Make command folding and output truncation in Codex CLI TUI configurable or disable-able

I would also like to see this enhancement. I would find it infinitely useful without any deeper layers adjustments.

is this planned? I'd love to see what's going on instead of just

<img width="274" height="33" alt="Image" src="https://github.com/user-attachments/assets/6b4733f8-2a75-4e19-91bb-5344bce8d5b8" />

I want to add that while this convenient feature is not implemented yet. I found I can view past run full command and its output when in the "double-tap-ESC history message editing" TUI.

I don't know why Google AI Mode and ChatGPT Pro all can't point that workaround out for me, but since they all direct me here for "existing issue", I guess I'll just "add this to the context". /s

Edit: The obvious disadvantage of this workaround is that I have to press ESC to get into this TUI, thus I can't inspect commands and outputs as they are getting run by the working Codex agent. I have to wait until it finishes.

2026, aka 10 years later in AI land, and this is still the current Codex UX for anything longer than a few lines:

```text
Ran ssh YYY "duckdb -c \"DESCRIBE SELECT * FROM read_parquet('XXX') LIMIT 0;\""
└ ┌───────────────────────────────────────────┐
│ Describe │
… +36 lines
└───────────────────────────────────────────┘

… +126 lines
└─────────────────────────────────────────┘
│ Describe │
… +19 lines
└─────────────────────────────────┘

  This makes safe development much harder because you cannot reliably inspect full tool output inline.

  Current workaround is exporting results to files and re-reading them, which is obviously silly for normal CLI flow.

  I understand external contribution load is high, but this is a critical UX issue and cannot be fixed downstream in user config alone because the truncation behavior is hardcoded in the renderer.

Maybe need to surface more of the UX so that community can extend via plugins and bypass the acceptance requirements if dev pipeline is so full that these sorts of things stay open for half a year?

  ## Accurate fix (current Codex layout)

  Note: current paths are under codex-rs/tui_app_server/src/* (not codex-rs/tui/src/*).

  ### 1) `codex-rs/core/src/config/types.rs`

  Add enum:

#[derive(Serialize, Deserialize, Debug, Clone, Copy, PartialEq, Eq, JsonSchema, Default)]
#[serde(rename_all = "lowercase")]
pub enum ToolOutputDisplay {
#[default]
Collapsed,
Full,
}

  Add field to Tui:

#[serde(default)]
pub tool_output_display: ToolOutputDisplay,

  ### 2) `codex-rs/core/src/config/mod.rs`

  Import and expose runtime config field:

use crate::config::types::ToolOutputDisplay;

pub tool_output_display: ToolOutputDisplay,

  Wire from TOML:

tool_output_display: cfg
.tui
.as_ref()
.map(|t| t.tool_output_display)
.unwrap_or_default(),

  ### 3) `codex-rs/tui_app_server/src/exec_cell/model.rs`

  Store mode on ExecCell and carry it through rebuilds:

use codex_core::config::types::ToolOutputDisplay;

pub(crate) struct ExecCell {
pub(crate) calls: Vec<ExecCall>,
animations_enabled: bool,
tool_output_display: ToolOutputDisplay,
}

  Constructor keeps default behavior:

pub(crate) fn new(call: ExecCall, animations_enabled: bool) -> Self {
Self {
calls: vec![call],
animations_enabled,
tool_output_display: ToolOutputDisplay::Collapsed,
}
}

  Accessors:

pub(crate) fn tool_output_display(&self) -> ToolOutputDisplay {
self.tool_output_display
}

pub(crate) fn set_tool_output_display(&mut self, tool_output_display: ToolOutputDisplay) {
self.tool_output_display = tool_output_display;
}

  And in `with_added_call`, `preserve tool_output_display`.

  ### 4) `codex-rs/tui_app_server/src/exec_cell/render.rs`

  Import:

use codex_core::config::types::ToolOutputDisplay;

  Update constructor path to receive mode and set it:

pub(crate) fn new_active_exec_command(
call_id: String,
command: Vec<String>,
parsed: Vec<ParsedCommand>,
source: ExecCommandSource,
interaction_input: Option<String>,
animations_enabled: bool,
tool_output_display: ToolOutputDisplay,
) -> ExecCell {
let mut cell = ExecCell::new(
ExecCall {
call_id,
command,
parsed,
output: None,
source,
start_time: Some(Instant::now()),
duration: None,
interaction_input,
