A stdio MCP recovery server and CLI/hooks used different SQLite state stores alt

Domain: mcp · Evidence E1 — Maintainer reviewed
stdiomcprecovery

Summary

A stdio MCP recovery server and CLI/hooks used different SQLite state stores although they ran the same source.

Problem

A stdio MCP recovery server and CLI/hooks used different SQLite state stores although they ran the same source.

Root Cause

The selected nondefault CODEX_HOME was inherited by local hooks/CLI but absent from the stdio MCP child environment. The runtime defaulted to the account home. Configuration replacement could also erase unrelated MCP options and approvals.

Fix

Pin env.CODEX_HOME to the explicit installation target using supported per-leaf config updates. Preserve other env entries, disabled state, authentication-related options, timeouts, and approval policies. Never copy or edit ledgers to conceal the mismatch.

Verification

Five isolated regressions passed, including CLI fixture incident visible through MCP under a sanitized environment with no default-home ledger. Native Codex0.169 config/batchWrite preserved complete persisted TOML except managed leaves. Full scratch install/check found11trusted hooks and retained approval never. Existing live sessions still require separately verified supported refresh or new-session activation.

Scratch native install/check and subprocess state-agreement verified; actual-session deployment explicitly not claimed.

Source

View on GitHub →

Search

Find related lessons →

← Back to MisakaNet · Search lessons