## Why
`codex-tools` already owns the shared tool input schema model and parser
from the first extraction step, but `core/src/tools/spec.rs` still owned
the MCP-specific adapter that normalizes `rmcp::model::Tool` schemas and
wraps `structuredContent` into the call result output schema.
Keeping that adapter in `codex-core` means the reusable MCP schema path
is still split across crates, and the unit tests for that logic stay
anchored in `codex-core` even though the runtime orchestration does not
need to move yet.
This change takes the next small step by moving the reusable MCP schema
adapter into `codex-tools` while leaving `ResponsesApiTool` assembly in
`codex-core`.
## What changed
- added `tools/src/mcp_tool.rs` and sibling
`tools/src/mcp_tool_tests.rs`
- introduced `ParsedMcpTool`, `parse_mcp_tool()`, and
`mcp_call_tool_result_output_schema()` in `codex-tools`
- updated `core/src/tools/spec.rs` to consume parsed MCP tool parts from
`codex-tools`
- removed the now-redundant MCP schema unit tests from
`core/src/tools/spec_tests.rs`
- expanded `codex-rs/tools/README.md` to describe this second migration
step
## Test plan
- `cargo test -p codex-tools`
- `cargo test -p codex-core --lib tools::spec::`
## Summary
This PR makes Windows sandbox proxying enforceable by routing proxy-only
runs through the existing `offline` sandbox user and reserving direct
network access for the existing `online` sandbox user.
In brief:
- if a Windows sandbox run should be proxy-enforced, we run it as the
`offline` user
- the `offline` user gets firewall rules that block direct outbound
traffic and only permit the configured localhost proxy path
- if a Windows sandbox run should have true direct network access, we
run it as the `online` user
- no new sandbox identity is introduced
This brings Windows in line with the intended model: proxy use is not
just env-based, it is backed by OS-level egress controls. Windows
already has two sandbox identities:
- `offline`: intended to have no direct network egress
- `online`: intended to have full network access
This PR makes proxy-enforced runs use that model directly.
### Proxy-enforced runs
When proxy enforcement is active:
- the run is assigned to the `offline` identity
- setup extracts the loopback proxy ports from the sandbox env
- Windows setup programs firewall rules for the `offline` user that:
- block all non-loopback outbound traffic
- block loopback UDP
- block loopback TCP except for the configured proxy ports
- optionally allow broader localhost access when `allow_local_binding=1`
So the sandboxed process can only talk to the local proxy. It cannot
open direct outbound sockets or do local UDP-based DNS on its own.The
proxy then performs the real outbound network access outside that
restricted sandbox identity.
### Direct-network runs
When proxy enforcement is not active and full network access is allowed:
- the run is assigned to the `online` identity
- no proxy-only firewall restrictions are applied
- the process gets normal direct network access
### Unelevated vs elevated
The restricted-token / unelevated path cannot enforce per-identity
firewall policy by itself.
So for Windows proxy-enforced runs, we transparently use the logon-user
sandbox path under the hood, even if the caller started from the
unelevated mode. That keeps enforcement real instead of best-effort.
---------
Co-authored-by: Codex <noreply@openai.com>
## Why
`PermissionProfile` should only describe the per-command permissions we
still want to grant dynamically. Keeping
`MacOsSeatbeltProfileExtensions` in that surface forced extra macOS-only
approval, protocol, schema, and TUI branches for a capability we no
longer want to expose.
## What changed
- Removed the macOS-specific permission-profile types from
`codex-protocol`, the app-server v2 API, and the generated
schema/TypeScript artifacts.
- Deleted the core and sandboxing plumbing that threaded
`MacOsSeatbeltProfileExtensions` through execution requests and seatbelt
construction.
- Simplified macOS seatbelt generation so it always includes the fixed
read-only preferences allowlist instead of carrying a configurable
profile extension.
- Removed the macOS additional-permissions UI/docs/test coverage and
deleted the obsolete macOS permission modules.
- Tightened `request_permissions` intersection handling so explicitly
empty requested read lists are preserved only when that field was
actually granted, avoiding zero-grant responses being stored as active
permissions.
