; without pagination, the marker comment on a later page is missed, and you get duplicates. The template uses github.paginate(...) and matches the marker with startsWith (not includes) so a human comment that merely quotes the marker isn't mistaken for the bot's and overwritten.
• Token: GH_TOKEN || github.token. The template prefers a GH_TOKEN secret (a PAT/bot you already use) and falls back to github.token — the canonical reference to the built-in GITHUB_TOKEN. Prefer github.token over secrets.GITHUB_TOKEN; the built-in token is enough for same-repo comments given the permissions above.
• COCOMO is illustrative. It's the basic COCOMO model (COCOMO I, Organic mode) on LOC — a fun, relative signal of size/effort, not a real valuation. Don't quote it as the literal worth of the code.

Provenance

Generalised from the working code-complexity.yml used in production repos (a production app, a production monorepo): same scc + COCOMO + sticky-comment approach, hardened here with version pinning + checksum verification, persist-credentials: false, comment pagination, a portable ubuntu-latest runner, and the fork-PR security note.