## Summary
Make blocking hook behavior reliable for tools invoked from code mode.
Previously, a `PostToolUse` hook could block a completed tool result,
but code mode would still return the original typed result to
JavaScript. The hook appeared blocked in hook telemetry while the
running script continued with the result.
This change:
- rejects the nested JavaScript tool promise when `PostToolUse` blocks
- normalizes `decision: "block"` and exit code 2 to the same blocking
behavior
- surfaces the hook feedback as the rejected promise's error
- adds end-to-end coverage for the relevant PreToolUse and PostToolUse
interactions
## Hook semantics in code mode
| Hook behavior | Code-mode result |
|---|---|
| PreToolUse block | Reject the promise before the tool executes |
| PreToolUse `updatedInput` | Execute the rewritten invocation and
return its result |
| PostToolUse `decision: "block"` | Execute the tool, then reject the
promise with the hook reason |
| PostToolUse exit code 2 | Same behavior as `decision: "block"` |
| PostToolUse `continue: false` | Preserve the existing feedback-only
behavior; do not reject the promise |
## Test coverage
Added or strengthened end-to-end coverage proving that:
- a PreToolUse block rejects the JavaScript promise before execution
- a PreToolUse input rewrite executes only the rewritten command
- JavaScript receives the rewritten command's result
- PostToolUse `decision: "block"` rejects after the command executes
- PostToolUse exit code 2 has the same behavior
- the hook observes the original completed tool response
- the blocked original result does not reach JavaScript
- existing direct-mode replacement behavior remains intact
- `continue: false` without a reason produces deterministic fallback
feedback
## Description
This PR adds an optional `metadata` field to `ResponseItem` for
Responses API calls. Only mechanical plumbing, no actual values
populated and sent yet. Turns out just adding a new field to
`ResponseItem` has quite a large blast radius already.
This change is backwards compatible because `metadata` is optional and
omitted when absent, so existing response items and rollout history
without it still deserialize and requests that do not set it keep the
same wire shape. For provider compatibility, we strip out `metadata`
before non-OpenAI Responses requests so Azure and AWS Bedrock never see
this field.
My followup PR here will actually make use of it to start storing and
passing along `turn_id`: https://github.com/openai/codex/pull/28360
## What changed
- Added `ResponseItemMetadata` with optional `turn_id`, plus optional
`metadata` on Responses API item variants and inter-agent communication.
- Preserved item metadata through response-item rewrites such as
truncation, missing tool-output synthesis, compaction history
rebuilding, visible-history conversion, rollout/resume, and generated
app-server schemas/types.
- Strip item metadata from non-OpenAI Responses requests while
preserving it for OpenAI-shaped requests.
- Updated the mechanical fixture/test construction churn required by the
new optional field.
## Why
Tool router construction rebuilds the deferred-tool BM25 index during
session initialization and before each sampling continuation, even when
the searchable tool metadata is unchanged. Local profiling measured
`append_tool_search_executor` at roughly 113 ms per continuation, making
repeated index construction the largest measured router-building cost.
## What changed
- Add a session-scoped `ToolSearchHandlerCache` so continuations and
user turns can reuse the existing handler.
- Key reuse on the complete ordered `Vec<ToolSearchInfo>`, rebuilding
when searchable text, loadable tool specs, source metadata, or ordering
changes.
- Build handlers outside the cache lock and recheck before publishing
them, avoiding holding the mutex during index construction.
## Verification
- `cache_reuses_identical_search_infos_and_rebuilds_changed_inputs`
covers exact cache reuse and invalidation when the ordered search
metadata changes.
- Local rollout profiling showed the initial router build populating the
cache and unchanged later continuations reusing it:
- uncached: 118 ms median across 14 spans from 3 rollouts
- cached: 4 ms median across 12 spans from 3 rollouts
## Why
`wait_agent` can block for a long timeout while waiting for sub-agent
mailbox activity. Although same-turn user steer is accepted during that
tool call, the input remains pending until the wait returns, so an
explicit request to change direction can appear unresponsive.
## What changed
- Notify active `wait_agent` calls when user input is steered into the
current turn.
- Check for already-pending steer input when subscribing so input that
races with tool startup is not missed.
- Distinguish mailbox activity, steered input, and timeout outcomes,
returning `Wait interrupted by new input.` for the steer path.
- Update the `wait_agent` tool description to document the early-return
behavior.
## Testing
- `just test -p codex-core input_queue_`
- `just test -p codex-core wait_agent`
The coverage includes steer notification before and after subscription,
plus an end-to-end test that verifies the interrupted wait result and
steered user input are both included exactly once in the follow-up model
request.
## Why
`main` fails dev-profile Cargo and Bazel Clippy builds because
`core/src/tools/runtimes/mod_tests.rs` imports `PathExt` after its last
use was removed. With warnings denied, that stale import prevents
`codex-core` test targets from compiling across platforms.
## What changed
Remove the unused `PathExt` import. Remaining `.abs()` calls in the
module operate on `PathBuf` and continue to use `PathBufExt`.
## Validation
- `just fmt`
- Focused `codex-core` test compile attempted; blocked locally by disk
exhaustion before compilation completed. The CI failure itself is the
unused-import diagnostic this change removes.
## Why
Memory read telemetry currently reconstructs the executable shell
command after a tool call finishes. That duplicates shell, login-policy,
and cwd resolution owned by the tool handlers, and can diverge from the
environment-specific command that unified exec actually ran.
## What changed
- Expose the existing restricted shell-script parser directly for raw
script text.
- Parse `shell_command` and `exec_command` input into plain command argv
before classifying memory reads.
- Preserve all-or-nothing safe-command validation for multi-command
scripts.
- Remove cwd resolution, shell selection, and the unnecessary async
boundary from memory read metric emission.
## Testing
- `just test -p codex-shell-command`
- `cargo check -p codex-core`
## Why
Shell snapshot lifecycle state was split between `Shell` and
`SessionServices`: `Shell` carried the receiver while session code
exposed and forwarded the raw sender. That coupled shell identity to
mutable snapshot state and made refresh, inheritance, and file lifetime
harder to reason about.
## What changed
- make each `Arc<ShellSnapshot>` represent one cwd-specific snapshot
generation
- store the active generation in `SessionServices` with `ArcSwapOption`
- have construction start the background build and expose only a
cwd-validated snapshot path
- use `ShellSnapshotFile` ownership to delete snapshot files
automatically
- pass snapshot paths explicitly to shell runtimes instead of storing
snapshot state on `Shell`
- preserve inherited and in-flight generations by pinning their `Arc`
while they are in use
## Test plan
- `cargo check -p codex-core --lib`
- `just test -p codex-core 'shell_snapshot::tests'`
- `just test -p codex-core
shell_command_snapshot_still_intercepts_apply_patch`
- `just test -p codex-core
shell_snapshot_deleted_after_shutdown_with_skills`
Follow-up to #27356.
