agent-team 1.3.1

• Works on Windows. The scripts print seat marks in UTF-8 whatever the console's code page, read
  git and their own output as UTF-8, print project paths with forward slashes, and hash rulebooks the
  same way on every system, so PYTHONUTF8 is no longer needed.
• The README and docs/facts.json (install_notes) say that the one-line install needs a recent
  Claude Code, and what Windows needs for python3.

agent-team 1.3.0

• Dynamic seats. team-setup scans the repo (scan.py: signals with the evidence that raised
  each, and a size class: tiny, small, medium, large) and offers three teams, lean, recommended and
  full, every seat with its reason (plan.py). The owner picks one in setup's single question and
  may edit it.
• Merged seats: eight groups one session can hold while the work is small (Growth, Product &
  Design, Engineering & Platform, Quality, Money, Legal & Compliance, Data & ML, Customer), composed
  from their members' rulebooks with every rule kept and recomposed on sync. A check seat joins only
  a check merge (Quality), and no merge proposed after setup holds the CEO or the Bar Raiser.
• New seats: roles no catalogue seat covers (hardware and firmware, smart contracts, desktop,
  browser extension) and one frontend seat per UI app are derived from a base rulebook and written
  for their role per references/seat-contract.md before they work (seatgen.py).
• Eight new catalogue seats, 29 in all: Mobile engineer, ML engineer, Data engineer, Security,
  Docs writer, Community & DevRel, Game designer, Localization.
• team-reevaluate, a new skill (reevaluate.py): measures each seat's commits, worker runs,
  inbox and blocked lines, re-scans the repo against the setup baseline, and proposes adding,
  retiring, merging, unmerging or splitting seats with the evidence for each; a retired seat is not
  proposed again without a new signal. Nothing changes without the owner's one answer.
• scaffold.py applies a plan (--plan), records the owner's words as a decision (--record),
  retires seats with their briefs kept (--retire), and records each rulebook's origin in
  docs/team/seat-skills.lock. gen-team-agents.py --check fails while any seated rulebook
  still carries AUTHORING NEEDED.
• Tests: python3 -m unittest discover -s tests runs the scan, planner, scaffold and
  re-evaluation end to end on invented fixture projects.
• Docs: how the team is sized and re-evaluated, the catalogue with merged seats and new roles, the
  walkthrough's three options and its month-later re-evaluation, and when the owner is asked.

