New capabilities: browser formula authoring, per-agent models, self-improving agents, and the fable agent family

[stempeck]

Summary

Adds a batch of new capabilities to the factory — headlined by things you couldn't do
before: author and edit formulas in the browser, assign a different model (or a local
endpoint) to each agent, turn on agents that improve their own formulas, a new family
of fixed-checklist "fable" agents, and headless visual verification for agents that
build UIs. It also enforces explicit, source-controlled agent installation, hardens command
safety across nested-factory checkouts, and adds a CLI formula validator.

Closes #82 — full capability breakdown, affected files, and acceptance criteria are in the
issue.

Highlights

• Browser formula authoring — a Formulas tab to browse, edit (comments/formatting
  preserved), and validate formulas, plus a wizard for new ones and a "Generate All Agents"
  button with live-streaming progress. Saves are conflict-checked and validated against the
  real engine.
• Per-agent model selection — each agent runs on its own model or a custom base
  URL + auth token; per-launch override available. Per-agent settings now survive redeploys
  instead of being wiped.
• Self-improving agents — opt-in continuous_improvement per agent (AND-gated with a
  factory switch, off by default); a finished agent refines its own formula and mails back a
  pass/fail verdict for a human to promote.
• Three new fable agents — fable-implement, fable-increment, fable-review, each
  running the same eleven-phase fixed checklist.
• Playwright visual verification — fail-soft install of headless Chromium + browser
  plugin so UI-building agents can screenshot and check their work; degrades to the human
  gate when unavailable.
• Explicit agent installation — af install --init seeds only manager + supervisor;
  every other agent enters agents.json only via an explicit install verb (ADR-020).
• Reliability — commands no longer silently act on the wrong factory in nested checkouts;
  new af formula validate returns a machine-readable pass/fail verdict, kept in sync with
  the browser validator via a conformance lane.

Testing

• make sync-formulas, make sync-skills, make build — all pass (formulas/skills in sync).
• make test (root module) — all packages green.
• cd web && go test ./... (web module) — all packages green.

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ stempeck
❌ agentfactory

---

agentfactory seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}