## Stack note
This PR changes Codex's internal dynamic-tool shape while leaving
`thread/start` unchanged. App-server therefore converts the existing
per-tool input into explicit functions and namespaces before passing it
to core.
[#27371](https://github.com/openai/codex/pull/27371) updates
`thread/start` to use the same explicit shape and removes this temporary
conversion.
## Why
Dynamic tools repeat namespace metadata on every function. Core should
keep one explicit namespace with its member tools so descriptions and
membership stay consistent across sessions and runtime planning.
## What changed
- Represent dynamic tools as top-level functions or explicit namespaces
in protocol and session state.
- Read old flat rollout metadata and write the canonical hierarchy.
- Flatten namespace members only when registering callable tools.
- Keep `thread/start.dynamicTools` flat for now and normalize it at the
app-server boundary.
New builds can read old rollout metadata. Older builds cannot read newly
written hierarchical metadata.
## Test plan
- `just test -p codex-app-server
thread_start_normalizes_legacy_dynamic_tools_into_model_request`
- `just test -p codex-protocol
session_meta_normalizes_legacy_dynamic_tools`
- `just test -p codex-core
resume_restores_dynamic_tools_from_rollout_with_sqlite_enabled`
- `just test -p codex-core
tool_search_returns_deferred_dynamic_tool_and_routes_follow_up_call`
- `just test -p codex-core code_mode_can_call_hidden_dynamic_tools`
- `just test -p codex-tools`
## Summary
- align the default multi-agent v2 root and subagent hints with the
evaluated prompt guidance for direct collaboration-tool calls, parallel
delegation, and shared workspaces
- keep the current `interrupt_agent` tool name and existing
concurrency-hint placement, with the explicit no-spawn instruction last
- document the context tradeoff between `fork_turns="none"` and
`fork_turns="all"` in the v2 `spawn_agent` description
- extend the focused prompt and tool-surface tests
## Why
The evaluated multi-agent prompt includes operational guidance that is
missing from the current Codex defaults. This applies that guidance to
the current tool surface without restoring stale `close_agent` or
duplicated concurrency wording.
## User impact
Multi-agent v2 receives clearer instructions about when and how to
parallelize work, how agent workspaces interact, and how `fork_turns`
affects subagent context. The existing default opt-out behavior remains
in place.
## Testing
- `just fmt`
- `just test -p codex-core
multi_agent_v2_default_usage_hints_use_configured_thread_cap`
- `just test -p codex-core
multi_agent_feature_selects_one_agent_tool_family`
## Why
Next slice needed to make progress on the `remote_env_windows` test is
to support passing a Windows cwd for the remote environment and using
that environment's native shell. This lets the test run a real Windows
process instead of only recording an early path or shell mismatch.
## What
- change `TurnEnvironmentSelection.cwd` from `AbsolutePathBuf` to
`PathUri`
- convert local cwd values to URIs when constructing selections
- preserve a remote primary cwd instead of replacing it with the local
legacy fallback
- prefer the selected environment's discovered shell for unified exec,
falling back to the session shell when unavailable
- convert back to a host-native absolute path at current native-only
consumer boundaries
- reject or deny unsupported foreign cwd values at the existing
request-permissions boundary, with TODOs for its future migration
- extend the hermetic Wine test to execute Windows PowerShell in
`C:\windows` and verify successful process completion
- record the current app-server rejection against the same Wine-backed
remote Windows fixture when its cwd is supplied as a native Windows path
## Why
`PathUri::from_abs_path` can fail for absolute paths that do not have a
normal `file:` URI representation, forcing filesystem call sites to
handle a conversion error even though the original path can be preserved
losslessly.
## What
Make `from_abs_path` infallible and migrate its callers. Unrepresentable
paths use `file:///%00/bad/path/<base64>`, encoding Unix bytes or
Windows UTF-16LE; `to_abs_path` validates and decodes that fallback. The
leading encoded null reserves a namespace that cannot collide with a
real Unix or Windows path, and fallback URIs remain opaque to lexical
path operations.
## Validation
Added path-URI coverage for Unix null and non-UTF-8 paths, Windows
device/verbatim and non-Unicode paths, serialization, malformed
fallbacks, opaque lexical operations, invalid native payloads, and
literal `/bad/path` collision resistance.
## Why
Sandbox cwd values can cross app-server and exec-server host boundaries.
They should retain URI semantics until the receiving host validates them
instead of being interpreted early as native paths.
## What
- Carry `PathUri` through filesystem sandbox contexts, sandbox commands,
and transform inputs.
- Convert command and policy cwd once in `SandboxManager::transform`,
then keep launch requests native.
- Preserve sandbox cwd over remote filesystem transport and reject
non-native URIs without fallback.
- Cache paired native/URI turn-environment cwd values during migration,
with immutable access to keep them synchronized.
- Extend existing protocol, forwarding, transform, and core runtime
tests.
## Why
`request_user_input` has direct blocking semantics when invoked by the
model. When it is exposed as a nested code-mode tool, the call has to
flow through code-mode waiting and continuation behavior instead, which
is not the behavior we want for this user-input request surface.
## What changed
- Mark `request_user_input` with `ToolExposure::DirectModelOnly` when
registering the core utility tool.
- Keep `request_user_input` direct-model visible, including in
code-mode-only planning.
- Add focused `spec_plan_tests` coverage that verifies
`request_user_input` remains visible and registered as
direct-model-only, while it is omitted from the nested code-mode tool
description.
No active goal suppression or runtime unavailability behavior is
included in this PR.
## Validation
- No new build/test run for this housekeeping pass, per maintainer
request.
- Earlier targeted run, confirmed from session context: `just test -p
codex-core request_user_input` passed.
## Why
`request_user_input` is moving beyond its original plan-mode-only
workflow, and future default/goal-mode usage needs a way for the model
to ask helpful but non-blocking questions without forcing the turn to
wait forever. This PR adds an explicit `autoResolutionMs` contract so a
later client/runtime change can auto-resolve unanswered prompts after a
bounded window while leaving truly blocking questions unchanged.
This is contract plumbing only; it does not implement the client-side
timer or auto-selection behavior, and the model-facing description
treats the field as reserved unless the current runtime explicitly
supports auto-resolution.
## What Changed
- Added optional `autoResolutionMs` to the model-facing
`request_user_input` args and core `RequestUserInputEvent`.
- Added model-facing schema text for `autoResolutionMs` while marking it
reserved for runtimes that explicitly support auto-resolution.
- Bounds `autoResolutionMs` to `60_000..=240_000` ms during argument
normalization by clamping out-of-range model-provided values.
- Propagated the field through app-server v2
`ToolRequestUserInputParams`, app-server request forwarding, generated
TypeScript, and JSON schema fixtures.
- Updated app-server, core, protocol, and TUI call sites/tests so
omitted values preserve existing `None`/`null` behavior and coverage
verifies a `Some(60_000)` round trip.
