Skip to main content
Parallel branches now share one checkout, and join_policy was removed. Remove join_policy from parallel nodes; fan-out now waits for every branch. Keep file-writing branches read-only or assign disjoint paths, and update fan-in logic to consume parallel.results instead of a selected branch workspace or parallel_results.json.
Sandbox commands no longer load login-shell profiles. Environment set in /etc/profile.d/*.sh, ~/.bash_profile, or nvm/rbenv/sdkman initializers is gone, so commands that relied on it now fail mid-run with command-not-found. Move those exports into your Dockerfile’s ENV (or your Daytona snapshot image).

​
Sandbox commands are Bash everywhere

Every sandbox now evaluates command strings as non-login Bash — bash -c, with no sh fallback and no ambient provider shell. Previously the three backends disagreed: Daytona ran commands through sh, and Docker’s streaming, stdio, and setup paths used a login shell. Bash-only syntax such as [[ ... ]] and arrays now behaves identically across local, Docker, and Daytona, through both buffered and streaming execution. The shell tool’s name, argument names and types, and required fields are unchanged; its descriptions now make explicit that command is Bash source. Fabro still adds no shell options of its own — there is no implicit errexit or pipefail, so false | true still succeeds. A workflow that wants different semantics writes them into the command (sh -c ..., a #!/bin/sh shebang, an explicit set -o pipefail). Both fresh initialization and resume now verify Bash before reporting the sandbox usable, so a missing or non-Bash interpreter fails at the lifecycle boundary with provider-specific remediation instead of on the first command. Fabro also clears BASH_ENV at each provider boundary so ambient worker or image configuration cannot execute a startup file before sandbox commands. Docker and Daytona require /bin/bash; local sandboxes resolve bash through the worker’s PATH, which keeps NixOS working. Prepare-step script values and sandbox MCP script values run under the same contract. Hooks, host-side stdio MCP scripts, and interactive terminals are separate executors and keep their existing behavior.

​
Shared-checkout parallel execution

Parallel branches now run concurrently in the run’s existing sandbox and working directory instead of creating branch-specific Git worktrees. Each branch still receives isolated context updates, returned in outgoing-edge order under parallel.results, while a prompted fan-in can synthesize the complete typed result array. Promptless fan-in nodes act as barriers and never select, restore, or merge workspace state.

​
Fireworks AI

Fireworks AI is now an opt-in built-in provider with a curated serverless catalog for Kimi, DeepSeek, GLM, MiniMax, Qwen, and GPT-OSS models. The integration uses the existing OpenAI-compatible path, supports cached-input accounting, and works with the normal provider login, secret storage, diagnostics, and model-testing flows.
settings.toml

​
Model-native Kimi agents and agent chat

Kimi K3 sessions now use a model-specific agent profile with Kimi Code tool names, descriptions, paging, and TODO behavior. The stage activity view also has a Chat tab that groups prompts, assistant messages, tool calls, and disclosed provider reasoning into a readable conversation.

​
More

  • GET /api/v1/runs/{id}/events now supports descending cursor pagination with order=desc and before_seq; fabro run events --tail uses it to fetch only the newest events
  • Completion usage now reports disjoint input, output, reasoning, cache-read, and cache-write token buckets
  • Model responses now advertise exact controls.reasoning_effort values, and completion requests—including structured completions—validate and forward reasoning effort with a 400 response for unsupported values
  • Event history scans now use seek-based pagination with bounded cursors, avoiding full-history work and incorrect results for oversized sequence values
  • Preflight now prefers providers that are actually ready while preserving explicit provider pins and useful diagnostics for unavailable offerings
  • OpenAI-compatible agent providers now receive compatible file-edit and tool schemas
  • Large values inside parallel branch results now stay available to fan-in prompts through normal artifact storage
  • Agent compaction now preserves history when a summary is empty or exceeds the model’s reasoning budget
  • Agent tool secrets now reach model-specific profiles, and OpenAI-compatible streaming usage is requested consistently
  • Resumed nodes now start a new execution such as work@2, preserving the earlier execution’s events, output, timing, billing, and terminal state
  • Added claude-opus-5 to the first-party Anthropic and optional OpenRouter model catalogs; the opus and claude-opus aliases now resolve to Opus 5
  • Added gpt-sol, gpt-terra, and gpt-luna aliases for GPT-5.6 offerings
  • Added portable glm, glm52, glm5.2, deepseek, and deepseek-flash aliases across direct and OpenRouter offerings