Skip to content

Releases

Turnloom 0.7.1

Choose a tag to compare

@github-actions github-actions released this 19 Nov 06:37

A release shaped by running Turnloom beside seven other agents on Terminal-Bench tasks, and by writing the reference documentation. Both turned up things to fix.

Security

  • A policy rule that asks is always put to you. With a provio.yaml in the workspace, a rule with the verdict ask was answered without a prompt when --auto-edit, an --allow pattern or an earlier "always" already covered the call. In the terminal, Studio and editors the question is now asked for that call, names the rule, and offers no "always". Under turnloom serve nobody is present to answer, so the call is refused, as it already was under --yes.

Changed

  • A file tool asked for a path outside the workspace now says that the file tools reach only the project root and, in a session that has a shell, that the shell is the tool for other paths. Before, the bare refusal left the model retrying or guessing.
  • turnloom-agent tells the model that the session is unattended: nobody will act on advice in its reply, so it has to run the commands and start the services itself. Before, some tasks ended with instructions for a person who was not there.
  • The shell tool's description says that a long-running process belongs in the background.
  • --budget without known prices says how to supply them: pass --input-rate and --output-rate, or run turnloom connect again.
  • turnloom-agent no longer insists on an API key, so it works with a local endpoint that needs none.
  • The GitHub Action installs the guard extra, so a repository with a provio.yaml runs under its policy instead of stopping with an error.
  • Help text for the flags that had none: --workspace, --reply, --deadline, --effort, --no-fold, --skills-dir, --tier, --search and the proxy's --port.

Fixed

  • An endpoint that refuses max_tokens and asks for max_completion_tokens, as OpenAI models do behind a gateway or proxy, is remembered for the rest of the process. Before, every model call was sent twice.
  • turnloom personas says "1 persona" when there is one.

Documentation

  • New pages: quick start, command reference, configuration, recipes and FAQ, with an index.
  • The README is rewritten around a first run, with copy-and-paste examples, and its links now work on PyPI.
  • Benchmarks: the method, scripts and results of the comparison with other agents.
  • Corrections where a page disagreed with the code: the order of the HARNESS_* and HLB_* variables, what an MCP server inherits from your environment, where the unwind key is accepted, and examples that asked a read-only session to review a Git diff it cannot see.

Project

  • A roadmap, a code of conduct, a contributor guide that walks through a first pull request, and starter issues.

Turnloom 0.7.0

Choose a tag to compare

@github-actions github-actions released this 18 Nov 19:24

The first public release. The agent gains real file tools, streaming, sessions that survive the process, and reversible edits. Studio is redesigned and rebuilt on those. Turnloom can now be driven from editors and from other agents.

Added

  • File tools in every session. list_files, read_file, search_files, edit_file and write_file are confined to the workspace. Edits show a diff for approval, keep line endings, and are written atomically. Before this release the terminal agent could only use the shell.
  • Undo for edits. turnloom undo, /undo in chat and Studio restore the files a turn edited. A file changed since is kept and reported.
  • Streaming. Replies appear as they are written in the terminal, in Studio and in editors, with tool activity shown while it runs.
  • Sessions. turnloom sessions lists them, --resume continues one with its tool results intact, and Studio can start, list and resume sessions.
  • Approval modes. --auto-edit, --read-only, --allow PATTERN, and an "always" answer at the prompt. Studio has the same choices.
  • Editors. turnloom acp serves the Agent Client Protocol, for Zed, JetBrains IDEs and other ACP clients.
  • Project instructions and skills. AGENTS.md (or CLAUDE.md) is read from the workspace root. Skills under .agents/skills and .claude/skills are offered to the model.
  • MCP over HTTP. Server entries may be a streamable HTTP url with a bearer_env, and process entries may set env.
  • Turnloom as an MCP server for other agents. turnloom serve adds transcript and read_result, and with --allow-run a run_task tool.
  • GitHub Action and container image definition for unattended runs.
  • turnloom studio, turnloom doctor, --max-steps, --events, and chat commands /diff, /undo, /new, /help.
  • Studio can create, rename and delete files, shows what a turn edited with a review and a revert, and has light and dark themes.
  • TURNLOOM_ENDPOINT, TURNLOOM_MODEL and TURNLOOM_API_KEY.

