dualmode

# Dual Mode: Luna executes, Sol reasons The core idea: one model running everything burns context and quota. Splitting the job at the handoff boundary costs a tiny 1-3K token task packet and keeps Sol out of the repo while Luna does the grunt work. ## Roles - **Luna (execution)**: reads the repo, searches, runs commands, edits code, runs tests, verifies, and reports exact commands and results. - **Sol (reasoning)**: receives only the task packet, makes architecture and judgment calls, analyzes tricky bugs, and returns a structured plan. Sol never opens the repo. Luna never invents architecture. ## Artifacts - [references/loop-prompt.md](references/loop-prompt.md) is the paste-into- ChatGPT artifact when the user is not in Codex. - [references/task-packet.md](references/task-packet.md) is the fill-in packet template. Use it, or read it, when a Sol handoff needs compression. - `scripts/install.sh` installs this skill into Codex. ## The loop 1. **Luna boot**: read the relevant code, tests, and configuration directly. Run commands and gather facts. Do not guess. 2. **Compress**: write a task packet of 1-3K tokens covering goal, key code, constraints, attempted solutions, and the one open question. 3. **Hand off to Sol**: send the packet, nothing else. Sol returns root cause, impact scope, fix plan, and verification steps. 4. **Luna executes**: modify code, add tests, run the narrowest relevant checks, and verify until the acceptance criteria are met. 5. **Check done**: if the outcome and proof are met, report. If not, compress the new state and repeat from step 3. Cap at 5 Sol handoffs per task unless the user says continue. ## When to exit - Exit when the acceptance criteria pass and exact commands/results are reported. Do not keep looping for polish. - Exit after 5 Sol handoffs per task unless the user says continue. Treat the cap as a decision point, not an automatic stop after one plan. - If two consecutive Sol plans disagree with each other, stop and ask the user instead of cycling. - If the plan grows beyond the task, apply minimal-scope discipline and ask before broadening. ## Paste-into-ChatGPT prompt When the user wants the loop outside Codex (or just the raw prompt), read [references/loop-prompt.md](references/loop-prompt.md) and hand it over verbatim as the artifact. ## Example packet Use [references/task-packet.md](references/task-packet.md) as the blank template. A filled example is in [references/example-packet.md](references/example-packet.md); read it when you have never built a packet before or the user asks to see one. ## Constraints - Never send full project context to Sol. Only the packet. - A Sol handoff with no packet, or a packet missing the open question, is a violation: stop and compress before reasoning. - If the plan grows beyond the task, apply minimal-scope discipline and ask before broadening. - Report every cycle as `PACKET -> SOL -> PLAN -> LUNA -> VERIFY` so the user can watch the loop consume quota deliberately.