@orvix-id/omp

Oh My Pi (omp) extension for the Orvix Coding Plan. omp runs the agent; this extension connects it to https://api.orvix.id/coding/v1 and keeps every session on the prompt cache.

Using OpenCode or Pi instead? See @orvix-id/opencode and @orvix-id/pi.

Setup

omp needs Bun 1.3.14 or newer.

1. Create a Coding key. In platform.orvix.id/api-keys, create a key and tick Orvix Coding (coding:invoke). The default AI Router scope (ai:invoke) is not enough; Coding endpoints answer 401 coding:invoke scope required.

2. Expose the key to omp:

   export ORVIX_CODING_API_KEY="orv-sk_live_..."

   Put it in your shell profile or secret manager rather than typing it into commands you share.

3. Install the extension, then restart omp or run /reload-plugins:

   omp plugin install @orvix-id/omp

4. Pick a model: omp --model orvix-coding/deepseek-v4-flash.

Models

The model list comes from GET /coding/v1/models when the key is available, so it always matches your plan. Without network access the extension falls back to the lineup it shipped with. Model ids drop the orvix/ prefix: orvix-coding/glm-5.2 calls orvix/glm-5.2.

What the extension does

| Piece | Why | | --- | --- | | Provider orvix-coding on omp's built-in openai-completions API | Message conversion, tools, and usage accounting stay omp's own. | | session_id added in before_provider_request | Orvix keeps a session on the same upstream route and prompt cache only when the request body carries session_id. omp's session id is stable across --continue. | | x-orvix-coding-client header | Reports the client name and extension version. Used for diagnostics only. |

Cache behaviour to know about

omp rewrites parts of its own prompt on purpose, and each rewrite costs one cache miss from that point on:

• Skill summaries. omp shortens skill descriptions with a model call and swaps the summary into the system prompt after the first turn. This happens once per new skill.
• Superseded reads. When a file is read again, omp replaces the older result with [Superseded by a newer read of this file]. It only does this while few tokens follow the old result, or after the session has been idle for a while, so the cost stays small.

omp's system prompt is also the largest of the supported harnesses. Expect the first session on a new machine or project to use more uncached tokens than OpenCode or Pi.

Troubleshooting

| Symptom | Cause | | --- | --- | | SyntaxError: Unexpected identifier at startup | Bun is older than 1.3.14. | | No orvix-coding models | The extension is not installed, or ORVIX_CODING_API_KEY is not set in omp's environment. | | 401 coding:invoke scope required | The key was created without the Orvix Coding scope. | | 429 coding_concurrency_exceeded | Too many Coding requests in flight at once. omp retries; retry_after is a few seconds. | | 429 coding_quota_exceeded | A 5-hour, weekly, or monthly window is used up. resets_at says when. |

Development

bun install
bun run check
PROBE_CLI=omp bun run probe   # 3 real omp turns against a local mock, reports prefix reuse per request

License

MIT