## Why
`parse_tool_input_schema` and the supporting `JsonSchema` model were
living in `core/src/tools/spec.rs`, but they already serve callers
outside `codex-core`.
Keeping that shared schema parsing logic inside `codex-core` makes the
crate boundary harder to reason about and works against the guidance in
`AGENTS.md` to avoid growing `codex-core` when reusable code can live
elsewhere.
This change takes the first extraction step by moving the schema parsing
primitive into its own crate while keeping the rest of the tool-spec
assembly in `codex-core`.
## What changed
- added a new `codex-tools` crate under `codex-rs/tools`
- moved the shared tool input schema model and sanitizer/parser into
`tools/src/json_schema.rs`
- kept `tools/src/lib.rs` exports-only, with the module-level unit tests
split into `json_schema_tests.rs`
- updated `codex-core` to use `codex-tools::JsonSchema` and re-export
`parse_tool_input_schema`
- updated `codex-app-server` dynamic tool validation to depend on
`codex-tools` directly instead of reaching through `codex-core`
- wired the new crate into the Cargo workspace and Bazel build graph
Replaced ci.bazelrc and v8-ci.bazelrc by custom configs inside the main
.bazelrc file. As a result, github workflows setup is simplified down to
a single '--config=<foo>' flag usage.
Moved the build metadata flags to config=ci.
Added custom tags metadata to help differentiate invocations based on
workflow (bazel vs v8) and os (linux/macos/windows).
Enabled users to override the default values in .bazelrc by using a
user.bazelrc file locally.
Added user.bazelrc to gitignore.
Highlights:
- Trimmed down to just the repository cache for faster upload / download
- Made the cache key only include files that affect external
dependencies (since that's what the repository cache caches) -
MODULE.bazel, codex-rs/Cargo.lock, codex-rs/Cargo.toml
- Split the caching action in to explicit restore / save steps (similar
to your rust CI) which allows us to skip uploads on cache hit, and not
fail the build if upload fails
This should get rid of 842 network fetches that are happening on every
Bazel CI run, while also reducing the Github flakiness @bolinfest
reported. Uploading should be faster (since we're not caching many small
files), and will only happen when MODULE.bazel or Cargo.lock /
Cargo.toml files change.
In my testing, it [took 3s to save the repository
cache](https://github.com/siggisim/codex/actions/runs/23014186143/job/66832859781).
## Summary
Fail closed when the network proxy's local/private IP pre-check hits a
DNS lookup error or timeout, instead of treating the hostname as public
and allowing the request.
## Root cause
`host_resolves_to_non_public_ip()` returned `false` on resolver failure,
which created a fail-open path in the `allow_local_binding = false`
boundary. The eventual connect path performs its own DNS resolution
later, so a transient pre-check failure is not evidence that the
destination is public.
## Changes
- Treat DNS lookup errors/timeouts as local/private for blocking
purposes
- Add a regression test for an allowlisted hostname that fails DNS
resolution
## Validation
- `cargo test -p codex-network-proxy`
- `cargo clippy -p codex-network-proxy --all-targets -- -D warnings`
- `just fmt`
- `just argument-comment-lint`
## Why
This is effectively a follow-up to
[#15812](https://github.com/openai/codex/pull/15812). That change
removed the special skill-script exec path, but `skill_metadata` was
still being threaded through command-approval payloads even though the
approval flow no longer uses it to render prompts or resolve decisions.
Keeping it around added extra protocol, schema, and client surface area
without changing behavior.
Removing it keeps the command-approval contract smaller and avoids
carrying a dead field through app-server, TUI, and MCP boundaries.
## What changed
- removed `ExecApprovalRequestSkillMetadata` and the corresponding
`skillMetadata` field from core approval events and the v2 app-server
protocol
- removed the generated JSON and TypeScript schema output for that field
- updated app-server, MCP server, TUI, and TUI app-server approval
plumbing to stop forwarding the field
- cleaned up tests that previously constructed or asserted
`skillMetadata`
## Testing
- `cargo test -p codex-app-server-protocol`
- `cargo test -p codex-protocol`
- `cargo test -p codex-app-server-test-client`
- `cargo test -p codex-mcp-server`
- `just argument-comment-lint`
## Summary
- move the bwrap PATH lookup and warning helpers out of config/mod.rs
- move the related tests into a dedicated bwrap_tests.rs file
## Validation
- git diff --check
- skipped heavier local tests per request
Follow-up to #15791.
