feat: agent tool parity, Tools panel and the tools-per-module rule (CLEAN-109)

[maksymhryzodub-prog]

Summary

Restores the Ranch philosophy that the console is a window and the chat is the hands: everything the admin console can do, the Ranch admin agent can now do through a tool, and a person can see those tools from the chat.

Ticket: CLEAN-109

Three parts (spec, plan and audit in specs/016-agent-tool-parity/):

1. Parity. 93 new tools across 20 slice files, each beside the controller it mirrors and calling the same services: agents (stop/start/delete/status/env/logs/MCP list/capacity), workspace (files, secrets, channels, share link), templates (create/delete/attach MCP servers/restart fleet/git install/export), skills, LLM credentials, MCP servers and OAuth, knowledge bases and sources, settings (the tool description now carries a catalogue of every key), paddock, users, API keys, chats, usage, integrations, upgrade, Rancher status. 152 tools in total.
2. Tools panel. A wrench beside the composer in the Rancher chat and on an agent's Chat tab opens a sheet: accordions per topic, search across topics, one click drops a starter prompt into the composer with the first «…» selected. GET /agents/:id/tools returns exactly what that agent's runtime would receive (one shared listing implementation with MCP tools/list). A new AgentToolListing row records what each pod was last told, so tools added after a pod started show "after restart" with a restart button.
3. The rule. @Tool now requires topic, title and template, and a destructive tool a boolean confirm; the API refuses to boot otherwise. docs/agent-tools.md is the canonical rule, CLAUDE.md states it, the graft skill carries a pointer that scripts/ensure-agent-tools-rule.mjs re-inserts after graft init, and the new PR template asks for the parity check.

Decisions from the ticket discussion: full parity including users, keys and deletes (operator-only, confirm: true on 30 destructive tools); the panel shows the live per-agent list; existing tool names unchanged.

Test plan

• cd api && NODE_OPTIONS=--experimental-vm-modules npx jest — 121 suites, 1592 tests green. Includes mcp/tool-metadata.spec.ts (loads every tool file and runs the boot-time validation) and mcp/tool-secrets.spec.ts (feeds every credential-taking tool a sentinel and asserts it never comes back; create_api_key is the documented exception).
• cd api && bun run build clean; bun run generate:swagger → cd admin && bun run build:api regenerated the SDK.
• cd admin && bun test slices (13 new util tests) and npx nuxt typecheck clean.
• Not run in this PR: the live walkthrough in specs/016-agent-tool-parity/quickstart.md §3–§6 (needs a running stack with a deployed Rancher agent). Please run it as part of review; §4 (after-restart marker) and §6 (parity prompts) are the two to watch.
• Migration: api/prisma/migrations/20260922120000_agent_tool_listing (additive).

Checklist

• Jira issue linked above; commits carry the CLEAN-<n> id
• Console capability added or changed → agent tool added or changed, with a spec (docs/agent-tools.md)
• Destructive tools take confirm; no secret is echoed in any tool result
• User-visible app strings went through en.json; admin stays English-only
• Client state follows docs/state.md (one entity, one store, render by id)

🤖 Generated with Claude Code