### Motivation
Today config.toml has three different OTEL knobs under `[otel]`:
- `exporter` controls where OTEL logs go
- `trace_exporter` controls where OTEL traces go
- `metrics_exporter` controls where metrics go
Those often (pretty much always?) serve different purposes.
For example, for OpenAI internal usage, the **log exporter** is already
being used for IT/security telemetry, and that use case is intentionally
content-rich: tool calls, arguments, outputs, MCP payloads, and in some
cases user content are all useful there. `log_user_prompt` is a good
example of that distinction. When it’s enabled, we include raw prompt
text in OTEL logs, which is acceptable for the security use case.
The **trace exporter** is a different story. The goal there is to give
OpenAI engineers visibility into latency and request behavior when they
run Codex locally, without sending sensitive prompt or tool data as
trace event data. In other words, traces should help answer “what was
slow?” or “where did time go?”, not “what did the user say?” or “what
did the tool return?”
The complication is that Rust’s `tracing` crate does not make a hard
distinction between “logs” and “trace events.” It gives us one
instrumentation API for logs and trace events (via `tracing::event!`),
and subscribers decide what gets treated as logs, trace events, or both.
Before this change, our OTEL trace layer was effectively attached to the
general tracing stream, which meant turning on `trace_exporter` could
pick up content-rich events that were originally written with logging
(and the `log_exporter`) in mind. That made it too easy for sensitive
data to end up in exported traces by accident.
### Concrete example
In `otel_manager.rs`, this `tracing::event!` call would be exported in
both logs AND traces (as a trace event).
```
pub fn user_prompt(&self, items: &[UserInput]) {
let prompt = items
.iter()
.flat_map(|item| match item {
UserInput::Text { text, .. } => Some(text.as_str()),
_ => None,
})
.collect::<String>();
let prompt_to_log = if self.metadata.log_user_prompts {
prompt.as_str()
} else {
"[REDACTED]"
};
tracing::event!(
tracing::Level::INFO,
event.name = "codex.user_prompt",
event.timestamp = %timestamp(),
// ...
prompt = %prompt_to_log,
);
}
```
Instead of `tracing::event!`, we should now be using `log_event!` and
`trace_event!` instead to more clearly indicate which sink (logs vs.
traces) that event should be exported to.
### What changed
This PR makes the log and trace export distinct instead of treating them
as two sinks for the same data.
On the provider side, OTEL logs and traces now have separate
routing/filtering policy. The log exporter keeps receiving the existing
`codex_otel` events, while trace export is limited to spans and trace
events.
On the event side, `OtelManager` now emits two flavors of telemetry
where needed:
- a log-only event with the current rich payloads
- a tracing-safe event with summaries only
It also has a convenience `log_and_trace_event!` macro for emitting to
both logs and traces when it's safe to do so, as well as log- and
trace-specific fields.
That means prompts, tool args, tool output, account email, MCP metadata,
and similar content stay in the log lane, while traces get the pieces
that are actually useful for performance work: durations, counts, sizes,
status, token counts, tool origin, and normalized error classes.
This preserves current IT/security logging behavior while making it safe
to turn on trace export for employees.
### Full list of things removed from trace export
- raw user prompt text from `codex.user_prompt`
- raw tool arguments and output from `codex.tool_result`
- MCP server metadata from `codex.tool_result` (mcp_server,
mcp_server_origin)
- account identity fields like `user.email` and `user.account_id` from
trace-safe OTEL events
- `host.name` from trace resources
- generic `codex.tool_decision` events from traces
- generic `codex.sse_event` events from traces
- the full ToolCall debug payload from the `handle_tool_call` span
What traces now keep instead is mostly:
- spans
- trace-safe OTEL events
- counts, lengths, durations, status, token counts, and tool origin
summaries
## Summary
This changes `custom_tool_call_output` to use the same output payload
shape as `function_call_output`, so freeform tools can return either
plain text or structured content items.
The main goal is to let `js_repl` return image content from nested
`view_image` calls in its own `custom_tool_call_output`, instead of
relying on a separate injected message.
## What changed
- Changed `custom_tool_call_output.output` from `string` to
`FunctionCallOutputPayload`
- Updated freeform tool plumbing to preserve structured output bodies
- Updated `js_repl` to aggregate nested tool content items and attach
them to the outer `js_repl` result
- Removed the old `js_repl` special case that injected `view_image`
results as a separate pending user image message
- Updated normalization/history/truncation paths to handle multimodal
`custom_tool_call_output`
- Regenerated app-server protocol schema artifacts
## Behavior
Direct `view_image` calls still return a `function_call_output` with
image content.