## Why
`codex-core` is already the largest crate in `codex-rs`, so defaulting
to it for new functionality makes it harder to keep the workspace
modular. The repo guidance should make it explicit that contributors are
expected to look for an existing non-`codex-core` crate, or introduce a
new crate, before growing `codex-core` further.
## What Changed
- Added a dedicated `The \`codex-core\` crate` section to `AGENTS.md`.
- Documented why `codex-core` should be treated as a last resort for new
functionality.
- Added concrete guidance for both implementation and review: prefer an
existing non-`codex-core` crate when possible, introduce a new workspace
crate when that is the cleaner boundary, and push back on PRs that grow
`codex-core` unnecessarily.
## Why
`SandboxCommand.program` represents an executable path, but keeping it
as `String` forced path-backed callers to run `to_string_lossy()` before
the sandbox layer ever touched the command. That loses fidelity earlier
than necessary and adds avoidable conversions in runtimes that already
have a `PathBuf`.
## What changed
- Changed `SandboxCommand.program` to `OsString`.
- Updated `SandboxManager::transform` to keep the program and argv in
`OsString` form until the `SandboxExecRequest` conversion boundary.
- Switched the path-backed `apply_patch` and `js_repl` runtimes to pass
`into_os_string()` instead of `to_string_lossy()`.
- Updated the remaining string-backed builders and tests to match the
new type while preserving the existing Linux helper `arg0` behavior.
## Verification
- `cargo test -p codex-sandboxing`
- `just argument-comment-lint -p codex-core -p codex-sandboxing`
- `cargo test -p codex-core` currently fails in unrelated existing
config tests: `config::tests::approvals_reviewer_*` and
`config::tests::smart_approvals_alias_*`
## Why
`token_data` is owned by `codex-login`, but `codex-core` was still
re-exporting it. That let callers pull auth token types through
`codex-core`, which keeps otherwise unrelated crates coupled to
`codex-core` and makes `codex-core` more of a build-graph bottleneck.
## What changed
- remove the `codex-core` re-export of `codex_login::token_data`
- update the remaining `codex-core` internals that used
`crate::token_data` to import `codex_login::token_data` directly
- update downstream callers in `codex-rs/chatgpt`,
`codex-rs/tui_app_server`, `codex-rs/app-server/tests/common`, and
`codex-rs/core/tests` to import `codex_login::token_data` directly
- add explicit `codex-login` workspace dependencies and refresh lock
metadata for crates that now depend on it directly
## Validation
- `cargo test -p codex-chatgpt --locked`
- `just argument-comment-lint`
- `just bazel-lock-update`
- `just bazel-lock-check`
## Notes
- attempted `cargo test -p codex-core --locked` and `cargo test -p
codex-core auth_refresh --locked`, but both ran out of disk while
linking `codex-core` test binaries in the local environment
## Problem
Codex already treated an existing top-level project `./.codex` directory
as protected, but there was a gap on first creation.
If `./.codex` did not exist yet, a turn could create files under it,
such as `./.codex/config.toml`, without going through the same approval
path as later modifications. That meant the initial write could bypass
the intended protection for project-local Codex state.
## What this changes
This PR closes that first-creation gap in the Unix enforcement layers:
- `codex-protocol`
- treat the top-level project `./.codex` path as a protected carveout
even when it does not exist yet
- avoid injecting the default carveout when the user already has an
explicit rule for that exact path
- macOS Seatbelt
- deny writes to both the exact protected path and anything beneath it,
so creating `./.codex` itself is blocked in addition to writes inside it
- Linux bubblewrap
- preserve the same protected-path behavior for first-time creation
under `./.codex`
- tests
- add protocol regressions for missing `./.codex` and explicit-rule
collisions
- add Unix sandbox coverage for blocking first-time `./.codex` creation
- tighten Seatbelt policy assertions around excluded subpaths
## Scope
This change is intentionally scoped to protecting the top-level project
`.codex` subtree from agent writes.
