Claude Code is the executor — the agent that takes a TaskSpec and turns it into a working PR. Unlike the other agents, Claude Code operates directly on the codebase: reading files, writing code, running tests, creating commits, and opening PRs. It's the only agent with file system access.
The night-shift execution model is one of Novact's most distinctive features. When the CEO approves specs during the day, Maren queues them with a night-shift-eligible flag. During off-hours, Claude Code picks up queued tasks, implements them, runs preflight checks (8 validation steps), creates draft PRs, and outputs TCP (Task Completion Protocol) bundles for Nova to verify at deploy time. The CEO wakes up to completed work, not a queue of requests.
Claude Code never marks its own tasks complete — that's Nova's responsibility after deploy verification. This separation of execution and verification is an architectural decision (ADR-0086): the builder doesn't grade its own work. Five safety gates (pytest, smoke test, health check, migration validation, merge conflict check) ensure that even autonomous overnight deploys maintain production quality.
Opus for complex architectural tasks, Sonnet for routine implementations. Model selection happens per-task based on spec complexity score.