Sol Project Workflow

Problem

Running a large software task end to end with a coding agent means covering a lot of ground consistently: inspect before touching anything, plan, implement, verify with real evidence, handle blockers, and close out without silently declaring victory. Sol Project Workflow packages that as one reusable Codex/Claude Code plugin instead of re-deriving the same steps every session.

Approach

The plugin is five companion skills — sol-project-workflow as orchestrator, plan-project-work, run-codex-goal-loop, validate-software-work, and report-workflow-state — each with its own SKILL.md and per-model agent adapters. The orchestrator drives a state machine (intake, planning, goal-init, implement, verify, QA/review, learning closeout, complete) under a single Codex goal per batch, delegating only bounded lanes with disjoint write ownership and reviewing every return before trusting it. Validation runs as separately observed commands rather than one pass/fail bundle. A closing learning pass records the outcome and can suggest, but never auto-apply, a process improvement. The contracts behind blockers, goals, plans, routing, and validation are written down as reference docs and backed by a dedicated workflow_contract.py, with its own test file alongside it.

Outcome

The repo is private. A single commit shipped the entire skill set on 2026-07-22, and it hasn't been touched since.