mobile-reality-skills

Distributable agent skills for the team — usable from Claude Code and Codex, on Windows / macOS / Ubuntu. Installed via npx skills; discovery is filesystem-based (every skills/<name>/SKILL.md is a skill).

Available skills

Skill	Description	Agents
coderabbit-fix	Fetch & fix CodeRabbit comments on a Bitbucket Cloud PR — classify by severity/effort, auto-fix Quick Wins, flag Heavy Lift for decision.	Claude Code, Codex

Install

Installs both agents in one command:

# macOS / Linux
npx skills add MobileReality/mobile-reality-skills -a claude-code -a codex

# Windows (symlinks can be blocked without Developer Mode/admin → use --copy)
npx skills add MobileReality/mobile-reality-skills -a claude-code -a codex --copy

• One specific skill: append -s coderabbit-fix.
• Install globally (all repos) instead of repo-local: append -g.

First-time setup

Each team member uses their own Bitbucket credentials — there is no shared secret.

1. Generate an Atlassian API token

https://id.atlassian.com/manage-profile/security/api-tokens → Create API token with scopes → app = Bitbucket → scopes: read:pullrequest, write:pullrequest, read:repository.

Use an API token, NOT an App Password (Atlassian is disabling App Passwords).

2. Set the env vars

# bash / zsh
export BITBUCKET_EMAIL="you@example.com"
export BITBUCKET_API_TOKEN="ATATT3xFfGF0..."

# PowerShell
$env:BITBUCKET_EMAIL    = "you@example.com"
$env:BITBUCKET_API_TOKEN = "ATATT3xFfGF0..."

Persist them in your shell profile (~/.bashrc, ~/.zshrc, or PowerShell $PROFILE) so they survive new sessions.

Fallback (any agent/OS): put both vars in a .env file at your repo root — the fetch script reads it when env vars are absent. Never commit .env.

3. Point it at your Bitbucket repo

The coderabbit-fix skill has no hardcoded workspace/repo — it works for any team. It resolves the target in this order:

1. --workspace <ws> --repo <repo> flags, then
2. BITBUCKET_WORKSPACE / BITBUCKET_REPO env vars, then
3. the origin git remote, when it points at bitbucket.org/<ws>/<repo>.

If you run the skill from inside your Bitbucket clone, (3) means zero config — it just works. Otherwise set the env vars (or pass the flags).

4. Codex extra setup

Codex's sandbox blocks network and strips *TOKEN* env vars by default. Add to ~/.codex/config.toml:

[sandbox_workspace_write]
network_access = true            # Linux/Windows; IGNORED on macOS

[shell_environment_policy]
ignore_default_excludes = true
include_only = ["PATH", "HOME", "BITBUCKET_EMAIL", "BITBUCKET_API_TOKEN"]

On macOS, network_access is silently ignored — run codex --sandbox danger-full-access, or prefer Claude Code. Full detail in each skill's SKILL.md ("Codex setup").

Update

npx skills update coderabbit-fix   # one skill
npx skills update                  # all installed skills

There is no lockfile / version pin — update pulls the latest from the default branch.

Platform / agent matrix

	Windows	macOS	Ubuntu
Claude Code	✓	✓	✓
Codex	✓ (config.toml)	⚠️ needs --sandbox danger-full-access; prefer Claude Code	✓ (config.toml)

Adding a new skill

The repo is multi-skill by design — no refactor needed to add one:

1. Create skills/<name>/SKILL.md (plus an optional scripts/ folder).
2. Keep the frontmatter minimal and portable: name, description, metadata.version only — no allowed-tools / Hooks / context: (Claude-Code-only; Codex ignores or chokes on them).
3. Save SKILL.md as UTF-8 without BOM — Codex won't detect the frontmatter otherwise.
4. Commit and push. Team members pick it up with npx skills update.

License

MIT.