## Why
MCP servers discovered on different executors can share a name. Resolving them by name alone exposes only one server and prevents tools and resources from being addressed independently.
## What changed
- Add the disabled-by-default `multi_environment_mcp` feature to expose environment-scoped servers, including multiple installations of the same plugin.
- Carry exact `McpServerKey` identities through catalogs, connections, tool calls, resources, and approval tracking. Give environment-scoped tools stable namespaces derived from their environment.
- Include `environment_id` in scoped resource listings and accept it in resource requests. Require it for environment-owned servers when multiple environments are selected.
- Preserve server ownership through hooks and use the owning environment's permissions for scoped elicitation review.
## Testing
Add coverage for same-named servers across environments, independent calls and approvals, resource selection, plugin discovery, and rejection of hook changes to the tool environment.
GitOrigin-RevId: 5593372b06a15c16aa3220faa3b19f5d3b20e67e
## Why
Search discovery must account for the active model's supported tools while honoring host search restrictions. Automatic fallback must also preserve the provider that owns existing browsing references.
## What changed
- Resolve search from model metadata, client configuration, and managed requirements, exposing the selected provider's tools.
- Discover hosted search with the model's supported tools and carry the discovered invocation context into calls.
- Allow native fallback for eligible discovery failures only when search policy permits it and hosted search has not already been used. Do not replay failed hosted invocations through native search.
- Keep native fallback references on native search across later turns, discovery recovery, and resume. Release unused fallback offers and coordinate background calls to prevent conflicting provider use.
## Testing
Add app-server and core scenario coverage for provider selection, discovery failures, search restrictions, reference continuity, resume, and concurrent background search. Include a snapshot covering discovery and call history across a model switch.
GitOrigin-RevId: 0f22605722daaa424c4fa34660654cde5a6db049
## Why
Sessions using the same MCP catalog can repeat schema normalization and retain duplicate tool specifications.
## What changed
Split immutable MCP tool metadata and normalized specifications from per-handler caches, and reuse matching definitions within each `McpManager`. Require matching tool metadata, agent-plugin mode, and schema budget before reuse.
Keep shared definitions behind weak references and prune expired entries on cache misses so completed sessions do not keep their schemas alive. Build definitions outside the shared index lock.
GitOrigin-RevId: 02345359a1b0d297ca56f2f469de88aba89106f1
## What changed
Remove the duplicate tool metadata stored in `CachedMcpHandler` and compare incoming metadata directly with `McpHandler::tool_info()`.
Preserve complete agent plugin namespace descriptions in the handler, applying the existing 1,000-byte limit when building tool specs and search metadata. Truncation continues to respect UTF-8 character boundaries.
## Testing
Extend the agent plugin namespace description test to verify that the handler retains the full description while search output remains bounded.
GitOrigin-RevId: d0a9d5f6796143315235e658e77e059657e85f13
## Why
Refreshing the MCP catalog with unchanged tool metadata recreated handlers and forced the tool search index to rebuild.
## What changed
Cache handlers by canonical tool name and reuse them when tool metadata, agent plugin status, and the input schema size limit are unchanged. Replace handlers when those inputs change and evict tools removed from the catalog, preserving search index reuse across equivalent bindings.
## Testing
Extend the deferred tool world state test to verify that an unchanged catalog refresh captures a new binding without rebuilding the search index, and that removing all deferred tools stops advertising `tool_search`.
GitOrigin-RevId: 150866f635e9ee817d7ca44fb76e724a3f7608a5
## Why
Fixed schema budgets can strip parameter descriptions from large MCP tools or render their Code Mode input types as `unknown`. Allow larger budgets to preserve that guidance.
## What changed
- Add `mcp_servers.<name>.tool_input_schema_max_bytes` to configure the UTF-8 byte threshold for best-effort input schema compaction, retaining the 5,000-byte default.
- Add `features.code_mode.tool_input_schema_max_bytes` for rendered input types. Use the largest of the 16,000-byte default, the Code Mode setting, and the ordinary MCP server's explicit setting.
- Apply rendering budgets to prompt declarations and the runtime tool catalog, and avoid reusing cached declarations after the Code Mode budget changes.
## Testing
Add coverage for UTF-8 compaction thresholds, budget precedence, configuration round trips, and preservation of parameter guidance in both the Code Mode prompt and `ALL_TOOLS`.
