Nerve
Blockchain Development

Nerve

Reflexive AI orchestration CLI that routes coding tasks through a lead/reviewer agent loop with hash-checked patches

GitHub
Rust Cargo Workspace Claude Code Codex Tokio

Problem

Single-agent coding has a predictable failure mode: the same model that writes a patch tends to rationalise it. Ad-hoc reviewers in a chat window are easy to skip, hard to audit, and produce no machine-readable trail of what was proposed, what was challenged, and what actually landed on disk. A reviewer needs to be a first-class step in the orchestrator, not a polite suggestion bolted on at the end.

Approach

  • Lead / reviewer split as the core abstraction. The lead proposes a patch; the reviewer returns LGTM / REQUEST_CHANGES / BLOCK; the orchestrator iterates until a configured policy accepts.
  • Generic ModelAdapter trait so any CLI or model — Claude Code, Codex, or a deterministic mock — plugs in identically.
  • Patches as data, not side effects. All changes are emitted as NvPatch values with SHA-256 checks; nothing touches the filesystem unless --apply is passed.
  • Strategies as a dispatch layer. consensus, pipeline, and tournament modes route a single task through different combinations of agents.
  • Auditable by default. Round records, event streams, and patch index live under .nerve/ so every session is replayable.

Implementation

Cargo Workspace and nv CLI

A multi-crate Rust workspace exposes the nv binary. Config loading, profile matching, and a deterministic mock lead/reviewer loop give a reliable test path that runs without any LLM calls (NERVE_ADAPTER=mock).

Real Subprocess Adapters

The Claude Code and Codex adapters spawn the upstream CLIs as subprocesses, parse JSONL string fields, and convert unified diffs into safe NvPatch values for create / modify / delete / rename operations.

Patch Apply / Rollback with Hash Checks

NvPatch apply and rollback are SHA-256 verified end-to-end and indexed so previously-applied patches can be reversed deterministically.

Strategies, Daemon, and TUI

consensus, pipeline, and tournament strategies dispatch the same task across configured adapter pairs. A line-oriented daemon mode and a terminal TUI (--tui) surface session summaries and per-round events. A scratch-file watcher feeds real-time cross-firing between agents during long runs.

Machine-readable Reports

Every session can be exported as --json for downstream tooling, and the persistent history under .nerve/ keeps the patch index and round records across runs.

Outcome

  • Phase 1 MVP plus the planned Phase 2/3 CLI execution features all implemented and pushed.
  • Mock adapter loop is fully deterministic and tested; real Claude Code / Codex subprocess paths convert unified diffs into typed patches.
  • Verification gates (fmt, clippy, full tests, config validate, mock doctor, daemon smoke, TUI smoke) all pass on the merged roadmap.
  • Three-strategy dispatch (consensus, pipeline, tournament), persistent .nerve/ history, daemon and TUI surfaces, and machine-readable session reports are live.

Technologies

  • Language / Runtime: Rust, Tokio, Cargo workspace
  • Agents: Claude Code, Codex CLI (via ModelAdapter subprocess)
  • Patch Format: NvPatch with SHA-256 integrity, unified-diff conversion
  • UX: nv CLI, --tui terminal interface, daemon line-oriented mode
  • Audit: .nerve/ history, patch index, per-round event stream, --json reports
← Back to Projects