FAQ¶
Why not just one big plugin?¶
Different teams want different surfaces. A library team wants review-pack + tdd-pack. A frontend team wants verification-pack + safety-pack + the Playwright MCP. Splitting into 7 functional packs (plus 1 umbrella) means selective install without "all or nothing."
Why hooks at all? Settings permissions cover most of the same ground.¶
permissions.deny is the cheaper, more reliable layer; we use it first (see the permissions model). Hooks add what permissions cannot pattern-match: high-entropy secret detection, branch lookup at runtime, MCP allowlisting, Stop-time verification. The two layers compose.
Why Python for guard hooks and bash for orchestration hooks?¶
Python has stdlib JSON parsing, deterministic exit codes, and a richer regex implementation. Bash is shorter for scripts that mostly orchestrate other tools (formatters, test runners). The shared plugins/_shared/hook-prelude.sh removes the boilerplate from bash hooks.
How is this different from codex-harnesses?¶
Same idea, different agent. Skills are largely shared in spirit; hooks differ because Codex and Claude Code expose different lifecycle primitives. See migrating from codex-harnesses.
Can I use this in a repository that already has hooks/skills/agents?¶
Yes. scripts/install.sh deep-merges into existing .claude/settings.json, concatenating per-event hook arrays rather than replacing them. Existing skills/agents/commands are left in place; the install adds new ones alongside.
Will the cost ceiling stop a productive long session?¶
The default ceiling is 5000 tool calls per 24 hours. A typical "long session" is 200–800 tool calls; multi-day work runs around 2000. The ceiling is a guardrail against runaway loops, not against productive work. Raise it with CLAUDE_HARNESSES_COST_CEILING=10000 if you're sure.
Are subagent context isolations actually enforced?¶
Honor system, not sandbox. The TDD subagents (tdd-test-writer, tdd-implementer, tdd-refactorer) are constrained by their prompts and their tools field; they do not see each other's output unless the parent agent passes it. We rely on the parent agent to honor the boundary. See the tdd pack page for the threat model.
Why aren't the MCP servers enabled by default?¶
MCP servers come from third parties and may change scope, auth, or stability between releases. Shipping them disabled (under _disabled in .mcp.json) means installing the pack does not surprise you with extra processes, network calls, or write access. Each server has a per-server doc with last verified date.
What does "harness" mean in this project?¶
A harness is the durable scaffolding around an autonomous agent: the durable guidance file (CLAUDE.md), reusable workflows (skills/), enforcement (settings.json + hooks), task memory (ledger/), and verification (scripts/verify.sh). The agent supplies intelligence; the harness supplies guardrails, memory, and reusable workflows. See harness model.
Can I contribute a new pack?¶
Yes — see CONTRIBUTING.md. New packs need: a plugin.json validated against schemas/plugin.schema.json, a marketplace entry, a docs page under docs/packs/, and tests where applicable.
How do I disable just one hook without uninstalling the pack?¶
Edit .claude/settings.json and remove the matching entry from the relevant event's array. Or for a single session, the global kill switch:
export CLAUDE_HARNESSES_DISABLE=1
What's the relationship between skills at the repo root and skills inside packs?¶
The repo-root skills/<name>/SKILL.md is the canonical home so gh skill install and npx skills add work directly against the repo. Packs reference the same skills via relative symlinks under plugins/<pack>/skills/, so plugin install picks them up too. CI verifies the symlinks resolve.