## Summary
- add a default-on `auto_compaction` feature flag as an internal escape
hatch
- skip pre-turn, model-switch/hash, and mid-turn automatic compaction
when the flag is disabled
- preserve manual `/compact` behavior and surface the existing
context-window error when the provider runs out of room
- add integration coverage for disabled pre-turn and mid-turn compaction
## Motivation
Long-running SPO optimization rollouts need the option to preserve their
full context and fail on context exhaustion instead of entering another
compaction window. This deliberately uses the existing feature-flag
mechanism rather than adding a dedicated public config or app-server
API.
Disable it with:
```sh
codex --disable auto_compaction
```
## Testing
- `just test -p codex-features` — 51 passed
- `just test -p codex-core auto_compaction_feature_disabled` — 2 passed
- `just fix -p codex-core -p codex-features`
- `just write-config-schema`
- `just test -p codex-core` — the new compaction tests passed; the
overall local run had 54 unrelated environment failures, primarily
missing first-party test binaries and shell-snapshot timeouts
## Summary
- expose `clock.curr_time` when current-time reminders are enabled
- query the session's configured time provider with the calling thread
id
- return the existing UTC reminder text for direct model calls
- return `{ "current_time": "YYYY-MM-DD HH:MM:SS UTC" }` in Code Mode
Clock lookup failures remain fatal, matching pre-inference reminder
behavior.
## Testing
- `just test -p codex-core current_time_tool_returns_the_latest_time`
- `just test -p codex-core
code_mode_current_time_returns_structured_result`
- `just fix -p codex-core`
## Why
Som tools, such as history and notes, must remain top-level when MCP
deferral is enabled while staying unavailable through code-mode `exec`.
## What changed
- Added `features.code_mode.direct_only_tool_namespaces`.
- Classified matching MCP tools as `DirectModelOnly`.
- Kept those tools top-level in `code_mode_only`.
- Excluded them from `tool_search` deferral and the nested `exec`
surface.
- Updated the generated config schema.
## Validation
- `code_mode_only_exposes_direct_model_only_mcp_namespaces`
- `load_config_resolves_code_mode_config`
## Why
`TurnContext` had accumulated dead fields and cached projections of
values already owned by its per-turn `Config` or `ModelInfo`. Keeping
both copies made ownership unclear and allowed artificial split-brain
states, such as a compatibility hash differing from the model metadata
it came from.
`Prompt` similarly carried a write-only personality after personality
selection had already been materialized into its base instructions.
This makes the canonical owner explicit: configuration-backed values
come from `config`, model-derived values come from `model_info`, and
prompts contain only data consumed by request construction.
## What changed
- Remove the unused `ghost_snapshot`, `codex_self_exe`, and
`thread_source` fields.
- Remove duplicate `comp_hash`, `truncation_policy`, `features`,
`shell_environment_policy`, `codex_linux_sandbox_exe`, `compact_prompt`,
and `tool_mode` fields.
- Read those values directly from `TurnContext::config` or
`TurnContext::model_info` at their consumers.
- Remove the write-only `Prompt::personality` field and its constructor
assignments.
- Preserve review-turn inheritance of the parent turn's shell policy,
Linux sandbox executable, and compact prompt through the review config.
## Testing
- `cargo check -p codex-core --tests`
Summary
- Add list-backed and developer-context presentations for plugin
suggestion candidates.
- Let tool planning, install validation, and request-tool copy follow
the selected presentation.
- Keep every production caller on the existing list-backed presentation,
preserving the current list tool, request schema, connector behavior,
and model-visible copy.
- Leave developer-context presentation latent until the final PR in the
stack.
Stack
- 2/3, based on #28399.
- Follow-up: #27704 activates endpoint recommendations.
Validation
- `just test -p codex-core request_plugin_install`
- `just test -p codex-core spec_plan`
- `just fix -p codex-core`
- `just fmt`
- `git diff --check`
## Why
Models sometimes need to pause briefly while waiting for external work,
but using a shell command for that delay ties the wait to a process and
does not naturally resume when new turn input arrives.
## What changed
- add a built-in `sleep` tool behind the under-development `sleep_tool`
feature
- accept a bounded `duration_ms` argument, matching the millisecond
convention used by unified exec
- end the sleep early when either steered user input or mailbox input
arrives
- include elapsed wall-clock time in completed and interrupted outputs
- emit a dedicated core `SleepItem` through `item/started` and
`item/completed`
- expose the sleep item as app-server v2 `ThreadItem::Sleep` and retain
it in reconstructed thread history
- regenerate the configuration schema for the new feature flag
- regenerate app-server JSON and TypeScript schema fixtures
## Test plan
- `just test -p codex-core sleep_tool_follows_feature_gate`
- `just test -p codex-core any_new_input_interrupts_sleep`
- `just test -p codex-app-server-protocol`
- `just test -p codex-app-server
sleep_emits_started_and_completed_items`
## 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
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`
## 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
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
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
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`
## 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
- 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
`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`
## 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.
## 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
Research and training setups need to control which tool namespaces
appear inside code mode's nested `tools` surface without disabling those
tools entirely. This makes it possible to train against a deliberately
reduced nested-tool setup while preserving the normal direct and
deferred tool paths.
