Skip to content

Customer-first dashboard; custom rules, re-ask, detector switches, Guardrails AI import - #74

Merged
architsharm merged 1 commit into
mainfrom
claude/agent-fox-ux-redesign-b53bd5
Oct 8, 2026
Merged

architsharm merged 1 commit into
mainfrom
claude/agent-fox-ux-redesign-b53bd5

Conversation

@architsharm

Copy link
Copy Markdown
Owner

What changes for a customer

  • Dashboard redesign: Home, Observe, Issues, Agents, Policies, Approvals, Test, Reports — tabs instead of text, one shared kit.
  • Tune any rule in place: action, on/off, sensitivity (Low/Medium/High), message to the end user, "ask the model to fix it" — each previewed against the last 7 days before applying.
  • Your own rules: words, patterns, a topic to avoid, allowed topics, tool sequences; try box before saving.
  • Detector switches per workspace (Policies → Library → Detectors).
  • Import from Guardrails AI: paste Python / .rail / to_dict() JSON; preview the plan, untick, import. The code is parsed, never executed.
  • agent-integrity pack (watching): task drift, insecure code in tool args, NLI grounding.

Design

See docs/adr/0002-customer-authored-rules.md: custom rules are data + a managed custom policy pack (same modes, simulation, versions, audit); importers only plan; new semantic checks raise risks and packs decide.

Deploy notes

  • New tables custom_rules, detector_settings (migration e2a6c4f8b131); init_db() creates them on startup.
  • Vendored wheels rebuilt.

Tests

Full backend suite green apart from the wheel checks (now rebuilt, 5/5); ruff, format, import-linter clean; dashboard typecheck + 67 vitest pass; verified in the browser on a seeded demo.

🤖 Generated with Claude Code

…d Guardrails AI import

Dashboard: rebuilt the logged-in app around what a customer does — Home,
Observe, Issues, Agents, Policies, Approvals, Test, Reports — with a shared
kit, query-param tabs and little text. Rules can be tuned in place (action,
on/off, sensitivity, end-user message, re-ask) with a 7-day replay first.

Gateway:
- Custom rules (words, patterns, topics, allowed topics, tool sequences)
  stored as data and enforced through a managed `custom` policy pack.
- Rule.message (shown to the end user) and Rule.on_block=reask (gateway asks
  the model once more instead of refusing).
- Per-workspace detector switches; opt-in topic and NLI grounding models.
- Guardrails AI importer: reads Python, .rail or to_dict() JSON without
  executing it and returns a plan; apply goes through the normal stores.
- agent-integrity pack: task drift, insecure code, model-based grounding.
- Metrics, access, library and rule-patch routes for the dashboard.

ADR 0002 records the patterns. Vendored wheels rebuilt.

Co-Authored-By: Claude <noreply@anthropic.com>
@vercel

vercel Bot commented Oct 8, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
guardrails-api Ready Ready Preview Oct 8, 2026 2:48pm UTC
guardrails-dashboard Ready Ready Preview Oct 8, 2026 2:48pm UTC
guardrails-redteam-lang Ready Ready Preview Oct 8, 2026 2:48pm UTC

This branch was successfully deployed

3 active deployments
Preview – guardrails-api — 4a7eeb56 Deployed Oct 8, 2026 by vercel[bot]
Preview – guardrails-redteam-lang — 4a7eeb56 Deployed Oct 8, 2026 by vercel[bot]
Preview – guardrails-dashboard — 4a7eeb56 Deployed Oct 8, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant