Adds the core runtime behavior for active goals on top of the model
tools from PR 3.
## Why
A long-running goal should be a core runtime concern, not something
every client has to implement. Core owns the turn lifecycle, tool
completion boundaries, interruptions, resume behavior, and token usage,
so it is the right place to account progress, enforce budgets, and
decide when to continue work.
## What changed
- Centralized goal lifecycle side effects behind
`Session::goal_runtime_apply(GoalRuntimeEvent::...)`.
- Starts goal continuation turns only when the session is idle; pending
user input and mailbox work take priority.
- Accounts token and wall-clock usage at turn, tool, mutation,
interrupt, and resume boundaries; `get_thread_goal` remains read-only.
- Preserves sub-second wall-clock remainder across accounting boundaries
so long-running goals do not drift downward over time.
- Treats token budget exhaustion as a soft stop by marking the goal
`budget_limited` and injecting wrap-up steering instead of aborting the
active turn.
- Suppresses budget steering when `update_goal` marks a goal complete.
- Pauses active goals on interrupt and auto-reactivates paused goals
when a thread resumes outside plan mode.
- Suppresses repeated automatic continuation when a continuation turn
makes no tool calls.
- Added continuation and budget-limit prompt templates.
## Verification
- Added focused core coverage for continuation scheduling, accounting
boundaries, budget-limit steering, completion accounting, interrupt
pause behavior, resume auto-activation, and wall-clock remainder
accounting.
Adds the model-facing goal tools on top of the app-server API from PR 2.
## Why
Once goals are persisted and exposed to clients, the model needs a
small, constrained tool surface for goal workflows. The tool contract
should let the model inspect goals, create them only when explicitly
requested, and mark them complete without giving it broad control over
user/runtime-owned state.
## What changed
- Added `get_goal`, `create_goal`, and `update_goal` tool specs behind
the `goals` feature flag.
- Added core goal tool handlers that validate objectives and token
budgets before mutating persisted state.
- Constrained `create_goal` to create only when no goal exists, with
optional `token_budget` only when a budget is explicitly provided.
- Tightened the `create_goal` instructions so the model does not infer
goals from ordinary task requests.
- Constrained `update_goal` to expose only goal completion; pause,
resume, clear, and budget-limited transitions remain user- or
runtime-controlled.
- Registered the goal tools in the tool registry and kept them out of
review contexts where they should not appear.
## Verification
- Added tool-registry coverage for feature gating and tool availability.
- Added core session tests for create/get/update behavior, duplicate
goal rejection, budget validation, and completion-only updates.
Adds the app-server v2 goal API on top of the persisted goal state from
PR 1.
## Why
Clients need a stable app-server surface for reading and controlling
materialized thread goals before the model tools and TUI can use them.
Goal changes also need to be observable by app-server clients, including
clients that resume an existing thread.
## What changed
- Added v2 `thread/goal/get`, `thread/goal/set`, and `thread/goal/clear`
RPCs for materialized threads.
- Added `thread/goal/updated` and `thread/goal/cleared` notifications so
clients can keep local goal state in sync.
- Added resume/snapshot wiring so reconnecting clients see the current
goal state for a thread.
- Added app-server handlers that reconcile persisted rollout state
before direct goal mutations.
- Updated the app-server README plus generated JSON and TypeScript
schema fixtures for the new API surface.
## Verification
- Added app-server v2 coverage for goal get/set/clear behavior,
notification emission, resume snapshots, and non-local thread-store
interactions.
Adds the persisted goal foundation for the rest of the stack. This PR is
intentionally limited to feature flag and state-layer behavior;
app-server APIs, model tools, runtime continuation, and TUI UX are
layered in later PRs.
## Why
Goal mode needs durable thread-level state before clients or model tools
can safely build on it. The state layer needs to know whether a goal
exists, what objective it tracks, whether it is active, paused,
budget-limited, or complete, and how much time/token usage has already
been accounted.
## What changed
- Added the `goals` feature flag and generated config schema entry.
- Added the `thread_goals` state table and Rust model for persisted
thread goals.
- Added state runtime APIs for creating, replacing, updating, deleting,
and accounting goal usage.
- Added `goal_id`-based stale update protection so an old goal update
cannot overwrite a replacement.
- Kept this PR scoped to persistence and state runtime behavior, with no
app-server, model-facing, continuation, or TUI behavior yet.
## Verification
- Added state runtime coverage for goal creation, replacement, stale
update protection, status transitions, token-budget behavior, and usage
accounting.
## Summary
Fix a Bazel-only path resolution bug in
`codex_utils_cargo_bin::cargo_bin`.
Under Bazel runfiles, `rlocation` can return a relative `bazel-out/...`
path even though `cargo_bin()` documents that it returns an absolute
path. That can break callers that store the returned binary path and
later spawn it after changing cwd, because the relative path is resolved
from the wrong directory.
This patch absolutizes the runfiles-resolved path before returning it.
## Why
`ReadOnlyAccess` was a transitional legacy shape on `SandboxPolicy`:
`FullAccess` meant the historical read-only/workspace-write modes could
read the full filesystem, while `Restricted` tried to carry partial
readable roots. The partial-read model now belongs in
`FileSystemSandboxPolicy` and `PermissionProfile`, so keeping it on
`SandboxPolicy` makes every legacy projection reintroduce lossy
read-root bookkeeping and creates unnecessary noise in the rest of the
permissions migration.
