Claude Code setup — parallel multi-agent starter kit

My Claude Code setup: an opinionated CLAUDE.md, plus a working parallel multi-agent system — a skill that orchestrates fan-out work and the role agents it spawns. Clone, run the installer, and you have a plan → execute → review → merge pipeline that fans out across background subagents (the default — in-process, isolated worktrees only where they write in parallel, no extra setup).

The default path uses ordinary background subagents and (optionally) Workflows — no experimental flags, no extra setup.

Fan-out uses significantly more tokens than a single session — use it for parallel research, review, and feature work, not routine tasks.

What's inside

Path	What it is
global/CLAUDE.md	Lean always-on layer: principles (think-before-coding, simplicity, surgical changes), the hard gates (push approval, codex gate, AFK-not-approval), and a pointer to the feature-workflow skill. Lives under global/ so working sessions in this repo don't load it twice alongside ~/.claude/CLAUDE.md
global/rules/	Path-scoped user rules, installed to ~/.claude/rules/ — load only when a matching file is touched, so they don't add to every session's always-on context
skills/feature-workflow/SKILL.md	The six-stage single-master feature pipeline, the parallel-multi-agent mechanism picker, and the token-discipline rules. Loads on demand when a pipeline or fan-out starts.
skills/agent-teams/SKILL.md	The orchestration playbook — when to fan out, how to pick the mechanism (subagents / Workflows), the pipeline, models, worktree/merge flow, the plan-approval gate. Loads on demand.
agents/explorer.md	Read-only codebase search on Sonnet at effort medium — the pinned stand-in for built-in Explore (Sonnet)
agents/plan-reviewer.md	Validates the plan against the code before the lead presents it via ExitPlanMode for your approval (the session's model and effort)
agents/team-executor.md	Parallel fan-out only. Implements one unit of the fan-out as a background subagent — carries isolation: worktree in its frontmatter, since concurrent writers merge later (Sonnet, high; Opus when the plan marks the unit)
agents/step-executor.md	Implements one sequential step on the session's own branch — no worktree, nothing to merge; the feature-workflow counterpart to team-executor (Sonnet, high; Opus when the plan marks the step)
agents/fixer.md	Fixes one review round's finding set (P0/P1 plus adjacent P2s) on the session's own branch, test-first red-then-green — a brownfield bug fix, so it runs deeper than a plan step (Sonnet, high; Opus when a finding's mechanism survived an earlier round)
agents/codex-triage.md	Reads one round's codex-challenge.sh output file(s) — all slices of a split round — verifies each finding against git show <head>: and git diff <base> <head>, and returns the single ≤2k deduped verdict; the run itself is a background Bash in the master, the only context the harness re-wakes on completion (Sonnet medium)
agents/spec-reviewer.md	At the final gate, checks the feature's whole diff against the approved plan file — missing requirements, scope creep, wrong-logic-vs-spec; gaps only, in parallel with the whole-range codex challenge (Sonnet medium)
agents/team-reviewer.md	Parallel fan-out only. Adversarially verifies each unit's diff before merge — read-only, no worktree (Opus)
agents/team-merger.md	Parallel fan-out only. Merges the worktrees the lead names into the base branch, removes each worktree + branch after landing, reports done (Sonnet)
settings.example.json	model: "opus" + modelSettings.claude-opus-5-5.effortLevel: "xhigh" (the recommended master — the session you open — the latest Opus at xhigh; the install wizard and /stack-update ask: recommended / keep current / don't ask again / another model, and install.sh --master=recommended applies it), modelSettings.claude-sonnet-5-5.effortLevel: "high" (the effort of Sonnet spawns — executors, fixer, unpinned; the agents carry no effort: line — filled in by install.sh only where you have none saved, never overriding, and not part of the master recommendation), worktree.baseRef: "head" so executor worktrees branch from your in-progress branch rather than the remote default, CLAUDE_CODE_ENABLE_TODO_TOOLS (the task-list feature), CLAUDE_CODE_SUBAGENT_MODEL (the Sonnet floor for unpinned spawns — see Model pinning), BASH_DEFAULT_TIMEOUT_MS: 900000 (a build or test run with no explicit timeout is no longer auto-backgrounded at 2 minutes; this is also the ceiling), CODEX_REVIEW_MODEL + CODEX_REVIEW_EFFORT (which codex model and reasoning effort the cross-review gate uses — set here, not in the script, because install.sh replaces the skill directory on every run; reinstalling never overrides an existing value, so change it by editing settings.json), bashOutputMaxChars: 30000 (a valid command result over 30k characters arrives as a file path plus a 2k preview instead of flooding the context), bashEditDiffEnabled: true (the transcript records which files each Bash command changed — /analyze-arcs reads that instead of parsing commands; on by default only in auto mode, so it is pinned for every mode; never shown to the model), the SessionStart update-check and session-name hooks, and the PreToolUse subagent-no-background hook
hooks/subagent-no-background.sh	PreToolUse on Bash: denies run_in_background inside any subagent (with fork mode on every spawn is a background subagent, whose background commands keep running past its final report — nobody stops them) and any until/while poll on a .output.done marker (the harness never writes one). Fail-open on anything it does not understand; tests in hooks/tests/
`hooks/ses