v0.2.2

Ordinus 0.2.2

Token usage tracking

Every run now records how many tokens it used: input, output, cached, and the run's real cost computed from the provider's raw counters. Codex reports cumulative session totals, Claude and Gemini report per-run totals; both are handled correctly. The data is stored per run and will feed the run inspector later.

Lower token consumption

• Resumed runs no longer repeat the workspace and output-format rules on every message. The session already has them; a one-line reminder is enough.
• When an agent continues its own work in the same session, its previous output is no longer sent back to it. It only receives a summary and a run id.
• These changes also improve provider cache hit rates.

Follow-up work can now see earlier results

Until now, a follow-up item added to an existing Work Request could not read the output of earlier items unless you linked them by hand. Three things fix that:

• The planner can now bind new items to existing completed runs on its own.
• The plan review dialog lists the request's earlier items, shows what the planner linked, and lets you change the links before starting.
• Each Work Request keeps a digest.md file in its working folder: a plain record of what was done, by whom, with run ids. Agents can fetch the full output of any listed run through a new read-only tool.

Better plans

The planning prompt was rebuilt. It now sizes plans from the actual boundaries in the request instead of leaning toward fewer items, writes its reasoning before producing the plan, checks itself before answering, and treats your agent selections as real choices instead of hints. A new example shows a healthy multi-agent plan.

Agent creation fix

When you described a new agent, the text you typed was saved as the agent's capabilities, overwriting the proper capabilities line the AI had generated. This misled the planner when assigning work. Fixed; the generated line is kept. New agents also get a short title as their role instead of a long sentence.

Security

The internal tool server now uses token-gated endpoints. Worker agents can only reach their own request's read-only tools and can no longer reach the assistant's full tool catalog by changing the URL.

Small fixes

• The @agent suggestion panel in the new request dialog now opens at the cursor instead of the bottom of the box, and the list scrolls along when you navigate with arrow keys. The same scroll fix applies in Conversations.
• Clicking an agent's name on a Work Item now opens the Agents page with that agent selected.
• Plan links to failed or cancelled runs are no longer offered, since they would block the new item forever.