agent-config

Claude Code configuration: coding standards, git hooks, slash commands, and skills.

What's here

• standards/ — Code quality rules loaded by /code-review: naming, testing, duplication, module depth, dependency discipline, UI preferences.
• hooks/ — Pre-commit secret scanning (filename, regex, gitleaks).
• commands/ — Slash commands for code review (code-review plus the layered multi-agent deep-code-review), release checklists, cloud agent briefs, Chrome extension scaffolding, pre-public-repo audits, and selective context compaction (context-collect).
• skills/ — Agent skills: idea (vague → grilled spec → plan), run-plan (subagent execution), to-issues, grill-me, debug (root-cause-first investigation).
• CLAUDE.md — Global conventions cloned in as ~/.claude/CLAUDE.md: scope discipline, two-strike rule, git workflow, permission hygiene.
• settings.json — Claude Code settings: default model, permission allowlist + default mode, and statusline wiring.
• statusline.py — Optional Claude Code status line script showing session usage metrics.
• gitleaks.toml — Secret pattern rules for the pre-commit scan.
• network-allowlist.conf — Squid ACL entries used by agent-sandbox.

Using this

Clone into ~/.claude:

git clone https://github.com/curtyo18/agent-config.git ~/.claude

Or let agent-sandbox do it automatically on container start.

Heads-up on defaults. settings.json ships permissive defaults (bypassPermissions, skipDangerousModePermissionPrompt) tuned for the agent-sandbox container, where the container is the sandbox. If you clone this standalone onto a host, review settings.json first — outside the sandbox you'll likely want a stricter permissions.defaultMode.

Status line

statusline.py renders a compact, colour-coded session readout in the Claude Code status bar:

Left to right:

• cwd — current directory (dimmed; ~ for home, long paths truncated) for orientation.
• ctx:N% (Mk) — context window used: percentage of the window plus the absolute token count (in thousands). Its colour tracks the absolute tokens, not the percentage — what degrades a session ("context rot") is how many tokens are in play, not how full the window is, so ctx:45% can already be red if that's ~80k tokens.
• 5hr:N% (Xh Ym) — Claude.ai Pro/Max 5-hour rate-limit usage and time to reset; falls back to session:$X.XX (session cost) when no rate-limit data is present.
• weekly:N% — Claude.ai Pro/Max 7-day rate-limit usage.

The context colour escalates with absolute size — red at ~75k, and bold red at ~100k, where context rot starts to bite (your cue to compact, summarise, or start a fresh session):

Context tokens	Colour
< 25k	green
25k–50k	yellow
50k–75k	orange
75k–100k	red
≥ 100k	bold red (rot territory)

(The 5-hour and weekly bars use a percentage scale instead: green < 40%, then yellow, orange, red ≥ 75%, and bold red ≥ 90%.)

Planning & execution skills

Self-contained — no plugin dependency; clone the repo and they work as-is. idea is the front door (vague idea → grilled spec → numbered plan → handoff menu); run-plan executes the plan via per-task subagent dispatch with two-stage review (or inline); to-issues converts it to GitHub issues instead; grill-me (a relentless-interview approach inspired by Matt Pocock) drives the questioning in idea's shaping and grilling phases and also runs standalone to stress-test a plan or design. They compose, but each stands alone. (Reimplemented from what the superpowers plugin used to provide.)

Adapting it

Fork and edit directly. The standards, hooks, and commands are designed to be modified — they represent opinions, not rules.