GitOrigin-RevId: 0ef2823a614d4794e636f518b09c0b03a5239b45
## Why
MCP tool handlers and their Code Mode definitions are immutable for the lifetime
of a stable MCP binding, so rebuilding them for every sampling step repeats schema
construction unnecessarily.
## What changed
- Cache MCP handlers per session and reuse them while the same binding remains
active, invalidating the cache when the binding changes.
- Lazily cache schema-free Code Mode definitions on MCP handlers while continuing
to rebuild definitions for step-local tools whose specifications may change.
- Reapply current app enablement, tool policy, exposure, and plugin schema budgets
whenever cached handlers are registered for a step.
## Testing
- Cover handler and Code Mode definition reuse, binding invalidation, changing app
policy, stable continuation schemas, and step-local custom tool updates.
GitOrigin-RevId: 69c5011fb97c08580ac476e7fb2286b1e92e0359
## What changed
- Track Agent Plugin manifests through plugin, skill, and MCP loading so their capabilities use format-specific behavior without changing legacy plugins.
- Discover only direct-child skills, exclude app and hook capabilities, isolate MCP data, and reject MCP configuration files that are non-regular or resolve outside the plugin root.
- Bound model-visible skill instructions, plugin instructions, MCP descriptions, schemas, individual tools, and the aggregate Agent Plugin MCP tool set.
- Stop MCP and OAuth redirects when Agent Plugins send configured or authorization headers, while retaining existing redirect behavior for legacy MCP servers.
## Testing
- Add coverage for capability filtering, skill discovery, isolated MCP data and reserved-path expansion, unsafe MCP configuration files, context limits, and redirect handling.
GitOrigin-RevId: c9af66b051269f3226628ca280a58d32c808c38f
## What changed
- Gate Codex Apps MCP tool registration on whether apps are enabled instead of requiring each tool's connector to appear in the accessible connector list.
- Continue enforcing model-visibility and app-tool policy checks, including the requirement for connector metadata.
- Use the merged connector catalog for tool-suggestion discovery.
## Testing
- Cover app-tool registration from catalog metadata, including synthetic links, source ordering, and the apps-disabled case.
GitOrigin-RevId: eef4eb03da9738d3a165dd6cfda0d3c1844d51e7
## Why
MCP tools can be exposed directly, discovered through tool search, or called
from Code Mode. Servers need to be able to opt out of any of these surfaces
without disabling their tools everywhere.
## What changed
- Add `omit_tools_from` to MCP server configuration, accepting any combination
of `direct`, `deferred`, and `code_mode`.
- Apply the exclusions independently when building direct, deferred, and Code
Mode tool surfaces while keeping omitted tools registered for permitted uses.
- Remove client-private `_meta` fields from MCP results returned to Code Mode.
- Include the new setting in the configuration schema and MCP config
serialization.
## Testing
- Cover every exposure combination across Code Mode, Code-Mode-only sessions,
tool search, direct-only namespaces, and prefixed and unprefixed MCP names.
- Verify direct and nested execution paths, parallel-call support, config
round-tripping, and `_meta` filtering.
GitOrigin-RevId: 12dfcb78bb5c5ecf4d70f38a8b5022792463a27f
## Why
`ToolExecutor::exposure()` describes a tool's preferred exposure, while the host
may apply step-specific policy when building a tool plan.
## What changed
- Store each runtime together with its effective exposure in `ToolRegistry`.
- Apply MCP, namespace, and collaboration-tool exposure overrides directly to
registry entries instead of wrapping runtimes.
- Use the registry-owned exposure when building model-visible specs, deferred
tool search, code-mode executors, and parallel-call support.
## Testing
- Cover that a hidden MCP tool remains routable but is not eligible for parallel
tool calls.
GitOrigin-RevId: 1bc2760add8d993f477954269fa13681bfcc2444
## Why
External MCP, extension, and dynamic tools can collide with host-owned tool
names. In particular, a dynamic `tool_search` definition must not replace the
host implementation used to discover and route deferred tools.
## What changed
- Build each turn's tools through a single ordered `ToolRegistry` that retains
the first runtime registered for a name.
- Treat duplicate external tools as skippable collisions while preserving
strict duplicate checks for trusted host tools.
- Reserve host-owned code-mode and `tool_search` runtimes before producing the
model-visible tool list.
- Base hosted web-search fallback behavior on the extension runtime that
actually won registration.
