This document explains the Execution composition: entry, routing, and handoff. Internal step-by-step workflows live in each Skill's SKILL.md and references/.
light-implement— user-invoked general-purpose bounded executoragent-config— model-invoked profile-driven cross-harness execution configuratorlight-tdd— model-invoked test-driven looplight-diagnosing-bugs— model-invoked diagnosis loopresolving-merge-conflicts— model-invoked merge/rebase resolver
All are first-party and self-contained; no install of mattpocock/skills or sol-advisor is required.
| Situation | Entry | Typical path | Handoff / Stop |
|---|---|---|---|
| One clear ticket / SPEC slice | light-implement — user-invoked |
light-implement → inspect context → optional agent-config offer (when profile routing, review isolation, or topology materially helps) → execute → verify → hand to review-loop with the right reviewer |
clear code diff + focused tests + verification evidence; stop at ticket scope |
| Need to configure model, effort, or execution topology | agent-config — model-invoked |
requires: bounded task description + acceptance criteria + current host evidence + confirmed profile; determines provider mode and task complexity, right-sizes model tier & effort across execution modes | execution plan, not direct execution; executor performs work per plan |
| Code feature/fix should be test-first | light-tdd — model-invoked |
red → green → refactor test-first cycle |
test suite + implementation slice |
| Hard bug / regression | light-diagnosing-bugs — model-invoked |
build a tight pass/fail signal → reproduce → hypothesize → instrument → fix → cleanup |
fix accompanied by diagnostic evidence |
| Merge/rebase conflict | resolving-merge-conflicts — model-invoked |
resolve conflicted files per git guidance | clean working tree ready for verification |
The execution configuration system cleanly separates policy reasoning from host persistence and mutation:
- Skill (
skills/engineering/agent-config): Installed from this repository (LightDevCoder/skills). Performs task difficulty assessment, tier selection, and execution topology planning. Runs locally in the agent conversation. - Companion MCP Runtime: Maintained in the independent repository LightDevCoder/agent-config. Provides 10 native host adapters (Codex, Claude Code, Antigravity / agy, DeepSeek Harness / DSH, OpenCode, ZCode, Cursor, Grok Build, Hermes, Pi) plus generic fallback.
- Optional Companion: Without the companion MCP server registered,
agent-configremains fully operational in session-local, plan-only mode without mutating host configuration. With the companion installed and registered, it enables authentic host capability inspection, profile persistence, configuration preview before apply, and health validation.
light-implement does not copy reviewer instructions. Coding work follows:
light-implement → light-tdd (when appropriate) → code changes + tests → review-loop → light-code-review
Non-coding work:
light-implement → artifact → review-loop → generic-review / domain reviewer
The reviewer is read-only; review-loop is the convergence engine; project-review (see review-system) owns final PASS/FAIL/BLOCKED when project acceptance is needed.
- Vague idea → clarification-system first.
- Approved SPEC →
project-ticketsbeforelight-implement. - Completed project →
project-review/ review-system. - Unknown entry →
ask-light.