- Make `ContextManager::for_prompt` modality-aware and strip input_image
content when the active model is text-only.
- Added a test for multi-model -> text-only model switch
## Summary
- Reduced repeated approvals for equivalent wrapper commands and fixed
execpolicy matching for heredoc-style shell invocations, with minimal
behavior change and fail-closed defaults.
## Fixes
1. Canonicalized approval matching for wrappers so equivalent commands
map to the same approval intent.
2. Added heredoc-aware prefix extraction for execpolicy so commands like
`python3 <<'PY' ... PY` match rules such as `prefix_rule(["python3"],
...)`.
3. Kept fallback behavior conservative: if parsing is ambiguous,
existing prompt behavior is preserved.
## Edge Cases Covered
- Wrapper path/name differences: `/bin/bash` vs `bash`, `/bin/zsh` vs
`zsh`.
- Shell modes: `-c` and `-lc`.
- Heredoc forms: quoted delimiter (`<<'PY'`) and unquoted delimiter (`<<
PY`).
- Multi-command heredoc scripts are rejected by the fallback
- Non-heredoc redirections (`>`, etc.) are not treated as heredoc prefix
matches.
- Complex scripts still fall back to prior behavior rather than
expanding permissions.
---------
Co-authored-by: Dylan Hurd <dylan.hurd@openai.com>
- 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.
- Keep `view_image` in the advertised tool list for all models.
- Return a clear error when the current model does not support image
inputs, and cover it with a unit test.
The `TODO` in `core/src/seatbelt.rs` claimed that `apply_patch` still needed to honor `SandboxPolicy`. That was true when the comment was added, but it is no longer true.
Analysis:
- The TODO was introduced in #1762, when seatbelt code was split out of `exec.rs`.
- `apply_patch` sandboxing was later implemented in #1705.
- Today, `apply_patch` calls are routed through the tool orchestrator and delegated to `ApplyPatchRuntime`, which executes via `execute_env()` using the active sandbox attempt policy.
- On macOS, the sandbox transform path for that execution still builds seatbelt args with `create_seatbelt_command_args(command, policy, sandbox_policy_cwd)`, so the same `SandboxPolicy` gates `apply_patch` writes and network behavior.
Because this behavior is already enforced, the TODO is stale and removing it avoids implying missing sandbox coverage where none exists.
No functional behavior change; comment-only cleanup.
## Summary
- keep wiremock MockServer handles alive through async assertions in
remote model suite tests
- assert /models request count in remote_models_hide_picker_only_models
- use a slightly higher parallel timing threshold on aarch64 while
keeping existing x86 threshold
## Validation
- just fmt
- targeted tests:
- cargo test -p codex-core --test all
suite::remote_models::remote_models_merge_replaces_overlapping_model --
--exact
- cargo test -p codex-core --test all
suite::remote_models::remote_models_hide_picker_only_models -- --exact
- cargo test -p codex-core --test all
suite::tool_parallelism::shell_tools_run_in_parallel -- --exact
- soak loop: 40 iterations of all three targeted tests
## Notes
- cargo test -p codex-core has one unrelated local-env failure in
shell_snapshot::tests::try_new_creates_and_deletes_snapshot_file from
exported certificate env content in this workspace.
- local bazel test //codex-rs/core:core-all-test failed to build due
missing rust-objcopy in this host toolchain.
During thread/fork, the new rollout includes the fork’s own session_meta
plus copied history that can contain older session_meta entries from the
source thread. thread/list was overwriting metadata on later
session_meta lines, so a fork could be reported with the source thread’s
thread_id. This fix only uses the first session_meta, so the fork keeps
its own ID.
This removes overly directed language about how the model should behave
when it's in `approval_policy=never` mode.
---------
Co-authored-by: Dylan Hurd <dylan.hurd@openai.com>
## Summary
- remove redundant user message wait that could time out and cause
flakiness
- rely on the existing turn-complete wait to ensure the follow-up
request is observed
## Testing
- Not run (not requested)
Summary
- move `core/src/hooks` implementation into a new `codex-hooks` crate
with its own manifest
- update `codex-rs` workspace and `codex-core` crate to depend on the
extracted `hooks` crate and wire up the shared APIs
- ensure references, modules, and lockfile reflect the new crate layout
Testing
- Not run (not requested)
## Align with the new phase-1 design
Basically we know run phase 1 in parallel by considering:
* Max 64 rollouts
* Max 1 month old
* Consider the most recent first
This PR also adds stronger parallelization capabilities by detecting
stale jobs, retry policies, ownership of computation to prevent double
computations etc etc
As of this PR, `SessionServices` retains a
`Option<StartedNetworkProxy>`, if appropriate.
Now the `network` field on `Config` is `Option<NetworkProxySpec>`
instead of `Option<NetworkProxy>`.
Over in `Session::new()`, we invoke `NetworkProxySpec::start_proxy()` to
create the `StartedNetworkProxy`, which is a new struct that retains the
`NetworkProxy` as well as the `NetworkProxyHandle`. (Note that `Drop` is
implemented for `NetworkProxyHandle` to ensure the proxies are shutdown
when it is dropped.)
The `NetworkProxy` from the `StartedNetworkProxy` is threaded through to
the appropriate places.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/11207).