## What
- Extend `features.code_mode` to accept structured configuration while
preserving the existing boolean syntax.
- Add an exact `excluded_tool_namespaces` list under
`[features.code_mode]`:
```toml
[features.code_mode]
enabled = true
excluded_tool_namespaces = ["mcp__codex_apps", "multi_agent_v1"]
```
- Filter matching canonical `ToolName` namespaces when constructing code
mode's nested router and code-mode-specific direct tool descriptions.
- Keep excluded tools registered, directly exposed in mixed code mode,
and discoverable through top-level `tool_search` when otherwise
eligible.
- Derive deferred nested-tool guidance after namespace filtering so the
`exec` description does not advertise excluded-only deferred tools.
- Preserve the boolean/table representation when materializing config
locks and update the generated config schema.
## Testing
- `just test -p codex-features`
- `just test -p codex-config`
- `just test -p codex-core load_config_resolves_code_mode_config`
- `just test -p codex-core
lock_contains_prompts_and_materializes_features`
- `just test -p codex-core
excluded_deferred_namespaces_do_not_enable_nested_tool_guidance`
- `just test -p codex-core
code_mode_excludes_configured_nested_tool_namespaces`
- `cargo check -p codex-thread-manager-sample`
## Why
`code_mode_only` moved ordinary runtime tools behind `exec`, but it also
hid hosted Responses tools. Hosted `web_search` and `image_generation`
do not have a nested `exec` runtime path, so code-only sessions lost
those capabilities entirely even when their existing provider, auth,
model, and configuration gates passed.
## What changed
- Keep hosted Responses tools top-level in `code_mode_only` sessions
after their existing gates pass.
- Preserve the existing nested-tool behavior for ordinary runtimes and
the direct-only behavior for multi-agent v2 tools.
- Add planner coverage for `code_mode_only` with default multi-agent v2
settings, hosted live web search, and hosted image generation.
## Verification
- Added focused regression coverage in
`codex-rs/core/src/tools/spec_plan_tests.rs`.
- Left execution to CI per repository workflow.
Stack split from #25708. Original PR intentionally left open. This third
PR resolves the effective per-thread multi-agent runtime from persisted
metadata, inherited runtime, and current model selection.
Deferred tools need to be searchable even when they are not implemented
inside `codex-core`. Extension-provided tools can be registered for
later discovery, but the search metadata path was still owned by
core-specific runtime hooks, which meant the shared `ToolExecutor`
abstraction could not describe how a deferred extension tool should
appear in `tool_search`.
## Changes
- Move `ToolSearchEntry` and `ToolSearchInfo` into `codex-tools` and
re-export them from the shared tools crate.
- Add a default `ToolExecutor::search_info` implementation that derives
loadable tool-search metadata from function and namespace specs.
- Forward search metadata through extension adapters and exposure
overrides while keeping custom search text/source metadata for dynamic,
MCP, and multi-agent tools.
- Remove the old core-local `tool_search_entry` module now that search
metadata lives with the shared executor APIs.
## Testing
- Added `deferred_extension_tools_are_discoverable_with_tool_search`
coverage in `core/src/tools/spec_plan_tests.rs`.
## Why
When unified exec is configured to launch through the zsh fork, local
commands should not let the model override the shell binary with the
`shell` parameter. The configured zsh fork is the mechanism that makes
`execv(2)` interception reliable, so exposing `shell` for local zsh-fork
execution would create a confusing API surface and undermine the
composition.
Remote environments are different: zsh-fork interception is local-only,
so remote unified-exec calls must keep direct unified-exec behavior and
still expose `shell` when a remote environment can be selected.
## What Changed
- Taught the `exec_command` schema builder to omit the `shell` parameter
when requested.
- Hid `shell` from the unified-exec tool schema only when zsh-fork
unified exec applies to all selectable environments.
- Kept `shell` visible when any remote environment can be targeted,
because those calls run through direct unified exec.
- Made unified exec choose the effective shell mode per selected
environment: local environments keep zsh-fork mode, remote environments
use direct mode.
- Left direct unified-exec behavior unchanged, including support for
model-specified shells there.
