v0.11.0

Hydaelyn v0.11.0

Status: Released

Released on 2026-07-18.

Context-aware agent compaction

Agent loops can now manage per-request context capacity independently from
cumulative run spend:

• agent.LoopPolicy.ContextTokenTarget declares the usable history allowance
  for one model request;
• agent.TargetContextManager adds token-targeted CompactTo preparation;
• targeted context preparation runs before every model request, including the
  first request and requests following tool results;
• complete tool turns and framework-owned skill context are validated before a
  compacted request is sent;
• existing ContextManager implementations and cumulative
  api.TaskBudget.MaxTokens behavior remain compatible.

Applications should derive ContextTokenTarget from the selected model's
context window after reserving capacity for output, reasoning, tool schemas, and
provider framing.

OpenAI Responses API

The OpenAI provider now supports the Responses API required by Codex models.
Set openai.Config.WireAPI to openai.WireResponses to use /responses;
Chat Completions remains the default. Responses streaming preserves encrypted
reasoning/provider state needed by subsequent turns and conforms to the shared
provider event contract.

Upgrade

go get github.com/Viking602/go-hydaelyn@v0.11.0
go install github.com/Viking602/go-hydaelyn/cmd/hydaelyn@v0.11.0

What's Changed

• feat(openai): support Responses API for Codex by @Viking602 in #38
• fix(agent): decouple context compaction from run budget by @Viking602 in #39

Full Changelog: v0.10.1...v0.11.0