v0.1.2

Added

• symbols — code navigation through a language server, or the text
• an approval reviewer — a model answers the prompts nobody can (M17)
• atelier acp — an Agent Client Protocol agent, for editors (M21)
• /loop — re-send a prompt at an interval while the session is idle
• advisory locks — instances say which paths they are working on
• an idle instance can be woken by its peers (opt-in)
• peer_ask — put a question to another instance's agent and get its answer
• atelier mcp — atelier's tools as an MCP server over stdio (M20)
• code mode — node scripts call the session's tools (M20)
• a GitHub Action — review pull requests, run prompts, answer @atelier
• worktree isolation for subagents — isolation: worktree (M18)
• background subagents — task with background: true (M18)
• model: in a command file's front matter, for that turn only
• MCP prompts as slash commands — /mcp::
• [instructions] files — instruction files named in the settings
• /init — have the agent write the repository's AGENTS.md
• atelier serve — a protocol to embed atelier, JSON-RPC over stdio (M21)
• project trust — atelier.toml runs and grants nothing until accepted (M17)
• sandboxed bash on nixvm, behind an opt-in sandbox feature
• /goal — keep working until a stated condition holds
• atelier doctor — a checklist of the setup
• atelier debug prompt — print the first request without sending it
• atelier review — a read-only review of a change
• a finished background job announces itself (M20)
• read-only tool calls of one model step run in parallel (M20)
• OAuth sign-in for remote MCP servers — /mcp login, /mcp logout (M20)
• deferred MCP tools — mcp_search and mcp_call past a spec budget (M20)
• connect configured MCP servers concurrently (M20)
• keep MCP tool annotations and show them in /tools (M20)
• pass MCP server instructions to the model as standing context (M20)
• MCP per-server tool filters, timeouts and enabled = false (M20)
• usage log, atelier stats, /usage
• named providers, a model per role, request variants
• permission modes — plan, ask-edits, auto-edit, yolo (M17)
• subagents — the task tool, agents as files, the step cap (M18)
• /context — where the context window goes
• compaction in stages — prune tool output, hand-off summary, /compact, retry on overflow
• save tool output past the budget to a file, return head and tail
• ask tool, a structured question answered on the input line
• hooks — user-configured commands on lifecycle events (M19)
• skills — SKILL.md directories, the skill tool, /skills (M19)
• custom slash commands from Markdown files (M19)
• permission rules — allow / ask / deny patterns per call (M17)
• web_search tool with searxng, brave, tavily and duckduckgo providers
• diagnostics for every language, reporting only what changed
• machine-usable --print (JSON output, stream input, step limit, schema)
• load all instruction files, with imports, a byte budget and nested files on demand
• durable memory — memory tool, ambient index, /memory (M15)

Fixed

• coordinator queues an answer under the lock its request ran under
• a subagent runs in its parent's current mode, and isn't compacted at its end
• never leave two user messages in a row in the conversation
• /trust revoke withdraws MCP instructions, prompts and deferral too
• gate a project's [sandbox] enabled behind project trust
• gate a project's [search] url — the user's search key was sent to it
• a command file's shell lines ask even while the sandbox is on
• permission rules and modes govern claude's built-in tools
• confine instruction imports to the repository; keep the memory index in every request

Other

• roadmap — M15–M22 status after the last merges
• a newer push cancels the older run on the same branch
• bring configuration, tools, hooks and protocol in line with the code
• fix two races in the peer-message and background-job tests
• roadmap — status of M15–M22
• keep the polled background job running until it is polled
• a cache-stable prompt — constant system message, context with the user turn
• roadmap — drop M16, sandbox on nixvm
• roadmap M15–M22, from a review against Claude Code, opencode and Codex