It does not make `.codex` unreadable, and it does not change the product
behavior around loading project skills from `.codex` when project config
is untrusted.
## Why this shape
The fix is pointed rather than broad:
- it preserves the current model of “project `.codex` is protected from
writes”
- it closes the security-relevant first-write hole
- it avoids folding a larger permissions-model redesign into this PR
## Validation
- `cargo test -p codex-protocol`
- `cargo test -p codex-sandboxing seatbelt`
- `cargo test -p codex-exec --test all
sandbox_blocks_first_time_dot_codex_creation -- --nocapture`
---------
Co-authored-by: Michael Bolin <mbolin@openai.com>
## Why
Skill metadata accepted a `permissions` block and stored the result on
`SkillMetadata`, but that data was never consumed by runtime behavior.
Leaving the dead parsing path in place makes it look like skills can
widen or otherwise influence execution permissions when, in practice,
declared skill permissions are ignored.
This change removes that misleading surface area so the skill metadata
model matches what the system actually uses.
## What changed
- removed `permission_profile` and `managed_network_override` from
`core-skills::SkillMetadata`
- stopped parsing `permissions` from skill metadata in
`core-skills/src/loader.rs`
- deleted the loader tests that only exercised the removed permissions
parsing path
- cleaned up dependent `SkillMetadata` constructors in tests and TUI
code that were only carrying `None` for those fields
## Testing
- `cargo test -p codex-core-skills`
- `cargo test -p codex-tui
submission_prefers_selected_duplicate_skill_path`
- `just argument-comment-lint`
## Summary
- resolve system bwrap from PATH instead of hardcoding /usr/bin/bwrap
- skip PATH entries that resolve inside the current workspace before
launching the sandbox helper
- keep the vendored bubblewrap fallback when no trusted system bwrap is
found
## Validation
- cargo test -p codex-core bwrap --lib
- cargo test -p codex-linux-sandbox
- just fix -p codex-core
- just fix -p codex-linux-sandbox
- just fmt
- just argument-comment-lint
- cargo clean
- [x] Polish tool suggest prompts to distinguish between missing
connectors and discoverable plugins, and be very precise about the
triggering conditions.
## TR;DR
Replicates the `/title` command from `tui` to `tui_app_server`.
## Problem
The classic `tui` crate supports customizing the terminal window/tab
title via `/title`, but the `tui_app_server` crate does not. Users on
the app-server path have no way to configure what their terminal title
shows (project name, status, spinner, thread, etc.), making it harder to
identify Codex sessions across tabs or windows.
## Mental model
The terminal title is a *status surface* -- conceptually parallel to the
footer status line. Both surfaces are configurable lists of items, both
share expensive inputs (git branch lookup, project root discovery), and
both must be refreshed at the same lifecycle points. This change ports
the classic `tui`'s design verbatim:
1. **`terminal_title.rs`** owns the low-level OSC write path and input
sanitization. It strips control characters and bidi/invisible codepoints
before placing untrusted text (model output, thread names, project
paths) inside an escape sequence.
2. **`title_setup.rs`** defines `TerminalTitleItem` (the 8 configurable
items) and `TerminalTitleSetupView` (the interactive picker that wraps
`MultiSelectPicker`).
3. **`status_surfaces.rs`** is the shared refresh pipeline. It parses
both surface configs once per refresh, warns about invalid items once
per session, synchronizes the git-branch cache, then renders each
surface from the same `StatusSurfaceSelections` snapshot.
4. **`chatwidget.rs`** sets `TerminalTitleStatusKind` at each state
transition (Working, Thinking, Undoing, WaitingForBackgroundTerminal)
and calls `refresh_terminal_title()` whenever relevant state changes.
5. **`app.rs`** handles the three setup events (confirm/preview/cancel),
persists config via `ConfigEditsBuilder`, and clears the managed title
on `Drop`.