Fixed in review, before release

• The lock records where each rulebook came from and when. Every entry in
  docs/team/seat-skills.lock carries its origin (catalogue, merge:…, derive:…, or project
  for a rulebook the project already had) and the date its seat was seated. A project's own rulebook
  is never synced, overwritten or deleted. Old bare-hash entries still read, as catalogue copies.
• Every team change is one decision, in the owner's words. Setup records the team as DEC-2;
  --record works with --plan, --add, --retire and --seats, and --question names what was
  asked. The row says what changed in seat names (seated, retired, merged, split) and what was
  declined. A run that changes nothing records nothing.
• Declines are remembered. scaffold.py --only KEY,... (or --only none) applies the proposals
  the owner approved; every other one goes to docs/team/declined.json and is not proposed again
  unless a repo signal it rests on is new since the baseline.
• A role is never held twice, and the CEO and the Bar Raiser always are. Adding a role a merged
  seat holds is refused unless the same change splits that seat; a new seat's mark must be free; a
  retire that would leave the CEO or the Bar Raiser held by no seat is refused, in plain words.
• No worker before a new seat's rulebook is written. The generator writes none for a rulebook
  still marked AUTHORING NEEDED and removes an old one; a derived rulebook now points at its own
  brief, inbox, run logs and worker, not its base's.
• Retiring archives, never deletes, a rulebook that is derived, the project's own, untracked or
  edited, to docs/team/retired/<skill>-<date>. Re-adding a seat removes its "Retired" line, and a
  derived seat comes back from its archived copy.
• Re-evaluation measures against what was chosen. It compares with the option picked at setup;
  holds back seats seated inside the window; never merges the CEO or the Bar Raiser, nor Engineering
  & Platform past 2,000 source files; lets each merge claim its members; keeps new roles, per-app
  frontend seats and the coordination seats while they are called for; treats a missing baseline as
  nothing new. Each proposal has a key and a title in seat names, and the plan carries the project's
  name.
• The planner's rules are shared and stated once. The Chief of Staff joins at five other seats
  (seven in lean), the Scope Trimmer at ten; split seats never reuse a catalogue mark, and past four
  UI apps the Frontend engineer holds the rest; plan.py --markdown prints the owner-ready question.
• The scan reads what it should. The team's own files never move a signal; Python dependencies
  come from dependency tables only; a mobile app needs Android or iOS evidence; a monorepo needs
  manifests in two folders or a workspace config.
• Sync walks the roster, never touches a derived or project-own rulebook, and says what it
  compared. Line endings of the team's files are kept as found.
• check.py checks more: every rulebook is 130 to 200 lines, its description at most 700
  characters, every section-2 shift carries its source, marks match the catalogue for short names
  too (🌱 Community), and the generated docs are current.
• Docs that cannot drift. docs/seats.md is generated by tools/gen_seats_md.py (29 rows, the
  merged seats, the new roles, and starter shapes measured on the test projects);
  tools/build_tally_example.py rebuilds the example's generated files, now numbered as the
  walkthrough (the team DEC-2, recurring invoices DEC-3, the switch DEC-4) and labelled as a team of
  eight the owner named; tools/measure_tally.py --check re-measures the walkthrough's figures.
• Plainer for the owner. Setup asks whether finished work is uploaded on its own or the owner is
  asked each time, instead of "who pushes", and ends by naming whom to open now (the CEO and the
  Chief of Staff). The Game design seat is renamed Game designer. install.sh links both skills.
• A change the owner names is made, not re-asked. "Add Sales", "retire Design": team-reevaluate
  measures, applies with the owner's words as the record, and asks only when the measurement argues
  against the request. A declined merge keeps its members, so "change nothing" is not asked again in
  another shape.
• Tests and CI: every test runs, none marked as expected to fail; CI runs them on Python 3.9 and 3.12,
  walks the skills' commands in order on a fresh project, and re-measures the walkthrough's figures.
• Second review pass. Seats ask the owner before each push unless the owner chose otherwise;
  the tracker and the owner-questions file are written for every team (the CEO keeps them when no
  Chief of Staff is seated); a merged seat's rulebook names only its own brief, inbox and worker; a
  re-evaluation's question names who takes each retiring seat's work; "apply all" passes every
  proposal put to the owner; the walkthrough check passes in CI; and the new seats' sources each say
  whether the page was read or found.
• Upgrading a 1.2 project: the project keeps its copy of scripts/gen-team-agents.py; copy the
  new one from skills/team-setup/scripts/ to get the no-worker-before-authoring rule.

Ready to maintain in the open

• A front door for GitHub. The README is written for a first visit: what agent-team is, who it
  is for, install and update, a quick start, what is inside, how the owner stays in charge, and where
  to read more. CITATION.cff says how to cite it, and plugin.json names the homepage, the
  repository and the author's profile.
• Contributor files. CONTRIBUTING.md rewritten for contributors from outside, a code of conduct
  (Contributor Covenant), SECURITY.md (private vulnerability reporting), SUPPORT.md,
  GOVERNANCE.md, issue forms for a bug, a feature and a seat proposal, a pull request template,
  CODEOWNERS, and the label set in .github/labels.json.
• One version, released from a tag. tools/bump_version.py moves the Unreleased notes under a new
  heading and keeps plugin.json, marketplace.json, this changelog and CITATION.cff on one version;
  check.py fails when they disagree. A vX.Y.Z tag runs every check again and publishes the GitHub
  Release with that version's notes. docs/maintaining.md is the maintainer's guide, and
  tools/github_setup.sh applies the repository's settings, labels and branch protection, printing
  every call unless given --apply.
• A fact sheet for websites. docs/facts.json, generated by tools/export_facts.py from the
  repo's own sources; check.py fails when it is stale.
• CI runs the private leak-word sweep from a repository secret, with read-only permissions, and
  Dependabot proposes updates to the workflows' actions.
• install.sh works from a fresh clone, run from any folder. It links both skills to the clone by
  absolute path (no loop when ~/.claude/skills is itself a link to ~/.agents/skills), prints each
  skill it linked, exits 1 when a real folder blocks a skill, and removes only the seat links an
  install before 1.1.0 made into this clone.