This PR makes the legacy policy model narrower and explicit:
`SandboxPolicy::ReadOnly` and `SandboxPolicy::WorkspaceWrite` represent
the old full-read sandbox modes only. Split readable roots, deny-read
globs, and platform-default/minimal read behavior stay in the runtime
permissions model.
## What changed
- Removes `ReadOnlyAccess` from
`codex_protocol::protocol::SandboxPolicy`, including the generated
`access` and `readOnlyAccess` API fields.
- Updates legacy policy/profile conversions so restricted filesystem
reads are represented only by `FileSystemSandboxPolicy` /
`PermissionProfile` entries.
- Keeps app-server v2 compatible with legacy `fullAccess` read-access
payloads by accepting and ignoring that no-op shape, while rejecting
legacy `restricted` read-access payloads instead of silently widening
them to full-read legacy policies.
- Carries Windows sandbox platform-default read behavior with an
explicit override flag instead of depending on
`ReadOnlyAccess::Restricted`.
- Refreshes generated app-server schema/types and updates tests/docs for
the simplified legacy policy shape.
## Verification
- `cargo check -p codex-app-server-protocol --tests`
- `cargo check -p codex-windows-sandbox --tests`
- `cargo test -p codex-app-server-protocol sandbox_policy_`
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/19449).
* #19395
* #19394
* #19393
* #19392
* #19391
* __->__ #19449
## Why
When using the Amazon Bedrock provider with `openai.gpt-5.4-cmb`, the
model picker allowed `xhigh` because the CMB catalog entry was derived
from the bundled `gpt-5.4` reasoning metadata. Bedrock rejects that
effort level, causing the request to fail before the turn can run:
```text
{"error":{"code":"validation_error","message":"Failed to deserialize the JSON body into the target type: Invalid 'reasoning': Invalid 'effort': unknown variant `xhigh`, expected one of `high`, `low`, `medium`, `minimal` at line 1 column 77239","param":null,"type":"invalid_request_error"}}
```
## What Changed
- Replace the runtime lookup of bundled `gpt-5.4` metadata for
`openai.gpt-5.4-cmb` with an explicit Bedrock CMB `ModelInfo` entry.
- Advertise only the Bedrock-supported CMB reasoning levels: `minimal`,
`low`, `medium`, and `high`.
- Keep the existing GPT OSS Bedrock model metadata and reasoning levels
unchanged.
- Add catalog coverage for the hardcoded CMB metadata and
Bedrock-compatible reasoning level list.
## Why
This prepares feedback log capture for a future remote app-server hook
sink without changing the current local SQLite upload path. The
important boundary is now intentionally small: a log sink is a tracing
`Layer` that can also flush entries it has accepted.
That keeps the existing SQLite implementation simple while giving the
upcoming gRPC sink a place to fit beside it. SQLite and gRPC have
different worker/write semantics, so this PR avoids introducing a shared
buffered-sink abstraction and instead lets each `LogWriter` own the
buffering mechanics it needs.
## What Changed
- Added `LogSinkQueueConfig` with the existing local defaults: queue
capacity `512`, batch size `128`, and flush interval `2s`.
- Added `LogDbLayer::start_with_config(...)` while preserving
`LogDbLayer::start(...)` and `log_db::start(...)` defaults.
- Introduced the `LogWriter` trait as the minimal shared interface:
`tracing_subscriber::Layer` plus `flush()`.
- Made `LogDbLayer` implement `LogWriter`.
- Kept tracing event formatting inside `LogDbLayer`; it still creates
one `LogEntry` per tracing event before queueing it for SQLite.
- Kept normal event capture best-effort and non-blocking via bounded
`try_send`.
## Behavior Notes
This does not change the SQLite schema, retention behavior,
`/feedback/upload`, or Sentry upload behavior. Normal log events still
drop when the queue is full; explicit `flush()` still waits for queue
capacity and receiver processing before returning.
## Verification
- `cargo test -p codex-state log_db`
- `cargo test -p codex-state`
- `just fix -p codex-state`
The added tests cover configured batch-size flushing, configured
interval flushing, queue-full drops, and the flush barrier semantics.
## Summary
- include the outer tool `call_id` in Codex Apps MCP request metadata
under `_meta._codex_apps.call_id`
- preserve existing Codex Apps metadata like `resource_uri` and
`contains_mcp_source`
- add request metadata coverage for both the existing-metadata and
no-existing-metadata cases
## Why
The paired backend change in
[openai/openai#850796](https://github.com/openai/openai/pull/850796)
updates MCP compliance logging to prefer `_meta._codex_apps.call_id`
instead of the JSON-RPC request id. This client change sends that outer
tool call id so the backend can record the model/tool call identifier
when it is available.
This is wire-compatible with older backends because `_meta._codex_apps`
is already reserved backend-only metadata. Backends that do not read
`call_id` will ignore the extra field.
## Testing
- `cargo test -p codex-core request_meta`
- `just fmt`
- `just fix -p codex-core`
- Add an integration test that guarantees nothing gets written to codex
home dir or sqlite when running a rollout with a non-local ThreadStore
- Add an in-memory "spy" ThreadStore for tests like this
Note I could not find a good way to also ensure there were no filesystem
_reads_ that didn't go through threadstore. I explored a more elaborate
sandboxed-subprocess approach but it isn't platform portable and felt
like it wasn't (yet) worth it.
