Claude Code Chat Mode

Claude Code Chat is the original Yeaft session backend — it wraps a local Claude Code CLI process in a Web chat surface, giving you the full Claude Code feature set (skills, MCP, subagents, /compact, /clear) without losing the underlying stream-json protocol.

This is the 1:1 chat mode powered by Claude Code CLI. To swap Claude for GitHub Copilot CLI, see Copilot Mode. For multi-VP parallel + cross-task memory, see Yeaft Sessions.

Prerequisites

1. Claude Code CLI installed: claude --version works on the Agent machine
2. Logged in: claude auth login has been run on the Agent machine (or ANTHROPIC_API_KEY is set)
3. Agent online: the sidebar shows at least one online Agent

Create a session

Two ways:

1. Welcome page — when no session is selected the main area shows "New session" (only when an Agent is online)
2. Sidebar — the + icon next to "Recent chats"

In the modal:

• Agent — which machine runs it
• Provider — pick Claude Code (default)
• Working directory — project path (affects cd, .claude/ lookup)
• Model (optional) — Claude defaults to whatever claude config says; override here if needed

Sending messages

• Type in the bottom input box
• Enter sends, Shift+Enter newline
• The send button turns into a stop button mid-execution — interrupt any time
• Draft auto-save — switching sessions doesn't lose your draft

File / image attachments

• Click the 📎 paperclip next to the input, or drag a file / screenshot in
• Paste images: Ctrl+V / Cmd+V
• Supported types: image/* , text, PDF, Word (doc/docx), Excel (xls/xlsx), JSON, Markdown, Python, JS, TS, CSS, HTML
• Thumbnails preview before sending; after sending they collapse into a "📎 2 images, 1 file" tag

Slash commands

Typing / opens an auto-complete menu:

Command	What it does
/compact	Compact the session context — fewer tokens, key points kept
/clear	Wipe all messages, reset context
/context	Show the current context-usage breakdown
/cost	Token usage and cost
/init	Initialize the project (generate CLAUDE.md)
/doctor	Run diagnostics
/memory	Manage Claude's project memory
/model	Switch model
/review	Code review
/mcp	MCP server management
/skills	List loaded skills
/btw	Side-question without interrupting the current task

Arrow keys navigate, Tab / Enter selects.

Compact vs Clear

Two buttons on the right side of the chat header:

Compact (↕)

• What: invokes Claude Code's /compact, replacing history with a concise summary
• When: context above 50% and you want to keep this conversation going
• Effect: history messages remain visible in the UI, but the next request to Claude carries the compressed context
• Input is disabled during compact; status bar shows "Compacting..."

Clear (🗑)

• What: invokes /clear, deletes all messages in this session
• When: you want a fresh start in the same session shell
• Confirmation: a confirm dialog prevents misclicks

Session resume

If the server restarts or the connection drops, sessions can resume:

• Click the ↻ refresh button in the chat header to re-sync the last 5 turns from the Agent
• Agents store sessions at ~/.claude/projects/<hash>/sessions/<sessionId>.jsonl
• The same sessionId resumes across restarts as long as the jsonl file is intact

Context usage indicator

The percent badge top-right of the chat header tells you how much of the context window is used:

• 🟢 Green (0–49%): healthy
• 🟡 Yellow (50–79%): consider compacting soon
• 🔴 Red (80%+): nearly full, compact or clear now

Hover for the precise number (Context: 45k / 200k).

Assistant reply rendering

Each reply renders as a Turn card:

• Markdown — syntax-highlighted code blocks
• Copy — per-reply and per-code-block copy buttons
• Tool calls — Read / Edit / Bash etc. visualized; latest stays expanded
• Todo progress — TodoWrite calls render as a checklist (✓ / ⏳ / ◯)
• AskUserQuestion — Claude-initiated questions become interactive cards (single/multi-select / freeform / submit)
• Sub-Agent nesting — Agent tool's sub-agents are inspectable

Differences vs Copilot / Yeaft Sessions

Capability	Claude Code	Copilot	Yeaft Sessions
/compact auto-compress	✓	—	✓ (H2-AMS)
/clear reset	✓	✓	✓
Model picker	✓	✓	✓ (per VP)
MCP tools	✓	✓	✓
Image / file attachments	✓	✓	✓
AskUser permission popup	✓	✓	✓
Subagent nested monitor	✓	—	✓
Expert Panel	✓	—	—
Cross-task persistent memory	—	—	✓
Multi-VP parallel reply	—	—	✓

Troubleshooting

Agent shows online but can't create sessions

• Confirm claude --version works on the Agent machine
• Check Agent logs: yeaft-agent logs
• Usually Claude CLI isn't installed or logged in

Claude keeps "thinking" forever

• Could be a Claude API timeout; hit stop, resend
• Look at server and agent logs for stack traces

/skills doesn't show my skill

• Skills are a Claude Code feature, living in ~/.claude/skills/ or project .claude/skills/
• Confirm claude /skills shows it on the Agent machine first