agent-team v0.5.0

The first release. agent-team is a configurable team of AI agents that plans, builds, reviews,
fixes, merges, releases and markets a software project through GitHub, with members on any
model and any machine. It grew out of the scripts that built todo-py v1.0.0 and v1.0.1.

The team

• Teams are described in .team.toml: any number of members and roles, files owned per member,
  per-member skills (Markdown files or Claude Code skill folders), and team memory: reviewers'
  LESSON: lines are kept and shown to every member in later work.
• Members can be on Claude (anthropic), Ollama (local or another machine), any OpenAI-compatible
  server, or another machine running agent-team serve (token-protected; nodes can be found on the
  local network with agent-team discover). Every kind can call tools, so any member can run the
  devops shift.
• Roles the workflows use: lead, builder, devops, marketing, fact-check, and optional QA (writes
  tests first), security (reviews every approved build) and docs (updates the docs).

Workflows

• discuss (a roadmap the members agree, reading the devops feedback digest), build (the lead
  plans, builders implement with search/replace edits, tests run, the lead reviews with tools to
  read files and run tests; --parallel splits a plan across builders), ship (one devops shift:
  triage, sign-offs, reviews, fixes, merges, releases), fix (apply a review to an open PR),
  setup, market and autopilot (work through a milestone a PR at a time).
• GitHub issues labelled team:build are the team's to-do list (agent-team issues sync creates
  them from the roadmap); autopilot builds them and its PRs close them.
• Scheduled shifts (agent-team schedule --every 6h, launchd or cron, never two at once) and
  background runs (agent-team jobs, the dashboard, the Claude panel).

Guardrails

• Merges need green CI, the lead's approval and every owner's sign-off on the exact head commit;
  paths in require_approval also need the owners' approval of that commit
  (agent-team approve <PR>); release = "ask" makes releases proposals the owners approve.
• Re-reviews block only on points still open or problems the new commit introduces.
• Every push is checked for secrets first; text from outside the team reaches the models marked
  as untrusted; members can post as their own GitHub accounts (github_token_env) so their
  reviews are real GitHub reviews.
• [sandbox] mode = "container" runs the code the team writes in a throwaway container with no
  network and none of your keys.
• Guardrail scenario tests run whole shifts against a stand-in GitHub.

Cost and models

• Every model call is recorded with its tokens and an estimated cost; spend limits per run, per
  day and per milestone stop a run before the next call; agent-team estimate predicts a run's
  cost from past runs; agent-team scoreboard shows approval rates, rounds and cost per member.
• Routing: a member hands its work to a stronger one after rejected rounds (escalate_to), and
  routine checks can use a cheaper model ([routing].light_model).
• Focused context for big projects: a ranked shortlist of relevant files and a map of the rest.
• agent-team eval runs fixed scenarios with your real models in a throwaway project.

Dashboard and Claude Code

• agent-team ui: Overview, Team (edit everything, add a machine, review the diff before saving),
  Runs (start and follow runs), Costs (with the scoreboard), Activity (live charts), Settings (easy
  switches and team templates) and Projects (every project on this Mac).
• A Claude Code plugin: /team-init, /team-doctor, /team-plan, /team-build, /team-fix,
  /team-ship, /team-autopilot, /team-issues, /team-schedule, /team-market, /team-ui,
  /team-run, and a side panel (/team-panel) with the team, health, runs, spend and approvals.
• [notify] posts to Slack or any webhook when PRs open or merge, approvals or releases wait,
  autopilot stops, a spend limit is hit or a run fails.

Setup

• agent-team ready lists what a machine and project still need, each with its fix; init detects
  Python, Node (Jest, Vitest, Mocha, node --test), Go, Rust and .NET projects, and Ollama on
  another machine; agent-team keys set keeps API keys where every command finds them.
• One-command installer for macOS and Linux (scripts/install.sh), and this release on PyPI
  (pipx install agent-team).