## Testing
- Cover ordered registry mutation and duplicate handling.
- Verify a client-provided `tool_search` cannot shadow the host tool.
- Verify hosted web-search fallback follows the registered browser runtime.
GitOrigin-RevId: acebe360de8ec69ad7da2f684cc2542cbf18a984
## What changed
- Convert filtered MCP tool metadata into `CoreToolRuntime` instances before constructing the tool router.
- Pass those runtimes through the shared tool-planning path instead of maintaining separate direct and deferred MCP tool lists.
- Preserve MCP visibility, app policy filtering, and direct versus deferred exposure behavior.
GitOrigin-RevId: 67f7fbeecc2992318c859e038cd97938833911db
## Why
MCP tools were only placed behind `tool_search` when a feature flag was
enabled or when there were at least 100 tools. That made the model's
tool flow depend on both rollout configuration and the number of
installed tools.
The searched-tool flow is now the intended behavior. Making it
unconditional when the model and provider support it gives every
supported setup the same behavior and lets us retire the feature flag
safely.
## What changed
- Defer all effective MCP tools when `tool_search` and namespaced tools
are supported.
- Keep exposing MCP tools directly when search cannot be used, so older
or unsupported model/provider combinations still work.
- Mark `tool_search_always_defer_mcp_tools` as removed and ignore old
configured values.
- Keep plugin filtering, app-only filtering, file handling, and MCP
calls working through the searched-tool flow.
## Why many tests changed
Many tests used to act as if the model could see MCP tools in its first
request and call them immediately. That is no longer the real flow: the
model first receives `tool_search`, searches for a tool, receives the
matching MCP tool, and then calls it in the next request.
The tests therefore needed an extra search step, and checks for tool
names, descriptions, and input fields had to move from the first request
to the search result. These are not separate product changes; they make
the tests follow what the model will actually see after this change. The
plugin tests still check which tools are allowed and where they came
from, the file tests still check upload fields and behavior, and the MCP
round-trip test still checks a successful call from start to finish.
## Tests
- `just test -p codex-features`
- Focused `codex-core` tests for MCP exposure and tool planning
- `just test -p codex-core explicit_plugin_mentions`
- `just test -p codex-core stdio_server_round_trip`
- Focused `codex-core` tests for tool search, app-only tools, and MCP
file uploads
## Summary
- move `AppToolPolicyEvaluator` and the Apps config/requirements policy
logic from `codex-core` into `codex-connectors`
- resolve one immutable policy snapshot per exposure build and reuse it
across every Codex Apps MCP tool
- keep core as a thin adapter from MCP metadata to connector-owned
policy input while preserving the call-time defense-in-depth check
## Why
`build_mcp_tool_exposure` evaluates every Codex Apps tool on each
sampling request. The old path rebuilt effective Apps configuration for
every tool, and the policy implementation lived in the already-large
core crate even though it is connector-specific.
The connector-owned evaluator keeps the expensive config merge/decode
out of the loop and gives core only the effective policy result it
needs.
## Performance
With the real 557-tool Apps corpus, `build_mcp_tool_exposure` measured
3.74 ms and 3.33 ms after the extraction (3.54 ms mean). The original
path measured 807 ms mean, so the final result retains the 99.6%
reduction.
## Validation
- `cargo check -p codex-connectors -p codex-core`
- `just test -p codex-connectors` — 15 passed
- `just test -p codex-core --lib connectors` — 35 passed
- `just test -p codex-core --lib mcp_tool_exposure` — 5 passed
- `just test -p codex-core --lib mcp_tool_call` — 72 passed
- `just bazel-lock-update`
- `just bazel-lock-check`
- `just fix -p codex-connectors`
- `just fix -p codex-core`
- `just fmt`
## Why
Codex app-server latency traces do not granularly cover turn
orchestration, sampling-request preparation, and tool-loading work.
These spans help separate local coordination/setup costs from model
streaming and tool execution.
## What changed
- Add `run_turn.*` spans around sampling-request input preparation and
post-sampling state collection
- Add function-level trace spans around turn setup, hook execution,
compaction, prompt construction, and MCP tool exposure
- Add `built_tools.*` spans around plugin loading and discoverable-tool
loading
## Verification
Trigger Codex rollout and observe new spans are included
## 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
This PR removes the synthetic `HashMap<String, ToolInfo>` keys from MCP
tool discovery. `McpConnectionManager::list_all_tools()` now returns
normalized `Vec<ToolInfo>`, and downstream code derives identity from
`ToolInfo::canonical_tool_name()`.