## Summary
- Mirrors the OpenAI Docs skill cleanup in the bundled Codex skill copy
- Clarifies reasoning-effort recommendation wording
- Replaces internal snake_case prompt block names with natural-language
guidance aligned to the prompting guide
## Test plan
- `git diff --check`
- Verified the old snake_case prompt block names no longer appear in the
bundled upgrade guide
## Summary
- Mirrors openai/skills#374 in the Codex bundled OpenAI Docs skill
- Adds `gpt-image-2` as the best image generation/edit model
- Updates `gpt-image-1.5` to less expensive image generation/edit
quality
## Test plan
- `git diff --check`
- Route cold thread/resume and thread/fork source loading through
ThreadStore reads instead of direct rollout path operations
- Keep lookups that explicitly specify a rollout-path using the local
thread store methods but return an invalid-request error for remote
ThreadStore configurations
- Add some additional unit tests for code path coverage
## Why
The profile conversion path still required a `cwd` even when it was only
translating a legacy `SandboxPolicy` into a `PermissionProfile`. That
made profile producers invent an ambient `cwd`, which is exactly the
anchoring we are trying to remove from permission-profile data. A legacy
workspace-write policy can be represented symbolically instead: `:cwd =
write` plus read-only `:project_roots` metadata subpaths.
This PR creates that cwd-free base so the rest of the stack can stop
threading cwd through profile construction. Callers that actually need a
concrete runtime filesystem policy for a specific cwd still have an
explicitly named cwd-bound conversion.
## What Changed
- `PermissionProfile::from_legacy_sandbox_policy` now takes only
`&SandboxPolicy`.
- `FileSystemSandboxPolicy::from_legacy_sandbox_policy` is now the
symbolic, cwd-free projection for profiles.
- The old concrete projection is retained as
`FileSystemSandboxPolicy::from_legacy_sandbox_policy_for_cwd` for
runtime/boundary code that must materialize legacy cwd behavior.
- Workspace-write profiles preserve `CurrentWorkingDirectory` and
`ProjectRoots` special entries instead of materializing cwd into
absolute paths.
## Verification
- `cargo check -p codex-protocol -p codex-core -p
codex-app-server-protocol -p codex-app-server -p codex-exec -p
codex-exec-server -p codex-tui -p codex-sandboxing -p
codex-linux-sandbox -p codex-analytics --tests`
- `just fix -p codex-protocol -p codex-core -p codex-app-server-protocol
-p codex-app-server -p codex-exec -p codex-exec-server -p codex-tui -p
codex-sandboxing -p codex-linux-sandbox -p codex-analytics`
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/19414).
* #19395
* #19394
* #19393
* #19392
* #19391
* __->__ #19414
Selection menus in the TUI currently let disabled rows interfere with
numbering and default focus. This makes mixed menus harder to read and
can land selection on rows that are not actionable. This change updates
the shared selection-menu behavior in list_selection_view so disabled
rows are not selected when these views open, and prevents them from
being numbered like selectable rows.
- Disabled rows no longer receive numeric labels
- Digit shortcuts map to enabled rows only
- Default selection moves to the first enabled row in mixed menus
- Updated affected snapshot
- Added snapshot coverage for a plugin detail error popup
- Added a focused unit test for shared selection-view behavior
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
- Switch Unix socket app-server connections to perform the standard
WebSocket HTTP Upgrade handshake
- Update the Unix socket test to exercise a real upgrade over the Unix
stream
- Refresh the app-server README to describe the new Unix socket behavior
## Testing
- `cargo test -p codex-app-server transport::unix_socket_tests`
- `just fmt`
- `git diff --check`
## Why
`thread/fork` responses intentionally include copied history so the
caller can render the fork immediately, but `thread/started` is a
lifecycle notification. The v2 `Thread` contract says notifications
should return `turns: []`, and the fork path was reusing the response
thread directly, causing copied turns to be emitted through
`thread/started` as well.
## What Changed
- Route app-server `thread/started` notification construction through a
helper that clears `thread.turns` before sending.
- Keep `thread/fork` responses unchanged so callers still receive copied
history.
- Add persistent and ephemeral fork coverage that asserts
`thread/started` emits an empty `turns` array while the response retains
fork history.
## Testing
- `just fmt`
- `cargo test -p codex-app-server`
## Why
`openai.gpt-5.4-cmb` is served through the Amazon Bedrock provider,
whose request validator currently accepts `function` and `mcp` tool
specs but rejects Responses `custom` tools. The CMB catalog entry reuses
the bundled `gpt-5.4` metadata, which marks `apply_patch_tool_type` as
`freeform`. That causes Codex to include an `apply_patch` tool with
`type: "custom"`, so even heavily disabled sessions can fail before the
model runs with:
```text
Invalid tools: unknown variant `custom`, expected `function` or `mcp`
```
This is provider-specific: the model should still expose `apply_patch`,
but for Bedrock it needs to use the JSON/function tool shape instead of
the freeform/custom shape.
## What Changed
- Override the `openai.gpt-5.4-cmb` static catalog entry to set
`apply_patch_tool_type` to `function` after inheriting the rest of the
`gpt-5.4` model metadata.
- Update the catalog test expectation so the CMB entry continues to
track `gpt-5.4` metadata except for this Bedrock-specific tool shape
override.
## Verification
- `cargo test -p codex-model-provider`
- `just fix -p codex-model-provider`
## Summary
This PR hardens package-manager usage across the repo to reduce
dependency supply-chain risk. It also removes the stale `codex-cli`
Docker path, which was already broken on `main`, instead of keeping a
bitrotted container workflow alive.