## Verification
- Added schema coverage showing `exec_command` can hide `shell`.
- Added planner coverage showing zsh-fork unified exec hides `shell` for
local-only execution while direct unified exec still exposes it.
- Added planner coverage showing `shell` remains visible when a remote
environment is available.
- Added handler coverage showing remote environments use direct
unified-exec shell mode instead of zsh-fork mode.
- Ran the focused `codex-core` shell-parameter and zsh-fork tests.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/24980).
* #24982
* #24981
* __->__ #24980
## Summary
Renames the MultiAgentV2 turn-triggering tool from `assign_task` to
`followup_task` so the exposed tool name better describes sending an
additional task to an existing agent.
This updates the tool spec, handler/module names, registry wiring,
default multi-agent v2 usage hints, and tests. Rollout trace
classification keeps accepting legacy `assign_task` events so older
traces still reduce correctly, while docs show the new tool name.
## Test plan
- `just test -p codex-core followup_task`
- `just test -p codex-core -E
'test(multi_agent_feature_selects_one_agent_tool_family) |
test(multi_agent_v2_can_use_configured_tool_namespace) |
test(code_mode_only_can_expose_namespaced_multi_agent_v2_as_normal_tools)'`
- `just test -p codex-rollout-trace`
- `just fix -p codex-core`
- `just fix -p codex-rollout-trace`
Notes: `just fmt` ran `cargo fmt` but failed in the Python ruff phase
because the local environment could not resolve `hatchling>=1.27.0` from
the configured internal registry. A full `just test -p codex-core` also
hit unrelated environment-sensitive integration failures involving
missing spawned test binaries/sandbox behavior; the changed multi-agent
spec/handler tests passed in the filtered runs above.
## Summary
- rename the multi-agent v2 follow-up task tool surface to assign_task
- update core tests and spec-plan expectations
- keep rollout-trace classification backward-compatible with legacy
followup_task
## Tests
- just fmt
- just test -p codex-core
multi_agents_spec::tests::assign_task_tool_requires_message_and_has_no_output_schema
- just test -p codex-rollout-trace
- just fix -p codex-core
- just fix -p codex-rollout-trace
Note: a broad just test -p codex-core run was attempted locally, but
this sandbox produced unrelated environment failures around
sandbox-exec, missing test_stdio_server, and realtime timeouts.
## Summary
- Keep the original `TOOL_SUGGEST_DISCOVERABLE_PLUGIN_ALLOWLIST` as a
fallback seed list, so users with no installed plugins still get initial
install suggestions.
- Allow additional install suggestions from trusted marketplaces:
`openai-curated` and `openai-bundled`.
- Require non-fallback, non-configured marketplace candidates to share
`.app.json` connector IDs with already installed plugins.
- Preserve explicit configured plugin discoverables as an override,
while still omitting installed, disabled, and `NOT_AVAILABLE` plugins.
## Context
`list_available_plugins_to_install` controls which plugins the model can
trigger via `request_plugin_install`. We want a small starter set for
empty/new users, but we also want installed workflow plugins to unlock
relevant source plugins without maintaining every source plugin ID by
hand.
This keeps the legacy plugin ID allowlist only as the starter fallback.
For everything else, the trusted marketplace is the candidate boundary,
and installed app connector overlap is the relevance filter. For
example, an installed Sales plugin can make HubSpot and Granola
suggestible when those source plugins are in `openai-curated` and share
Sales app connector IDs, while an unrelated test-source plugin with an
app connector not declared by Sales stays hidden.
## Test Coverage
- Empty/no-installed-plugin case: returns the fallback seed plugins from
the original allowlist.
- Installed-app expansion: returns non-fallback marketplace plugins only
when their app connector IDs overlap with an installed plugin.
- Sales workflow case: installed Sales declares HubSpot and Granola
apps, so `hubspot@openai-curated` and `granola@openai-curated` are
returned.
- Sales negative case: `test-source@openai-curated` has an app connector
not declared by Sales, so it is not returned.
- Existing guardrails: installed plugins, disabled suggestions, and
`NOT_AVAILABLE` plugins remain omitted; explicit configured
discoverables still work as an override.
## Validation
- `just fmt`
- `just test -p codex-core plugins::discoverable::tests`
- `just test -p codex-core` was attempted earlier, but current `main` /
local env failed with unrelated existing failures around missing
`test_stdio_server`, CLI/code-mode MCP tool setup, and
unified_exec/shell snapshot flakes/timeouts. The touched discoverable
tests pass.
## Why
Some models need to select their code-execution behavior through model
catalog metadata. Models without that metadata must continue to follow
the existing `CodeMode` and `CodeModeOnly` feature flags, including when
a newer server sends an enum value this client does not recognize.
