Render `UserPromptSubmit.additionalContext` as multiline hook context in the TUI

Start in codex-rs/tui/src/chatwidget.rs at the HookOutputEntryKind::Context rendering branch, then review the hook data flow in codex-rs/hooks/src/engine/output_parser.rs and codex-rs/core/src/hook_runtime.rs. Verify how multiline entry.text is displayed in the TUI. Done means Context notifications preserve line breaks and render a separate hook context label without changing the injected string.

What version of Codex CLI is running?

0.118.0

What subscription do you have?

Plus

Which model were you using?

No response

What platform is your computer?

No response

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

No response

What issue are you seeing?

UserPromptSubmit.hookSpecificOutput.additionalContext is already treated as the model-visible hook context surface, and the core pipeline preserves it as a raw string, including newlines.

Relevant code paths:

• codex-rs/hooks/src/schema.rs

  pub(crate) struct UserPromptSubmitHookSpecificOutputWire {
      pub hook_event_name: HookEventNameWire,
      #[serde(default)]
      pub additional_context: Option<String>,
  }
  

• codex-rs/hooks/src/engine/output_parser.rs

  let additional_context = wire
      .hook_specific_output
      .and_then(|output| output.additional_context);
  

• codex-rs/core/src/hook_runtime.rs

  fn additional_context_messages(additional_contexts: Vec<String>) -> Vec<ResponseItem> {
      additional_contexts
          .into_iter()
          .map(|additional_context| DeveloperInstructions::new(additional_context).into())
          .collect()
  }
  

So the injected content path already supports structured multiline text.

The TUI side does not present that structure cleanly. It currently prefixes the whole entry with hook context: and renders it as one formatted line:

• codex-rs/tui/src/chatwidget.rs

  let prefix = match entry.kind {
      codex_protocol::protocol::HookOutputEntryKind::Context => "hook context: ",
      // ...
  };
  lines.push(format!("  {prefix}{}", entry.text).into());
  

This makes structured hook context hard to read in the TUI. In practice, hook authors are pushed to flatten rich context into synthetic one-line encodings such as:

[Action now]: ...; [Authority]: ...; [Wait rule]: ...

instead of emitting the more natural multiline content they actually want the model to see, for example:

Action now
- Keep replying with plan preview only.
- Do not emit the wrapped full plan.

Authority
- Managed planning is active.

Wait rule
- If managed evidence is needed, spawn a managed sub-agent first.

What steps can reproduce the bug?

1. Register a UserPromptSubmit hook that returns multiline additionalContext.
2. Example hook output:

{
  "hookSpecificOutput": {
    "hookEventName": "UserPromptSubmit",
    "additionalContext": "Action now\n- Keep replying with plan preview only.\n- Do not emit the wrapped full plan.\n\nAuthority\n- Managed planning is active."
  }
}

3. Trigger the hook in the Codex TUI.
4. Observe the rendered hook notification in the transcript/history area.

Minimal behavior to inspect:

• model-side injection path preserves the raw string
• TUI-side hook notification path does not render the structure as a readable multiline block

What is the expected behavior?

The TUI should render HookOutputEntryKind::Context using the exact text structure carried in additionalContext, including line breaks.

For example:

UserPromptSubmit hook (completed)
  hook context:
    Action now
    - Keep replying with plan preview only.
    - Do not emit the wrapped full plan.

    Authority
    - Managed planning is active.

At minimum:

• multiline additionalContext should not force hook authors to collapse content into a single line
• the TUI-visible content should preserve the same structure that gets injected as developer instructions
• only the host prefix should differ; the body should stay semantically identical

Additional information

Why this matters:

• Hook authors often need to send structured policy or coordination context to the model.
• The cleanest machine/human shape for that content is often multiline Markdown-like text.
• The current TUI rendering strongly encourages ugly one-line serialization purely for display.
• That makes hook debugging harder, because the operator sees a degraded version of the intended context shape.
• It also pushes hook implementations to optimize for host rendering instead of clear model guidance.

Suggested implementation direction:

• In codex-rs/tui/src/chatwidget.rs, special-case HookOutputEntryKind::Context when entry.text contains \n.
• Render a standalone hook context: label line, then append the body as indented lines rather than format!(" {prefix}{}", entry.text).
• Keep the underlying injected additionalContext string unchanged in core/hooks.

Duplicate-issue search:

• I searched open issues in openai/codex for combinations of UserPromptSubmit, additionalContext, hook context, multiline, and TUI.
• I did not find a clear duplicate for this exact topic.
• Closest but different issues:
  • #5253 about missing interrupted tool-call context, not hook context rendering: https://github.com/openai/codex/issues/5253
  • #4790 about prior tool calls disappearing from the TUI transcript, not multiline hook rendering: https://github.com/openai/codex/issues/4790

Generated with Codex CLI.