## What changed
- Updated pnpm package manager pins and workspace install settings.
- Removed stale `codex-cli` Docker assets instead of trying to keep a
broken local container path alive.
- Added uv settings and lockfiles for the Python SDK packages.
- Updated Python SDK setup docs to use `uv sync`.
## Why
This is primarily a security hardening change. It reduces
package-install and supply-chain risk by ensuring dependency installs go
through pinned package managers, committed lockfiles, release-age
settings, and reviewed build-script controls.
For `codex-cli`, the right follow-up was to remove the local Docker path
rather than keep patching it:
- `codex-cli/Dockerfile` installed `codex.tgz` with `npm install -g`,
which bypassed the repo lockfile and age-gated pnpm settings.
- The local `codex-cli/scripts/build_container.sh` helper was already
broken on `main`: it called `pnpm run build`, but
`codex-cli/package.json` does not define a `build` script.
- The container path itself had bitrotted enough that keeping it would
require extra packaging-specific behavior that was not otherwise needed
by the repo.
## Gaps addressed
- Global npm installs bypassed the repo lockfile in Docker and CLI
reinstall paths, including `codex-cli/Dockerfile` and
`codex-cli/bin/codex.js`.
- CI and Docker pnpm installs used `--frozen-lockfile`, but the repo was
missing stricter pnpm workspace settings for dependency build scripts.
- Python SDK projects had `pyproject.toml` metadata but no committed
`uv.lock` coverage or uv age/index settings in `sdk/python` and
`sdk/python-runtime`.
- The secure devcontainer install path used npm/global install behavior
without a local locked package-manager boundary.
- The local `codex-cli` Docker helper was already broken on `main`, so
this PR removes that stale Docker path instead of preserving a broken
surface.
- pnpm was already pinned, but not to the current repo-wide pnpm version
target.
## Verification
- `pnpm install --frozen-lockfile`
- `.devcontainer/codex-install`: `pnpm install --prod --frozen-lockfile`
- `.devcontainer/codex-install`: `./node_modules/.bin/codex --version`
- `sdk/python`: `uv lock --check`, `uv sync --locked --all-extras
--dry-run`, `uv build`
- `sdk/python-runtime`: `uv lock --check`, `uv sync --locked --dry-run`,
`uv build --wheel`
- `pnpm -r --filter ./sdk/typescript run build`
- `pnpm -r --filter ./sdk/typescript run lint`
- `pnpm -r --filter ./sdk/typescript run test`
- `node --check codex-cli/bin/codex.js`
- `docker build -f .devcontainer/Dockerfile.secure -t codex-secure-test
.`
- `cargo build -p codex-cli`
- repo-wide package-manager audit
## Summary
Updates the bundled OpenAI Docs system skill for GPT-5.5.
## Changes
- Updates the bundled latest-model fallback
- Replaces bundled upgrade guidance with GPT-5.5 migration guidance
- Replaces bundled prompting guidance with GPT-5.5 prompting guidance
## Test plan
- Ran `node scripts/resolve-latest-model-info.js`
- Verified bundled files match the OpenAI Docs skill fallback content
## Why
The elevated Windows command runner currently trusts the first process
that connects to its parent-created named pipes. Tightening the pipe ACL
already narrows who can reach that boundary, but verifying the connected
client PID gives the parent one more fail-closed check: it only accepts
the exact runner process it just spawned.
## What changed
- validate `GetNamedPipeClientProcessId` after `ConnectNamedPipe` and
reject clients whose PID does not match the spawned runner
- also did some code de-duplication to route the one-shot elevated
capture flow in `windows-sandbox-rs/src/elevated_impl.rs` through
`spawn_runner_transport()` so both elevated codepaths use the same pipe
bootstrap and PID validation
Using the transport unification here also reduces duplication in the
elevated Windows IPC bootstrap, so future hardening to the runner
handshake only needs to land in one place.
## Validation
- `cargo test -p codex-windows-sandbox`
- manual testing: one-shot elevated path via `target/debug/codex.exe
exec` running a randomized shell command and confirming captured output
- manual testing: elevated session path via `target/debug/codex.exe -c
'windows.sandbox="elevated"' sandbox windows -- python -u -c ...` with
stdin/stdout round-trips (`READY`, then `GOT:...` for two input lines)
---------
Co-authored-by: viyatb-oai <viyatb@openai.com>
Fix a bug where the `turn/interrupt` RPC hangs when interrupting a turn
that has already completed.
Before this change, `turn/interrupt` requests were queued in app-server
and only answered when a later TurnAborted event arrived. If the target
turn was already complete, core treated Op::Interrupt as a no-op, so no
abort event was emitted and the RPC could hang indefinitely.
This change fixes that in two places:
* Reject turn/interrupt immediately with `INVALID_REQUEST` when the
requested turn is no longer the active turn.
* Resolve any already-accepted pending interrupt requests when the turn
reaches TurnComplete, covering the case where a turn finishes naturally
after the interrupt request is accepted but before it aborts.
I tested this by adding a failing test in
707487c063. You may view the results here:
https://github.com/openai/codex/actions/runs/24585182419/
<img width="1512" height="310" alt="CleanShot 2026-04-17 at 16 33 30@2x"
src="https://github.com/user-attachments/assets/f4a88228-b2a4-41f4-9aaa-ec82814096af"
/>
## Why
Agent interruptions currently always persist a model-visible
interrupted-turn marker before emitting `TurnAborted`. That marker is
useful by default because it gives the next model turn context about a
deliberately interrupted task, but some deployments need to suppress
that history injection entirely while still keeping the client-visible
interruption event.