## Verification
- `just test -p codex-app-server-protocol`
- `just test -p codex-core request_user_input`
- `just test -p codex-app-server request_user_input_round_trip`
- `just test -p codex-tui request_user_input`
- `just test -p codex-protocol`
## Why
First-party async traits should expose their `Send` contracts explicitly
without requiring `async_trait`. This completes the migration pattern
established in #27303 and #27304.
## What changed
- Replaced the remaining first-party `async_trait` traits with native
return-position `impl Future + Send` where statically dispatched and
explicit boxed `Send` futures where object safety is required.
- Kept implementations behavior-preserving, outlining existing async
bodies into inherent methods where that keeps the diff reviewable.
- Removed all direct first-party `async-trait` dependencies and the
workspace dependency declaration.
- Added a cargo-deny policy that permits `async-trait` only through the
remaining transitive wrapper crates.
- Updated `rand` from 0.8.5 to 0.8.6 to resolve RUSTSEC-2026-0097 and
keep the full cargo-deny check passing.
## Validation
- `just test -p codex-exec-server`: 216 passed, 2 skipped.
- `just test -p codex-model-provider`: 39 passed.
- `just test -p codex-core` and `just test`: changed tests passed;
remaining failures are environment-sensitive suites unrelated to this
migration.
- `cargo deny check`
- `just fix`
- `just fmt`
- `cargo shear`
- `just bazel-lock-check`
## Why
Image generation used `std::fs::read` for referenced image paths, which
did not support environment-backed filesystems or their sandbox context.
## What changed
- Expose optional turn environments to extension tool calls.
- Include each environment’s ID, working directory, filesystem, and
sandbox context.
- Read referenced images through the selected environment filesystem.
- Keep sandbox usage at the extension call site so extensions can choose
the appropriate access mode.
- Consolidate image request construction into one async function.
- Add coverage for successful environment reads and read failures.
## Validation
- `cargo check -p codex-image-generation-extension --tests`
- `just fmt`
- `just bazel-lock-update`
- `just bazel-lock-check`
`just test -p codex-image-generation-extension` could not complete
because the build exhausted available disk space.
## Why
The token budget full-context fragment identifies the current context
window, but not the thread that owns that window. Including the thread
id makes the initial context-window metadata self-contained, and
`get_context_remaining` also needs to be usable from Code Mode without
forcing callers to parse the model-facing fragment string.
## What changed
- Include the session thread id in the initial `<token_budget>` context
fragment.
- Expose `get_context_remaining` as a Code Mode nested tool while
keeping `new_context` direct-model-only.
- Keep direct model-facing `get_context_remaining` output as the
existing `<token_budget>` text fragment.
- Return only `tokens_left` from the Code Mode structured result for
`get_context_remaining`.
- Update token-budget integration tests and add Code Mode coverage for
the structured result.
## Verification
- `just test -p codex-core token_budget`
- `just test -p codex-core
code_mode_get_context_remaining_returns_structured_result`
- `just test -p core_test_support redacted_text_mode_normalizes_uuids`
## Why
We want to remove implicit use of `$CODEX_HOME` from `codex-core` and
make embedders responsible for supplying user-level instructions. This
also ensures user instructions load when no primary environment is
selected.
## What changed
Stacked on #27415, which makes `codex exec` surface thread-scoped
runtime warnings.
- Added `UserInstructionsProvider` to `codex-extension-api`, with
absolute source attribution and recoverable loading warnings.
- Added `codex-home` with the filesystem-backed provider for
`AGENTS.override.md` and `AGENTS.md`, preserving precedence, fallback,
trimming, lossy UTF-8 handling, and the existing uncapped global
instruction size.
- Removed global instruction loading from `Config` and require
`ThreadManager` callers to inject a provider.
- Load provider instructions once for each fresh root runtime, including
runtimes without a primary environment. Running sessions retain their
snapshot, while child agents inherit the parent snapshot without
invoking the provider.
- Keep provider instructions separate while loading project `AGENTS.md`,
then assemble the model-visible instructions with the existing ordering,
source attribution, warning, and turn-context behavior.
- Wired the Codex home provider through the CLI, app server, MCP server,
core facade, and thread-manager sample.
## Validation
- `just test -p codex-home -p codex-extension-api`
- `just test -p codex-core agents_md`
- `just test -p codex-core guardian`
- `just test -p codex-app-server
thread_start_without_selected_environment_includes_only_global_instruction_source`
- `just test -p codex-exec warning`
- `just bazel-lock-check`
## Why
We're moving exec-server to use PathUri for its internal path
representations.
## What
Move `ExecutorFileSystem` APIs to use `PathUri` instead of
`AbsolutePathBuf`. Future changes will convert higher-level parts of
exec-server.
## Why
Native Codex currently teaches multi-agent concurrency through the
`spawn_agent` tool description, while bridge-driven evals frame the same
limit as a shared pool of active agent slots. That mismatch makes the
model-facing story harder to reason about, especially because the
tool-level wording does not make it explicit that the limit covers the
whole agent team, including the current agent.
This change gives native Codex the same mental model: tell the root
agent and subagents how many active slots exist, and remove the separate
`spawn_agent` limit wording.
## What changed
- Extend the built-in `multi_agent_v2` root and subagent usage hints
with shared-slot guidance derived from the resolved
`max_concurrent_threads_per_session` value.
- Keep the complete default hints in `MultiAgentV2Config` so initial
context and forked histories consume the same canonical strings.
- Drop the redundant `spawn_agent` description text and remove the
now-unused limit plumbing from the tool spec path.
## Testing
- `just test -p codex-core usage_hint`
- `just test -p codex-core
multi_agent_v2_default_session_thread_cap_counts_root`
- `just test -p codex-core
multi_agent_v2_default_usage_hints_use_configured_thread_cap`
- `just test -p codex-core
spawn_agent_tool_v2_requires_task_name_and_lists_visible_models`
- `just test -p codex-core
multi_agent_feature_selects_one_agent_tool_family`
## Why
Dynamic tools maintained a separate search-text builder even though the
shared tool search path already derives the same metadata from
`ToolSpec`. Using the shared path removes duplicate behavior before
adding explicit namespaces.
## What changed
- Build dynamic-tool search entries with
`ToolSearchInfo::from_tool_spec`.
- Remove the custom search-text state and its implementation-only unit
test.
The old search text included the tool name, its space-separated form,
description, namespace, and top-level parameter names. The shared
builder preserves all of those terms and also indexes namespace
descriptions and nested schema metadata.
## Test plan
- `just test -p codex-core
tool_search_returns_deferred_dynamic_tool_and_routes_follow_up_call`
## Why
The token budget feature can inject remaining-context notices into
model-visible context, but the model does not have a direct way to ask
for that same remaining-token fragment on demand.
This PR adds a small model tool for the token budget feature so the
model can request the current remaining context window message without
duplicating the fragment format.
## What changed
- Adds a `get_context_remaining` direct-model tool behind
`Feature::TokenBudget`.
- Renders the tool output through `TokenBudgetRemainingContext`,
matching the existing budget message shape.
- Registers the tool alongside `new_context` in the token budget tool
set.
