Agent harness for OpenAI/Anthropic tool-calling loops: model providers, sessions, tools, permissions, hooks, MCP, subagents, compaction, and resilience.
Headline additions in this tag
Effort.nonemaps to OpenAI'sreasoning_effort: "none"— required for function-tool calls with gpt-5.6 on Chat Completions (any other effort is rejected; an omitted effort defaults tomediumand also fails).OpenAIModelProvider.extraHeaders— custom request headers for gateway integrations (e.g. Cloudflare AI Gateway'scf-aig-byok-alias).
This is the first public tag, so it also includes the full release-hardening body of work: provider retries with backoff, per-turn session checkpoints, concurrent same-turn tool execution, fail-closed permission modes (breaking), AgentSession.query/MCP renames (breaking), security hardening, and more — see CHANGELOG.md.
Validation: full suite of 1036 tests, 0 failures on macOS (Swift 6).