Commit Graph
31 Commits
Author SHA1 Message Date
Matthew Zeng 4c7a1d800f update 2026-03-10 00:13:15 -07:00
Charley CunninghamandCodex f23fcd6ced guardian initial feedback / tweaks (#13897)
## Summary
- remove the remaining model-visible guardian-specific `on-request`
prompt additions so enabling the feature does not change the main
approval-policy instructions
- neutralize user-facing guardian wording to talk about automatic
approval review / approval requests rather than a second reviewer or
only sandbox escalations
- tighten guardian retry-context handling so agent-authored
`justification` stays in the structured action JSON and is not also
injected as raw retry context
- simplify guardian review plumbing in core by deleting dead
prompt-append paths and trimming some request/transcript setup code

## Notable Changes
- delete the dead `permissions/approval_policy/guardian.md` append path
and stop threading `guardian_approval_enabled` through model-facing
developer-instruction builders
- rename the experimental feature copy to `Automatic approval review`
and update the `/experimental` snapshot text accordingly
- make approval-review status strings generic across shell, patch,
network, and MCP review types
- forward real sandbox/network retry reasons for shell and unified-exec
guardian review, but do not pass agent-authored justification as raw
retry context
- simplify `guardian.rs` by removing the one-field request wrapper,
deduping reasoning-effort selection, and cleaning up transcript entry
collection

## Testing
- `just fmt`
- full validation left to CI

---------

Co-authored-by: Codex <noreply@openai.com>
2026-03-09 09:25:24 -07:00
Matthew Zeng a4a9536fd7 [elicitations] Support always allow option for mcp tool calls. (#13807)
- [x] Support always allow option for mcp tool calls, writes to
config.toml.
- [x] Fix config hot-reload after starting a new thread for TUI.
2026-03-08 01:46:40 +00:00
Charley CunninghamandCodex e84ee33cc0 Add guardian approval MVP (#13692)
## Summary
- add the guardian reviewer flow for `on-request` approvals in command,
patch, sandbox-retry, and managed-network approval paths
- keep guardian behind `features.guardian_approval` instead of exposing
a public `approval_policy = guardian` mode
- route ordinary `OnRequest` approvals to the guardian subagent when the
feature is enabled, without changing the public approval-mode surface

## Public model
- public approval modes stay unchanged
- guardian is enabled via `features.guardian_approval`
- when that feature is on, `approval_policy = on-request` keeps the same
approval boundaries but sends those approval requests to the guardian
reviewer instead of the user
- `/experimental` only persists the feature flag; it does not rewrite
`approval_policy`
- CLI and app-server no longer expose a separate `guardian` approval
mode in this PR

## Guardian reviewer
- the reviewer runs as a normal subagent and reuses the existing
subagent/thread machinery
- it is locked to a read-only sandbox and `approval_policy = never`
- it does not inherit user/project exec-policy rules
- it prefers `gpt-5.4` when the current provider exposes it, otherwise
falls back to the parent turn's active model
- it fail-closes on timeout, startup failure, malformed output, or any
other review error
- it currently auto-approves only when `risk_score < 80`

## Review context and policy
- guardian mirrors `OnRequest` approval semantics rather than
introducing a separate approval policy
- explicit `require_escalated` requests follow the same approval surface
as `OnRequest`; the difference is only who reviews them
- managed-network allowlist misses that enter the approval flow are also
reviewed by guardian
- the review prompt includes bounded recent transcript history plus
recent tool call/result evidence
- transcript entries and planned-action strings are truncated with
explicit `<guardian_truncated ... />` markers so large payloads stay
bounded
- apply-patch reviews include the full patch content (without
duplicating the structured `changes` payload)
- the guardian request layout is snapshot-tested using the same
model-visible Responses request formatter used elsewhere in core

## Guardian network behavior
- the guardian subagent inherits the parent session's managed-network
allowlist when one exists, so it can use the same approved network
surface while reviewing
- exact session-scoped network approvals are copied into the guardian
session with protocol/port scope preserved
- those copied approvals are now seeded before the guardian's first turn
is submitted, so inherited approvals are available during any immediate
review-time checks

## Out of scope / follow-ups
- the sandbox-permission validation split was pulled into a separate PR
and is not part of this diff
- a future follow-up can enable `serde_json` preserve-order in
`codex-core` and then simplify the guardian action rendering further

---------

Co-authored-by: Codex <noreply@openai.com>
2026-03-07 05:40:10 -08:00
Owen Lin 289ed549cf chore(otel): rename OtelManager to SessionTelemetry (#13808)
## Summary
This is a purely mechanical refactor of `OtelManager` ->
`SessionTelemetry` to better convey what the struct is doing. No
behavior change.

