feat(chat): Claude Code as local chat provider

[vibegui]

Summary

• Adds Claude Code (Claude Agent SDK) as a first-class chat model provider in Decopilot
• Auto-wires MCP tools via ephemeral API keys (created per session, revoked after completion)
• Supports vision (image analysis), reasoning blocks (inline), and plan mode (tool-free planning)
• Maps three model variants: Opus (1M context), Sonnet (1M context), Haiku (200k context)
• Enhanced system prompt with comprehensive tool reference (shared-prompts.ts)
• Chat store improvements: planMode state, resumeStream message seeding, transport URL fix

Files Changed

• apps/mesh/src/api/routes/decopilot/claude-code-provider.ts — NEW SDK adapter (1,274 lines)
• apps/mesh/src/api/routes/decopilot/stream-core.ts — Claude Code branch with ephemeral API key lifecycle
• apps/mesh/src/api/routes/decopilot/schemas.ts — Add claude-code provider + planMode
• apps/mesh/src/api/routes/decopilot/routes.ts — Skip model permissions for Claude Code
• apps/mesh/src/api/routes/decopilot/constants.ts — Enhanced system prompt
• apps/mesh/src/api/routes/decopilot/types.ts — Connection auth + prompt suggestion data parts
• apps/mesh/src/tools/shared-prompts.ts — NEW shared tool reference prompt
• apps/mesh/src/web/components/chat/store/ — planMode state + transport fixes
• apps/mesh/src/web/components/chat/select-model.tsx — Claude Code tier mappings
• apps/mesh/src/web/components/chat/types.ts — planMode metadata field

Test plan

• Install deps: bun install
• Start dev server: bun run dev
• Select "Claude Code" model in chat model selector
• Send a message — verify response streams via Claude Agent SDK
• Verify tool calls work (ephemeral API key auto-wires MCP)
• Test plan mode toggle (if UI available)
• Verify bun run lint and bun run check pass (no new errors)

🤖 Generated with Claude Code

---

Summary by cubic

Adds Claude Code as a local chat provider using @anthropic-ai/claude-agent-sdk. Streams text, reasoning, tools, and vision into chat with per-session MCP auto-wiring and an optional plan mode.

• New Features

  • New provider with model IDs claude-code:opus, claude-code:sonnet, claude-code:haiku; mapped to tiers; uses local auth and skips model permission checks.
  • Ephemeral API key auto-wires Mesh MCP over HTTP per session and is revoked after; streams tool inputs/results, prompt suggestions, subtask updates, and tool latency; persists streamed tool parts for accurate history.
  • Vision reads images from the last user message.
  • Plan mode flows from UI to backend; toggles Claude Code between planning (plan) and execution (bypassPermissions).
  • System prompt is enriched with server agent instructions for consistent tool guidance.
  • Records LLM usage and cost; simple thread title is derived from the user prompt.

• Bug Fixes

  • Fixed chat reconnect by seeding existing messages before resume and correcting the transport URL.

^{Written for commit 043219c. Summary will update on new commits.}

[github-actions]

🧪 Benchmark

Should we run the Virtual MCP strategy benchmark for this PR?

React with 👍 to run the benchmark.

Reaction	Action
👍	Run quick benchmark (10 & 128 tools)

Benchmark will run on the next push after you react.

[github-actions]

Release Options

Should a new version be published when this PR is merged?

React with an emoji to vote on the release type:

Reaction	Type	Next Version
👍	Prerelease	2.181.4-alpha.1
🎉	Patch	2.181.4
❤️	Minor	2.182.0
🚀	Major	3.0.0

Current version: 2.181.3

Deployment

• Deploy to production (triggers ArgoCD sync after Docker image is published)

[cubic-dev-ai]

1 issue found across 12 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="apps/mesh/src/api/routes/decopilot/stream-core.ts">

<violation number="1" location="apps/mesh/src/api/routes/decopilot/stream-core.ts:360">
P2: Tool execution metrics are recorded with `durationMs: 0`, which provides no meaningful timing data for monitoring dashboards. The `claude-code-provider.ts` tracks tool timing internally but doesn't expose it. Consider either passing actual durations through `ccResult.parts` or recording metrics at the actual call site in the provider.

(Based on your team's feedback about recording tool execution metrics at the call site with actual timing.) [FEEDBACK_USED]</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}