- Adds integration coverage that verifies the tool is exposed and
returns the same `<token_budget>` remaining fragment already present in
context.
## Validation
- `just test -p codex-core token_budget`
## Why
The token budget feature tells the model how much room remains in the
current context window. When the model decides the current window is no
longer useful, it needs a way to ask Codex to start over with a fresh
context window without spending tokens on a compaction summary.
This PR adds that model-requestable escape hatch on top of #27438.
## What changed
- Added a direct-model-only `new_context` tool behind
`Feature::TokenBudget`.
- Stores the tool request on `AutoCompactWindow` and consumes it after
sampling so the next follow-up request in the same turn starts in the
new window.
- Starts the new window as a no-summary compaction checkpoint that
contains only fresh initial context, not preserved conversation history.
- Keeps the new window aligned with token-budget startup context,
including the `Current context window Z` message.
- Added integration coverage and a snapshot showing the same-turn
`new_context` flow into a fresh full-context follow-up request.
## Validation
- `just test -p codex-core token_budget`
## Summary
Adds complete client-side image preparation behind the default-off
`resize_all_images` feature flag.
When enabled, local image producers defer decoding and resizing. Images
are prepared centrally before insertion into conversation history,
covering user input, `view_image`, and structured tool-output images.
## Behavior
- Processes base64 `data:` images in messages and function/custom tool
outputs.
- Leaves non-data URLs, including HTTP(S) URLs, unchanged.
- Applies image-detail budgets:
- `high` and omitted: 2048px maximum dimension and 2.5K 32px patches.
- `original`: 6000px maximum dimension and 10K 32px patches.
- `auto`: uses the same 2048px / 2.5K-patch budget as high.
- `low`: unsupported and replaced with an actionable placeholder.
- Preserves original image bytes when no resize or format conversion is
needed.
- Enforces the shared 1 GiB encoded and decoded data-URL sanity limits.
- Replaces only an image that fails preparation, preserving sibling
content and tool-output metadata.
- Uses bounded placeholders distinguishing generic processing failures,
oversized images, and unsupported `low` detail.
- Prepares resumed and forked history before installing it as live
history without modifying persisted rollouts.
## Flag-Off Behavior
When `resize_all_images` is disabled:
- Existing local user-input and `view_image` processing remains
unchanged.
- Existing decoding and error behavior remains unchanged.
- Arbitrary tool-output images are not processed.
- HTTP(S) image URLs continue to be forwarded unchanged.
#### [git stack](https://github.com/magus/git-stack-cli)
- ✅ `1` https://github.com/openai/codex/pull/27245
- 👉 `2` https://github.com/openai/codex/pull/27247
- ⏳ `3` https://github.com/openai/codex/pull/27246
- ⏳ `4` https://github.com/openai/codex/pull/27266
## Summary
Turn diff updates repeatedly rendered and serialized the entire
accumulated diff after every `apply_patch`. The event path also rendered
once before updating the tracker solely to test whether a diff existed.
In production feedback CODEX-20PW, 2,589 patches across 72 paths
produced 401 notifications totaling 441 MB, with the hottest paths
patched 518 and 495 times.
This change:
- replaces the pre-update render with a cheap cached-state check
- caches each rendered file diff by path and content revision, so an
update only invokes Myers for affected paths
- caches the deterministic aggregate diff so event emission and turn
completion reuse it without recomputation
- preserves invalidation and net-zero clear notifications
- applies a 100 ms per-file `similar` timeout; ordinary files complete
far below this threshold, while pathological rewrites fall back to a
coarse unified hunk that still represents the exact final contents
The 100 ms deadline bounds synchronous tool-completion latency while
leaving substantial headroom for normal diffs. The regression test
applies the fallback diff through the repository's patch parser and
verifies byte-for-byte final contents.
## Validation
- `cargo test -p codex-core turn_diff_tracker::tests` (14 passed)
- `cargo test -p codex-core tools::events::tests` (4 passed)
- `just fix -p codex-core`
- `just fmt`
Focused coverage verifies that 42 updates across two files perform 42
file renders rather than repeatedly rendering the accumulated set,
unchanged paths are not re-diffed, clear events remain correct, and a
48,000-line near-total rewrite returns promptly and applies to the exact
expected result. The full `codex-core` suite was not used as the final
gate because an unrelated existing multi-agent test hit a stack overflow
when run during investigation.
## Bug context
- Sentry feedback: CODEX-20PW
- Correlation IDs: `019eb2a9-13d2-74e0-b690-27ee224ffb6d`,
`019e9ad7-09c3-7cb2-b728-ee3acba103ab`
## Summary
- skip the local `openai-curated` marketplace before marketplace loading
when tool-suggest discovery uses remote plugins
- preserve existing marketplace listing behavior for all other callers
and when remote plugins are disabled
- add regression coverage proving the curated marketplace is excluded
before its malformed manifest can be read
## Why
Tool-suggest discovery previously loaded every local `openai-curated`
plugin manifest and only discarded that marketplace afterward when
remote plugins were enabled. The remote catalog is used in that mode, so
the local scan consumed CPU without contributing discoverable plugins.
## Impact
Remote-plugin tool suggestion discovery no longer reads the local
curated marketplace and its plugin manifests. `openai-bundled`,
configured marketplaces, normal `plugin/list` behavior, and local
curated discovery when remote plugins are disabled are unchanged.
## Validation
- `just test -p codex-core-plugins
list_marketplaces_can_skip_openai_curated_before_loading`
- `just test -p codex-core
list_tool_suggest_discoverable_plugins_omits_openai_curated_when_remote_enabled`
- `just fmt`
- `git diff --check`
Summary
- Propagate cached remote plugin IDs through Codex plugin discovery.
- Inject `remote_plugin_id` and connector IDs into
`request_plugin_install` elicitation `_meta` from the resolved plugin.
- Keep the remote plugin ID out of the model-facing tool schema,
arguments, and result.
Validation
- `just test -p codex-tools`
- `just test -p codex-core-plugins`
- `just test -p codex-core
list_tool_suggest_discoverable_plugins_includes_cached_remote_global_plugins`
- `just fix -p codex-tools`
- `just fix -p codex-core-plugins`
- `just fix -p codex-core`
- `git diff --check`
- `just test -p codex-core` was also attempted: 2,581 passed, 55 failed,
and 1 timed out across unrelated sandbox/environment-sensitive
integration tests.
## Why
We're now [discouraging use of
`async_trait`](https://github.com/openai/codex/pull/20242).
Removing use of `async_trait` from `ToolExecutor` yields a `codex_core`
debug test build speedup of ~78% (from 227.5s to 50.3s) on my machine.
Stacked on #27299, this PR applies the trait change after the handler
bodies have been outlined.
## What
Changed `ToolExecutor::handle` to return an explicit boxed
`ToolExecutorFuture` instead of using `async_trait`.
Updated ToolExecutor implementors to return `Box::pin(...)`, reexported
the future alias through `codex-tools` and `codex-extension-api`, and
removed `codex-tools` direct `async-trait` dependency.