## Why

`OtelManager` ended up sounding much broader than what this type
actually does. It doesn't manage OTEL globally; it's the session-scoped
telemetry surface for emitting log/trace events and recording metrics
with consistent session metadata (`app_version`, `model`, `slug`,
`originator`, etc.).

`SessionTelemetry` is a more accurate name, and updating the call sites
makes that boundary a lot easier to follow.

## Validation

- `just fmt`
- `cargo test -p codex-otel`
- `cargo test -p codex-core`
2026-03-06 16:23:30 -08:00
Matthew Zeng 98dca99db7 [elicitations] Switch to use MCP style elicitation payload for mcp tool approvals. (#13621)
- [x] Switch to use MCP style elicitation payload for mcp tool
approvals.
- [ ] TODO: Update the UI to support the full spec.
2026-03-06 01:50:26 -08:00
jif-oai b649953845 feat: polluted memories (#13008)
Add a feature flag to disable memory creation for "polluted"
2026-03-02 11:57:32 +00:00
alexsong-oai 09a82f364f Support implicit skill invocation analytics events (#12049)
- use `skills_for_cwd` lookup to scope allowed skills and build
invocation context for downstream processing
- add detection in `stream_events_utils` to classify tool calls as
implicit skill invocations per the proposal (script runners, extensions,
`scripts` dirs, and SKILL.md reads)
- deduplicate invocations per turn and emit analytics/OTEL events on the
same background queue as explicit invokes
2026-02-23 21:55:49 -08:00
Michael Bolin 66d5d34e6e core: preserve constrained approval/sandbox policies in TurnContext (#12473) 2026-02-21 14:40:24 -08:00
colby-oai d3cf8bd0fa fix(core): require approval for destructive MCP tool calls (#12353)
Summary
- ensure destructive tool annotations short-circuit to require approval
- simplify approval logic to only require read/write + open-world when
destructive is false
- update the unit test to cover the new destructive behavior

Testing
- Not run (not requested)
2026-02-20 12:12:16 -08:00
Matthew Zeng aa121a115e [apps] Implement apps configs. (#12086)
- [x] Implement apps configs.
2026-02-20 12:05:21 -08:00
colby-oai f6fd4cb3f5 Adjust MCP tool approval handling for custom servers (#11787)
Summary
This PR expands MCP client-side approval behavior beyond codex_apps and
tightens elicitation capability signaling.

- Removed the codex_apps-only gate in MCP tool approval checks, so
local/custom MCP servers are now eligible for the same client-side
approval prompt flow when tool annotations indicate side effects.
- Updated approval memory keying to support tools without a connector ID
(connector_id: Option<String>), allowing “Approve this Session” to be
remembered even when connector metadata is missing.
- Updated prompt text for non-codex_apps tools to identify origin as The
<server> MCP server instead of This app.
- Added MCP initialization capability policy so only codex_apps
advertises MCP elicitation capability; other servers advertise no
elicitation support.
- Added regression tests for:
server-specific prompt copy behavior
codex-apps-only elicitation capability advertisement

Testing
- Not run (not requested)
2026-02-19 12:52:42 -05:00
alexsong-oai e71760fc64 support app usage analytics (#11687)
Emit app mentioned and app used events. Dedup by (turn_id, connector_id)