## Non-goals
- **Restoring the previous terminal title on exit.** There is no
portable way to read the terminal's current title, so `Drop` clears the
managed title rather than restoring it.
- **Sharing code between `tui` and `tui_app_server`.** The
implementation is a parallel copy, matching the existing pattern for the
status-line feature. Extracting a shared crate is future work.
## Tradeoffs
- **Duplicate code across crates.** The three core files
(`terminal_title.rs`, `title_setup.rs`, `status_surfaces.rs`) are
byte-for-byte copies from the classic `tui`. This was chosen for
consistency with the existing status-line port and to avoid coupling the
two crates at the dependency level. Future changes must be applied in
both places.
- **`status_surfaces.rs` is large (~660 lines).** It absorbs logic that
previously lived inline in `chatwidget.rs` (status-line refresh, git
branch management, project root discovery) plus all new terminal-title
logic. This consolidation trades file size for a single place where both
surfaces are coordinated.
- **Spinner scheduling on every refresh.** The terminal title spinner
(when active) schedules a frame every 100ms. This is the same pattern
the status-indicator spinner already uses; the overhead is a timer
registration, not a redraw.
## Architecture
```
/title command
-> SlashCommand::Title
-> open_terminal_title_setup()
-> TerminalTitleSetupView (MultiSelectPicker)
-> on_change: AppEvent::TerminalTitleSetupPreview -> preview_terminal_title()
-> on_confirm: AppEvent::TerminalTitleSetup -> ConfigEditsBuilder + setup_terminal_title()
-> on_cancel: AppEvent::TerminalTitleSetupCancelled -> cancel_terminal_title_setup()
Runtime title refresh:
state change (turn start, reasoning, undo, plan update, thread rename, ...)
-> set terminal_title_status_kind
-> refresh_terminal_title()
-> status_surface_selections() (parse configs, collect invalids)
-> refresh_terminal_title_from_selections()
-> terminal_title_value_for_item() for each configured item
-> assemble title string with separators
-> skip if identical to last_terminal_title (dedup OSC writes)
-> set_terminal_title() (sanitize + OSC 0 write)
-> schedule spinner frame if animating
Widget replacement:
replace_chat_widget_with_app_server_thread()
-> transfer last_terminal_title from old widget to new
-> avoids redundant OSC clear+rewrite on session switch
```
## Observability
- Invalid terminal-title item IDs in config emit a one-per-session
warning via `on_warning()` (gated by
`terminal_title_invalid_items_warned` `AtomicBool`).
- OSC write failures are logged at `tracing::debug` level.
- Config persistence failures are logged at `tracing::error` and
surfaced to the user via `add_error_message()`.
## Tests
- `terminal_title.rs`: 4 unit tests covering sanitization (control
chars, bidi codepoints, truncation) and OSC output format.
- `title_setup.rs`: 3 tests covering setup view snapshot rendering,
parse order preservation, and invalid-ID rejection.
- `chatwidget/tests.rs`: Updated test helpers with new fields; existing
tests continue to pass.
---------
Co-authored-by: Eric Traut <etraut@openai.com>
## Summary
Add a focused codex network proxy unit test for the denylist pattern
with wildcard in the middle `region*.some.malicious.tunnel.com`. This
does not change how existing code works, just ensure that behavior stays
the same and we got CI guards to guard existin behavior.
## Why
The managed Codex denylist update relies on this mid label glob form,
and the existing tests only covered exact hosts, `*.` subdomains, and
`**.` apex plus subdomains.
## Validation
`cargo test -p codex-network-proxy
compile_globset_supports_mid_label_wildcards`
`cargo test -p codex-network-proxy`
`./tools/argument-comment-lint/run-prebuilt-linter.sh -p
codex-network-proxy`
## Summary
- block git global options that can redirect config, repository, or
helper lookup from being auto-approved as safe
- share the unsafe global-option predicate across the Unix and Windows
git safety checks
- add regression coverage for inline and split forms, including `bash
-lc` and PowerShell wrappers
## Root cause
The Unix safe-command gate only rejected `-c` and `--config-env`, even
though the shared git parser already knew how to skip additional
pre-subcommand globals such as `--git-dir`, `--work-tree`,
`--exec-path`, `--namespace`, and `--super-prefix`. That let those
arguments slip through safe-command classification on otherwise
read-only git invocations and bypass approval. The Windows-specific
safe-command path had the same trust-boundary gap for git global
options.