## What changed
- Add `[agents] interrupt_message = false` to disable the model-visible
interrupted-turn marker.
- Resolve the setting into `Config::agent_interrupt_message_enabled`,
defaulting to `true` so existing behavior is unchanged.
- Apply the setting to both live interrupted turns and interrupted fork
snapshots.
- Keep emitting `TurnAborted` even when the history marker is disabled.
- Regenerate `core/config.schema.json` for the new
`agents.interrupt_message` field.
## Testing
- `cargo test -p codex-core load_config_resolves_agent_interrupt_message
-- --nocapture`
- `cargo test -p codex-core
disabled_interrupted_fork_snapshot_appends_only_interrupt_event --
--nocapture`
- `cargo test -p codex-core
multi_agent_v2_interrupted_marker_uses_developer_input_message --
--nocapture`
- `cargo test -p codex-core
multi_agent_v2_followup_task_can_disable_interrupted_marker --
--nocapture`
- `cargo test -p codex-core
multi_agent_v2_followup_task_interrupts_busy_child_without_losing_message
-- --nocapture`
- `cargo check -p codex-core`
## Summary
- Thread `agent_max_threads` into `ToolsConfig` and
`SpawnAgentToolOptions`.
- Render the configured `max_concurrent_threads_per_session` value in
the MultiAgentV2 `spawn_agent` description.
- Cover the description text in `codex-tools` unit tests and
`codex-core` tool spec tests.
## Validation
- `just fmt`
- `cargo test -p codex-tools`
- `cargo test -p codex-core spawn_agent_description`
- `git diff --check`
## Notes
- `cargo test -p codex-core` was also attempted, but unrelated
environment-sensitive tests failed with the active local environment.
Examples: approvals reviewer defaults observed `AutoReview` instead of
`User`, request-permissions event tests did not emit events, and
proxy-env tests saw `http://127.0.0.1:50604` from the active proxy
environment.
Co-authored-by: Codex <noreply@openai.com>
## Why
`MultiAgentV2` follow-up messages are delivered to agents as
assistant-authored `InterAgentCommunication` envelopes. When
`followup_task` used `interrupt: true`, the interrupted-turn guidance
was still persisted as a contextual user message, so model-visible
history made a system-generated interruption boundary look
user-authored.
This keeps interruption guidance consistent with the rest of the v2
inter-agent message stream while preserving the legacy marker shape for
non-v2 sessions.
## What changed
- Make `interrupted_turn_history_marker` feature-aware.
- Record the interrupted-turn marker as an assistant `OutputText`
message when `Feature::MultiAgentV2` is enabled.
- Keep the existing user contextual fragment for non-v2 sessions.
- Apply the same feature-aware marker to interrupted fork snapshots.
- Add coverage for the live `followup_task` interrupt path and the
helper-level v2 marker shape.
## Testing
- `cargo test -p codex-core
multi_agent_v2_followup_task_interrupts_busy_child_without_losing_message
-- --nocapture`
- `cargo test -p codex-core
multi_agent_v2_interrupted_marker_uses_assistant_output_message --
--nocapture`
- `cargo test -p codex-core interrupted_fork_snapshot -- --nocapture`
Supersedes #18735.
The scheduled rust-release-prepare workflow force-pushed
`bot/update-models-json` back to the generated models.json-only diff,
which dropped the test and snapshot updates needed for CI.
This PR keeps the latest generated `models.json` from #18735 and adds
the corresponding fixture updates:
- preserve model availability NUX in the app-server model cache fixture
- update core/TUI expectations for the new `gpt-5.4` `xhigh` default
reasoning
- refresh affected TUI chatwidget snapshots for the `gpt-5.5`
default/model copy changes
Validation run locally while preparing the fix:
- `just fmt`
- `cargo test -p codex-app-server model_list`
- `cargo test -p codex-core includes_no_effort_in_request`
- `cargo test -p codex-core
includes_default_reasoning_effort_in_request_when_defined_by_model_info`
- `cargo test -p codex-tui --lib chatwidget::tests`
- `cargo insta pending-snapshots`
---------
Co-authored-by: aibrahim-oai <219906144+aibrahim-oai@users.noreply.github.com>
## Summary
Fixes#19022.
`codex exec --json` currently emits `turn.completed.usage` with input,
cached input, and output token counts, but drops the reasoning-token
split that Codex already receives through thread token usage updates.
Programmatic consumers that rely on the JSON stream, especially
ephemeral runs that do not write rollout files, need this field to
accurately display reasoning-model usage.
This PR adds `reasoning_output_tokens` to the public exec JSON `Usage`
payload and maps it from the existing `ThreadTokenUsageUpdated` total
token usage data.
## Verification
- Added coverage to
`event_processor_with_json_output::token_usage_update_is_emitted_on_turn_completion`
so `turn.completed.usage.reasoning_output_tokens` is asserted.
- Updated SDK expectations for `run()` and `runStreamed()` so TypeScript
consumers see the new usage field.
- Ran `cargo test -p codex-exec`.
- Ran `pnpm --filter ./sdk/typescript run build`.
- Ran `pnpm --filter ./sdk/typescript run lint`.
- Ran `pnpm --filter ./sdk/typescript exec jest --runInBand
--testTimeout=30000`.
