You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Forward daemon-related CBM environment variables in generated Codex configuration #1664
Version 0.10.3 fixed forwarding CBM_CACHE_DIR to the Codex MCP subprocess. However, other CBM environment variables may also affect the daemon started or used by that subprocess. Codex sanitizes the subprocess environment unless variables are explicitly included in env_vars.
Proposed change
Include all supported daemon-relevant CBM variables in the generated Codex MCP configuration:
The rendezvous-critical half of this is in PR #1683: CBM_RUNTIME_DIR (which exists since #1645) is now forwarded alongside CBM_CACHE_DIR — without it a relocated-runtime setup gets a Codex subprocess that looks for the daemon in the default location and never finds it, the same silent split as #1562. Both names decide WHICH daemon a process talks to, so they are listed unconditionally. The broader list you proposed (LOG_LEVEL, WORKERS, MEM_BUDGET, DIAGNOSTICS, ALLOWED_ROOT, DUMP_VERIFY_MIN_RATIO) is a real enhancement question we are deliberately not deciding in a bugfix: forwarding behavioural knobs means a shell-scoped experiment variable silently reconfigures the agent's long-lived daemon, which cuts both ways. Leaving this open for that decision.
Isn't that reconfiguring daemon behavior already true for all other agents?
Codex is the one behaving differently in this case. And first level issue
is that you can't troubleshoot mcp started by codex, because it strips env
like diagnostics mode and log level.
—
Kind regards,
Anton
On Sun, Aug 16, 2026 at 8:41 PM Martin Vogel ***@***.***> wrote:
*DeusData* left a comment (DeusData/codebase-memory-mcp#1664)
<#1664 (comment)>
The rendezvous-critical half of this is in PR #1683
<#1683>:
CBM_RUNTIME_DIR (which exists since #1645
<#1645>) is now
forwarded alongside CBM_CACHE_DIR — without it a relocated-runtime setup
gets a Codex subprocess that looks for the daemon in the default location
and never finds it, the same silent split as #1562
<#1562>. Both names
decide WHICH daemon a process talks to, so they are listed unconditionally.
The broader list you proposed (LOG_LEVEL, WORKERS, MEM_BUDGET, DIAGNOSTICS,
ALLOWED_ROOT, DUMP_VERIFY_MIN_RATIO) is a real enhancement question we are
deliberately not deciding in a bugfix: forwarding behavioural knobs means a
shell-scoped experiment variable silently reconfigures the agent's
long-lived daemon, which cuts both ways. Leaving this open for that
decision.
—
Reply to this email directly, view it on GitHub
<#1664?email_source=notifications&email_token=AAZPJXKJDCEIUQPYSDV4U7D5KJ5FTA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZRGE2DSNRQHE2KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-5311496094>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAZPJXIBWEFYV554KKGRN7L5KJ5FTAVCNFSNUABGKJSXA33TNF2G64TZHMYTCNRWGEYDEMJUHA5US43TOVSTWNJRGYYTMNBQGAYDDILWAI>
.
You are receiving this because you authored the thread.Message ID:
***@***.***>
Thank you, Anton — the inability to pass diagnostic settings for troubleshooting is a useful distinction from changing normal daemon behavior. Your point about consistency with other clients is also noted. The broader forwarding policy still needs a maintainer decision; I do not have a new implementation outcome to report here.
Thank you for explaining how missing diagnostic environment settings affect troubleshooting in Codex. The supported forwarding policy remains a maintainer decision. Sorry for the delay; this is waiting on maintainer follow-up, and I’ll follow up soon.
Activity
fix(cli): forward CBM_RUNTIME_DIR in the generated Codex configuration (
DeusData commentedon Aug 17, 2026
The rendezvous-critical half of this is in PR #1683: CBM_RUNTIME_DIR (which exists since #1645) is now forwarded alongside CBM_CACHE_DIR — without it a relocated-runtime setup gets a Codex subprocess that looks for the daemon in the default location and never finds it, the same silent split as #1562. Both names decide WHICH daemon a process talks to, so they are listed unconditionally. The broader list you proposed (LOG_LEVEL, WORKERS, MEM_BUDGET, DIAGNOSTICS, ALLOWED_ROOT, DUMP_VERIFY_MIN_RATIO) is a real enhancement question we are deliberately not deciding in a bugfix: forwarding behavioural knobs means a shell-scoped experiment variable silently reconfigures the agent's long-lived daemon, which cuts both ways. Leaving this open for that decision.
anton-matosov commentedon Aug 17, 2026
fix(install): adopt the Codex MCP table after Codex rewrites it (#1720)
DeusData commentedon Oct 19, 2026
Thank you, Anton — the inability to pass diagnostic settings for troubleshooting is a useful distinction from changing normal daemon behavior. Your point about consistency with other clients is also noted. The broader forwarding policy still needs a maintainer decision; I do not have a new implementation outcome to report here.
DeusData commentedon Oct 25, 2026
Thank you for explaining how missing diagnostic environment settings affect troubleshooting in Codex. The supported forwarding policy remains a maintainer decision. Sorry for the delay; this is waiting on maintainer follow-up, and I’ll follow up soon.