Changed

  • Commands and edits are declined unless something approves them. A caller that sets no approval callback must opt in with auto_approve. Previously a missing callback meant "approve everything".
  • The policy gate covers file edits as fs.write, alongside shell commands, reads and MCP calls. Searches return only lines from files the policy allows, delegated tasks through turnloom serve are gated, and a rule that asks is put to the operator once, with an edit's diff in view. Studio refuses to run ungated when a policy file exists and Provio is missing.
  • Programs are never taken from the project folder. ripgrep, Git, Provio and taskkill are resolved from PATH and the interpreter's scripts folder only. On Windows a bare program name used to resolve in the current directory first.
  • Directory junctions leading out of the workspace are no longer listed or searched. Read-only sessions no longer offer MCP tools.
  • Studio remembers an enabled MCP server together with its definition and switches it off when the definition changes. Choosing "always" on an edit lasts for the session and is not saved.
  • Context overflow archives instead of deleting. Exchanges removed to fit the window stay recoverable through recall, and an oversized newest result is shortened in place rather than failing the turn.
  • Tool results the model has not read are never folded. A batch of parallel calls used to lose all but its last three results before the model saw them.
  • Full command output is archived when the prompt only receives its head and tail.
  • A timeout or a stop ends a command's whole process tree. Stopping a turn interrupts a running command and an in-flight model reply.
  • Chat continues after a failed turn instead of exiting, and Ctrl+C ends the turn rather than the program.
  • Retries honor Retry-After, cover dropped connections, and retry a timed-out request once instead of four times.
  • A truncated reply is returned as partial. An empty reply is retried once. Limits explain why they stopped the turn.
  • Three identical tool calls in a row end tool use everywhere, not only in Studio. Repeating a check between edits is no longer treated as a loop.
  • turnloom team runs without LangGraph installed, and planning and review stages are read-only by construction. Studio's team stages gain a read-only switch.
  • Session traces, the usage ledger and the undo database get a .gitignore, so they stay out of the user's repository. .turnloom/mcp.json and .turnloom/personas/ remain visible to Git so a team can share them.
  • Studio settings moved out of the workspace into the operator's configuration folder. A .turnloom/studio.json shipped in a repository could previously turn on auto-approve or enable MCP servers; that file is no longer read.
  • Studio no longer shows sample data. It keeps the conversation when a turn fails or the persona changes, waits for approvals without a timer, stops by itself when left unused in a browser, and opens without putting its token on a browser's command line. Closing Studio stops a running turn.
  • turnloom-studio is a console command, so its help and errors are visible. turnloom-studio-gui is the windowless variant. --no-open and --port support remote shells.
  • The proxy refuses requests that carry a browser Origin or a foreign Host.
  • Command output is decoded as UTF-8 first. Redirected terminal output is UTF-8 on Windows.
  • The evaluation interval resamples attempts as well as tasks, checks end their process tree at the timeout, and the example manifest is a task that runs.
  • The desktop extra no longer installs PyInstaller; build-app does. studio and full are composed from the smaller extras.

Removed

  • turnloom demo, which folded a hard-coded string. turnloom doctor is the offline check.
  • Benchmark-specific rules from the default system prompt. turnloom-agent and the evaluation runner add them to their own sessions.
  • Thirteen provider presets that need a vendor-specific client or a token exchange.
  • The Cline UI components and the Tailwind-dependent styling in Studio.

Upgrading from 0.6

  • Scripts that used turnloom run without a terminal must pass --read-only, --auto-edit or --yes.
  • HARNESS_ENDPOINT, HARNESS_MODEL and HARNESS_API_KEY still work. Prefer the TURNLOOM_ names.
  • Code that builds a Config directly: approve now receives (kind, detail), and unattended callers set auto_approve=True. workspace_tool is replaced by the built-in file tools.
  • A PyInstaller build installs .[build-app] instead of .[studio,full,desktop].
  • Studio settings saved by 0.6 in .turnloom/studio.json are not carried over. Set them again once; the old file can be deleted.
  • MCP clients of turnloom serve: the tools are now sessions, receipt, transcript, read_result, archive_output, recall and optionally run_task.