## Why
We're now [discouraging use of
`async_trait`](https://github.com/openai/codex/pull/20242).
Removing use of `async_trait` from `ToolExecutor` yields a `codex_core`
debug test build speedup of ~78% (from 227.5s to 50.3s) on my machine.
For ease of reviewing, this is a prefactor to extract trait method
implementations to inherent methods. This will prevent changing
indentation from creating a huge diff.
## What
Outlined existing `ToolExecutor::handle` bodies into inherent async
`handle_call` methods across core and extension tool handlers.
The trait methods still use `async_trait` and now delegate to
`self.handle_call(...).await`; handler behavior is unchanged.
## Summary
We originally addressed startup prewarming holding the read side of
`RwLock<McpConnectionManager>` by snapshotting tool-list state. Review
feedback identified the broader ownership problem: the outer
synchronization should only publish or retrieve the current manager,
while MCP operations rely on the manager's internal synchronization. A
follow-up preserved operation retirement with a separate gate, but
further review questioned whether that synchronization was actually
required and whether we could support latest-wins replacement instead.
This PR now stores the current MCP manager in `ArcSwap`. Each operation
uses `load_full()` to obtain an owned `Arc<McpConnectionManager>`, then
performs MCP I/O without retaining the publication mechanism. Refresh
cancels obsolete startup work, constructs a replacement, and atomically
publishes it. New operations see the latest manager, while operations
that already loaded the previous manager retain a valid handle. Refresh
happens at a turn boundary, so there should be no active user tool calls
to drain.
Git history supports dropping the outer `RwLock`. It was introduced in
`03ffe4d595` on November 17, 2025 for non-blocking MCP startup: the
session published an empty manager, startup initialized that same object
while holding the write lock, and readers waited for initialization.
`7cd2e84026` on February 19, 2026 removed that two-phase initialization
in favor of constructing a fresh manager and swapping it in, explicitly
noting that `Option` or `OnceCell` could replace the placeholder design.
Hot reload later reused the existing lock to publish a replacement, but
I found no indication that the lock was introduced to guarantee
in-flight tool calls finish before refresh or shutdown.
Terminal shutdown remains separate from refresh: it aborts startup
prewarming and active tasks before shutting down the current manager, so
tool calls may be interrupted and no model WebSocket work continues
after shutdown. Focused regression coverage exercises pending tool-list
cancellation, deferred refresh, and startup-prewarm shutdown.
## Why
Extension contributors are registered behind `dyn Trait` objects, so
native `async fn`/RPITIT methods would make these traits
non-object-safe. Spell out the boxed, `Send` future contract directly so
`extension-api` no longer needs `async-trait` while retaining the
existing runtime model.
## What changed
- add a shared `ExtensionFuture` alias and use it for asynchronous
contributor methods
- migrate production and test implementations to return `Box::pin(async
move { ... })`
- remove `async-trait` dependencies where they are no longer used,
keeping it dev-only where unrelated test executors still require it
## Behavior
No behavior change is intended. Contributor futures remain boxed,
`Send`, dynamically dispatched, and lazily executed; cancellation and
callback ordering stay unchanged.
## Testing
- `just test -p codex-extension-api` (11 passed)
- affected extension crates (64 passed)
- targeted `codex-core` contributor tests (14 passed)
- `just fmt`
- `just bazel-lock-update`
- `just bazel-lock-check`
A broad local `codex-core` run compiled successfully but encountered
unrelated sandbox and missing test-binary fixture failures; CI will run
the full checks.
## Summary
- tag legacy multi-agent spawn metrics with `version=v1`
- tag multi-agent v2 spawn metrics with `version=v2`
## Why
`codex.multi_agent.spawn` is emitted by both runtimes, so the existing
metric cannot distinguish v2 adoption from aggregate multi-agent
spawning. The bounded version tag makes that breakdown directly
queryable without changing the counter's success-only semantics.
## Validation
- `just fmt`
- `git diff --check`
- Tests and Clippy were intentionally left to CI.
## Stack
- Base: #27191
- This PR is the third vertical and should be reviewed against
`jif/external-plugins-2`, not `main`.
## Why
#27191 moves the host-owned Apps MCP registration behind an extension
contributor, but deliberately preserves the existing endpoint-selection
feature while that contribution contract lands. App-server can therefore
resolve the server through extensions, yet the hosted plugin endpoint is
still selected through temporary `apps_mcp_path_override` plumbing.
That is not the long-term plugin model. A plugin can bundle skills,
connectors, MCP servers, and hooks, and those components do not all need
the same source or execution environment. In particular, an
authenticated HTTP MCP server can expose plugin capabilities directly
from a backend without an executor or an orchestrator filesystem.
This PR completes that hosted vertical. App-server's MCP extension now
owns the aggregate hosted plugin runtime at `/ps/mcp`. Connector actions
continue to arrive as MCP tools, while backend-provided skills arrive as
MCP resources and use Codex's existing resource list/read paths. No
second backend client, skill filesystem, or generic plugin activation
framework is introduced.
The backend route remains the hosted implementation. This change
replaces Codex's temporary endpoint-selection mechanism, not the service
behind the endpoint.
## What changed
### Hosted plugin runtime
The MCP extension now contributes `codex_apps` as the hosted plugin
runtime rather than as a configurable Apps endpoint:
- `https://chatgpt.com` resolves to
`https://chatgpt.com/backend-api/ps/mcp`;
- a bare custom ChatGPT base resolves to `/api/codex/ps/mcp`;
- the existing product-SKU header and ChatGPT authentication behavior
are preserved;
- executor availability is never consulted for this streamable HTTP
transport.
The same MCP connection carries both component shapes supported by the
hosted endpoint:
- connector actions are discovered and invoked as MCP tools;
- hosted skills are enumerated and read as MCP resources through the
existing `list_mcp_resources` and `read_mcp_resource` paths.
This keeps component access in the subsystem that already owns the
protocol instead of downloading backend skills into an orchestrator
filesystem or inventing a parallel hosted-skill client.
### Explicit runtime ordering
`McpManager` now resolves the reserved `codex_apps` entry in three
ordered phases:
1. install the legacy Apps fallback for compatibility;
2. apply ordered extension `Set` or `Remove` overlays;
3. apply the final ChatGPT-auth gate without synthesizing the server
again.
This ordering is important:
- an ordinary configured or plugin MCP server cannot claim the
auth-bearing `codex_apps` name;
- an extension-contributed hosted runtime wins over the fallback;
- an extension `Remove` remains authoritative;
- a host without the MCP extension retains the legacy Apps endpoint and
current local-only behavior.
The temporary `legacy_apps_mcp_loader_enabled` coordination flag is no
longer needed.
### Remove the path override
The `apps_mcp_path_override` feature and its runtime plumbing are
removed, including:
- the feature registry entry and structured feature config;
- `Config` and `McpConfig` fields;
- config schema output;
- config-lock materialization;
- URL override handling in `codex-mcp`.
