KiteBot on CopilotKit Intelligence — Phase 1 (channels migration + Intelligence Gateway)

[jerelvelarde]

KiteBot on CopilotKit Intelligence — Phase 1 (foundation)

Migrates OpenTag off the deprecated @copilotkit/bot* packages onto the renamed
@copilotkit/channels*, renames the bot to KiteBot, drops Redis, and adds an
Intelligence Gateway run mode. Spec + plan under docs/superpowers/.

What changed

• Package migration @copilotkit/bot* → @copilotkit/channels* (pure rename; API
  identifiers unchanged). All packages are now published on npm, so standalone
  pnpm install from this repo works — the old "monorepo-only" caveat is gone.
• app/managed.ts — a new Intelligence Gateway channel host (pnpm channel) using the
  published two-process pattern: createBot + startChannelsOverRealtimeGateway
  (@copilotkit/channels-intelligence). Same bot, no Slack tokens — Intelligence owns the
  Slack edge. The agent brain is reached over AG-UI at AGENT_URL (the existing
  runtime.ts triage backend for now; a LangGraph deep agent in Phase 2).
• KiteBot identity — Slack display name + agent persona + channel name kitebot. The
  OpenTag project name is unchanged.
• Redis dropped — demo-restart.tsx, docker-compose.yml, demo:restart, and the
  redis-store dep removed (matches upstream examples/slack).
• Docs — README/setup.md lead with the Intelligence Gateway deploy as recommended,
  self-hosted as the alternative; .env.example gains the INTELLIGENCE_* block.

Deviations from the original plan (surfaced during execution)

1. The unified CopilotRuntime({ channels }) / createChannel API (from the initial
   snippet) is main-only and unpublished on npm — so this uses the published
   two-process createBot + startChannelsOverRealtimeGateway pattern instead.
2. A duplicate-rxjs skew across the channels packages broke the AbstractAgent type; fixed
   with a pnpm.overrides.rxjs ^7.8.2.
3. @copilotkit/channels-discord/-telegram/-whatsapp are only published at 0.0.x; pinned
   to their latest published versions.

Code review

Ran a multi-agent CR (Opus). The pr-review-toolkit plugin wasn't available in-session, so
a fallback 7-slot general-purpose review was used. Two fix cycles landed genuine fixes:
PORT validation, onThreadStarted guard, error-wrapping for slash-command runAgent,
render-table truncation-notes surfacing, page-list block cap, priority rendering, HITL
no-response vs decline, a notion-mcp secret-in-argv hygiene fix, the OpenAI-only docs
correction, and ~29 new tests (65 → 94).

Accepted as by-design / not changed (documented rather than "fixed"):

• AGENT_MODEL cast to the OpenAI model union — no runtime union to validate against;
  the caveat is documented in-code.
• as ModalView in file-issue.tsx — necessary because channels-ui JSX types as BotNode
  (verified); commented.
• "unchecked thread.post" — a false positive: post throws on failure and the run-loop
  wraps throws into { error } for the agent (verified against the installed .d.ts).
• The intentional post-timeout MCP-connect catch in runtime.ts.
• The fallback whole-file review does not converge to zero (it audits the whole
  pre-existing app, not the diff). The canonical diff-scoped pr-review-toolkit CR should
  be run in a fresh session as the authoritative gate — a handful of low-value pre-existing
  coverage nitpicks (e.g. page-list editedBy, render-table center alignment) remain
  deferred.

Verification

• pnpm check-types clean; pnpm test 94/94.
• Not yet done (needs live creds): the gateway smoke test — set the real INTELLIGENCE_*
  values + confirm the channel slug (kitebot) and dev apiUrl/wsUrl, run pnpm runtime
  • pnpm channel, and confirm the KiteBot dashboard flips Waiting for runtime → live.

Out of scope (later phases)

• Phase 2: LangGraph deep-agent backend behind AGENT_URL (co-marketing).
• Phase 3: Railway one-click template.

🤖 Generated with Claude Code