},
animations_enabled,
);
cell.set_tool_output_display(tool_output_display);
cell
}

  Overflow-safe ellipsis math:

let show_ellipsis = total > line_limit.saturating_mul(2);

  Gate truncation:

let full_output_mode =
self.tool_output_display() == ToolOutputDisplay::Full && !call.is_user_shell_command();

let line_limit = if call.is_user_shell_command() {
USER_SHELL_TOOL_CALL_MAX_LINES
} else if full_output_mode {
usize::MAX
} else {
TOOL_CALL_MAX_LINES
};

let display_limit = if call.is_user_shell_command() {
USER_SHELL_TOOL_CALL_MAX_LINES
} else if full_output_mode {
usize::MAX
} else {
layout.output_max_lines
};

let rendered_output = if full_output_mode {
prefixed_output
} else {
Self::truncate_lines_middle(
&prefixed_output,
display_limit,
width,
raw_output.omitted,
Some(Line::from(Span::from(layout.output_block.subsequent_prefix).dim())),
)
};

  ### 5) `codex-rs/tui_app_server/src/chatwidget.rs`

  Pass the new arg at all new_active_exec_command(...) call sites:

self.config.tool_output_display

  ### 6) `codex-rs/tui_app_server/src/pager_overlay.rs`

  Update test call site with explicit mode:

codex_core::config::types::ToolOutputDisplay::Collapsed

  ### 7) `codex-rs/core/src/config/config_tests.rs`

  Update expected structs (Tui and Config) to include:

tool_output_display: ToolOutputDisplay::Collapsed

  (and import ToolOutputDisplay in the test module).

  ### 8) User config

  `~/.codex/config.toml:`

[tui]
tool_output_display = "full" # full | collapsed

  collapsed preserves current behavior, full disables the collapse path for tool output rendering.

  ———

  In the meantime, anyone blocked by this can apply the patch and rebuild.

  Repo (script: fetch latest, apply patch, build/install):
  https://github.com/TheReaperJay/codex-full-tool-output-installer/tree/master

  Current patch target is pinned to the layout used by Codex CLI 0.117.0 / commit e39ddc6, so expect refreshes as upstream files move.

To make it clear why this is a critical UX/security issue:
In the past few weeks we are dealing with increased security attacks through AI-assisted malware that is using all sorts of bleeding edge obfuscation tactics.

The fact that codex is truncating tool response with no way to force a review means this is yet another surface vector to slip things past even manual tool use - not to mention unattended agents.

@TheReaperJay Bro thanks for the patch, but I just want to note that even if they don't accept non-contributor PRs and close them, the best way to get people try your patch is still to file a PR and mention it here rather than in a long and unorganized comment...

Cool thanks @escape0707 - I read through the contributing guideline and saw that they just auto close the PR and assume they just straight up ignore it and move to a graveyard but you're right I'll give that a shot anyway!

Fwiw codex found this one for me and I can see there are lots of other issues opened for it too so I think you're right and it might surface it more.

@TheReaperJay great lead on this issue!

Make a PR and link it here! We should take any chance to surface this issue just as @escape0707 said.

absolutely deranged UX!
on par with while i am typing, giving a random 'approval' focus-stealing style. (why not a M- or C- required to approve if recent keystrokes)

Still an issue in Codex CLI v0.145.0. Tool calls, especially MCP calls returning large JSON payloads, overwhelm the transcript and make the actual conversation difficult to follow.

The VS Code Codex interface handles this much better by keeping tool details collapsed and available on demand. The CLI should offer equivalent behaviour.

Suggested configuration:

[tui]
tool_output_display = "collapsed" # hidden | collapsed | full

"Collapsed" should show one concise status line per call, such as Tool call completed, without arguments or output. Users should be able to expand an individual call when debugging. This should apply consistently to shell commands, MCP tools, searches, patches, and other tool types.

This is separate from reasoning visibility. hide_agent_reasoning and model_verbosity do not address tool call transcript noise. Please consider matching the VS Code default in the CLI.

Thanks all!