A run's context per agent and run limit come from the profile in use, not the run's #313

Closed
opened 2026-10-29 16:38:50 -04:00 by cmoriarty · 1 comment
Owner

What happens
Since #305 a run keeps the backend it was created with, and its server is set up from that profile. A few settings are still read from the profile in use while the run goes: the context per agent (which sizes sessions and picks the apply mode), the run limit, mediation_max, and the engine probe, which checks only the backend in use. On 2026-10-29, with Free in use, the Breakout project (run 146, on vLLM) queued behind a Free run under Free's run limit of 1 (main's is 3), and a step it starts then is sized by the flat 150,000 tokens per agent, not main's 400,000.

Fix
Read these from the run's recorded profile. Count the run limit per backend, so runs on different backends don't wait for each other (with #306). Probe each backend a going run uses.

Done when
A run started on one profile while another is in use is admitted, sized and probed by its own profile, in a test and on production.

**What happens** Since #305 a run keeps the backend it was created with, and its server is set up from that profile. A few settings are still read from the profile in use while the run goes: the context per agent (which sizes sessions and picks the apply mode), the run limit, `mediation_max`, and the engine probe, which checks only the backend in use. On 2026-10-29, with Free in use, the Breakout project (run 146, on vLLM) queued behind a Free run under Free's run limit of 1 (main's is 3), and a step it starts then is sized by the flat 150,000 tokens per agent, not main's 400,000. **Fix** Read these from the run's recorded profile. Count the run limit per backend, so runs on different backends don't wait for each other (with #306). Probe each backend a going run uses. **Done when** A run started on one profile while another is in use is admitted, sized and probed by its own profile, in a test and on production.
Author
Owner

Confirmed on production (1fc0f22, PR #319); closing.

  • Item runs record the project's profile. Runs 152 and 156 of the Breakout project on breakout-planned-3 recorded main.
  • A running vLLM run is unaffected by the profile in use. With Free in use (the profile in use then read 150,000 tokens per agent and a run limit of 2), run 156's openspec.propose kept running on vLLM, holding its slot, with no hold.
  • The run limit counts per backend. The project and two of the seeded-fault runs went under main's limit of 3, and the third fault run queued on main alone.
  • Main's own settings are untouched: 400,000 tokens per agent and a run limit of 3 after switching back.
**Confirmed on production (1fc0f22, PR #319); closing.** - **Item runs record the project's profile.** Runs 152 and 156 of the Breakout project on `breakout-planned-3` recorded `main`. - **A running vLLM run is unaffected by the profile in use.** With Free in use (the profile in use then read 150,000 tokens per agent and a run limit of 2), run 156's `openspec.propose` kept running on vLLM, holding its slot, with no hold. - **The run limit counts per backend.** The project and two of the seeded-fault runs went under main's limit of 3, and the third fault run queued on main alone. - **Main's own settings are untouched:** 400,000 tokens per agent and a run limit of 3 after switching back.
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid#313
No description provided.