MCP reload

The log MCP server picks up new code without restarting Claude Code: a reload proxy that never exits owns the client's pipe and swaps the real server underneath it when it's safe. Rule 167.

REQ-RELOAD-01: logMcp is a proxy that never exits, with the server as its child

bin/logMcp.ts runs runReloadProxy (src/mcp/reload/client.ts) unless it's given --child or --http, or LOG_MCP_RELOAD=off. The proxy spawns bin/logMcp.ts --child with Bun and passes every newline-delimited JSON-RPC line both ways. It keeps the client's initialize request. It tracks each request it forwards until the child answers it. It exits only when the client closes its stdin.

Checked by: code and SubQ

REQ-RELOAD-02: A reload waits until nothing is running, then replays initialize

ReloadProxy.reloadDue is true only when codeStamp is newer than the child's stamp, that stamp didn't already fail, no request is in flight and jobsRunning (jobs in listJobs whose pid is the child's, still running) is empty. Otherwise it records what it waits on (waitingOn). It runs on every request and on a timer (tick).

• reload starts a new child and writes the saved initialize to it with its own id.
• The child's answer to that id isn't forwarded.
• When it comes, the proxy sends notifications/initialized to the new child, stops the old one (killed, so its exit isn't a crash), sends notifications/tools/list_changed to the client and passes on the lines it queued meanwhile.

Checked by: code and SubQ

REQ-RELOAD-03: A crash or a failed start never leaves the client hanging

When the child exits on its own, every request in flight gets a JSON-RPC error with logMcpRestarted (restartedMidCall), and the child is started again with initialize replayed. More than 5 crashes in a minute stop the restarts until the code changes again. A new child that exits or doesn't answer initialize within 30 s is stopped. The old child keeps serving, lastError says why, and the same stamp isn't tried again. A request that comes while no child is running is answered with the same error.

Checked by: code and SubQ

REQ-RELOAD-04: Its state is recorded and the hook reads it

runReloadProxy writes the proxy's state to reload.json (reloadStateFile, LOG_MCP_RELOAD_FILE): the proxy and child pids, the child's stamp, the reloads, waitingOn and lastError. Each reload is logged to stderr and the live log. staleServerLine reads it (readProxyState). While the proxy is alive, it never tells matt to restart, only what the reload waits on or why the new code failed.

Checked by: code and SubQ