Skip to content

Changelog

What changed for someone who uses Loopcaptain. This file holds the first version; every later version has its notes on the releases page, written from the commits.

0.1.0 (2026-11-15)

The first public version.

Security

  1. Start passes a started loop only PATH, HOME, LANG, TERM and LOOPCAPTAIN_INBOX_KEY, not the board's whole environment, and asks for a tick the first time a given command of a given repository is run.
  2. The board signs every inbox line (sig, an optional field of protocol 1) with a key outside every repository; the helper and the MCP tool inbox_read say whether an entry is signed, so a loop can refuse a forged stop or reply. Each line names its loop, and the board's own status ignores a line whose signature fails.
  3. The file viewer refuses credential files (.netrc, *.pem, id_rsa* and more) and shows content only for known text files.
  4. The board follows no symbolic link in a loop's folder, and the MCP tools write only under their working folder or LOOPCAPTAIN_MCP_ROOTS.

Changed

  1. The demo is one example loop that builds a dashboard on loop engineering: it asks once with buttons, converges over four rounds, and honours only a signed stop. loopcaptain demo writes it from an installed package; its history is the loop itself run with past times.
  2. loopcaptain validate also checks that artifact paths exist, that time does not run backwards and that a round has one outcome; it warns of rounds written after the fact and of loops that never read their inbox. Problems and warnings are printed apart; --strict fails on warnings too.
  3. loopcaptain doctor <repository> <loop-id> sends a running loop a signed stop and passes once the loop took it and ended: the one way to know a loop listens.
  4. loopcaptain watch <folder> and unwatch <folder> edit the board's config, so the board needs no checkout.

Added

  1. A documentation site (Zensical): get started, connect or build a loop, every agent, the board, the protocol, a reference written from the code, security, architecture. make docs builds it and joins make check; llms.txt points agents at the standard.
  2. The README opens with two prompts to copy into any coding agent (build a loop, watch the loops on this machine), kept once in standard/prompts/; AGENTS.md section A is the runbook the second one points at, and standard/AGENTS.md a guide to building a loop.
  3. loopcaptain mcp-config <agent> and docs/agents.md: how each coding agent adds the MCP server and runs a loop unattended.
  4. Protocol 1 grows, every field optional: at for the real time of an event, a round's whole cost, cache tokens and turns, metrics, attempt, error (needs_person waits for you), session, outcome (kept or discarded), next_at, question id and options, workspace and state in loop.json, and the events question_closed, loop_stopped and heartbeat.
  5. A loop that said when it is back shows as sleeping, not stalled.
  6. The board flags "no progress?" before the loop does: the same material findings open two rounds running, a measure flat for three rounds, open findings that do not fall.
  7. The overview groups loops by repository, with a "Waiting" tab for questions; worktrees of one repository show once, from the one written last.
  8. The loops protocol, version 1: a loop writes loop.json and appends to events.jsonl in .loops/<loop-id>/; the board appends to inbox.jsonl. Eight events: round_started, step_started, step_finished, finding_closed, round_finished, question, check, inbox_read.
  9. standard/: the text, a helper of 50 lines to paste, three loop.json templates, the JSON Schema, and an AGENTS.md for the agent that connects a loop.
  10. The board: all loops with what waits for you on top; one loop with its brief, the graph of a round, every round against every step, a round in detail, the items it evaluates; a page to copy the standard from.
  11. Steering from the board: answer a question, send a note, stop after the current round, start the command a loop declares.
  12. The loopcaptain command: serve, validate, emit, schema, spec, mcp.
  13. An MCP server with the protocol as five tools (protocol, loop_register, emit, inbox_read, loop_status), for loops an agent session drives.
  14. make demo: three invented loops, one of them startable, that cost nothing.
  15. The overview draws each loop's latest twelve rounds beside its round count: a square per round, darker for more findings left open, a green dot for none, a ring while it runs. A loop that converges is seen without opening it.
  16. The documentation has an Installation page (one command from the repository, with uv, pipx or pip), a five-minute Get started with the example loop, and a Setup page: what the board watches, how to check a loop, and what to look at when a loop does not show.
  17. The overview shows where to look first: the tiles for what runs and for what stalled or failed take colour when their count is not zero, every row carries a stripe in its status colour, and a row with a question or a failure is tinted.
  18. The README and the documentation show the board with an invented set of eight loops in every status: the front page large, four more views in a grid. make shots retakes them.