## What changed
- add optional `ModelInfo.tool_mode` metadata with `direct`,
`code_mode`, and `code_mode_only`
- treat omitted and unknown wire values as `None`
- resolve `None` from the existing feature flags
- carry the resolved `ToolMode` directly on `TurnContext`, outside
`Config`
- use the resolved value for turn creation, model switches, review
turns, tool planning, and code execution
## Coverage
- add protocol coverage for omitted, known, and unknown enum values
- add focused coverage for flag fallback and explicit metadata
overriding feature flags
- add core integration coverage that fetches remote model metadata
through `/v1/models` and verifies the outbound `/responses` tools for
explicit `direct` and `code_mode_only` selectors
## Stack
- followed by #25032
## Why
Add a standalone image generation path that can be exercised
independently of hosted Responses image generation, while retaining the
hosted tool as fallback unless the extension is actually available to
the model.
## What changed
- Added the `codex-image-generation-extension` crate with standalone
generate/edit execution, prior-image selection for edits, model-visible
image output, and local generated-image persistence.
- Installed the extension in app-server behind the disabled-by-default
`imagegenext` feature and backend eligibility checks.
- Updated core tool planning so eligible `image_gen.imagegen` exposure
replaces hosted `image_generation`, while unavailable configurations
retain hosted fallback.
- Added coverage for extension behavior, edit history reuse, feature
gating, auth eligibility, and hosted-tool replacement.
- The extension is installed through app-server only in this PR; other
execution paths retain hosted image generation because hosted
replacement occurs only when the standalone executor is actually
registered and model-visible.
- The initial extension contract intentionally fixes the image model to
`gpt-image-2` and uses automatic image parameters.
- Native generated-image history/card parity and rollout persistence
cleanup are intentionally deferred follow-up work.
## Validation
- `just test -p codex-image-generation-extension`
- `just test -p codex-features`
- `just test -p codex-core
hosted_tools_follow_provider_auth_model_and_config_gates`
- `just test -p codex-app-server`
- `just fix -p codex-image-generation-extension -p codex-features -p
codex-core -p codex-app-server`
- `just fmt`
- `just bazel-lock-update`
- `just bazel-lock-check`
---------
Co-authored-by: jif-oai <jif@openai.com>
## Summary
Add the extension-backed standalone `web.run` tool so Codex can call the
standalone search endpoint through the `codex-api` search client and
return its encrypted output to Responses.
- gate the new tool behind `standalone_web_search`
- install the extension in the app-server thread registry and hide
hosted `web_search` when standalone search is enabled for OpenAI
providers so the two paths stay mutually exclusive
- build search context from persisted history using a small tail
heuristic: previous user message, assistant text between the last two
user turns capped at about 1k tokens, and current user message
## Test Plan
- `cargo test -p codex-web-search-extension`
- `cargo test -p codex-api`
- `cargo test -p codex-core
hosted_tools_follow_provider_auth_model_and_config_gates`
## Why
Tool exposure is a planning concern, but the deferred MCP path and
dispatch-only legacy shell path were carrying those decisions in handler
constructors and a shell-only tool-family builder. Keeping those
decisions in `spec_plan` makes the core tool plan easier to follow and
keeps handlers focused on runtime behavior.
## What changed
- add `PlannedTools` helpers for ordinary runtimes, exposure overrides,
dispatch-only runtimes, and hosted specs
- inline shell tool assembly into `core/src/tools/spec_plan.rs` and
remove the shell-only `tool_family` module
- remove exposure state and special exposure constructors from
`McpHandler` and `ShellCommandHandler`
- keep hidden runtime behavior centralized in `ExposureOverride`,
including disabling parallel tool calls for hidden handlers
## Testing
- Not run (refactor only)
## Why
`ToolExecutor` is the runtime contract that keeps a callable tool and
its model-visible spec together. Leaving `spec()` optional lets a
registered runtime silently omit that half of the contract, and it also
overloads a missing spec as an exposure decision for tools that should
stay dispatchable without being shown to the model.
## What
- Make `ToolExecutor::spec()` required and update core, extension, and
test tool executors to return a concrete `ToolSpec`.
- Add `ToolExposure::Hidden` for dispatch-only tools. The legacy
`shell_command` runtime in unified-exec sessions now uses that explicit
exposure instead of hiding itself by omitting a spec.
- Build MCP tool specs when `McpHandler` is constructed so invalid MCP
specs are skipped before the handler is registered.
- Keep tool planning aligned with the new contract for direct, deferred,
hidden, code-mode, dynamic, and namespaced tool paths.
## Testing
- Added tool-plan coverage that invalid MCP tool specs are not
registered.