## Summary
Fixes#19275.
Codex runtime rejects inline MCP `bearer_token` config entries and asks
users to configure `bearer_token_env_var` instead, but the generated
config schema still advertised `mcp_servers.<name>.bearer_token` as a
supported field. That made editor/schema validation disagree with
runtime validation.
This keeps `bearer_token` in `RawMcpServerConfig` so Codex can continue
producing the targeted runtime error for recent or existing configs, but
skips the field during schemars generation. The checked-in
`core/config.schema.json` fixture now exposes `bearer_token_env_var`
without exposing unsupported inline `bearer_token`.
## Verification
- Added `config_schema_hides_unsupported_inline_mcp_bearer_token` to
assert the generated schema hides `bearer_token` while preserving
`bearer_token_env_var`.
- Ran `cargo test -p codex-config`.
- Ran `cargo test -p codex-core config_schema`.
we were not respecting turn's `truncation_policy` to clamp output tokens
for `unified_exec` and `write_stdin`.
this meant truncation was only being applied by `ContextManager` before
the output was stored in-memory (so it _was_ being truncated from
model-visible context), but the full output was persisted to rollout on
disk.
now we respect that `truncation_policy` and `ContextManager`-level
truncation remains a backup.
### Tests
added tests, tested locally.
## Summary
`codex.emitImage` accepted arbitrary image MIME types for byte payloads
and data URLs. That allowed a value like `image/rgba` to be wrapped as
an `input_image`, even though it is not a supported encoded image
format, so the invalid image could reach the model-input path and
trigger output sanitization.
This results in a panic in debug builds because the output sanitization
is meant as a final safety net, not a primary means of rejecting invalid
image types. I've hit this case multiple times when executing certain
long-running tasks.
This PR rejects unsupported image MIME types before they are emitted
from `js_repl`.
## Changes
- Validate `codex.emitImage({ bytes, mimeType })` in the JS kernel so
only encoded PNG, JPEG, WebP, or GIF payloads are accepted.
- Apply the same MIME allowlist to direct image data URLs, including the
Rust host-side validation path.
- Clarify the JS REPL instructions so agents know byte payloads must
already be encoded as PNG/JPEG/WebP/GIF.
Fixes#19257.
## Summary
Agent roles declared in config layers can set `config_file` to a
relative path, but deserializing the layer-local `[agents.*]` table
happened without an `AbsolutePathBuf` base path. That caused configs
like `config_file = "agents/my-role.toml"` to fail with `AbsolutePathBuf
deserialized without a base path`.
This updates agent role layer loading to deserialize `[agents.*]` while
the layer config folder is active as the path base, matching the
behavior documented for `AgentRoleToml.config_file`. It also adds
coverage for a user config layer with a relative agent role
`config_file`.
## Why
`PermissionProfile` is becoming the canonical permissions abstraction,
but the old shape only carried optional filesystem and network fields.
It could describe allowed access, but not who is responsible for
enforcing it. That made `DangerFullAccess` and `ExternalSandbox` lossy
when profiles were exported, cached, or round-tripped through app-server
APIs.
The important model change is that active permissions are now a disjoint
union over the enforcement mode. Conceptually:
```rust
pub enum PermissionProfile {
Managed {
file_system: FileSystemSandboxPolicy,
network: NetworkSandboxPolicy,
},
Disabled,
External {
network: NetworkSandboxPolicy,
},
}
```
This distinction matters because `Disabled` means Codex should apply no
outer sandbox at all, while `External` means filesystem isolation is
owned by an outside caller. Those are not equivalent to a broad managed
sandbox. For example, macOS cannot nest Seatbelt inside Seatbelt, so an
inner sandbox may require the outer Codex layer to use no sandbox rather
than a permissive one.
## How Existing Modeling Maps
Legacy `SandboxPolicy` remains a boundary projection, but it now maps
into the higher-fidelity profile model:
- `ReadOnly` and `WorkspaceWrite` map to `PermissionProfile::Managed`
with restricted filesystem entries plus the corresponding network
policy.
- `DangerFullAccess` maps to `PermissionProfile::Disabled`, preserving
the “no outer sandbox” intent instead of treating it as a lax managed
sandbox.
- `ExternalSandbox { network_access }` maps to
`PermissionProfile::External { network }`, preserving external
filesystem enforcement while still carrying the active network policy.
- Split runtime policies that legacy `SandboxPolicy` cannot faithfully
express, such as managed unrestricted filesystem plus restricted
network, stay `Managed` instead of being collapsed into
`ExternalSandbox`.
- Per-command/session/turn grants remain partial overlays via
`AdditionalPermissionProfile`; full `PermissionProfile` is reserved for
complete active runtime permissions.
## What Changed
- Change active `PermissionProfile` into a tagged union: `managed`,
`disabled`, and `external`.
- Keep partial permission grants separate with
`AdditionalPermissionProfile` for command/session/turn overlays.
- Represent managed filesystem permissions as either `restricted`
entries or `unrestricted`; `glob_scan_max_depth` is non-zero when
present.
- Preserve old rollout compatibility by accepting the pre-tagged `{
network, file_system }` profile shape during deserialization.
- Preserve fidelity for important edge cases: `DangerFullAccess`
round-trips as `disabled`, `ExternalSandbox` round-trips as `external`,
and managed unrestricted filesystem + restricted network stays managed
instead of being mistaken for external enforcement.
