Claude Code Configuration

Personal Claude Code config for Farista (@latuconsinafr) — hooks, skills, agents, and memory system extending Claude Code with principal-engineer workflows.

Published at https://github.com/latuconsinafr/claude-code-config

Quick start

Clone and symlink:

git clone https://github.com/latuconsinafr/claude-code-config ~/.claude
# or use directly without symlinking if Claude Code has already discovered ~/.claude

Install system dependencies:

brew install jq git curl vjeantet/tap/alerter

Set up Anthropic API OAuth token in macOS Keychain (optional, for quota display and session summaries):

security add-generic-password -a "Claude Code-credentials" -s "<token_json>" \
  -T /usr/bin/security

Directory structure

Directory	Purpose
scripts/	5 hooks + status line for safety, notifications, context tracking; details
skills/	14 slash-command orchestrators for common workflows (index)
agents/	8 specialist subagents for deep work in isolation (index)
projects/	File-based memory per project — behavioral rules, decisions, context
session-log.jsonl	Cumulative session history with AI-generated summaries
guard-blocked.log	Audit log of blocked bash commands by guard-bash.sh
CLAUDE.md	Master reference: principles, rules, autonomous invocation triggers

Skills (slash commands)

14 workflow orchestrators. Invoke with /skillname or trigger automatically based on rules in CLAUDE.md.

Skill	What it does
/plan	Structured implementation plan before coding — spawns explorer then architect if needed
/spec	Full technical specification (data model, API contract, edge cases) for complex features
/issue	Fetch GitHub/Jira issue → explore codebase → produce ready-to-execute plan
/commit	Generate conventional commit with ticket detection, dirty-diff check, breaking change flag
/review	Principal engineer code review (correctness, security, perf, tests) — spawns reviewer
/pr	Create GitHub PR with semver-prefixed title and structured body
/debug	Systematic bug investigation (reproduce → isolate → fix → regression test) — spawns debugger + qa
/research	Research unfamiliar libraries, architectural decisions, or technology choices — parallel subagents
/simplify	Remove over-engineering and dead code without changing behavior — spawns refactoring
/audit-deps	Scan all lockfiles (npm, pip, go, cargo, etc.) for CVEs, outdated majors, license violations
/tech-debt	Surface TODOs, deprecated APIs, dead code with age and priority
/security	Dedicated OWASP Top 10 security review (deeper than /review security section)
/env-audit	Audit all env var references vs. sources across code, Docker, CI, config files
/onboard	Session-start context primer — architecture, recent work, conventions, open issues

See skills/README.md for invocation details and agent chains.

Agents (specialist subprocesses)

8 principal-engineer-level agents, isolated context, chained by skills.

Agent	Model	Purpose
architect	opus	Design review before implementation — correctness, scalability, reversibility, system implications
debugger	sonnet	Root cause analysis (reproduce → isolate → hypothesis → verify → generalize)
reviewer	sonnet	Code review — correctness, security, performance, pattern precedent, PE lens
qa	sonnet	Quality assurance — risk-based test strategy, edge cases, behavior verification
refactoring	sonnet	Restructure code without behavior change — blast radius check, verify per step
performance	haiku	Bottleneck analysis — N+1 queries, hot paths, algorithmic complexity
explorer	haiku	Read-only codebase investigation — structure, usage, context gathering
docs	haiku	Documentation — keep docs in sync, capture decisions (ADRs), explain why

See agents/README.md for full chain map and frontmatter format.

Hooks & status line

Full details in scripts/README.md

Hook	Trigger	Purpose
guard-bash.sh	PreToolUse (Bash)	Block dangerous commands: secrets access, destructive SQL, rm -rf, force-push, etc.
on-permission.sh	PreToolUse (all tools)	macOS desktop notification when Claude is about to use a permission-required tool
on-stop.sh	Stop	macOS notification with task summary, turn/tool-call stats, inferred stop reason
on-file-change.sh	PostToolUse (Edit/Write/MultiEdit)	Track modified files per session for on-session-end.sh
on-session-end.sh	SessionEnd	Log session stats + AI-generated summary to ~/.claude/session-log.jsonl
status-line.sh	Custom status line	Display user, project, git branch, model, context %, quota (5h/7d) with color thresholds

All scripts require jq. on-permission.sh, on-stop.sh require alerter (brew install).

Memory system (dual)

Two complementary systems for knowledge retention.

File-based (projects/<encoded-path>/memory/)

• Human-readable markdown with MEMORY.md index
• YAML frontmatter: name, description, type (user/feedback/project/reference)
• Best for: behavioral rules with why, rich context, nuanced guidance
• Write when the memory needs explanation to be applied correctly

MCP memory (structured knowledge graph via mcp__memory__* tools)

• Entities, observations, relations
• Best for: named facts requiring fast recall (stack, key people, decisions, relationships)
• Write when the memory is a fact about a named entity — person, project, technology, concept

At session start:

1. Check MEMORY.md in file-based system
2. Run mcp__memory__search_nodes for current project/task

MCP dependencies

Server	Usage
context7	Fetch current docs for any library/framework before answering tech questions
sequentialthinking	Enable for complex multi-step problems, architectural decisions, debugging
mcp__memory__*	Structured knowledge graph — entities, observations, relations

System dependencies

All scripts tested on macOS with zsh.

Tool	Used by	Installation
jq	all scripts	brew install jq
git	guard-bash, status-line	pre-installed
curl	on-session-end, status-line	pre-installed
alerter	on-permission, on-stop	brew install vjeantet/tap/alerter

Autonomous skill invocation

Skills are automatically invoked per CLAUDE.md rules. Key triggers:

• Start of unfamiliar project → /onboard
• From a ticket/issue → /issue (not /plan)
• Complex/high-stakes feature → /spec then /plan
• Any non-trivial change → /plan before coding
• Before commit → /commit (not git commit)
• Before PR → /review then /pr
• Bug/error → /debug (not guessing)
• Technology choice unclear → /research

See CLAUDE.md for full rules and conditions.

Project reference

• Owner: Farista (latuconsinafr)
• License: Check repository for license file
• Audience: Personal use; shared for reference and inspiration
• Status: Active — hooks and skills in regular use across project work

---

For detailed information on hooks, see scripts/README.md For skill invocation and chains, see skills/README.md For agent roles and specialization, see agents/README.md For principles and rules, see CLAUDE.md