agentfuse-sdk v0.3.0

Two correctness fixes that keep the fuse honest where it was quietly failing.

Fixed

• Streamed responses now advance the cumulative ledger. A stream=True call returns a wrapper with no .usage until consumed, so post-call metering used to commit $0 and the cumulative USD/token ceilings could never trip on a streamed run — the dominant agent call mode. AgentFuse now meters the stream on exhaustion (real usage when the provider emits it, else the pre-call upper-bound estimate), so the ledger advances and the fuse still trips. Sync + async covered.

Changed

• Fuse(max_total_tokens=...) replaces the confusingly-named Fuse(max_tokens=...) for the cumulative whole-task token ceiling (max_tokens universally means a single call's completion length). max_tokens is kept as a deprecated alias for one release.

Full changelog: see CHANGELOG.md.

agentfuse-sdk v0.2.0

AgentFuse — per-task spend circuit-breaker for LLM agents

v0.1.1

agentfuse-sdk v0.1.1 — house-style docs + release artifacts

AgentFuse v0.1.0

First release: per-task spend circuit-breaker SDK for autonomous LLM agents.

• Pre-call budget gate: wraps litellm.completion/acompletion and halts a run before an over-budget call is ever sent (not a passive cost dashboard).
• Per-task ledger + cost meter (tiktoken + litellm model pricing); post-call metering via litellm success callback.
• Ergonomic API: Fuse(max_spend_usd=...) / task(...) context managers, @fuse decorator, install()/uninstall() monkeypatch.
• agentfuse CLI (status, demo) + offline runaway-agent demo.
• 30 tests passing; Apache-2.0.

See CHANGELOG.md for details.