OMNI v0.7.12: Which Project Spent It

0.7.11 showed what your agent was billed in tokens. This release says what that cost, and where it went.

Spend by model, project and session

omni stats --view bill now prices each request from the model that served it, at API list price, and splits the total three ways:

At API list price (2026-10-06): $1078.32 over 7,214 requests, subagents included.

Spend by model                            requests         tokens        USD   share
claude-opus-5                                2,125    953,466,343     559.51   51.9%
claude-opus-5-5                              5,061  1,383,803,492     516.31   47.9%
claude-fable-5-1                                20      2,117,625       2.22    0.2%
claude-sonnet-5-5                                8        515,991       0.28    0.0%

--by project and --by session give the same table for each repository and each session. Cache writes are priced by how long they live, and a model missing from the price table shows as unpriced with its tokens, never as $0. The figure is what the same work costs on the API; a subscription bills differently.

Fixed

sqlite3, psql and mysql output reaches your agent as printed. A query result was cut to three sample rows under a first row labelled as the column list, and a successful query whose rows contained the text error: was reported as a database error. The database summariser is gone.

A failing bun test keeps the lines that say where it failed: the file:line, the Expected and Received header, and the code frame.

A token inside YAML or other structured output is now hidden where it stands, and the rest of the document comes back unchanged. A comparison like token === "py" is no longer redacted as if it were a secret.

rustls and webpki move past six security advisories.

One limit stated plainly: a command the host reports as failed is still handed back untouched, credentials included. That is the next piece of #866.

brew upgrade omni