Existing boolean and structured forms still deserialize as ignored
compatibility input. They are omitted from new serialized config, and
config-lock comparison normalizes the removed input so older locks
remain replayable.
### App-server coverage
App-server MCP fixtures now serve the hosted route at
`/api/codex/ps/mcp`. Existing resource-read and tool/elicitation flows
therefore exercise the extension-owned endpoint rather than succeeding
through the legacy fallback.
The stack also adds the missing `codex_chatgpt::connectors` re-export
for the manager-backed connector helper introduced in #27191.
## Compatibility
- App-server installs the extension and uses `/ps/mcp` for the hosted
runtime.
- CLI and other hosts that do not install the extension retain the
legacy Apps endpoint.
- Apps disabled or non-ChatGPT authentication removes `codex_apps` from
the effective runtime view.
- Existing local plugins, local skills, executor-selected skills,
configured MCP servers, and MCP OAuth behavior are otherwise unchanged.
- Backend plugin enablement remains account/workspace state owned by the
hosted endpoint; this PR does not add thread-local backend plugin
selection.
## Architectural fit
The stack now proves two independent runtime shapes:
1. #27184 resolves filesystem-backed skills through the executor that
owns a selected root.
2. #27191 and this PR resolve a backend-hosted HTTP MCP through an
extension with no executor.
Together they preserve the intended separation:
- selection identifies a plugin/root when explicit selection is needed;
- each component's owning extension resolves its concrete access
mechanism;
- execution stays with the runtime required by that component;
- existing skills, MCP, connector, and hook subsystems remain the
downstream consumers.
## Planned follow-ups
1. **Executor stdio MCP:** selecting an executor plugin registers a
manifest-declared stdio MCP server and executes it in the environment
that owns the plugin.
2. **Optional backend selection:** only if CCA needs thread-local
selection distinct from backend account/workspace enablement, add a
concrete backend-owned capability location and surface those selected
skills through the skills catalog.
3. **Connector metadata and hooks:** activate those plugin components
through their existing owning subsystems, with executor hooks remaining
environment-bound.
4. **Propagation and persistence:** define explicit resume, fork,
subagent, refresh, and environment-removal semantics once selected roots
have multiple real consumers.
5. **Local convergence:** migrate legacy local skill, MCP, connector,
and hook paths behind their owning extensions one vertical at a time,
then remove duplicate core managers and compatibility plumbing after
parity.
## Verification
Coverage in this change exercises:
- extension-owned `/backend-api/ps/mcp` registration without an
executor;
- preservation of the legacy endpoint in hosts without the extension;
- extension `Set` and `Remove` precedence over the legacy fallback;
- ChatGPT-auth gating for the reserved server;
- hosted MCP resource reads with and without an active thread;
- connector tool invocation and MCP elicitation through the hosted
route;
- ignored boolean and structured forms of the removed path override;
- config-lock replay compatibility for the removed feature.
`cargo check -p codex-features -p codex-mcp-extension -p
codex-app-server` passes. Tests and Clippy were not run locally under
the current development instruction; CI provides the full validation
pass.
## Why
- Local profiling shows `append_tool_search_executor` averages ~113ms
per call. Adding a span lets us track this cost as we optimize in
follow-up PRs, either by reducing the work or avoiding repeated rebuilds
when inputs have not changed.
- While we're here, we can add spans to `build_tool_router` and other
sub-calls which code analysis shows may have additional opportunities
for improvement.
## What changed
Add function-level trace spans around `build_tool_router`,
`build_tool_specs_and_registry`, `add_tool_sources`,
`append_tool_search_executor`, and
`build_model_visible_specs_and_registry`
## Verification
Trigger Codex rollout and observe new spans are included
## Why
Multi-agent v2 identifies agents by canonical paths, but its tool
handlers still emitted the larger legacy collaboration begin/end events
built around nickname and role metadata. App-server, rollout-trace,
analytics, and TUI consumers therefore lacked one compact path-based
completion signal that behaved consistently across live events and
replay.
The TUI also needs a bounded `/agent` status surface for v2 agents. It
should use recent local activity for previews, refresh liveness without
loading full histories, and keep the legacy picker available when no
path-backed v2 agent is known.
## What changed
- Replace the v2 `spawn_agent`, `send_message`, `followup_task`, and
`interrupt_agent` legacy lifecycle emissions with a success-only
`SubAgentActivity` event. The event records the tool call ID, occurrence
time, affected thread, canonical agent path, and `started`,
`interacted`, or `interrupted` kind.
- Expose the activity as a completion-only app-server v2
`subAgentActivity` thread item in live notifications and reconstructed
history, regenerate the protocol schemas, and count it in sub-agent tool
analytics.
- Track canonical paths from live activity and loaded-thread metadata in
the TUI, and render the activity in live and replayed transcripts.
- Make `/agent` list running path-backed agents with summaries from
bounded local event buffers. Each summary is capped at 240 graphemes,
the scan is capped at six recent items, only the last three wrapped
lines are shown, and command output is omitted. Liveness falls back to
metadata-only `thread/read` when local turn state is unavailable.
- Persist the activity as a terminal rollout-trace runtime payload and
reduce it to the corresponding spawn, send, follow-up, or close
interaction edge. `interrupt_agent` is classified as a close-edge
operation.
- Preserve the legacy picker when no path-backed v2 agent is known.
## Compatibility
App-server v2 clients that consumed `collabAgentToolCall` begin/end
pairs for these tools must handle the new completion-only
`subAgentActivity` item. Legacy v1 collaboration behavior is unchanged.
## Screenshot
<img width="684" height="288" alt="Screenshot 2026-06-08 at 15 40 47"
src="https://github.com/user-attachments/assets/194b3cd0-619d-45fb-b587-cf3e2b1b8a1d"
/>
## Testing
- `just test -p codex-app-server-protocol`
- `just test -p codex-rollout-trace`
- Added focused coverage for activity analytics, terminal trace
serialization, spawn-edge reduction, `interrupt_agent` classification,
TUI status rendering without aggregated command output, and clearing
stale running state after a completed turn.
## Why
Thread cwd and environment selections are a single logical setting in
core: updating one without the other can silently desynchronize the
next-turn execution context. This change makes that relationship
explicit in the internal thread settings flow while preserving the
existing app-server public API shape.
## What changed
- Moved the cwd/environment pair through internal
`ThreadSettingsOverrides.environment_settings` instead of a top-level
internal `cwd` field.
- Kept `thread/settings/update` public params unchanged, with app-server
translating top-level `cwd` into the paired internal settings shape.
- Moved `Op::UserInput` environment overrides into thread settings so
user turns and settings updates use the same core path.
- Updated core, app-server, MCP, memories, sample, and test callsites to
construct the paired settings shape.
## Verification
- `git diff --check`
- Local test run starting after PR creation.