Example event params:
{
    "event_type": "codex_app_used",
    "connector_id": "asdk_app_xxx",
    "thread_id": "019c5527-36d4-xxx",
    "turn_id": "019c552c-cd17-xxx",
    "app_name": "Slack (OpenAI Internal)",
    "product_client_id": "codex_cli_rs",
    "invoke_type": "explicit",
    "model_slug": "gpt-5.3-codex"
}
2026-02-13 12:00:16 -08:00
Ahmed Ibrahim 9c4656000f Sanitize MCP image output for text-only models (#11346)
- Replace image blocks in MCP tool results with a text placeholder when
the active model does not accept image input.
- Add an e2e rmcp test to verify sanitized tool output is what gets sent
back to the model.
2026-02-10 11:25:32 -08:00
Owen Lin 5ea107a088 feat(app-server, core): allow text + image content items for dynamic tool outputs (#10567)
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
}
```
2026-02-04 16:12:47 -08:00
canvrno-oai 282f42c0ce Add option to approve and remember MCP/Apps tool usage (#10584)
This PR adds a new approval option for app/MCP tool calls: “Allow and
remember” (session-scoped).
When selected, Codex stores a temporary approval and auto-approves
matching future calls for the rest of the session.

Added a session-scoped approval key (`server`, `connector_id`,
`tool_name`) and persisted it in `tool_approvals` as
`ApprovedForSession`.
On subsequent matching calls, approval is skipped and treated as
accepted.
- Updated the approval question options to conditionally include:
- Accept
- Allow and remember (conditional)
- Decline
- Cancel

The new “Allow and remember” option is only shown when all of these are
true:

1. The call is routed through the Codex Apps MCP server (codex_apps).
2. The tool requires approval based on annotations:
- read_only_hint == false, and
- destructive_hint == true or open_world_hint == true.
3. The tool includes a connector_id in metadata (used to build the
remembered approval key).

If no `connector_id` is present, the prompt still appears (when approval
is required), but only with the existing choices (Accept / Decline /
Cancel). Approval prompting in this path has an explicit early return
unless server == `codex_apps`.
2026-02-04 09:38:41 -08:00
pakrym-oai 7f20357611 Stop client from being state carrier (#10595)
I'd like to make client session wide. This requires shedding all random
state it has to carry.
2026-02-04 09:05:37 -08:00
Michael Bolin 66447d5d2c feat: replace custom mcp-types crate with equivalents from rmcp (#10349)
We started working with MCP in Codex before
https://crates.io/crates/rmcp was mature, so we had our own crate for
MCP types that was generated from the MCP schema:


https://github.com/openai/codex/blob/8b95d3e082376f4cb23e92641705a22afb28a9da/codex-rs/mcp-types/README.md

Now that `rmcp` is more mature, it makes more sense to use their MCP
types in Rust, as they handle details (like the `_meta` field) that our
custom version ignored. Though one advantage that our custom types had
is that our generated types implemented `JsonSchema` and `ts_rs::TS`,
whereas the types in `rmcp` do not. As such, part of the work of this PR
is leveraging the adapters between `rmcp` types and the serializable
types that are API for us (app server and MCP) introduced in #10356.

Note this PR results in a number of changes to
`codex-rs/app-server-protocol/schema`, which merit special attention
during review. We must ensure that these changes are still
backwards-compatible, which is possible because we have:

```diff
- export type CallToolResult = { content: Array<ContentBlock>, isError?: boolean, structuredContent?: JsonValue, };
+ export type CallToolResult = { content: Array<JsonValue>, structuredContent?: JsonValue, isError?: boolean, _meta?: JsonValue, };
```

so `ContentBlock` has been replaced with the more general `JsonValue`.
Note that `ContentBlock` was defined as:

```typescript
export type ContentBlock = TextContent | ImageContent | AudioContent | ResourceLink | EmbeddedResource;
```

so the deletion of those individual variants should not be a cause of
great concern.

Similarly, we have the following change in
`codex-rs/app-server-protocol/schema/typescript/Tool.ts`:

```
- export type Tool = { annotations?: ToolAnnotations, description?: string, inputSchema: ToolInputSchema, name: string, outputSchema?: ToolOutputSchema, title?: string, };
+ export type Tool = { name: string, title?: string, description?: string, inputSchema: JsonValue, outputSchema?: JsonValue, annotations?: JsonValue, icons?: Array<JsonValue>, _meta?: JsonValue, };
```

so:

- `annotations?: ToolAnnotations` ➡️ `JsonValue`
- `inputSchema: ToolInputSchema` ➡️ `JsonValue`
- `outputSchema?: ToolOutputSchema` ➡️ `JsonValue`

and two new fields: `icons?: Array<JsonValue>, _meta?: JsonValue`

---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/10349).
* #10357
* __->__ #10349
* #10356
2026-02-02 17:41:55 -08:00
Matthew Zeng 34f89b12d0 MCP tool call approval (simplified version) (#10200)
Add elicitation approval request for MCP tool call requests.
2026-01-29 23:40:32 -08:00
jif-oai 225614d7fb chore: add mcp call metric (#8973) 2026-01-09 13:09:21 +00:00
Gabriel Peal b0bdc04c30 [MCP] Render MCP tool call result images to the model (#5600)
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
2025-10-27 17:55:57 -04:00
pakrym-oai 789e65b9d2 Pass TurnContext around instead of sub_id (#5421)
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.
2025-10-21 08:04:16 -07:00
Gabriel Peal 6b0c486861 [MCP] Render full MCP errors to the model (#5298)
Previously, the model couldn't see why MCP tool calls failed, many of
which were the model using the parameters incorrectly. A common failure
is the model stringifying the json for the notion-update-page tool which
it then couldn't correct.

I want to do some system prompt massaging around this as well. However,
it is crucial that the model sees the error so it can fix it.

Before:
<img width="2984" height="832" alt="CleanShot 2025-10-17 at 13 02 36"
src="https://github.com/user-attachments/assets/709a3d27-b71b-4d8d-87b6-9b2d7fe4e6f2"
/>

After:
<img width="2488" height="1550" alt="CleanShot 2025-10-17 at 13 01 18"
src="https://github.com/user-attachments/assets/13a0b7dc-fdad-4996-bf2d-0772872c34fc"
/>

🎉 
<img width="1078" height="568" alt="CleanShot 2025-10-17 at 13 09 30"
src="https://github.com/user-attachments/assets/64cde8be-9e6c-4e61-b971-c2ba22504292"
/>


Fixes #4707
2025-10-17 17:47:50 -04:00
Jeremy Rose 19f46439ae timeouts for mcp tool calls (#3959)
defaults to 60sec, overridable with MCP_TOOL_TIMEOUT or on a per-server
basis in the config.
2025-09-22 10:30:59 -07:00
Ahmed Ibrahim 097782c775 Move models.rs to protocol (#2595)
Moving models.rs to protocol so we can use them in `Codex` operations
2025-08-22 22:18:54 +00:00
Jeremy Rose 347c81ad00 remove conversation history widget (#1727)
this widget is no longer used.
2025-07-30 10:05:40 -07:00
Michael Bolin a768a6a41d fix: introduce ResponseInputItem::McpToolCallOutput variant (#1151)
The output of an MCP server tool call can be one of several types, but
to date, we treated all outputs as text by showing the serialized JSON
as the "tool output" in Codex:


https://github.com/openai/codex/blob/25a9949c49194d5a64de54a11bcc5b4724ac9bd5/codex-rs/mcp-types/src/lib.rs#L96-L101

This PR adds support for the `ImageContent` variant so we can now
display an image output from an MCP tool call.

In making this change, we introduce a new
`ResponseInputItem::McpToolCallOutput` variant so that we can work with
the `mcp_types::CallToolResult` directly when the function call is made
to an MCP server.

Though arguably the more significant change is the introduction of
`HistoryCell::CompletedMcpToolCallWithImageOutput`, which is a cell that
uses `ratatui_image` to render an image into the terminal. To support
this, we introduce `ImageRenderCache`, cache a
`ratatui_image::picker::Picker`, and `ensure_image_cache()` to cache the
appropriate scaled image data and dimensions based on the current
terminal size.

To test, I created a minimal `package.json`:

```json
{
  "name": "kitty-mcp",
  "version": "1.0.0",
  "type": "module",
  "description": "MCP that returns image of kitty",
  "main": "index.js",
  "dependencies": {
    "@modelcontextprotocol/sdk": "^1.12.0"
  }
}
```

with the following `index.js` to define the MCP server:

```js
#!/usr/bin/env node

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { readFile } from "node:fs/promises";
import { join } from "node:path";

const IMAGE_URI = "image://Ada.png";

const server = new McpServer({
  name: "Demo",
  version: "1.0.0",
});

server.tool(
  "get-cat-image",
  "If you need a cat image, this tool will provide one.",
  async () => ({
    content: [
      { type: "image", data: await getAdaPngBase64(), mimeType: "image/png" },
    ],
  })
);

server.resource("Ada the Cat", IMAGE_URI, async (uri) => {
  const base64Image = await getAdaPngBase64();
  return {
    contents: [
      {
        uri: uri.href,
        mimeType: "image/png",
        blob: base64Image,
      },
    ],
  };
});

async function getAdaPngBase64() {
  const __dirname = new URL(".", import.meta.url).pathname;
  // From https://github.com/benjajaja/ratatui-image/blob/9705ce2c59ec669abbce2924cbfd1f5ae22c9860/assets/Ada.png
  const filePath = join(__dirname, "Ada.png");
  const imageData = await readFile(filePath);
  const base64Image = imageData.toString("base64");
  return base64Image;
}

const transport = new StdioServerTransport();
await server.connect(transport);
```

With the local changes from this PR, I added the following to my
`config.toml`:

