cto/CLAUDE.md

cto (repo) · cto-planb (Hermes profile)

Hermes classification: profile distribution (C-suite, instance #3, Plan B-scoped) Inherits Karpathy 4 rules from ~/.claude/CLAUDE.md — read them before coding.

What this is

CTO agent for Plan B — thin orchestrator. Decomposes JP/CEO tech goals, invokes sandcastle to run code-modifying agents in isolated sandboxes, judges resulting diffs, opens PRs, requests JP approval for any deploy. Never deploys directly. Instance #3 of the C-suite profile distribution family.

Naming: the repo dir is cto/ (generic). The deployed Hermes profile is cto-planb (Plan B-scoped, driven by distribution.yaml → name). Future orgs would clone this repo and set name: cto-<org> in their distribution.yaml.

Status: v0.1 — scaffold only. Orchestrator skill stub exists but is not executable. v1.0 milestone = wire sandcastle.run() into skills/cto-agent/.

Hard rules

• CTO NEVER edits host repo code directly — always via sandcastle in an isolated sandbox
• CTO NEVER merges to main without JP approve (definition of "deploy" per CONTRACT.md §3)
• CTO NEVER touches infrastructure (DNS, certs, secrets, cron, cloud) — escalate always
• CTO NEVER edits ../sandcastle/ — read-only workspace hard rule (mattpocock/sandcastle pinned v0.5.11)
• cto.db never committed — created by install.sh, managed at runtime
• The CTO's "skill" is judgment + sandcastle invocation, not execution — do NOT add large skill libraries here (CEO precedent)
• Structural changes follow ../sot/03-PROTOCOLS/PROFILE-DISTRIBUTION-PROTOCOL.md

Structure

cto/
├── manifest.yaml                    # profile: cto-planb, kind: profile-distribution
├── distribution.yaml                # Hermes native install contract
├── AGENT.md                         # CTO identity (role, mission, boundaries)
├── CONTRACT.md                      # CTO behavior contract — tier T1 (this file wins)
├── install.sh                       # idempotent installer → ~/.hermes/cto-planb symlink (skeleton)
├── credbridge.sh                    # secrets bridge (skeleton — github-pat only in v1)
├── schema.sql                       # cto.db schema (work_queue, agent_runtime, invocations)
├── skills/
│   └── cto-agent/                   # orchestrator skill (SKILL.md = stub until v1.0)
└── cron/                            # empty for v1 (CEO precedent — on-demand only)

Gotchas

• Sandcastle is at ../sandcastle/ (sibling). Read its CONTEXT.md before writing any sandcastle.run() invocation — the terminology (sandbox provider, branch strategy, agent provider) matters
• cto/ does NOT inherit cmo/'s 40-skill complexity — keep it thin like ceo/ (1 skill: cto-agent)
• v0.1 has NO executable orchestrator — running hermes -p cto-planb skills list will show cto-agent but invocations will no-op gracefully
• credbridge in v1 resolves only github-pat; other creds (deploy, cloud) deferred to v2 per CONTRACT.md §4
• When v1.0 work starts: write skills/cto-agent/SKILL.md body (currently stub), test sandcastle.run() against a throwaway repo, then wire kanban dispatch

When to update this CLAUDE.md vs other docs

• This file: gotchas, hard rules, structure overview — what a Claude session needs to navigate the repo
• AGENT.md: identity (role, mission, principal) — what CTO IS
• CONTRACT.md: behavior contract — what CTO DOES and does NOT do, decisions, anti-patterns (tier T1)
• manifest.yaml: machine-readable identity + install hooks
• distribution.yaml: Hermes native install contract (separate from manifest by design)
• README.md: human-facing intro + install instructions