Skip to content

v2.3.0

Latest

Choose a tag to compare

@henomis henomis released this 28 Oct 13:58
3cd32db

Building blocks for long-running coding and work agents: session compaction, a sandboxed shell, rule-based permissions, and sturdier file, MCP and memory tools.

Breaking:

  • tool/bash: a background command no longer stops at the foreground timeout. It runs until it exits, kill_shell or Close, unless WithBackgroundTimeout is set.
  • tool/file: the read tool returns the cat -n text as a plain string, not {"content": ...}. ReadOutput is deprecated.
  • vectorstore/qdrant: writes wait until applied. WithWait(false) restores the old behaviour.

Deprecated: WithSummarization in memory/simple, jsonfile, psql and nats. Use agent.SetCompaction instead. bash.MaxOutputChars is an alias of bash.DefaultMaxOutputChars.

Highlights:

  • Compaction: agent.SetCompaction with agent.NewSummaryCompactor summarizes older history when the prompt passes a token threshold, or after a context-length error. Agent.Compact compacts on demand. Snapshots are stored by every memory backend, so the next run starts from the summary. Compaction events are in RunStream, and the summary model's usage counts in Result.Summary.
  • New tools: tool/permission (allow, ask and deny rules for tool calls, with bash command lines parsed by mvdan.cc/sh), tool/web (web_fetch with an SSRF guard) and tool/todo (a task list for long runs).
  • Bash: a bubblewrap sandbox runner, a pluggable Runner, environment options, opt-in persistent shell state, streamed output, and background shells with their own lifetime and output cap.
  • File: multi_edit, a read-before-write guard, read output capped and paged, WithReadOnlyRoots, glob and grep that fit real repositories.
  • Agent: caps on parallel tool calls with exclusive tools run alone, recovery from tool panics, the session saved after each iteration, tool calls cut at the output limit not run, SetLLM.
  • LLM: context-length and retryable provider errors are classified, llm.WithMaxOutputTokens, llm.EstimateTokens, and streaming kept through every middleware. Retry now retries only errors a retry can fix.
  • Providers: Anthropic caches the whole conversation and has a higher default max_tokens. OpenAI keeps reasoning_content and sends tool-result images correctly.
  • MCP: tool errors are reported, images are passed through, tool names are made valid for providers, and WithMaxOutputChars caps results.

What's Changed

Full Changelog: v2.2.0...v2.3.0