spec-development

Drive a single unit of work — one spec, ticket, finding, or change request — from spec to a committed, reviewed result with a dev team of coordinating agents: Dev to QA to Reviewer in a bounded cycle. Usable standalone, or as the per-subtask engine behind implement-full-spec.

Autonomous, self-re-prompting agent loops for Claude Code — packaged as independently-installable plugins (spec-development, dev-team, pr-autopilot, triangulated-code-review, multi-llm-convergence).

(beta/experimental) Direct-CLI, configurable, N-model variant of multi-llm-convergence. Reaches every model through its own CLI (claude, codex, gemini, grok) instead of the Codex companion script, driven by a JSON adapter registry you extend with a data edit. The operator picks which models converge (≥2 distinct families) and the host is a pure orchestrator. Keeps the original's guarantees — local grounding, a liveness watchdog, per-round commits, and a hard preflight stop (reachable + read-only proven by a negative probe) when cross-model consensus is impossible. Ships parallel to multi-llm-convergence (untouched) until promotion.

Triangulated multi-reviewer code review orchestrator. Borrows from research methodology — checks each finding against multiple independent reviewers (comprehensive, security, codex, codex adversarial) to reduce blind spots, then runs a QA analyst pass that substantiates every finding (verifying any third-party library claims via the context7 MCP) and demotes unsubstantiated ones into a dedicated 'Invalidated Findings' section of the prioritized, timestamped report.

Autonomously watch a single pull request and drive its review feedback to completion — each round fetch unaddressed reviews and comments from bots and humans, address the actionable ones, commit, push, reply/resolve threads, and re-request review, looping until a quiet round or a safety cap. Includes a --single-pass mode (one autonomous round, no loop). Owns the shared pr-review-mechanics reference.

Drives any artifact (plan, PRD, design doc, spec, or implementation diff) to convergence by alternating two genuinely different LLM reviewers — a Codex reviewer and an independent Claude review subagent — applying each round's findings and looping until both independently agree it clears the bar. Grounds offline reviewers in local source-of-truth clones, runs a liveness watchdog so a silent reviewer can't hang the loop, and commits per round for an auditable trail.