codex - 💡(How to fix) Fix codex exec --output-schema does not apply only to final output [1 comments, 2 participants]

What version of Codex CLI is running?

v0.125.0

What subscription do you have?

Pro

Which model were you using?

gpt-5.5

What platform is your computer?

Linux

What terminal emulator and version are you using (if applicable)?

No response

What issue are you seeing?

codex exec --output-schema does not appear to apply only to the final output.

When an agent run emits intermediate progress messages before tool calls, those intermediate agent_message items are also forced into the provided output schema. This can break workflows that consume the first schema-valid JSON message as the final result, because intermediate messages can look like valid final outputs.

In my case, this means automated workflows using --json --output-schema can intermittently fail or consume the wrong result depending on whether the model emits progress messages during the run.

What steps can reproduce the bug?

Create schema.json:

{
  "type": "object",
  "properties": {
    "final": {
      "type": "string"
    }
  },
  "required": [
    "final"
  ],
  "additionalProperties": false
}

Run:

codex exec --json --output-schema schema.json \
  "Strictly follow these three steps: First, say in one sentence that you will run pwd, then run pwd. Second, say in one sentence that you will run ls -1 | head -5, then run ls -1 | head -5. Third, say in one sentence that you will run git status --short, then run git status --short. Finally, summarize the results of all three commands. You must output the one-sentence explanation before running each command."

Actual output includes intermediate messages that already satisfy the final schema:

{"type":"item.completed","item":{"id":"item_0","type":"agent_message","text":"{\"final\": \"I will run `pwd` to confirm the current working directory.\"}"}}

{"type":"item.completed","item":{"id":"item_2","type":"agent_message","text":"{\"final\":\"I will run `ls -1 | head -5` to list the first five directory entries.\"}"}}

{"type":"item.completed","item":{"id":"item_4","type":"agent_message","text":"{\"final\":\"I will run `git status --short` to check the concise worktree status.\"}"}}

Only the last message is actually the final result:

{"type":"item.completed","item":{"id":"item_6","type":"agent_message","text":"{\"final\":\"Results: `pwd` returned `/home/sunday/Projects/codex`. `ls -1 | head -5` returned `AGENTS.md`, `BUILD.bazel`, `CHANGELOG.md`, `LICENSE`, and `MODULE.bazel`. `git status --short` shows one untracked file: `schema.json`.\"}"}}

What is the expected behavior?

--output-schema should constrain only the final output of codex exec, not intermediate progress messages in the agent loop.

Intermediate agent_message items should remain normal progress messages, and only the final returned result should be schema-compliant.

Additional information

From a quick source check, final_output_json_schema appears to be copied into each prompt built during the sampling loop, so every Responses API request in the turn may include text.format.schema.

Related issue: #15451 discusses schema/tool behavior, but this report is specifically about intermediate agent messages being shaped as final schema output.