- Preserve configured deny-read entries and bounded glob scan depth when
full profiles are projected back into runtime policies, including
unrestricted replacements that now become `:root = write` plus deny
entries.
- Regenerate the experimental app-server v2 JSON/TypeScript schema and
update the `command/exec` README example for the tagged
`permissionProfile` shape.
## Compatibility
Legacy `SandboxPolicy` remains available at config/API boundaries as the
compatibility projection. Existing rollout lines with the old
`PermissionProfile` shape continue to load. The app-server
`permissionProfile` field is experimental, so its v2 wire shape is
intentionally updated to match the higher-fidelity model.
## Verification
- `just write-app-server-schema`
- `cargo check --tests`
- `cargo test -p codex-protocol permission_profile`
- `cargo test -p codex-protocol
preserving_deny_entries_keeps_unrestricted_policy_enforceable`
- `cargo test -p codex-app-server-protocol
permission_profile_file_system_permissions`
- `cargo test -p codex-app-server-protocol serialize_client_response`
- `cargo test -p codex-core
session_configured_reports_permission_profile_for_external_sandbox`
- `just fix`
- `just fix -p codex-protocol`
- `just fix -p codex-app-server-protocol`
- `just fix -p codex-core`
- `just fix -p codex-app-server`
## Summary
- Add a remote plugin install write call that POSTs the selected remote
plugin to the ChatGPT cloud plugin API.
- Align remote install with the latest remote read contract:
`pluginName` carries the backend remote plugin id directly, for example
`plugins~Plugin_linear`, and install no longer synthesizes
`<name>@<marketplace>` ids.
- Validate remote install ids with the same character rules as remote
read, return the same install response shape as local installs, and
include mocked app-server coverage for the write path.
## Validation
- `just fmt`
- `cargo test -p codex-app-server --test all plugin_install`
- `cargo test -p codex-core-plugins`
- `just fix -p codex-app-server`
- `just fix -p codex-core-plugins`
## Why
Device-key providers should only own platform key material. The
account/client binding used to authorize a signing payload is app-server
state, and keeping that state in provider-specific metadata makes the
same check harder to audit and harder to share across platform
implementations.
Persisting the binding in the shared state database gives the device-key
crate a platform-neutral source of truth before it asks a provider to
sign. It also lets app-server move potentially blocking key operations
off the main message processor path, which matters once providers may
wait for OS authentication prompts.
## What changed
- Add a `device_key_bindings` state migration plus `StateRuntime`
helpers keyed by `key_id`.
- Add an async `DeviceKeyBindingStore` abstraction to `codex-device-key`
and use it from `DeviceKeyStore::create` and `DeviceKeyStore::sign`.
- Keep provider calls behind async store methods and run the synchronous
provider work through `spawn_blocking`.
- Wire app-server device-key RPC handling to the SQLite-backed binding
store and spawn response/error delivery tasks for device-key requests.
- Run the turn-start tracing test on the existing larger current-thread
test harness after the larger async surface made the default test stack
too small locally.
## Validation
- `cargo test -p codex-device-key`
- `cargo test -p codex-state device_key`
- `cargo test -p codex-state`
- `cargo test -p codex-app-server device_key`
- `cargo test -p codex-app-server
message_processor::tracing_tests::turn_start_jsonrpc_span_parents_core_turn_spans`
- `cargo test -p codex-app-server`
- `just fix -p codex-device-key`
- `just fix -p codex-state`
- `just fix -p codex-app-server`
- `just bazel-lock-update`
- `just bazel-lock-check`
- `git diff --check`
## Why
`codex-models-manager` had grown to own provider-specific concerns:
constructing OpenAI-compatible `/models` requests, resolving provider
auth, emitting request telemetry, and deciding how provider catalogs
should be sourced. That made the manager harder to reuse for providers
whose model catalog is not fetched from the OpenAI `/models` endpoint,
such as Amazon Bedrock.
This change moves provider-specific model discovery behind
provider-owned implementations, so the models manager can focus on
refresh policy, cache behavior, picker ordering, and model metadata
merging.
## What Changed
- Introduced a `ModelsManager` trait with separate `OpenAiModelsManager`
and `StaticModelsManager` implementations.
- Added `ModelsEndpointClient` so OpenAI-compatible HTTP fetching lives
outside `codex-models-manager`.
- Moved `/models` request construction, provider auth resolution,
timeout handling, and request telemetry into `codex-model-provider` via
`OpenAiModelsEndpoint`.
- Added provider-owned `models_manager(...)` construction so configured
OpenAI-compatible providers use `OpenAiModelsManager`, while
static/catalog-backed providers can return `StaticModelsManager`.
- Added an Amazon Bedrock static model catalog for the GPT OSS Bedrock
model IDs.
- Updated core/session/thread manager code and tests to depend on
`Arc<dyn ModelsManager>`.
- Moved offline model test helpers into
`codex_models_manager::test_support`.
## Metadata References
The Bedrock catalog metadata is based on the official Amazon Bedrock
OpenAI model documentation:
- [Amazon Bedrock OpenAI
models](https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-openai.html)
lists the Bedrock model IDs, text input/output modalities, and `128,000`
token context window for `gpt-oss-20b` and `gpt-oss-120b`.
- [Amazon Bedrock `gpt-oss-120b` model
card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-oss-120b.html)
lists the `bedrock-runtime` model ID `openai.gpt-oss-120b-1:0`, the
`bedrock-mantle` model ID `openai.gpt-oss-120b`, text-only modalities,
and `128K` context window.
