Security Remediate (Atmos)

Proactively fix GitHub-reported security alerts on the branch you're already working on, instead of letting them accumulate until a manual sweep (see commit 0e32451913 for what that manual sweep looked like).

Fork / permission prerequisite (read first)

This repo is public, so a plain gh token with repo scope (the default from gh auth login) is sufficient to read both alert endpoints — security_events is only required on private repos (see GitHub's Dependabot alerts API docs: private repos need security_events, public repos only need public_repo/repo). Don't chase a security_events scope refresh for this repo; it's unnecessary and, per cli/cli#12073 and independent reports (e.g. microsoft/sarif-sdk#3081), gh auth refresh -s security_events can report success while GitHub silently drops the scope anyway — a dead end that's easy to mistake for a real permission gap.

If gh api calls below still 404/403 even with -X GET present, that's a genuine permission gap (e.g. a fork contributor without repo access, or a future private-repo use of this skill lacking security_events) — stop and report it instead of attempting a workaround. .claude/hooks/security-remediate-trigger.sh treats that case exactly like "no alerts open" and silently does nothing.

Workflow

1. Resolve context.

   owner_repo="$(gh repo view --json owner,name --jq '.owner.login + "/" + .name')"
   branch="$(git rev-parse --abbrev-ref HEAD)"
   

2. Query both alert sources live (the hook's cached alert set was only a trigger signal — treat it as stale):

   gh api "repos/${owner_repo}/dependabot/alerts" -X GET --paginate -f state=open
   gh api "repos/${owner_repo}/code-scanning/alerts" -X GET --paginate -f state=open
   

   The -X GET is required: gh api implicitly sends a POST when -f/--paginate params are given without an explicit method, and these list endpoints 404 on POST — indistinguishable from a real permission error unless you notice the verb. If either call still 403/404s with -X GET present, stop and report the permission gap — do not proceed partway.

3. Filter to high-confidence fixes, following the precedent set by 0e32451913:

   • Dependabot alerts: only fix if a patched version exists within what .github/dependabot.yml's ignore rules allow (that file currently ignores all major-version bumps across gomod, github-actions, and npm/website). If the only fix requires a major bump the policy blocks, do not bump it — report it as "blocked by dependabot ignore policy" instead.
   • CodeQL alerts: only fix rule IDs with an established safe pattern in this repo. Currently that's go/allocation-size-overflow, fixed by avoiding summed len() capacity hints (e.g., replacing make([]T, len(a)+len(b)) with a pattern that doesn't sum lengths as a capacity hint). Extend this list cautiously — an unfamiliar rule ID with an ambiguous fix goes in the "not auto-fixable" bucket, not a guessed fix.
   • GitHub Action pins: bump to the patched SHA/tag referenced by the alert.
   • npm/pnpm alerts (in website/): bump via pnpm.overrides in website/package.json, then regenerate NOTICE:

     go tool mage notice:generate
     

4. Apply fixes in the working tree, then verify before committing anything:

   make lint
   go test ./... # or scope to the touched packages
   

   If either fails for a given fix, drop that fix from the commit and move it to the "not auto-fixable" report — don't commit something that doesn't pass.

5. Commit directly onto the current branch. No new branch, no new PR, no issue — this branch already has an open PR, and the fix belongs on it:

   git add <only the files you actually changed>
   git commit -m "fix(security): remediate N CodeQL/Dependabot alerts"
   

6. Report a summary back in chat (not as a GitHub comment/issue): which alerts were fixed (numbers + URLs), and which weren't, with a reason for each (major-bump blocked by policy, ambiguous/unfamiliar rule ID, no patch available yet, permission gap). Never silently drop an alert without accounting for it in this summary.

What this skill must never do

• Open a new pull request.
• Open a new GitHub issue (per standing project guidance: never open issues without explicit per-issue user authorization).
• Push past a dependabot.yml ignore rule.
• Commit a fix that fails make lint or the targeted tests.