## Why
`#[large_stack_test]` made the `apply_patch_cli` tests pass by giving
them more stack, but it did not address why those tests needed the extra
stack in the first place.
The real problem is the async state built by the `apply_patch_cli`
harness path. Those tests await three helper boundaries directly:
harness construction, turn submission, and apply-patch output
collection. If those helpers inline their full child futures, the test
future grows to include the whole harness startup and request/response
path.
This change replaces the workaround from #12768 with the same basic
approach used in #13429, but keeps the fix narrower: only the helper
boundaries awaited directly by `apply_patch_cli` stay boxed.
## What Changed
- removed `#[large_stack_test]` from
`core/tests/suite/apply_patch_cli.rs`
- restored ordinary `#[tokio::test(flavor = "multi_thread",
worker_threads = 2)]` annotations in that suite
- deleted the now-unused `codex-test-macros` crate and removed its
workspace wiring
- boxed only the three helper boundaries that the suite awaits directly:
- `apply_patch_harness_with(...)`
- `TestCodexHarness::submit(...)`
- `TestCodexHarness::apply_patch_output(...)`
- added comments at those boxed boundaries explaining why they remain
boxed
## Testing
- `cargo test -p codex-core --test all suite::apply_patch_cli --
--nocapture`
## References
- #12768
- #13429
## Summary
- enrich `codex.mcp.call` with `tool`, `connector_id`, and sanitized
`connector_name` for actual MCP executions
- record `codex.mcp.call.duration_ms` for actual MCP executions so
connector-level latency is visible in metrics
- keep skipped, blocked, declined, and cancelled paths on the plain
status-only `codex.mcp.call` counter
## Included Changes
- `codex-rs/core/src/mcp_tool_call.rs`: add connector-sliced MCP count
and duration metrics only for executed tool calls, while leaving
non-executed outcomes as status-only counts
- `codex-rs/core/src/mcp_tool_call_tests.rs`: cover metric tag shaping,
connector-name sanitization, and the new duration metric tags
## Testing
- `cargo test -p codex-core`
- `just fix -p codex-core`
- `just fmt`
## Notes
- `cargo test -p codex-core` still hits existing unrelated failures in
approvals-reviewer config tests and the sandboxed JS REPL `mktemp` test
- full workspace `cargo test` was not run
---------
Co-authored-by: Codex <noreply@openai.com>
## Symptoms
When `/review` ran through `tui_app_server`, the TUI could show
duplicate review content:
- the `>> Code review started: ... <<` banner appeared twice
- the final review body could also appear twice
## Problem
`tui_app_server` was treating review lifecycle items as renderable
content on more than one delivery path.
Specifically:
- `EnteredReviewMode` was rendered both when the item started and again
when it completed
- `ExitedReviewMode` rendered the review text itself, even though the
same review text was also delivered later as the assistant message item
That meant the same logical review event was committed into history
multiple times.
## Solution
Make review lifecycle items control state transitions only once, and
keep the final review body sourced from the assistant message item:
- render the review-start banner from the live `ItemStarted` path, while
still allowing replay to restore it once
- treat `ExitedReviewMode` as a mode-exit/finish-banner event instead of
rendering the review body from it
- preserve the existing assistant-message rendering path as the single
source of final review text
This PR partially rebase `unified_exec` on the `exec-server` and adapt
the `exec-server` accordingly.
## What changed in `exec-server`
1. Replaced the old "broadcast-driven; process-global" event model with
process-scoped session events. The goal is to be able to have dedicated
handler for each process.
2. Add to protocol contract to support explicit lifecycle status and
stream ordering:
- `WriteResponse` now returns `WriteStatus` (Accepted, UnknownProcess,
StdinClosed, Starting) instead of a bool.
- Added seq fields to output/exited notifications.
- Added terminal process/closed notification.