## Summary
- tighten the default multi-agent v2 root and subagent usage hints to
bias toward local work
- add a pre-call gate to the v2 spawn_agent description for independent,
bounded, parallelizable subtasks
## Validation
- just fmt
- started just test -p codex-core, but it was interrupted before
completion per follow-up request to commit and push immediately
## Why
Auto Review should remain the effective approval reviewer when settings
cross runtime boundaries. A config or app-server round trip must not
change the reviewer identity, and delegated work must not silently fall
back to user review.
This requires both a stable canonical serialized value and propagation
of the effective setting. `auto_review` is the canonical value across
protocol and app-server output, while `guardian_subagent` remains
accepted as backward-compatible input.
## What changed
- serialize `ApprovalsReviewer::AutoReview` consistently as
`auto_review` across core protocol and app-server v2
- continue accepting `guardian_subagent` when reading existing config or
client requests
- carry the active turn's approval reviewer into spawned agents
- update config/debug expectations and add delegated-task regression
coverage
## Scope
This does not change Guardian policy or remove compatibility with
existing `guardian_subagent` inputs. It preserves the selected reviewer
across serialization, config reloads, app-server settings, and delegated
task setup.
Related Guardian changes are split independently:
- #26231 adds denials and soft denials
- #26334 retries transient reviewer failures
- #26333 reuses narrowly scoped low-risk approvals
- #26232 adds TUI denial recovery
## Validation
- `just test -p codex-app-server-protocol` (224 passed)
- regression coverage for delegated task reviewer propagation
- serialization coverage for canonical `auto_review` output and legacy
`guardian_subagent` input
---------
Co-authored-by: saud-oai <saud@openai.com>
## Summary
- add contains_external_context() to tool output so other tools can be
opted out of influencing memory when disable_on_external_context=true
- Classify standalone web-search output as external context (to match
behavior as hosted web search)
- Verify with integration test
## Why
`close_agent` is the wrong model-facing name for the v2 operation after
the residency changes. V2 agents remain reusable by task name, and
residency/unloading owns capacity management; the exposed tool should
describe the action it actually performs: interrupt the target agent's
current turn without making the agent unavailable for future messages or
follow-up tasks.
## What changed
- Rename the multi-agent v2 tool from `close_agent` to
`interrupt_agent`.
- Keep the v1 `close_agent` surface unchanged.
- Update the v2 handler to send `Op::Interrupt`, keep interrupted agents
registered, and reject root/self targets with interrupt-specific errors.
- Route interrupt delivery through the existing dead-thread cleanup path
so stale resident entries do not keep consuming capacity.
- Update tool planning and handler tests for the new v2 surface and
semantics.
## Verification
Added focused coverage in:
- `core/src/tools/spec_plan_tests.rs`
- `core/src/tools/handlers/multi_agents_tests.rs`
## Why
This PR fixes approval sandbox semantics in the unified-exec path. The
zsh-fork runtime exposed the bug because the shell can do meaningful
work before any intercepted child `execv(2)` exists: redirections,
builtins, globbing, and pipeline setup all happen in the launch process.
If the model requested `sandbox_permissions=require_escalated`, or an
exec-policy `allow` rule explicitly bypassed the sandbox, that approved
sandbox decision needs to be preserved for the launch path and for
intercepted execs that use the same approval machinery.
The behavior is not only about zsh fork. The production changes are in
shared approval/escalation code, so they also affect non-zsh-fork
intercepted exec paths that go through the same sandbox decision logic.
The narrow intent is to preserve the approval decision while still
keeping denied-read profiles and bounded additional-permission requests
sandboxed.
## Production Changes
- `codex-rs/core/src/tools/runtimes/unified_exec.rs`: derives a
`launch_sandbox_permissions` value from the requested sandbox
permissions and the runtime filesystem policy, then uses that value for
managed-network/env setup and launch sandbox selection. This keeps full
approval or policy-bypass decisions visible to the first unified-exec
attempt, while still preventing a full sandbox override from discarding
denied-read restrictions. Direct unified exec keeps the same decision
surface; the important difference is that zsh-fork launch setup no
longer accidentally loses the approved parent sandbox decision.
- `codex-rs/core/src/tools/runtimes/shell/unix_escalation.rs`: makes
intercepted-exec escalation selection explicit for the three sandbox
permission modes. `UseDefault` only escalates when an exec-policy
decision allows sandbox bypass, `RequireEscalated` escalates when
unsandboxed execution is allowed, and `WithAdditionalPermissions`
escalates through the bounded additional-permissions path instead of
being treated as a full unsandboxed override. Unsandboxed intercepted
execs now also rebuild the environment as `RequireEscalated`, which
strips managed-network proxy variables consistently with other
unsandboxed execution.
## Test Coverage
Most of the PR is tests. The new coverage verifies:
- unified exec preserves parent approval and exec-policy sandbox
decisions for zsh-fork launch selection;
- bounded `with_additional_permissions` remains sandboxed and
permission-profile based;
- denied-read profiles are not weakened by parent approval;
- explicit prompt rules still prompt for intercepted execs after the
parent command is approved;
- unsandboxed intercepted execs strip managed-network env vars.
No documentation update is needed; this is an internal approval/sandbox
correctness fix.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/24981).
* #24982
* __->__ #24981
## Summary
Responses Lite does not execute hosted Responses tools, so models using
it must route web search and image generation through Codex-owned
executors & standalone Response's API endpoints.
This PR is stacked on #26487.
## Validation
- `cargo test -p codex-core responses_lite_ --lib`
- `cargo test -p codex-core
standalone_executors_remain_hidden_without_flags_or_responses_lite
--lib`
- `cargo test -p codex-core
hosted_tools_follow_provider_auth_model_and_config_gates --lib`
- `cargo test -p codex-web-search-extension -p
codex-image-generation-extension`
- `cargo test -p codex-app-server --test all standalone_`
- `cargo fmt --all -- --check`
## Stack
1. [#26547](https://github.com/openai/codex/pull/26547) - [1 of 2] Align
goal extension with core behavior
2. [#26548](https://github.com/openai/codex/pull/26548) - [2 of 2] Move
goal runtime to extension
## Why
This PR completes the switch of the goal behavior to the
extension-backed runtime and removes the old core goal implementation.
## What Changed
- Installs the goal extension for app-server `ThreadManager` sessions.
- Routes app-server thread goal `get`, `set`, and `clear` through
`GoalService`.
- Uses thread-idle lifecycle emission after goal resume and snapshot
ordering so the extension can decide whether to continue the goal.
- Forwards extension goal updates through a FIFO async app-server
notification path so backpressure does not drop them or reorder updates.
- Keeps review turns from enabling goal runtime behavior.
- Plans extension tools before dynamic tools so built-in goal tool names
keep their old precedence when goals are enabled.
- Removes the old core goal runtime, core goal tool handlers, and core
goal tool specs.
- Updates tests that were coupled to the core-owned goal runtime while
leaving the legacy `<goal_context>` compatibility path in core for old
threads.
- Removes the stale cargo-shear ignore now that `codex-goal-extension`
is used by the workspace.
