brain-compilerlisted
Install: claude install-skill sandeep84397/agent-brain
# Brain compiler
Read `{{CONTRACT_PATH}}` completely before generation. With a task-end lease,
retrieve it via `san_refresh(action="contract", refresh_token=..., repo=...)`
instead of filesystem/shell commands. Treat it as normative;
do not redefine SAN semantics in this skill.
Use only Agent Brain MCP tools inherited from the current host:
1. Reuse current repo context or call `get_roadmap(repo=...)`; call
`pre_check(repo=..., provider="openai", ...)` and `log_decision(repo=..., ...)`
for one bounded batch. Routing advice does not prove which model is running.
2. For a task-end lease, use the supplied `refresh_token` and file batch;
read each source with `san_refresh(action="read", refresh_token=..., source_path=...)`.
Do not re-plan, use shell commands, edit source, or add files to this batch.
Otherwise call `plan_san_refresh`; read only missing, stale or malformed sources.
3. Apply all contract parity checks before publication.
4. Call `publish_san(provider="codex", ...)` once per validated source and never
write `.san` directly. Publication uses host identifier `codex`; the `openai`
identifier above is for routing advice only. Supply the configured model and
reasoning effort exactly; do not infer execution settings from routing advice.
5. With a lease, pass `refresh_token` on every publication. Stop on the first
read/publication error; no retry in this continuation. Return its status to
the parent, which calls `san_refresh(action="finish",