Repository navigation
implement planning/claude-code-channel.md #112
Description
Implement the design in planning/claude-code-channel.md: pepper agents claude-channel, a command that serves a sponsored agent's room events to a Claude Code session as a channel (a stdio MCP server built on the mcp SDK, pushing notifications/claude/channel), plus MCP tools for the agent-token commands.
Phases (one PR each, an issue per phase as it is picked up) are in §13.2 of the doc. The maintainer's decisions are in §14 (2026-10-24). §15 lists what is still to verify, chiefly the P1 smoke test of the SDK's handshake-era behaviour against real Claude Code.
The design doc lands in its own docs PR: #113
Activity
okgrit commentedon Oct 24, 2026
The explicit P1 smoke gate in §13.1/§15 is useful. I maintain dotline, which has a separate small stdio channel, and reading this plan exposed a negotiation bug in ours: it echoed unsupported protocol strings. Our correction and regression now report the implemented revision instead. Your planned unsupported-version test already covers that boundary.
Would a small backend-neutral case card help cross-check P1 while keeping your selected SDK design?
A shared card could stay local and synthetic, with no Pepper PAT, room contents or transcript exchange. My candidate has protocol/HTTP regression evidence, but I have not run Pepper or established a successful native Claude channel trial. It is not a known-good native control.
If that comparison would be useful, happy to agree the case format here before implementing anything. Your interactive development test and later allowlisted-Team checks remain separate; I am not taking ownership of a P1 phase.
Prepared with Codex.
docs(planning): Claude Code channel design for agents claude-channel (V…