- Updated shell-family coverage for the hidden legacy `shell_command`
runtime and the affected tool executor test fixtures.
## Why
`code_mode_only_guides_all_tools_search_and_calls_deferred_app_tools`
was failing because code-mode prompt generation used the same nested
tool spec list for both the model-visible `exec` guide and the runtime
`ALL_TOOLS` surface. That allowed deferred MCP/app tools, such as
`calendar_timezone_option_99`, to leak into the `exec` description even
though they should only be discoverable through `ALL_TOOLS` at runtime.
## What changed
Split code-mode nested tool planning into two sets in
`core/src/tools/spec_plan.rs`:
- runtime nested tool specs still include deferred tools, so
`tools[...]` and `ALL_TOOLS` can call them
- `exec` prompt docs only render non-deferred tools, so deferred app
tools stay out of the model-visible guide
## Validation
- `cargo test -p codex-core --test all
code_mode_only_guides_all_tools_search_and_calls_deferred_app_tools --
--nocapture`
- looped the same focused test 5 additional times with `cargo test -q -p
codex-core --test all
code_mode_only_guides_all_tools_search_and_calls_deferred_app_tools`
## Summary
- Add `list_available_plugins_to_install` as the inventory step for
plugin and connector install suggestions.
- Slim `request_plugin_install` so it only handles the actual
elicitation, instead of carrying the full discoverable list in its
prompt.
- Emit send-time telemetry when an install elicitation is dispatched,
including requested tool identity in the event payload.
- Emit install-result telemetry through `SessionTelemetry`, including
tool type, user response action, and completion status.
- Update registration and tests to cover the new two-step flow while
keeping the existing `tool_suggest` feature gate unchanged.
## Testing
- `just fmt`
- `cargo test -p codex-tools`
- `cargo test -p codex-core request_plugin_install`
- `cargo test -p codex-core list_available_plugins_to_install`
- `cargo test -p codex-core
install_suggestion_tools_can_be_registered_without_search_tool`
- `cargo test -p codex-otel
manager_records_plugin_install_suggestion_metric`
- `cargo test -p codex-otel
manager_records_plugin_install_elicitation_sent_metric`
- `just fix -p codex-core`
- `just fix -p codex-tools`
- `just fix -p codex-otel`
- `cargo check -p codex-core`
Summary: defer v1 multi-agent tools when tool_search and namespace tools
are available; keep concise searchable descriptions and move the v1
usage guidance into developer instructions; add targeted coverage.
Testing: not run per request; ran just fmt.
## Why
`codex-tools` is meant to hold reusable tool primitives, but
`ToolsConfig` had become a second copy of core runtime decisions instead
of a small shared contract. It carried provider capabilities, auth/model
gates, permission and environment state, web/search/image feature gates,
multi-agent settings, and goal availability from core into `codex-tools`
([definition](https://github.com/openai/codex/blob/22dd9ad3929253ed24d7ee4f10f238e95ab25f37/codex-rs/tools/src/tool_config.rs#L97),
[stored on each
`TurnContext`](https://github.com/openai/codex/blob/22dd9ad3929253ed24d7ee4f10f238e95ab25f37/codex-rs/core/src/session/turn_context.rs#L87)).
Every session/context variant then had to build and mutate that snapshot
before assembling tools.
This PR removes that master object instead of renaming it. Tool planning
now reads the live `TurnContext`, where `codex-core` already owns those
decisions, while `codex-tools` keeps only reusable primitives and a
generic `ToolSetBuilder`/`ToolSet` accumulator.
## What Changed
- Removed `ToolsConfig` / `ToolsConfigParams` from `codex-tools`; the
crate keeps the shared helpers that still belong there, including
request-user-input mode selection, shell backend/type resolution,
`UnifiedExecShellMode`, and `ToolEnvironmentMode`.
- Replaced config-snapshot planning with `ToolRouter::from_turn_context`
and a `spec_plan` pipeline over `CoreToolPlanContext`, deriving provider
capabilities, auth gates, model support, feature gates, environment
count, goal support, multi-agent options, web search, and image
generation from the authoritative turn state.
- Added generic `codex_tools::ToolSetBuilder` / `ToolSet`, plus the
small core adapter needed to accumulate `CoreToolRuntime` values and
hosted model specs.
- Added the `tool_family::shell` registration module and moved
shell/unified-exec/memory accounting call sites to read the narrow
per-turn fields directly.
- Narrowed `TurnContext` to the remaining explicit per-turn fields
needed by planning: `available_models`, `unified_exec_shell_mode`, and
`goal_tools_supported`.
- Reworked MCP exposure and tool-search setup so deferred/direct MCP
behavior is driven by the current turn rather than a precomputed config
snapshot.
