What variant of Codex are you using?
CLI
What feature would you like to see?
Problem
codex exec resume does not support the -o (output file) flag, unlike codex exec which does. This means the only way to capture output from a resumed session is piping stdout (e.g. 2>&1 | tail -80), which mixes the actual model response with tool-call noise (file reads, thinking steps, exec output).
Impact
When building automated workflows that use codex exec for an initial call (with -o for clean output capture) and then codex exec resume for follow-up rounds, the resume rounds produce significantly noisier output that's hard to parse programmatically.
Example workflow
# Round 1 — clean output via -o
codex exec -m gpt-5.3-codex -s read-only -o /tmp/review.md "Review these files..."
# Round 2 — no -o support, forced to pipe stdout
codex exec resume $SESSION_ID "I've revised based on your feedback..." 2>&1 | tail -80
The tail -80 approach captures Codex's internal tool calls (file reads, shell execs) alongside the actual review response, making it difficult to extract just the model's final answer.
Expected Behavior
codex exec resume should accept -o to write only the model's final response to a file, consistent with codex exec.
codex exec resume $SESSION_ID -o /tmp/review-round2.md "Re-review the changes..."
Workaround
Currently piping through tail or grep to try to isolate the response, but this is fragile since the number of tool-call output lines varies.
Will PR this as time allows, hopefully can get some eyes on it.
What variant of Codex are you using?
CLI
What feature would you like to see?
Problem
codex exec resume does not support the -o (output file) flag, unlike codex exec which does. This means the only way to capture output from a resumed session is piping stdout (e.g. 2>&1 | tail -80), which mixes the actual model response with tool-call noise (file reads, thinking steps, exec output).
Impact
When building automated workflows that use codex exec for an initial call (with -o for clean output capture) and then codex exec resume for follow-up rounds, the resume rounds produce significantly noisier output that's hard to parse programmatically.
Example workflow
The tail -80 approach captures Codex's internal tool calls (file reads, shell execs) alongside the actual review response, making it difficult to extract just the model's final answer.
Expected Behavior
codex exec resume should accept -o to write only the model's final response to a file, consistent with codex exec.
codex exec resume $SESSION_ID -o /tmp/review-round2.md "Re-review the changes..."Workaround
Currently piping through tail or grep to try to isolate the response, but this is fragile since the number of tool-call output lines varies.
Will PR this as time allows, hopefully can get some eyes on it.