3. Demultiplexed remote notifications into per-process channels. Same as
for the event sys
4. Local and remote backends now both implement ExecBackend.
5. Local backend wraps internal process ID/operations into per-process
ExecProcess objects.
6. Remote backend registers a session channel before launch and
unregisters on failed launch.
## What changed in `unified_exec`
1. Added unified process-state model and backend-neutral process
wrapper. This will probably disappear in the future, but it makes it
easier to keep the work flowing on both side.
- `UnifiedExecProcess` now handles both local PTY sessions and remote
exec-server processes through a shared `ProcessHandle`.
- Added `ProcessState` to track has_exited, exit_code, and terminal
failure message consistently across backends.
2. Routed write and lifecycle handling through process-level methods.
## Some rationals
1. The change centralizes execution transport in exec-server while
preserving policy and orchestration ownership in core, avoiding
duplicated launch approval logic. This comes from internal discussion.
2. Session-scoped events remove coupling/cross-talk between processes
and make stream ordering and terminal state explicit (seq, closed,
failed).
3. The failure-path surfacing (remote launch failures, write failures,
transport disconnects) makes command tool output and cleanup behavior
deterministic
## Follow-ups:
* Unify the concept of thread ID behind an obfuscated struct
* FD handling
* Full zsh-fork compatibility
* Full network sandboxing compatibility
* Handle ws disconnection
Fixes#15283.
## Summary
Older system bubblewrap builds reject `--argv0`, which makes our Linux
sandbox fail before the helper can re-exec. This PR keeps using system
`/usr/bin/bwrap` whenever it exists and only falls back to vendored
bwrap when the system binary is missing. That matters on stricter
AppArmor hosts, where the distro bwrap package also provides the policy
setup needed for user namespaces.
For old system bwrap, we avoid `--argv0` instead of switching binaries:
- pass the sandbox helper a full-path `argv0`,
- keep the existing `current_exe() + --argv0` path when the selected
launcher supports it,
- otherwise omit `--argv0` and re-exec through the helper's own
`argv[0]` path, whose basename still dispatches as
`codex-linux-sandbox`.
Also updates the launcher/warning tests and docs so they match the new
behavior: present-but-old system bwrap uses the compatibility path, and
only absent system bwrap falls back to vendored.
### Validation
1. Install Ubuntu 20.04 in a VM
2. Compile codex and run without bubblewrap installed - see a warning
about falling back to the vendored bwrap
3. Install bwrap and verify version is 0.4.0 without `argv0` support
4. run codex and use apply_patch tool without errors
<img width="802" height="631" alt="Screenshot 2026-03-25 at 11 48 36 PM"
src="https://github.com/user-attachments/assets/77248a29-aa38-4d7c-9833-496ec6a458b8"
/>
<img width="807" height="634" alt="Screenshot 2026-03-25 at 11 47 32 PM"
src="https://github.com/user-attachments/assets/5af8b850-a466-489b-95a6-455b76b5050f"
/>
<img width="812" height="635" alt="Screenshot 2026-03-25 at 11 45 45 PM"
src="https://github.com/user-attachments/assets/438074f0-8435-4274-a667-332efdd5cb57"
/>
<img width="801" height="623" alt="Screenshot 2026-03-25 at 11 43 56 PM"
src="https://github.com/user-attachments/assets/0dc8d3f5-e8cf-4218-b4b4-a4f7d9bf02e3"
/>
---------
Co-authored-by: Michael Bolin <mbolin@openai.com>
For app-server websocket auth, support the two server-side mechanisms
from
PR #14847:
- `--ws-auth capability-token --ws-token-file /abs/path`
- `--ws-auth signed-bearer-token --ws-shared-secret-file /abs/path`
with optional `--ws-issuer`, `--ws-audience`, and
`--ws-max-clock-skew-seconds`
On the client side, add interactive remote support via:
- `--remote ws://host:port` or `--remote wss://host:port`
- `--remote-auth-token-env <ENV_VAR>`
Codex reads the bearer token from the named environment variable and
sends it
as `Authorization: Bearer <token>` during the websocket handshake.
