FAQ

FAQ

General

What is code-conductor?

A governance layer for Claude Code that installs slash commands, hooks, guardrails, and a persistent memory system. It enforces a spec-first workflow and cuts token usage.

Does it modify my project's code?

No. It only creates/manages files inside .claude/ and optionally .conductor/ (gitignored). Your source code is untouched.

What stacks does it support?

Any stack. /cc-stack auto-detects from package.json, go.mod, Cargo.toml, pom.xml, build.gradle, pyproject.toml, Gemfile, and others.

Is it a security sandbox?

No. The guards are process guardrails for cooperative agents, not a security boundary. They assume agent_type is reported honestly and do not cover every write path (e.g., Bash can still write files). See the Known Limits section in the README.

---

Installation

I get errors about CLAUDE.md being damaged

Install @latest (minimum 1.34.4). Older versions could strip or damage existing CLAUDE.md files.

The installer prints nothing — is that normal?

Yes. Silent exit 0 is by design and is asserted by a test.

How do I update?

npx @yeison.restrepo.r/code-conductor@latest --project

The installer is idempotent: it merges settings, preserves your content, and only updates managed sections.

---

Guards

Guard 3 blocked a command I need — what do I do?

Three options, in order of preference:

1. Use the targeted alternative shown in the denial message (grep, glob, bounded read)
2. Add an allowlist entry in .claude/memory/bash-scan-allowlist.txt with a comment explaining why
3. Set CC_GUARD3_WARN=1 for triage (ask instead of deny) — this is a triage aid, not a permanent setting

How do I debug a guard denial?

CC_HOOK_DEBUG=1   # shows diagnostic lines the hook normally swallows
CC_HOOK_ALLOW=1   # overrides an unparseable-input denial only

Can I disable all guards?

Don't. If a specific guard is wrong, report it as an issue. The guards are the value proposition — disabling them returns you to unguarded Claude Code.

---

Orchestrator

What's the difference between /cc-implement and /cc-orchestrate?

/cc-implement is manual: you drive each phase yourself. /cc-orchestrate is automated: it dispatches role agents through Define → Build → Verify with validation at every boundary.

The orchestrator halted — how do I recover?

# End the current run
node .claude/scripts/orchestrate.mjs end

# Start fresh
node .claude/scripts/orchestrate.mjs start

Do I still need to approve things in --auto mode?

Yes. --auto skips pauses between role dispatches, but spec and plan approvals always require your explicit approval.

---

Memory

Where is my project memory?

.claude/memory/project.md — committed to git, shared with your team.

Where are my personal preferences?

~/.claude/memory/personal.md — local, never committed.

How often should I run /cc-checkpoint?

Before /compact, after completing a feature, and after key architectural decisions. The more you checkpoint, the less you re-explain.

---

Token Usage

How much does code-conductor save?

Depends on your usage pattern. The biggest savings come from:

• Guard 1 blocking unbounded file reads
• Guard 3 blocking mass content dumps (47 denials measured in one real session)
• Memory-first lookup chain avoiding file reads entirely
• SNAP handoffs being 15%+ smaller than markdown
• Verbosity MIN producing one-sentence responses

Does code-conductor itself consume tokens?

The hooks are lightweight (pre-tool-use.mjs is pure Node, no model calls). Commands like /cc-spec and /cc-plan do consume tokens for their model work, but they prevent the much larger waste of building the wrong thing.