- reduce the server-side compaction test matrix to the highest-signal cases
- add comments around the deferred checkpoint rewrite and inline/preflight split
Co-authored-by: Codex <noreply@openai.com>
Ignore compact_prompt for OpenAI inline auto-compaction, remove the legacy compat downgrade path, and keep /compact on the point-in-time endpoint. Also skip previous-model preflight remote compaction when inline server-side compaction is available.\n\nCo-authored-by: Codex <noreply@openai.com>
Keep current-turn inputs in local inline compaction checkpoints and remember known backend incompatibilities after a compat downgrade so later turns skip the failed inline request path.
Co-authored-by: Codex <noreply@openai.com>
Move normal auto-compaction onto inline Responses API compaction behind a feature flag, keep the legacy path for manual and compatibility cases, and add observability plus integration coverage.
Co-authored-by: Codex <noreply@openai.com>
## What changed
- Drop failed websocket connections immediately after a terminal stream
error instead of awaiting a graceful close handshake before forwarding
the error to the caller.
- Keep the success path and the closed-connection guard behavior
unchanged.
## Why this fixes the flake
- The failing integration test waits for the second websocket stream to
surface the model error before issuing a follow-up request.
- On slower runners, the old error path awaited
`ws_stream.close().await` before sending the error downstream. If that
close handshake stalled, the test kept waiting for an error that had
already happened server-side and nextest timed it out.
- Dropping the failed websocket immediately makes the terminal error
observable right away and marks the session closed so the next request
reconnects cleanly instead of depending on a best-effort close
handshake.
## Code or test?
- This is a production logic fix in `codex-api`. The existing websocket
integration test already exercises the regression path.
Summary
- update the code-mode handler, runner, instructions, and error text to
refer to the `exec` tool name everywhere that used to say `code_mode`
- ensure generated documentation strings and tool specs describe `exec`
and rely on the shared `PUBLIC_TOOL_NAME`
- refresh the suite tests so they invoke `exec` instead of the old name
Testing
- Not run (not requested)
## Summary
- add `skill_approval` to `RejectConfig` and the app-server v2
`AskForApproval::Reject` payload so skill-script prompts can be
configured independently from sandbox and rule-based prompts
- update Unix shell escalation to reject prompts based on the actual
decision source, keeping prefix rules tied to `rules`, unmatched command
fallbacks tied to `sandbox_approval`, and skill scripts tied to
`skill_approval`
- regenerate the affected protocol/config schemas and expand
unit/integration coverage for the new flag and skill approval behavior
Pass more params to /compact. This should give us parity with the
/responses endpoint to improve caching.
I'm torn about the MCP await. Blocking will give us parity but it seems
like we explicitly don't block on MCPs. Happy either way
Summary
- document how code-mode can import `output_text`/`output_image` and
ensure `add_content` stays compatible
- add a synthetic `@openai/code_mode` module that appends content items
and validates inputs
- cover the new behavior with integration tests for structured text and
image outputs
Testing
- Not run (not requested)
Summary
- document that `@openai/code_mode` exposes
`set_max_output_tokens_per_exec_call` and that `code_mode` truncates the
final Rust-side output when the budget is exceeded
- enforce the configured budget in the Rust tool runner, reusing
truncation helpers so text-only outputs follow the unified-exec wrapper
and mixed outputs still fit within the limit
- ensure the new behavior is covered by a code-mode integration test and
string spec update
Testing
- Not run (not requested)
Summary
- drop `McpToolOutput` in favor of `CallToolResult`, moving its helpers
to keep MCP tooling focused on the final result shape
- wire the new schema definitions through code mode, context, handlers,
and spec modules so MCP tools serialize the exact output shape expected
by the model
- extend code mode tests to cover multiple MCP call scenarios and ensure
the serialized data matches the new schema
- refresh JS runner helpers and protocol models alongside the schema
changes
Testing
- Not run (not requested)
- add `model` and `reasoning_effort` to the `spawn_agent` schema so the
values pass through
- validate requested models against `model.model` and only check that
the selected model supports the requested reasoning effort
---------
Co-authored-by: Codex <noreply@openai.com>
Summary
- document output types for the various tool handlers and registry so
the API exposes richer descriptions
- update unified execution helpers and client tests to align with the
new output metadata
- clean up unused helpers across tool dispatch paths
Testing
- Not run (not requested)
Addresses #13586
This doesn't affect our CI scripts. It was user-reported.
Summary
- add `wiremock::ResponseTemplate` and `body_string_contains` imports
behind `#[cfg(not(debug_assertions))]` in
`codex-rs/core/tests/suite/view_image.rs` so release builds only pull
the helpers they actually use
## What changed
- The retry test now uses the same streaming SSE test server used by
production-style tests instead of a wiremock sequence.
- The fixture is resolved via `find_resource!`, and the test asserts
that exactly two outbound requests were sent.
## Why this fixes the flake
- The old wiremock sequence approximated early-close behavior, but it
did not reproduce the same streaming semantics the real client sees.
- That meant the retry path depended on mock implementation details
instead of on the actual transport behavior we care about.
- Switching to the streaming SSE helper makes the test exercise the real
early-close/retry contract, and counting requests directly verifies that
we retried exactly once rather than merely hoping the sequence aligned.
## Scope
- Test-only change.
