Evidence record / omnigent
A dated record of one change, kept so the writing that cites it can be checked. Compiled from the sources listed below by the research run, not written for reading. The judgment lives in the issues that cite it, below.
2026-08-17-omnigent-omnigent-s-max-cost-usd-is-still-a-downgrade-gate-not-a-hard-stop
Omnigent's max_cost_usd is still a downgrade gate, not a hard stop, unless you write expensive_models: [].
Nothing, and that is the finding. Across v0.8.0, v0.8.1, v0.8.2 and v0.9.0 the semantics of max_cost_usd are unchanged. _resolve_expensive_models at v0.9.0 lines 380-414 returns block_all_models=True when expensive_models is None or [] -- that path is a true hard stop, and it was established by #1631 on 2026-06-30, before this window. Supply a non-empty expensive_models list and you get the downgrade gate: over budget, tool calls are DENYed while the session sits on a listed model and ALLOWed again once it moves to a cheaper one. The cap is evaluated at the request phase as well as the tool-call phase, and it fails closed when the active model is undeterminable.
Channel: tagged-release. Ancestry: Read the policy source at https://raw.githubusercontent.com/omnigent-ai/omnigent/v0.9.0/omnigent/policies/builtins/cost.py and diffed it against the same file at v0.7.0: the only differences across the whole window are type-annotation and isinstance hardening from the pyrefly lint sweep (commit 1262652a039c, 2026-08-03, #3972). No semantic change. gh api "repos/omnigent-ai/omnigent/commits?path=omnigent/policies/builtins/cost.py&since=2026-08-01" returns exactly one commit, that lint sweep. The docstring at v0.9.0 lines 21-28 still reads: "once spend reaches this, the policy forces a model downgrade. Rather than stopping the session, it DENYs while the session is still on an expensive model ... the budget becomes a 'downgrade gate,' not a hard stop."
Operator consequence: Read your own config rather than the feature name. If you wrote cost_budget with an explicit expensive_models list expecting the number to stop the session, it does not -- it forces a model switch and the session keeps spending on the cheaper arm. Omitting expensive_models entirely, or setting it to [], is the hard stop. Nothing in this window changed that, so anyone who checked in July does not need to re-check.
Receipt
Finding metadata
Run: 2026-08-17-weekly-digest-2026-08-10_2026-08-17-frontier-v0
Finding ID: 2026-08-17-omnigent-omnigent-s-max-cost-usd-is-still-a-downgrade-gate-not-a-hard-stop
Source links
Primary links, including exact changelog lines when available.
Versioned source: run artifact