- Keeps realtime event matching exhaustive after removing the old
goal-specific realtime text path.
## Validation
- Ran manual `/goal` runs in TUI. Validated time accounting matched
wall-clock time and goal lifecycle state transitions.
## Why
Windows managed filesystem permissions can now be backed by a real
Windows sandbox. `exec-policy` was still treating the managed read-only
policy shape as if there were never a sandbox backend, so benign
unmatched commands such as PowerShell directory listings could be
rejected with `blocked by policy` even when `windows.sandbox` was
enabled.
The inverse case still needs to stay conservative: when the Windows
sandbox backend is disabled, managed filesystem restrictions are only
configuration intent, not an enforced filesystem boundary. That applies
to writable-root restricted profiles too, not just read-only profiles.
## What Changed
- Thread the effective `WindowsSandboxLevel` into exec-policy approval
decisions for shell, unified exec, and intercepted shell exec paths.
- Treat managed restricted filesystem profiles as lacking sandbox
protection only on Windows when `WindowsSandboxLevel::Disabled`.
- Exclude full-disk-write profiles from that no-backend path because
they do not rely on filesystem sandbox enforcement.
- Remove the cwd-sensitive read-only heuristic and the now-stale cwd
plumbing from exec-policy approval contexts.
- Add Windows coverage for both enabled-sandbox and disabled-backend
behavior, including a writable-root managed profile.
## Validation
- Added/updated `exec_policy` coverage for managed filesystem
restrictions, full-disk-write exclusion, enabled Windows sandbox
behavior, and disabled-backend read-only/writable-root behavior.
- `just test -p codex-core exec_policy` — 100 passed, 10 leaky
- Empirical local `codex exec` probe with `--sandbox read-only -c
'windows.sandbox="unelevated"'`: PowerShell directory listing completed
successfully.
- Disabled-backend control with Windows sandbox cleared: the same
command was rejected with `blocked by policy`.
## Why
Turn diffs were tracked as one flat set of absolute paths. In
multi-environment turns, local and remote environments can report the
same path while representing different filesystems, so a single path key
can collapse distinct changes or attribute them to the wrong
environment.
The environment name is **NOT** included in the generated unified diff.
This can come later.
## Summary
This is the first small step toward making multi-agent v2 agents durable
logical agents whose `ThreadManager` residency is only an implementation
detail.
This PR adds a narrow v2 reload-on-delivery hook:
- If a known v2 agent target is already loaded, delivery is unchanged.
- If the target is still registered but missing from `ThreadManager`,
delivery reloads that exact v2 thread from durable rollout history
before submitting the message.
- If the target is unknown, closed, missing from storage, or not a v2
thread, delivery still fails as not found.
The reload is wired only into existing-agent delivery paths: v2
`send_message` / `followup_task`, and legacy `send_input` when its
target is a known v2 agent.
## Stack
1. **Reload on delivery**: load known unloaded v2 agents before
`followup_task`, `send_message`, or `send_input` delivery. This PR.
2. **Residency LRU**: unload idle resident v2 agents from
`ThreadManager` without making them closed or unreachable.
3. **Execution concurrency**: count active non-root turns, not logical
agents or resident idle threads.
4. **Close semantics**: make v2 close interrupt-only and leave durable
agent identity intact.
5. **Resume cleanup**: remove user-facing v2 resume semantics;
addressing an unloaded durable agent reloads it implicitly.
## Validation
- Ran `just fmt`.
- Left broader tests and clippy to CI.
## Summary
- keep the legacy v1 `spawn_agent` role and model selectors visible
- add regression coverage for the default v1 tool plan
## Why
`hide_spawn_agent_metadata` is a multi-agent v2 setting, but the v1
planning branch also consumed it. After the default changed to `true`,
v1 stopped advertising `agent_type`, `model`, `reasoning_effort`, and
`service_tier`, preventing configured agents from being selected.
This keeps the hidden-metadata default for v2 while opting v1 out of
that behavior.
Fixes#26363.
## Validation
Not run locally, per request; CI will validate the change.
## Why
Multi-agent v2 currently routes agent instructions through normal tool
arguments and inter-agent context. That means the parent model can emit
plaintext task text, Codex can persist it in history/rollouts, and the
recipient can receive it as ordinary assistant-message JSON.
This changes the v2 path so agent instructions stay encrypted between
model calls: Responses encrypts the `message` argument returned by the
model, Codex forwards only that ciphertext, and Responses decrypts it
internally for the recipient model.
## What changed
- Mark the v2 `message` parameter as encrypted for `spawn_agent`,
`send_message`, and `followup_task`.
- Treat multi-agent v2 tool `message` values as ciphertext
unconditionally.
- Store v2 inter-agent task text in
`InterAgentCommunication.encrypted_content` with empty plaintext
`content`.
- Convert encrypted inter-agent communications into the Responses
`agent_message` input item before sending the child request.
- Preserve `agent_message` items across history, rollout, compaction,
telemetry, and app-server schema paths.
- Leave multi-agent v1 unchanged.
## Message shape
The model still calls the v2 tools with a `message` argument, but that
value is now ciphertext:
```json
{
"name": "spawn_agent",
"arguments": {
"task_name": "worker",
"message": "<ciphertext>"
}
}
```
Codex stores the task as encrypted inter-agent communication:
```json
{
"author": "/root",
"recipient": "/root/worker",
"content": "",
"encrypted_content": "<ciphertext>",
"trigger_turn": true
}
```
When Codex builds the recipient request, it forwards the ciphertext
using the new Responses input item:
```json
{
"type": "agent_message",
"author": "/root",
"recipient": "/root/worker",
"content": [
{
"type": "encrypted_content",
"encrypted_content": "<ciphertext>"
}
]
}
```
Responses decrypts that item internally for the recipient model.
## Context impact
- Parent context no longer carries plaintext v2 agent task instructions
from these tool arguments.
- Codex rollout/history stores ciphertext for v2 agent instructions.
- Recipient requests receive an `agent_message` item instead of
assistant commentary JSON for encrypted task delivery.
- Plaintext completion/status notifications are still plaintext because
they are Codex-generated status messages, not encrypted model tool
arguments.
## Validation
- `just test -p codex-tools`
- `just test -p codex-protocol`
- `just test -p codex-rollout`
- `just test -p codex-rollout-trace`
- `just test -p codex-otel`
- `just write-app-server-schema`
AbhinavandGitHub
Owen LinandGitHub
mchen-oaiandGitHub
jifandGitHub
pakrym-oaiandGitHub
sayan-oaiandGitHub
Adam Perry @ OpenAIandGitHub
Shijie RaoandGitHub
Won ParkandGitHub
Curtis 'Fjord' HawthorneandGitHub
Jeremy RoseandGitHub
xl-openaiandGitHub
Alex DaleyandGitHub
Charlie MarshandGitHub
rka-oaiandGitHub
Michael BolinandGitHub
Eric TrautandGitHub
iceweasel-oaiandGitHub