v0.1.9

What's new in 0.1.9

The first twenty-four items of the improvements report.

Your subscription's allowance

• A meter for the weekly allowance. Anthropic meters non-interactive Claude Code use — the way Concentus runs the CLI — in its own allowance. Set your plan's figure under Settings → Usage and the Usage page shows how far this machine's runs are from it; a run that starts past 80% says so in its first line.
• A fallback when it is spent. On the Coordinator: run on the API key, or run every agent on a local model. Applied at start when the meter says spent, and mid-run when the CLI refuses for the allowance — the run ends as a failure and continues as a new run on the fallback, which says where it came from.
• A hard ceiling. The monthly budget now stops a billed run that reaches it mid-run, not only the next one.

Fan-out

• Items that wait for others. A plan item may list dependsOn; it starts when those items finished and gets their reports in its prompt. Loops and unknown ids are refused; ordered items may share a file.
• Repositories inside the workers. A repository wired to a worker is cloned into that worker's own workspace; its changes reach the merge step as patches, and the merge — the one process with a shell — applies, verifies, commits and opens the pull request.
• ask_worker. A worker can put a question to a named sibling; it is answered from that sibling's workspace and report by a short read-only process, without interrupting it. Five per worker.
• Retries per block, and a block's on error / on rejected fires the moment the block settles instead of after the merge.

Runs

• Resume. Start again keeping every worker that passed; only the rest runs. Runs the backend lost in flight say so, with Resume as the way back.
• Telegram approvals and questions: a bot set up once under Settings → Approvals posts every request with ✅ / ❌ buttons and every question as a message you reply to. The app polls the bot, so the answer comes back with no public URL.
• Changes tab. The diffs of what the agents did to the repositories, per block and per run, persisted with the run.
• Evaluations. Datasets per flow — cases with contains, regex, exact or an LLM judge — and results per flow version.

Blocks and triggers

• Send mail node: wire it to a block's output and the run mails what that output carries — the error and its log, the verification report, the final answer.
• Folder watch trigger, and provider presets (GitHub, GitLab, Linear) for the webhook form.
• Publish a flow as an endpoint: POST /api/public/flows/{id}/run with a bearer token, a synchronous answer or 202 + polling, and a minimal chat page.
• Notes and groups on the canvas, "Go to block" in Ctrl+K, a minimap.
• Library agents as live references: link a block to a library agent, versioned; the doctor warns when the library moved on.

Knowledge

• OCR for scanned PDFs and images through the Tesseract program when it is installed (nothing bundled; the message says how to get it).
• The retrieval budget is in tokens (estimated); characters stay as the fallback.
• pgvector does ship in the desktop app — the message that said otherwise is gone, and tool search ranks with the built-in embedder.

Managed sessions

• A remote MCP server's credential reaches the managed session as authorization_token — declared, not yet verified against a live session; the run log says per server when a credential left the machine.

Not exercised against the real world in this release

SMTP send, the Telegram bot, Tesseract, the managed-session MCP token and the LLM judge were tested through fakes only. If one of them misbehaves for you, the run log says so — open an issue with that line.

Full Changelog: v0.1.8...v0.1.9