v1.4.0

One shared limit

Two agents should not both grab the last call.

AgentGuard 1.4.0 reserves capacity before supported API calls. Workers using the same local store and key share the limit.

• Reserve before send for sync OpenAI Chat Completions and OpenAI/Anthropic streams.
• Record final usage once. Keep capacity on hold when a stream drops or usage is unknown.
• Try the offline demo without an API key or account.

Install: python -m pip install agentguard47==1.4.0

Run: agentguard demo

Requires a shared StateStore. Token and cost reservations also need max_tokens. Cost is an estimate, not a cap on your provider invoice. Async non-stream and Anthropic non-stream calls still check recorded usage.

Supported paths and limits

---

What's Changed

Other changes

• Email AgentGuard subscribers after each stable release by @bmdhodl in #726
• Record AgentGuard 1.3.2 public install and publication proof by @bmdhodl in #727
• Document tested enforcement bounds and repair overbroad claims (AG-01) by @bmdhodl in #758
• Establish activation evidence and a voluntary demo feedback path (AG-02) by @bmdhodl in #759
• Design the local reservation and reconciliation contract (AG-03) by @bmdhodl in #761
• Reserve one OpenAI dispatch on a shared local store (AG-04) by @bmdhodl in #763
• Reserve store-backed streams before send (AG-05) by @bmdhodl in #765
• Prepare the 1.4.0 release candidate by @bmdhodl in #766

Full Changelog: v1.3.2...v1.4.0