```toml
[mcp_servers.kitty]
command = "node"
args = ["/Users/mbolin/code/kitty-mcp/index.js"]
```

Running the TUI from source:

```
cargo run --bin codex -- --model o3 'I need a picture of a cat'
```

I get:

<img width="732" alt="image"
src="https://github.com/user-attachments/assets/bf80b721-9ca0-4d81-aec7-77d6899e2869"
/>

Now, that said, I have only tested in iTerm and there is definitely some
funny business with getting an accurate character-to-pixel ratio
(sometimes the `CompletedMcpToolCallWithImageOutput` thinks it needs 10
rows to render instead of 4), so there is still work to be done here.
2025-05-28 19:03:17 -07:00
Michael Bolin a5f3a34827 fix: change EventMsg enum so every variant takes a single struct (#925)
https://github.com/openai/codex/pull/922 did this for the
`SessionConfigured` enum variant, and I think it is generally helpful to
be able to work with the values as each enum variant as their own type,
so this converts the remaining variants and updates all of the
callsites.

Added a simple unit test to verify that the JSON-serialized version of
`Event` does not have any unexpected nesting.
2025-05-13 20:44:42 -07:00
Michael Bolin 9da6ebef3f fix: add optional timeout to McpClient::send_request() (#852)
We now impose a 10s timeout on the initial `tools/list` request to an
MCP server. We do not apply a timeout for other types of requests yet,
but we should start enforcing those, as well.
2025-05-07 12:56:38 -07:00
Michael Bolin aa36a15f9f fix: make all fields of Session struct private again (#840)
https://github.com/openai/codex/pull/829 noted it introduced a circular
dep between `codex.rs` and `mcp_tool_call.rs`. This attempts to clean
things up: the circular dep still exists, but at least all the fields of
`Session` are private again.
2025-05-06 16:21:35 -07:00
Michael Bolin 147a940449 feat: support mcp_servers in config.toml (#829)
This adds initial support for MCP servers in the style of Claude Desktop
and Cursor. Note this PR is the bare minimum to get things working end
to end: all configured MCP servers are launched every time Codex is run,
there is no recovery for MCP servers that crash, etc.

(Also, I took some shortcuts to change some fields of `Session` to be
`pub(crate)`, which also means there are circular deps between
`codex.rs` and `mcp_tool_call.rs`, but I will clean that up in a
subsequent PR.)

`codex-rs/README.md` is updated as part of this PR to explain how to use
this feature. There is a bit of plumbing to route the new settings from
`Config` to the business logic in `codex.rs`. The most significant
chunks for new code are in `mcp_connection_manager.rs` (which defines
the `McpConnectionManager` struct) and `mcp_tool_call.rs`, which is
responsible for tool calls.

This PR also introduces new `McpToolCallBegin` and `McpToolCallEnd`
event types to the protocol, but does not add any handlers for them.
(See https://github.com/openai/codex/pull/836 for initial usage.)

To test, I added the following to my `~/.codex/config.toml`:

```toml
# Local build of https://github.com/hideya/mcp-server-weather-js
[mcp_servers.weather]
command = "/Users/mbolin/code/mcp-server-weather-js/dist/index.js"
args = []
```

And then I ran the following:

```
codex-rs$ cargo run --bin codex exec 'what is the weather in san francisco'
[2025-05-06T22:40:05] Task started: 1
[2025-05-06T22:40:18] Agent message: Here’s the latest National Weather Service forecast for San Francisco (downtown, near 37.77° N, 122.42° W):

This Afternoon (Tue):
• Sunny, high near 69 °F
• West-southwest wind around 12 mph

Tonight:
• Partly cloudy, low around 52 °F
• SW wind 7–10 mph
...
```

Note that Codex itself is not able to make network calls, so it would
not normally be able to get live weather information like this. However,
the weather MCP is [currently] not run under the Codex sandbox, so it is
able to hit `api.weather.gov` and fetch current weather information.

---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/829).
* #836
* __->__ #829
2025-05-06 15:47:59 -07:00