Remote auth
tokens are only allowed for `wss://` URLs or loopback `ws://` URLs.
Testing:
- tested both auth methods manually to confirm connection success and
rejection for both auth types
When `tui_app_server` is enabled, shell commands in the transcript
render as fully quoted invocations like `/bin/zsh -lc "..."`. The
non-app-server TUI correctly shows the parsed command body.
Root cause:
The app-server stores `ThreadItem::CommandExecution.command` as a
shell-quoted string. When `tui_app_server` bridges that item back into
the exec renderer, it was passing `vec![command]` unchanged instead of
splitting the string back into argv. That prevented
`strip_bash_lc_and_escape()` from recognizing the shell wrapper, so the
renderer displayed the wrapper literally.
Solution:
Add a shared command-string splitter that round-trips shell-quoted
commands back into argv when it is safe to do so, while preserving
non-roundtrippable inputs as a single string. Use that helper everywhere
`tui_app_server` reconstructs exec commands from app-server payloads,
including live command-execution items, replayed thread items, and exec
approval requests. This restores the same command display behavior as
the direct TUI path without breaking Windows-style commands that cannot
be safely round-tripped.
CHAINED PR - note that base is eternal/hooks-pretooluse-bash, not main
-- so the following PR should be first
Matching post-tool hook to the pre-tool functionality here:
https://github.com/openai/codex/pull/15211
So, PreToolUse calls for plain shell calls, allows blocking. This
PostToolUse call runs after the command executed
example run:
```
› as a test, run in parallel the following commands:
- echo 'one'
- echo '[block-pre-tool-use]'
- echo '[block-post-tool-use]'
⚠ MCP startup incomplete (failed: notion, linear)
• Cruising through those three commands in parallel now, and I’ll share the exact outputs right after
they land.
• Running PreToolUse hook: checking the observatory runes
• Running PreToolUse hook: checking the observatory runes
• Running PreToolUse hook: checking the observatory runes
PreToolUse hook (blocked)
warning: wizard-tower PreToolUse demo blocked a Bash command on purpose.
feedback: PreToolUse demo blocked the command. Remove [block-pre-tool-use] to continue.
PreToolUse hook (completed)
warning: wizard-tower PreToolUse demo inspected Bash: echo 'one'
PreToolUse hook (completed)
warning: wizard-tower PreToolUse demo inspected Bash: echo '[block-post-tool-use]'
• Ran echo '[block-post-tool-use]'
└ [block-post-tool-use]
• Ran echo 'one'
└ one
• Running PostToolUse hook: reading the observatory aftermath
• Running PostToolUse hook: reading the observatory aftermath
PostToolUse hook (blocked)
warning: wizard-tower PostToolUse demo blocked the Bash follow-through on purpose.
feedback: PostToolUse demo blocked the result after execution. Remove [block-post-tool-use] to
continue.
PostToolUse hook (completed)
warning: wizard-tower PostToolUse demo inspected Bash: echo 'one'
hook context: PostToolUse demo saw the Bash result. Response preview: one
──────────────────────────────────────────────────────────────────────────────────────────────────────
• Ran them in parallel. Little wave report:
- echo 'one' completed and returned one.
- echo '[block-pre-tool-use]' was blocked before execution by the PreToolUse hook.
- echo '[block-post-tool-use]' ran, but its result was blocked afterward by the PostToolUse hook.
The hook messages were:
- PreToolUse demo blocked the command. Remove [block-pre-tool-use] to continue.
- PostToolUse demo blocked the result after execution. Remove [block-post-tool-use] to continue.
› for the blocked messages (correctly blocked in this test) did the tool output land back into
context?
• Yeah, from this run:
- PreToolUse block: the command itself did not execute, so no command output landed in context. Only
the block message did.
- PostToolUse block: the command did execute, but the actual stdout did not come back into context.
Only the post-tool block message did.
So the blocked notifications landed in context, but the blocked command payloads themselves did not.
Nice clean guardrail, aloha.
```
Add environment manager that is a singleton and is created early in
app-server (before skill manager, before config loading).
Use an environment variable to point to a running exec server.
