Executor Adds Evidence-Gated Workflows for Coding Agents

Summary

Executor is an initiative-scoped workflow system for coding agents that takes work from intake through architecture, specification, planning, execution, review, verification, and handoff. It gives each initiative a folder and strict ID namespace, separates tracked reasoning documents under docs/executor/ from execution records under .executor/, and uses POSIX shell scripts to enforce state-transition contracts. The system can load skills from Claude Code, Codex, omp, Cursor, Aider, or similar harnesses, and the repository provides both a manual installation path and a one-paste installation prompt; users can pin the v0.3.0 release tag instead of using the default branch. Its router selects phase skills, while scripts allocate IDs, lint plans, create execution workspaces, generate task briefs and context, package reviews and fixes, audit run state, and check store integrity. Each task receives a brief, context, implementer report, and mandatory review verdict, including for documentation-only work. Reviews separate specification compliance from code quality, use severity-labelled findings, and support limited fix rounds before escalation or a recorded ruling. Verification reruns each requirement against the current commit and records PROVEN, FAILED, NOT-RUN, or UNAVAILABLE; a NOT-RUN result prevents the work from being called complete. The branch model uses one initiative branch and one branch per plan, with review-gated non-fast-forward merges and a human decision at handoff. Generated artifacts preserve identity, chronology, evidence citations, and provenance so a controller can resume from the ledger after context loss. The project also requires credential-shaped content scans and treats the execution record as potentially public.