- Replaced the large expected-spec fixture tests with focused
behavior-level coverage for shell tools, environments, goal and
agent-job gates, MCP direct/deferred exposure, tool search,
request-plugin-install, code mode, multi-agent mode, hosted tools, and
extension executor dispatch.
## Verification
- `cargo check -p codex-tools`
- `cargo check -p codex-core --lib`
- `cargo test -p codex-tools`
- `cargo test -p codex-core spec_plan --lib`
- `cargo test -p codex-core router --lib`
## Summary
- Add `features.multi_agent_v2.tool_namespace` with config/schema
validation for Responses-compatible namespace values.
- Thread the resolved namespace into `ToolsConfig` for normal turns and
review turns.
- Wrap MultiAgentV2 tool specs and registry names in the configured
namespace when namespace tools are supported, while falling back to the
plain tool names when they are not.
## Validation
- `just fmt`
- `just write-config-schema`
- `cargo test -p codex-features multi_agent_v2_feature_config --
--nocapture`
- `cargo test -p codex-core test_build_specs_multi_agent_v2 --
--nocapture`
- `cargo test -p codex-core multi_agent_v2_config -- --nocapture`
- `cargo test -p codex-core
multi_agent_v2_rejects_invalid_tool_namespace -- --nocapture`
- `cargo test -p codex-tools`
- `git diff --check`
## Summary
- move tool_user_shell_type out of the old tools::spec module and call
it from tools directly
- attach the remaining spec planning model tests under spec_plan
- delete core/src/tools/spec.rs
## Tests
- just fmt
- cargo test -p codex-core tools::spec_plan
Note: a broader cargo test -p codex-core run on the earlier PR-head
worktree still hit the pre-existing stack overflow in
agent::control::tests::spawn_agent_fork_last_n_turns_keeps_only_recent_turns.
## Why
The tool runtime path still had a typed output associated type on
`ToolExecutor`, plus a core-only `RegisteredTool` adapter and
extension-only executor aliases. That made every new shared tool runtime
carry extra adapter plumbing before it could participate in core
dispatch, extension tools, hook payloads, telemetry, and model-visible
spec generation.
This PR moves output erasure to the shared executor boundary so core and
extension tools can use the same execution contract directly.
## What Changed
- Changed `codex_tools::ToolExecutor` to return `Box<dyn ToolOutput>`
instead of an associated `Output` type.
- Removed the extension-specific `ExtensionToolExecutor` /
`ExtensionToolOutput` aliases and exposed `ToolExecutor<ToolCall>` plus
`ToolOutput` through `codex-extension-api`.
- Reworked core tool registration around `CoreToolRuntime` and
`ToolRegistry::from_tools`, removing the extra `RegisteredTool` /
`ToolRegistryBuilder` layer.
- Consolidated model-visible spec planning and registry construction in
`core/src/tools/spec_plan.rs`, including deferred tool search and
code-mode-only filtering.
- Added `ToolOutput` helpers for post-tool-use hook ids and inputs so
MCP, unified exec, extension, and other boxed outputs preserve the same
hook payload behavior.
- Updated core handlers, memories tools, and the related
registry/spec/router tests to use the simplified contract.
## Test Coverage
- Updated coverage for tool spec planning, registry lookup, deferred
tool search registration, extension tool routing, post-tool-use hook
payloads, dispatch tracing, guardian output extraction, and memories
extension tool execution.
## Why
`codex_tools::ToolExecutor` keeps a tool spec attached to its runtime
handler, but extension tools still carried a parallel
`ExtensionToolFuture` / `ExtensionToolExecutor` shape. That made
extension-owned tools look different from host tools even though
routing, registration, and execution need the same abstraction.
This PR makes the shared executor contract directly async and lets
extension tools implement it too, so host tools and extension tools can
move through the same registration path.
## What changed
- Changed `ToolExecutor::handle` to an `async fn` using `async-trait`,
and updated built-in tool handlers to implement the async trait
directly.
- Replaced the bespoke `ExtensionToolFuture` contract with a marker
`ExtensionToolExecutor` over `ToolExecutor<ToolCall, Output =
JsonToolOutput>`, re-exporting `ToolExecutor` from
`codex-extension-api`.
- Updated the memories extension tools to implement the shared executor
trait.
- Split tool-router construction into collected executors plus hosted
model specs, keeping hosted tools like web search and image generation
separate from executable handlers.
- Updated spec/router tests and extension-tool stubs for the new
executor shape.
## Verification
- Not run locally.
## Why
`code_mode_only` filters code-mode nested tools out of the top-level
tool list. For multi-agent v2, we need a rollout shape where the
collaboration tools remain callable as normal model tools without also
being embedded into the code-mode `exec` tool declaration.
