- name
- preview
- description
- Run a dry run of a Rubicon programming write batch via programming_write_dry_run to get a confirmationToken and a human-readable preview of the change before anything is published. Use this as stage 3 of the Rubicon programming pipeline, after programming-inference and before publish.
# Preview
Stage 3 of 4 in the Rubicon coach programming pipeline:
`sheet-read` → `programming-inference` → `preview` → `publish`
Input: the operation batch built in `programming-inference` (≤25
`create_node` / `patch_node` / `create_edge` / `patch_edge` operations).
Output: a `confirmationToken` and a diff-style preview of what would change
in Rubicon. **Nothing is written to Rubicon by this step.**
## What to do
1. Call the Coach MCP Surface tool `programming_write_dry_run` with the
exact operation batch from `programming-inference`, unmodified.
2. This is a protected tool — expect an OAuth challenge the first time in a
session (or after token expiry). Let the coach complete Codex's OAuth
flow as themselves; do not attempt to pre-supply or hardcode any token.
3. The tool returns:
- a `confirmationToken` — a SHA-256-based token computed server-side,
deterministic for the exact operation set + requester + track slug.
This plugin does not compute, cache, or modify this token. Treat it as
opaque.
- a preview of the resulting change (which nodes/edges would be
created or patched, and how).
4. Show the coach the preview in a clear, readable form — group by node vs.
edge, and call out anything that looks like it might be wrong (e.g. a
`patch_node` touching a node the coach didn't expect, or a movement edge
that resolved to a surprising `nodeId`).
5. Stop here. Do not proceed to `publish` until the coach explicitly
confirms the preview looks correct.
## Why the token matters
The `confirmationToken` is a hard safety gate the Rubicon server already
enforces (not something this plugin adds): `publish` will only succeed if
the operations sent to `programming_write_publish` are byte-for-byte the
same set that produced this token. If the coach wants to change anything
after seeing the preview, go back to `programming-inference`, rebuild the
batch, and re-run `preview` to get a fresh, matching token. Never try to
reuse a token against a modified operation set — the server will reject it,
and that rejection is working as intended.
## Handoff
Once the coach approves the preview, hand off to the `publish` skill with:
- the exact same operation batch used for this dry run
- the `confirmationToken` returned here
GitHub에서 보기