- [OpenAI `gpt-oss-120b` model
docs](https://developers.openai.com/api/docs/models/gpt-oss-120b)
document configurable reasoning effort with `low`, `medium`, and `high`,
plus text input/output modality.
The display names, default reasoning effort, and priority ordering are
Codex-local catalog choices.
## Test Plan
- Manually verified app-server model listing with an AWS profile:
```shell
CODEX_HOME="$(mktemp -d)" cargo run -p codex-app-server-test-client -- \
--codex-bin ./target/debug/codex \
-c 'model_provider="amazon-bedrock"' \
-c 'model_providers.amazon-bedrock.aws.profile="codex-bedrock"' \
-c 'model_providers.amazon-bedrock.aws.region="us-west-2"' \
model-list
```
The response returned the Bedrock catalog with `openai.gpt-oss-120b-1:0`
as the default model and `openai.gpt-oss-20b-1:0` as the second listed
model, both text-only and supporting low/medium/high reasoning effort.
## Summary
Adds the debug CLI entry point for reducing recorded rollout traces.
This gives developers a direct way to inspect whether the emitted trace
stream reduces into the expected conversation/runtime model.
## Stack
This is PR 5/5 in the rollout trace stack.
- [#18876](https://github.com/openai/codex/pull/18876): Add rollout
trace crate
- [#18877](https://github.com/openai/codex/pull/18877): Record core
session rollout traces
- [#18878](https://github.com/openai/codex/pull/18878): Trace tool and
code-mode boundaries
- [#18879](https://github.com/openai/codex/pull/18879): Trace sessions
and multi-agent edges
- [#18880](https://github.com/openai/codex/pull/18880): Add debug trace
reduction command
## Review Notes
This PR is intentionally last: it depends on the trace crate, core
recorder, runtime/tool events, and session/agent edge data all existing.
The command should remain a debug/developer tool and avoid adding new
runtime behavior.
The useful review question is whether the CLI exposes the reducer in the
smallest practical way for local inspection without turning the debug
command into a supported user-facing workflow.
## Why
AWS/Bedrock mode currently reports `account: null` with
`requiresOpenaiAuth: false` from `account/read`. That suppresses the
OpenAI-auth requirement, but it does not let app clients distinguish AWS
auth from any other non-OpenAI custom provider. For the prototype AWS
provider UX, clients need a simple provider-derived signal so they can
suppress ChatGPT/API-key login and token-refresh paths without
hardcoding Bedrock checks.
## What changed
- Adds an `aws` variant to the v2 `Account` protocol union.
- Adds `ProviderAccountKind` to `codex-model-provider` so the runtime
provider owns the app-visible account classification.
- Makes Amazon Bedrock return `ProviderAccountKind::Aws` from the
model-provider layer.
- Updates app-server `account/read` to map `ProviderAccountKind` to the
existing `GetAccountResponse` wire shape.
- Preserves the existing `account: null, requiresOpenaiAuth: false`
behavior for other non-OpenAI providers.
- Regenerates the app-server protocol schema fixtures.
- Adds coverage for provider account classification and for the Amazon
Bedrock `account/read` response.
## Testing
- `cargo test -p codex-model-provider`
- `cargo test -p codex-app-server-protocol`
- `cargo test -p codex-app-server get_account_with_aws_provider`
## Notes
I attempted `just bazel-lock-update` and `just bazel-lock-check`, but
both are blocked in my local environment because `bazel` is not
installed.
Fixes#18203.
## Why
Remote TUI clients connected through `codex app-server --listen
ws://...` can receive short bursts of outbound turn and tool-output
notifications. The WebSocket transport previously used the shared
128-message channel capacity for its outbound writer queue, so a healthy
client that briefly lagged during normal output streaming could fill the
queue and be disconnected immediately.
This is a smaller mitigation than #18265: instead of adding a new
overflow/backpressure pipeline, keep the existing non-blocking router
behavior and give WebSocket clients enough bounded headroom for
realistic bursts.
## What Changed
- Added a WebSocket-only outbound writer capacity of `64 * 1024`
messages.
- Used that larger capacity only for the WebSocket data writer queue in
`codex-rs/app-server/src/transport/websocket.rs`.
- Left the shared `CHANNEL_CAPACITY` and the existing disconnect-on-full
behavior unchanged for internal/control channels and genuinely stuck
clients.
## Verification
- `cargo test -p codex-app-server
transport::tests::broadcast_does_not_block_on_slow_connection`
- Manually retried the #18203 repro prompt against the remote TUI and
confirmed it stayed connected.
We used to attempt a read-ACL on the same dir as `codex.exe` to grant
the sandbox user the ability to invoke `codex-command-runner.exe`. That
worked for the CLI case but it always fails for the installed desktop
app.
We have another solution already in place that copies
`codex-command-runner.exe` to `CODEX_HOME/.sandbox-bin` so we don't even
need this anymore. It causes a scary looking error in the logs that is a
non-issue and is therefore confusing
Sometimes codex runs `Start-Process` to start up a service or something
similar, which launches a user-visible powershell window that probably
doesn't get cleaned up. This instruction change encourages it to do so
using a hidden window.
This was reported in
https://openai.slack.com/archives/C09K6H5DZC4/p1776741272870519
One caveat is that this change won't do anything to cleanup these
processes, but it will stop them from polluting the user's visible
workspace
---------
Co-authored-by: Codex <noreply@openai.com>
