# OpenCode through Yabanin

Use Yabanin as the only model provider (`yabanin/glm-5.3-flash` preferred; `yabanin/glm-5.3` and a local model stay selectable). Ordinary turns, file reads, edits, and bash go through the OpenAI-compatible gateway.

Call the MCP tool `tr_delegate` only for hard coding work that needs a Codex or Cursor agent:

- `mode=analyze` for read-only review
- `mode=implement` for workspace writes
- `provider=auto` unless the user named Codex or Cursor

Cursor runs through the local `agent` CLI and the signed-in Cursor subscription (`agent login`). Codex uses `~/.codex` login. Never ask the user to paste API keys or cookies into chat.

## Model routing policy

Pick the route by task shape (all routes go through the Yabanin gateway; `yabanin/glm-5.3-flash` is the preferred model everywhere):

- Primary agent `yabanin` (glm-5.3-flash): leading the session, planning, complex reasoning.
- Primary agent `luna` (glm-5.3-flash): release-program executor lane — one leased card per worktree, contract scope only, no master merge.
- Primary agent `sol-review` (glm-5.3-flash): independent reviewer session for one delivered diff; verdict and handoff notes only.
- `@reason` (glm-5.3-flash): deep analysis, architecture, trade-offs, research synthesis — read-only by intent.
- `@code` (glm-5.3-flash): clear, well-specified code edits — templates, tests, docs, instruction-driven refactors.
- `@complex-code` (glm-5.3-flash; grok-4.6 when configured): heavy algorithmic codegen and complex multi-file implementation.
- `tr_delegate` (MCP): only when the routes above are not enough — Codex/Cursor for the hardest tasks.

The release fleet (`plans/release-2026-10-31/execution-policy.json`, owner decision 2026-10-12) runs 30–50 parallel lanes through the public Yabanin stand (`https://yabanin.ai/gw`; wire agents with `make opencode-remote-wire` — the local dev stand `make stand-opencode` is retired). Cards execute ONLY as internal OpenCode tasks (Task-tool subagents of the orchestrator session, observable in OpenCode); external headless process spawning is prohibited. `scripts/agents/fleet.sh` manages worktrees only (create/status/remove). Reviews run as independent `sol-review` sessions; master merge stays a single serial queue owned by Sol.