When `view_image` is called inside `js_repl`, the outer `js_repl`
`custom_tool_call_output` now carries:
- an `input_text` item if the JS produced text output
- one or more `input_image` items from nested tool results
So the nested image result now stays inside the `js_repl` tool output
instead of being injected as a separate message.
## Compatibility
This is intended to be backward-compatible for resumed conversations.
Older histories that stored `custom_tool_call_output.output` as a plain
string still deserialize correctly, and older histories that used the
previous injected-image-message flow also continue to resume.
Added regression coverage for resuming a pre-change rollout containing:
- string-valued `custom_tool_call_output`
- legacy injected image message history
#### [git stack](https://github.com/magus/git-stack-cli)
- 👉 `1` https://github.com/openai/codex/pull/12948
Took over the work that @aaronl-openai started here:
https://github.com/openai/codex/pull/10397
Now that app-server clients are able to set up custom tools (called
`dynamic_tools` in app-server), we should expose a way for clients to
pass in not just text, but also image outputs. This is something the
Responses API already supports for function call outputs, where you can
pass in either a string or an array of content outputs (text, image,
file):
https://platform.openai.com/docs/api-reference/responses/create#responses_create-input-input_item_list-item-function_tool_call_output-output-array-input_image
So let's just plumb it through in Codex (with the caveat that we only
support text and image for now). This is implemented end-to-end across
app-server v2 protocol types and core tool handling.
## Breaking API change
NOTE: This introduces a breaking change with dynamic tools, but I think
it's ok since this concept was only recently introduced
(https://github.com/openai/codex/pull/9539) and it's better to get the
API contract correct. I don't think there are any real consumers of this
yet (not even the Codex App).
Old shape:
`{ "output": "dynamic-ok", "success": true }`
New shape:
```
{
"contentItems": [
{ "type": "inputText", "text": "dynamic-ok" },
{ "type": "inputImage", "imageUrl": "data:image/png;base64,AAA" }
]
"success": true
}
```
- Inline response recording during streaming: `run_turn` now records
items as they arrive instead of building a `ProcessedResponseItem` list
and post‑processing via `process_items`.
- Simplify turn handling: `handle_output_item_done` returns the
follow‑up signal + optional tool future; `needs_follow_up` is set only
there, and in‑flight tool futures are drained once at the end (errors
logged, no extra state writes).
- Flattened stream loop: removed `process_items` indirection and the
extra output queue
- - Tests: relaxed `tool_parallelism::tool_results_grouped` to allow any
completion order while still requiring matching call/output IDs.
It's pretty amazing we have gotten here without the ability for the
model to see image content from MCP tool calls.
This PR builds off of 4391 and fixes#4819. I would like @KKcorps to get
adequete credit here but I also want to get this fix in ASAP so I gave
him a week to update it and haven't gotten a response so I'm going to
take it across the finish line.
This test highlights how absured the current situation is. I asked the
model to read this image using the Chrome MCP
<img width="2378" height="674" alt="image"
src="https://github.com/user-attachments/assets/9ef52608-72a2-4423-9f5e-7ae36b2b56e0"
/>
After this change, it correctly outputs:
> Captured the page: image dhows a dark terminal-style UI labeled
`OpenAI Codex (v0.0.0)` with prompt `model: gpt-5-codex medium` and
working directory `/codex/codex-rs`
(and more)
Before this change, it said:
> Took the full-page screenshot you asked for. It shows a long,
horizontally repeating pattern of stylized people in orange, light-blue,
and mustard clothing, holding hands in alternating poses against a white
background. No text or other graphics-just rows of flat illustration
stretching off to the right.
Without this change, the Figma, Playwright, Chrome, and other visual MCP
servers are pretty much entirely useless.
I tested this change with the openai respones api as well as a third
party completions api
Currently we collect all all turn items in a vector, then we add it to
the history on success. This result in losing those items on errors
including aborting `ctrl+c`.
This PR:
- Adds the ability for the tool call to handle cancellation
- bubble the turn items up to where we are recording this info
Admittedly, this logic is an ad-hoc logic that doesn't handle a lot of
error edge cases. The right thing to do is recording to the history on
the spot as `items`/`tool calls output` come. However, this isn't
possible because of having different `task_kind` that has different
`conversation_histories`. The `try_run_turn` has no idea what thread are
we using. We cannot also pass an `arc` to the `conversation_histories`
because it's a private element of `state`.
That's said, `abort` is the most common case and we should cover it
until we remove `task kind`
Today `sub_id` is an ID of a single incoming Codex Op submition. We then
associate all events triggered by this operation using the same
`sub_id`.
At the same time we are also creating a TurnContext per submission and
we'd like to start associating some events (item added/item completed)
with an entire turn instead of just the operation that started it.
Using turn context when sending events give us flexibility to change
notification scheme.