Related to this:
https://openai-corpws.slack.com/archives/C0AQLHB4U75/p1778660267922549
## What Changed
- Adds `features.multi_agent_v2.non_code_mode_only`, including config
resolution, profile override handling, and generated schema coverage.
- Introduces `ToolExposure::DirectModelOnly` so a tool can be included
in the initial model-visible list while staying out of the nested
code-mode tool surface.
- Applies that exposure to the multi-agent v2 tools when the new flag is
set: `spawn_agent`, `send_message`, `followup_task`, `wait_agent`,
`close_agent`, and `list_agents`.
- Updates code-mode-only filtering so direct-model-only tools remain
visible while ordinary nested code-mode tools are still hidden.
## Verification
- Added config parsing/profile tests for `non_code_mode_only`.
- Added tool spec coverage for the code-mode-only multi-agent v2
exposure behavior.
## Why
Recent session history showed no active use of the raw `shell`,
`local_shell`, or `container.exec` execution surfaces. Keeping those
handlers/specs wired into core leaves duplicate shell execution paths
alongside the supported `shell_command` and unified exec tools.
## What changed
- Removed the raw `shell` handler/spec and its `ShellToolCallParams`
protocol helper.
- Removed the legacy `local_shell` and `container.exec` handler/spec
plumbing while preserving persisted-history compatibility for old
response items.
- Normalized model/config `default` and `local` shell selections to
`shell_command`.
- Pruned tests that exercised removed raw-shell/local-shell/apply-patch
variants and kept coverage on `shell_command`, unified exec, and
freeform `apply_patch`.
## Verification
- `git diff --check`
- `cargo test -p codex-protocol`
- `cargo test -p codex-tools`
- `cargo test -p codex-core tools::handlers::shell`
- `cargo test -p codex-core tools::spec`
- `cargo test -p codex-core tools::router`
- `cargo test -p codex-core
active_call_preserves_triggering_command_context`
- `cargo test -p codex-core guardian_tests`
- `cargo test -p codex-core --test all shell_serialization`
- `cargo test -p codex-core --test all apply_patch_cli`
- `cargo test -p codex-core --test all shell_command_`
- `cargo test -p codex-core --test all local_shell`
- `cargo test -p codex-core --test all otel::`
- `cargo test -p codex-core --test all hooks::`
- `just fix -p codex-core`
- `just fix -p codex-tools`
## Why
Deferred tools were tracked with separate side-channel filtering after
tool specs had already been assembled. That made the registry
responsible for executing tools while the router/spec planner separately
decided whether those same tools should be exposed to the model up
front.
This PR makes exposure part of the tool handler contract so direct
versus deferred availability travels with the executable tool
registration.
Next step will be to simplify registration
## What Changed
- Adds `ToolExposure` to `codex-tools` and exposes it through
`ToolExecutor`, defaulting tools to `Direct`.
- Teaches dynamic tools and MCP handlers to mark deferred tools as
`Deferred` at construction time.
- Renames the registry object-safe wrapper from `AnyToolHandler` to
`RegisteredTool` and uses `ToolExposure` when deciding whether to
include a handler's spec in the initial model-visible tool list.
- Refactors tool spec planning to derive direct specs and deferred
search entries from registered handlers, removing the router's
special-case deferred dynamic tool filtering.
## Verification
- Not run.
## Why
Extension tools were split across two public runtime contracts:
`codex-tool-api` exposed `ToolBundle` plus its own call/spec/error
types, while core native tools used `codex_tools::ToolExecutor`. That
made contributed tool specs and execution behavior easy to drift apart
and added another crate boundary for what should be one executable-tool
seam.
This PR makes `ToolExecutor` the single runtime contract and keeps
extension-specific pinning in `codex-extension-api`.
## Remaining todo
https://github.com/openai/codex/pull/22369/changes#diff-b935ea8245c3ce568a30cff660175fa6390b66b872ae409e1e2e965738250741R5
Either generic `Invocation` or sub-extract the `ToolCall` and clean
`ToolInvocation`
## What changed
- Removed the `codex-tool-api` workspace crate and its dependencies from
core and `codex-extension-api`.
- Made `codex_tools::ToolExecutor` object-safe with `async_trait` so
extension contributors can return a dyn executor.
- Added the extension-facing aliases under
`ext/extension-api/src/contributors/tools.rs`, including
`ExtensionToolExecutor = dyn ToolExecutor<ToolCall, Output =
ExtensionToolOutput>`.
- Changed `ToolContributor::tools` to return extension executors
directly instead of `ToolBundle`s.
- Updated core’s extension tool handler/registry/router path to adapt
those extension executors into the existing native `ToolInvocation`
runtime path.