- collect input/output transcript deltas into active handoff transcript
state
- attach and clear that transcript on each handoff, and regenerate
schema/tests
**Summary**
- allow `code_mode` to pass enabled tools metadata to the runner and
expose them via `tools.js`
- import tools inside JavaScript rather than relying only on globals or
proxies for nested tool calls
- update specs, docs, and tests to exercise the new bridge and explain
the tooling changes
**Testing**
- Not run (not requested)
## Summary
We need to support allowing request_permissions calls when using
`Reject` policy
<img width="1133" height="588" alt="Screenshot 2026-03-09 at 12 06
40 PM"
src="https://github.com/user-attachments/assets/a8df987f-c225-4866-b8ab-5590960daec5"
/>
Note that this is a backwards-incompatible change for Reject policy. I'm
not sure if we need to add a default based on our current use/setup
## Testing
- [x] Added tests
- [x] Tested locally
## Summary
The apply_patch tool should also respect AdditionalPermissions
## Testing
- [x] Added unit tests
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
request_permissions flows should support persisting results for the
session.
Open Question: Still deciding if we need within-turn approvals - this
adds complexity but I could see it being useful
## Testing
- [x] Updated unit tests
---------
Co-authored-by: Codex <noreply@openai.com>
## What changed
- The RMCP streamable HTTP tests now wait for metadata and tool
readiness before issuing tool calls.
- OAuth state is isolated per test home.
- The helper server startup path now uses bounded bind retries so
transient `AddrInUse` collisions do not fail the test immediately.
## Why this fixes the flake
- The old tests could begin issuing tool requests before the helper
server had finished advertising its metadata and tools, so the first
request sometimes raced the server startup sequence.
- On top of that, shared OAuth state and occasional bind collisions on
CI runners introduced cross-test environmental noise unrelated to the
functionality under test.
- Readiness polling makes the client wait for an observable “server is
ready” signal, while isolated state and bounded bind retries remove
external contention that was causing intermittent failures.
## Scope
- Test-only change.
## Summary
Alternative to #14061 - we need to use a child process on windows to
correctly validate Powershell behavior.
## Testing
- [x] These are tests
## What changed
- add a bounded `resume_until_initial_messages` helper in
`core/tests/suite/resume.rs`
- retry the resume call until `initial_messages` contains the fully
persisted final turn shape before asserting
## Why this fixes flakiness
The old test resumed once immediately after `TurnComplete` and sometimes
read rollout state before the final turn had been persisted. That made
the assertion race persistence timing instead of checking the resumed
message shape. The new helper polls for up to two seconds in 10ms steps
and only asserts once the expected message sequence is actually present,
so the test waits for the real readiness condition instead of depending
on a lucky timing window.
## Scope
- test-only
- no production logic change
## What changed
- The realtime startup-context tests no longer assume the interesting
websocket payload is always `connection 1 / request 0`.
- Instead, they now wait for the first outbound websocket request that
actually carries `session.instructions`, regardless of which websocket
connection won the accept-order race on the runner.
- The env-key fallback test stays serialized because it mutates process
environment.
## Why this fixes the flake
- The old test synchronized on the mirrored `session.updated` client
event and then inspected a fixed websocket slot.
- On CI, the response websocket and the realtime websocket can race each
other during startup. When the response websocket wins that race, the
fixed slot can contain `response.create` instead of the
startup-context-bearing `session.update` request the test actually cares
about.
- That made the test fail nondeterministically by inspecting the wrong
request, or by timing out waiting on a secondary event even though the
real outbound request path was correct.
- Waiting directly on the first request whose payload includes
`session.instructions` removes both ordering assumptions and makes the
assertion line up with the actual contract under test.
- Separately, serializing the environment-mutating fallback case
prevents unrelated tests from seeing partially updated auth state.
## Scope
- Test-only change.
## 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>
Adds a built-in `request_permissions` tool and wires it through the
Codex core, protocol, and app-server layers so a running turn can ask
the client for additional permissions instead of relying on a static
session policy.
The new flow emits a `RequestPermissions` event from core, tracks the
pending request by call ID, forwards it through app-server v2 as an
`item/permissions/requestApproval` request, and resumes the tool call
once the client returns an approved subset of the requested permission
profile.
## Summary
This test was missing the turn completion event in the responses stream,
so it was hanging. This PR fixes the issue
## Testing
- [x] This does update the test
## What changed
- The duration-recording fixture sleep was reduced from a large
artificial delay to `0.2s`, and the assertion floor was lowered to
`0.1s`.
- The shell tool fixtures now force `login = false` so they do not
invoke login-shell startup paths.
## Why this fixes the flake
- The old tests were paying for two kinds of noise that had nothing to
do with the feature being validated: oversized sleep time and variable
shell initialization cost.
- Login shells can pick up runner-specific startup files and incur
inconsistent startup latency.
- The test only needs to prove that we record a nontrivial duration and
preserve shell output. A shorter fixture delay plus a non-login shell
keeps that coverage while removing runner-dependent wall-clock variance.
## Scope
- Test-only change.
## Why
After `#13440` and `#13445`, macOS Seatbelt policy generation was still
deriving filesystem and network behavior from the legacy `SandboxPolicy`
projection.
That projection loses explicit unreadable carveouts and conflates split
network decisions, so the generated Seatbelt policy could still be wider
than the split policy that Codex had already computed.
## What changed
- added Seatbelt entrypoints that accept `FileSystemSandboxPolicy` and
`NetworkSandboxPolicy` directly
- built read and write policy stanzas from access roots plus excluded
subpaths so explicit unreadable carveouts survive into the generated
Seatbelt policy
- switched network policy generation to consult `NetworkSandboxPolicy`
directly
- failed closed when managed-network or proxy-constrained sessions do
not yield usable loopback proxy endpoints
- updated the macOS callers and test helpers that now need to carry the
split policies explicitly
## Verification
- added regression coverage in `core/src/seatbelt.rs` for unreadable
carveouts under both full-disk and scoped-readable policies
- verified the current PR state with `just clippy`
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/13448).
* #13453
* #13452
* #13451
* #13449
* __->__ #13448
* #13445
* #13440
* #13439
---------
Co-authored-by: viyatb-oai <viyatb@openai.com>
## 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>
