v0.12.308

maxCostPerRun: 0 now means nothing gets paid for

Behaviour change: a cap of 0 used to switch the cap off, while startup logged
Cost cap: $0.00 per session. It now refuses every paid call. If you set 0 to
mean "no limit", remove the key.

The refusal covers every route on both rails: chat, /v1/messages, image, video,
music and speech generation, the partner and data APIs, and the /imagegen and
/img2img chat commands. These answer 429 cost_cap_exceeded, with the
X-ClawRouter-Cost-Cap-Exceeded: 1 header.

Free models still answer, but routed profiles need care: pin a free model. Strict
mode refuses auto and premium, and graceful mode refuses them on tool calls.
Graceful mode no longer adds a budget warning to replies at a cap of 0; the
X-ClawRouter-Budget-Mode: downgraded header still reports the downgrade.

On the API-key rail nothing shows what a call costs, so a cap of 0 lets through
only free models on the chat routes and reads that can't bill (catalogs, health,
usage, balance). A free partner call is refused there too.

A positive cap still covers /v1/chat/completions only.

Thanks @0xcfloki, who found the bug and fixed the chat path (#553).

Cost caps are harder to slip past

These apply at any cap:

• A chat request with no x-session-id and no user message used to skip the cap.
  It now counts against a shared budget.
• In graceful mode, a model with no local price is refused once the budget has run
  out. It used to be paid for. This also applies in strict mode on the API-key rail.
• startProxy ignores a cap that isn't a finite number, and the plugin config does
  too. A NaN cap could make every paid request fail and an identical retry hang.
• A routed request whose free primary survives the budget filter no longer gets a
  false "raise your cap" notice.

Free tier: Bolt and Super 120B in, Llama out

• Added: free/nemotron-3.5-bolt (about a second to answer, reads images) and
  free/nemotron-3-super-120b (reasoning). Try /model bolt or /model super-120b.
• Free vision works again, on one model. Bolt got 8 of 8 right on 64x64
  solid-colour images across Base and Solana, and is the only free model flagged
  for vision. /model vision-free now pins it.
• Retired: free/llama-3.2-11b-vision, which NVIDIA deprovisioned on
  November 2. Pins still route and are answered by a live free model.
  llama-free now points at free/muse-glimmer-30b, and nemotron-super at
  free/nemotron-3-super-120b.

Video: Sora 2 removed

azure/sora-2 is gone from local pricing and the README: its provider shut it down
on November 2, and the gateway refuses new requests with no substitute. The README
video table now lists all seven live models.