- Added focused coverage for extension tools being registered,
model-visible, dispatchable, and not replacing built-in tools.
## Verification
- `cargo test -p codex-tools`
- `cargo test -p codex-extension-api`
## Why
This builds on the handler-owned spec refactor by moving deferred
tool-search metadata to the same handlers that already own tool specs.
The registry builder no longer needs a separate prebuilt
`tool_search_entries` path; it can collect searchable entries from
deferred handlers directly.
## What changed
- Added `search_info()` to tool handlers and implemented it for MCP and
dynamic handlers.
- Reused handler `spec()` output when constructing tool-search entries,
adapting it into the deferred `LoadableToolSpec` shape expected by
`tool_search`.
- Simplified `build_tool_registry_builder(...)` so `tool_search`
registration is based on deferred handlers with search info.
- Removed the old standalone search-entry builders and now-unused
`codex-tools` discovery helper exports.
## Verification
- `cargo test -p codex-core tools::handlers::tool_search::tests:: --
--nocapture`
- `cargo test -p codex-core tools::spec_plan::tests::search_tool --
--nocapture`
- `cargo test -p codex-core tools::spec::tests:: -- --nocapture`
- `cargo test -p codex-core tools::spec_plan::tests:: -- --nocapture`
- `cargo test -p codex-tools`
- `just fix -p codex-core`
- `just fix -p codex-tools`
## Why
Code mode already builds the merged nested `ToolSpec`s that feed the
`exec` prompt. Keeping a separate `tool_namespaces` map in the planning
path duplicated that metadata and left extra wrapper plumbing in
`spec.rs`.
## What changed
- derive code-mode namespace descriptions from the merged
`ToolSpec::Namespace` entries before building the code-mode handlers
- extract `build_code_mode_handlers(...)` so the code-mode-specific
planning stays in one place
- remove `tool_namespaces` from `ToolRegistryBuildParams`
- delete the now-unused `McpToolPlanInputs` wrapper and related test
helper plumbing
## Testing
- `cargo test -p codex-core spec_plan`
## Summary
This refactor makes tool handlers the owner of the specs they can
publish, so registry construction can register handlers once and
separately publish only the specs that should be model-visible.
The main motivation is deferred tools: MCP and dynamic tools still need
handlers registered up front, but deferred tools should be discoverable
through `tool_search` rather than emitted in the initial tool spec list.
## What changed
- `McpHandler` and `DynamicToolHandler` can return their own `ToolSpec`.
- `build_tool_registry_builder` now collects handlers, registers them
through the no-spec path, and publishes only non-deferred handler specs.
- Deferred MCP and dynamic tool names are combined into one
`all_deferred_tools` set that drives spec filtering, code-mode
deferred-tool signaling, and `tool_search` registration.
- `tool_search` registration now requires both deferred tools and
`namespace_tools`.
- Namespace specs are merged in `spec_plan`, preserving top-level spec
order, sorting tools within each namespace, and backfilling empty
namespace descriptions.
- Hosted web search and image-generation specs are included in the
collected spec vector before namespace merge/publication, and tool-name
tests that should not care about hosted relative order now compare sets.
## Testing
- `cargo test -p codex-core tools::spec::tests:: -- --nocapture`
- `cargo test -p codex-core tools::spec_plan::tests:: -- --nocapture`
- `cargo test -p codex-core
tools::router::tests::specs_filter_deferred_dynamic_tools --
--nocapture`
- `cargo test -p codex-core
suite::prompt_caching::prompt_tools_are_consistent_across_requests --
--nocapture`
- `just fmt`
- `just fix -p codex-core`
- `cargo test -p codex-core -- --skip
tools::handlers::multi_agents::tests::tool_handlers_cascade_close_and_resume_and_keep_explicitly_closed_subtrees_closed`
passed the library suite after skipping the known stack-overflowing unit
test.
Full `cargo test -p codex-core` currently hits a stack overflow in
`tools::handlers::multi_agents::tests::tool_handlers_cascade_close_and_resume_and_keep_explicitly_closed_subtrees_closed`;
the same focused test reproduces on `origin/main`.
rhan-oaiandGitHub
rka-oaiandGitHub
Won ParkandGitHub
Owen LinandGitHub
pakrym-oaiandGitHub
Alex DaleyandGitHub
mchen-oaiandGitHub
sayan-oaiandGitHub
Shijie RaoandGitHub
jifandGitHub
Adam Perry @ OpenAIandGitHub
Eric TrautandGitHub
Ahmed IbrahimandGitHub
Michael BolinandGitHub
Dylan HurdandGitHub
Noah MacCallumandGitHub
Matthew ZengandGitHub