* #11285
* __->__ #11207
The dynamic model refresh feature (`https://api.openai.com/v1/models`
endpoint) is currently gated on a runtime check for an auth method other
than API Key. It should be gated on a check specifically for ChatGPT
Auth because some custom model providers (e.g. for local models) use no
auth mechanism. A call to `self.auth_manager.auth_mode()` will return
`None` in this case.
Addresses #11213
…ount_id and chatgpt_plan_type
### Summary
Following up on external auth mode which was introduced here:
https://github.com/openai/codex/pull/10012
Turns out some clients have a differently shaped ID token and don't have
a chosen workspace (aka chatgpt_account_id) encoded in their ID token.
So, let's replace `id_token` param with `chatgpt_account_id` and
`chatgpt_plan_type` (optional) when initializing the external ChatGPT
auth mode (`account/login/start` with `chatgptAuthTokens`).
The client was able to test end-to-end with a Codex build from this
branch and verified it worked!
Summary
- add platform-aware defaults for shell command timeouts so Windows
tests get longer waits
- keep medium timeout longer on Windows to ensure flakiness is reduced
Testing
- Not run (not requested)
# External (non-OpenAI) Pull Request Requirements
Before opening this Pull Request, please read the dedicated
"Contributing" markdown file or your PR may be closed:
https://github.com/openai/codex/blob/main/docs/contributing.md
If your PR conforms to our contribution guidelines, replace this text
with a detailed and high quality description of your changes.
Include a link to a bug report or enhancement request.
Match model metadata by longest matching remote slug prefix before local
fallback.
- Update `get_model_info` to prefer the most specific remote slug prefix
for the requested model.
- Add an integration test to assert `gpt-5.3-codex-test` resolves to
`gpt-5.3-codex` over `gpt-5.3`.
**Why We Did This**
- The goal is to reduce MCP tool context pollution by not exposing the
full MCP tool list up front
- It forces an explicit discovery step (`search_tool_bm25`) so the model
narrows tool scope before making MCP calls, which helps relevance and
lowers prompt/tool clutter.
**What It Changed**
- Added a new experimental feature flag `search_tool` in
`core/src/features.rs:90` and `core/src/features.rs:430`.
- Added config/schema support for that flag in
`core/config.schema.json:214` and `core/config.schema.json:1235`.
- Added BM25 dependency (`bm25`) in `Cargo.toml:129` and
`core/Cargo.toml:23`.
- Added new tool handler `search_tool_bm25` in
`core/src/tools/handlers/search_tool_bm25.rs:18`.
- Registered the handler and tool spec in
`core/src/tools/handlers/mod.rs:11` and `core/src/tools/spec.rs:780` and
`core/src/tools/spec.rs:1344`.
- Extended `ToolsConfig` to carry `search_tool` enablement in
`core/src/tools/spec.rs:32` and `core/src/tools/spec.rs:56`.
- Injected dedicated developer instructions for tool-discovery workflow
in `core/src/codex.rs:483` and `core/src/codex.rs:1976`, using
`core/templates/search_tool/developer_instructions.md:1`.
- Added session state to store one-shot selected MCP tools in
`core/src/state/session.rs:27` and `core/src/state/session.rs:131`.
- Added filtering so when feature is enabled, only selected MCP tools
are exposed on the next request (then consumed) in
`core/src/codex.rs:3800` and `core/src/codex.rs:3843`.
- Added E2E suite coverage for
enablement/instructions/hide-until-search/one-turn-selection in
`core/tests/suite/search_tool.rs:72`,
`core/tests/suite/search_tool.rs:109`,
`core/tests/suite/search_tool.rs:147`, and
`core/tests/suite/search_tool.rs:218`.
- Refactored test helper utilities to support config-driven tool
collection in `core/tests/suite/tools.rs:281`.
**Net Behavioral Effect**
- With `search_tool` **off**: existing MCP behavior (tools exposed
normally).
- With `search_tool` **on**: MCP tools start hidden, model must call
`search_tool_bm25`, and only returned `selected_tools` are available for
the next model call.
## Summary
- add targeted remote-compaction failure diagnostics in compact_remote
logging
- log the specific values needed to explain overflow timing:
- last_api_response_total_tokens
- estimated_tokens_of_items_added_since_last_successful_api_response
- estimated_bytes_of_items_added_since_last_successful_api_response
- failing_compaction_request_body_bytes
- simplify breakdown naming and remove
last_api_response_total_bytes_estimate (it was an approximation and not
useful for debugging)
## Why
When compaction fails with context_length_exceeded, we need concrete,
low-ambiguity numbers that map directly to:
1) what the API most recently reported, and
2) what local history added since then.
This keeps the failure logs actionable without adding broad, noisy
metrics.
## Testing
- just fmt
- cargo test -p codex-core
Instead of storing a special connection on the client level make the
regular task responsible for establishing a normal client session and
open a connection on it.
Then when the turn is started we pass in a pre-established session.
On some platforms, the "notify" file watcher library emits events for
file opens and reads, not just file modifications or deletes. The
previous implementation didn't take this into account.
Furthermore, the `tracing.info!` call that I previously added was
emitting a lot of logs. I had assumed incorrectly that `info` level
logging was disabled by default, but it's apparently enabled for this
crate. This is resulting in large logs (hundreds of MB) for some users.
