[santoshkumarradha]

What

The person's own signed-in Claude Code, driven the way T3 Code drives it (claude over stream-json with the control protocol, CLAUDE_CODE_ENTRYPOINT), in three places:

• Task program claude (internal/claudecode): a task handed to Claude Code on a branch; messages typed on the task reach it mid-run; outward commands (push, PR, release, publish) refused on the wire via --disallowedTools, over the person's own hooks.
• Provider claude-code (internal/claudemodel, kind agent, no key): a conversation or crew member on claude-code/opus|sonnet|haiku is a live Claude Code session in its folder. Toolless small jobs (titles, summaries) are one-shot claude -p, so codeaf with only Claude Code connected works with no OpenRouter key.
• Worker seat: a Worker pinned to Claude Code hands tasks to the claude program.

T3 Code parity, checked live:

• Mid-turn messages: codeaf cuts the request on a steer; the transport now keeps the cut turn running for 3s and writes the correction into it. Live: a steer 7s into sleep 12 && … did not stop it; both files made, one answer.
• Resume: the session id is kept in the conversation folder and --resumed after an engine restart (live: code word recalled across two engine processes).
• Switching models: moving off Claude Code and back hands it what it missed.

Wire: providers.list gains kind agent; keys.check reports Found Claude Code <version> · signed in · <plan>; keys.set connects with no key.

Desk: Agent-Field/codeaf-desk#325.

Verified live (this laptop, Claude Max plan)

connect → models.list → conversation turn (file written) → crew member started on claude-code/haiku reports finished → steer mid-turn → resume across restart → Claude Code calls mcp__codeaf__manual and mcp__codeaf__propose_task; the task is proposed and queued.

codeaf's tools on Claude Code

codeaf's belt is served to the Claude Code session as an in-process SDK MCP server (--mcp-config type sdk, answered over the control protocol's mcp_message), less the tools it has its own of. A call becomes a tool_calls delta, so codeaf's turn loop runs it with its own row; the next request answers the waiting MCP call and the same Claude turn continues. Pictures on the newest message go as image blocks (unit-tested, not yet live).

Claude Code's own steps, in codeaf's format

Each step Claude Code runs itself (Bash, Read, Edit, MultiEdit, Write, Grep, Glob, LS, WebFetch, WebSearch) is handed up the stream when its result arrives, as a codeaf_steps delta → provider.StreamStepsRan → session.recordStepsRan, and recorded as codeaf's own call and result (bash, read, edit, …) with its took line. So the page draws rows, a reopened chat redraws them, and the next model reads them in its history. Nothing is run twice. Ids carry claude_code_, so the transport never re-tells Claude Code its own history.

Canvas and skills

Conversations and tasks on Claude Code load the skills kept for codeaf and other programs (.codeaf/skills, .agents/skills, Codex, Cursor…) as a session plugin (--plugin-dir, links, not copies), shown as codeaf:<name>. The turn's request context carries CODEAF_AGENT_DID to the session, so a canvas it makes belongs to the conversation. Live: "mock up a landing page for Ember on a canvas" → codeaf:workaf-canvas loaded → canvas new/frame add/put/look → canvas member is the conversation's did:key:….

Fixed along the way (found live)

• Every turn after the first was mis-addressed: the last user-role message (codeaf's session note) was sent as "the person now says". Now every user-role message since the last answer goes, notes as notes.
• False compaction: Claude Code's result usage is the turn's sum (4 steps ≈ 199k) and codeaf read it as context fill. Now the last call's context is reported.
• A correction delivered twice: the 10s first-reply watch cut the adopting request during a long command and re-asked. An adopting stream now speaks at once, and a re-ask with nothing new only watches.
• codeaf connect claude-code / chat /connect treated it as a key service (bare row, 128k window, local chat refused turns). Both now run the app's connect.
• First-run setup in the chat lists Claude Code only where claude is installed.

Verified live

• App wire: steps rows + turn 2 reads them; canvas with DID.
• Terminal (codeaf chat): multi-turn, steering lands once (queued_command ×1), file-change line.
• Relay (codeaf serve + chat --at, own relay): turns 1–3, steering, manual row, context ≈ 51k.

Not done / limits

• No approval UI: edits are accepted; only outward commands are refused.
• Effort setting does not reach Claude Code; its tool and skill lists are fixed at session start.
• /rewind does not put back files Claude Code changed (stated in the manual).
• --host over ssh not exercised (the box's ssh codeaf predates this).

Checks

make pr-ready BASE=origin/app/v3 — pass on c6b458a22 (laws + touched packages incl. tui3, session, provider, codeaf).

🤖 Generated with Claude Code