The motivation is to keep model-visible tool identity on
`ToolName`/`ToolInfo` instead of parallel string map keys, so future
namespace changes do not have to preserve otherwise-unused lookup keys.
## Changes
- Rename the MCP normalization path from `qualify_tools` to
`normalize_tools_for_model` and return tool values directly.
- Flow MCP tool lists through connectors, plugin injection, router/spec
building, code mode, and tool search as vectors/slices.
- Keep direct/deferred subtraction local to `mcp_tool_exposure`, using
`ToolName` values.
- Update tests to compare `ToolName` instances where MCP identity
matters.
## Validation
- `cargo test -p codex-mcp test_normalize_tools`
- `cargo test -p codex-core mcp_tool_exposure`
- `cargo test -p codex-core
direct_mcp_tools_register_namespaced_handlers`
- `cargo test -p codex-core
search_tool_registers_namespaced_mcp_tool_aliases`
- `just fix -p codex-mcp`
- `just fix -p codex-core`
add new `tool_search_always_defer_mcp_tools` feature flag that always
defers all mcp tools rather than deferring once > 100 deferrable tools.
add new tests, also move `mcp_exposure` tests into dedicated file rather
than polluting `codex_tests`.
stacked on #17402.
MCP tools returned by `tool_search` (deferred tools) get registered in
our `ToolRegistry` with a different format than directly available
tools. this leads to two different ways of accessing MCP tools from our
tool catalog, only one of which works for each. fix this by registering
all MCP tools with the namespace format, since this info is already
available.
also, direct MCP tools are registered to responsesapi without a
namespace, while deferred MCP tools have a namespace. this means we can
receive MCP `FunctionCall`s in both formats from namespaces. fix this by
always registering MCP tools with namespace, regardless of deferral
status.
make code mode track `ToolName` provenance of tools so it can map the
literal JS function name string to the correct `ToolName` for
invocation, rather than supporting both in core.
this lets us unify to a single canonical `ToolName` representation for
each MCP tool and force everywhere to use that one, without supporting
fallbacks.
- [x] Expand tool search to custom MCPs.
- [x] Rename several variables/fields to be more generic.
Updated tool & server name lifecycles:
**Raw Identity**
ToolInfo.server_name is raw MCP server name.
ToolInfo.tool.name is raw MCP tool name.
MCP calls route back to raw via parse_tool_name() returning
(tool.server_name, tool.tool.name).
mcpServerStatus/list now groups by raw server and keys tools by
Tool.name: mod.rs:599
App-server just forwards that grouped raw snapshot:
codex_message_processor.rs:5245
**Callable Names**
On list-tools, we create provisional callable_namespace / callable_name:
mcp_connection_manager.rs:1556
For non-app MCP, provisional callable name starts as raw tool name.
For codex-apps, provisional callable name is sanitized and strips
connector name/id prefix; namespace includes connector name.
Then qualify_tools() sanitizes callable namespace + name to ASCII alnum
/ _ only: mcp_tool_names.rs:128
Note: this is stricter than Responses API. Hyphen is currently replaced
with _ for code-mode compatibility.
**Collision Handling**
We do initially collapse example-server and example_server to the same
base.
Then qualify_tools() detects distinct raw namespace identities behind
the same sanitized namespace and appends a hash to the callable
namespace: mcp_tool_names.rs:137
Same idea for tool-name collisions: hash suffix goes on callable tool
name.
Final list_all_tools() map key is callable_namespace + callable_name:
mcp_connection_manager.rs:769
**Direct Model Tools**
Direct MCP tool declarations use the full qualified sanitized key as the
Responses function name.
The raw rmcp Tool is converted but renamed for model exposure.
**Tool Search / Deferred**
Tool search result namespace = final ToolInfo.callable_namespace:
tool_search.rs:85
Tool search result nested name = final ToolInfo.callable_name:
tool_search.rs:86
Deferred tool handler is registered as "{namespace}:{name}":
tool_registry_plan.rs:248
When a function call comes back, core recombines namespace + name, looks
up the full qualified key, and gets the raw server/tool for MCP
execution: codex.rs:4353
**Separate Legacy Snapshot**
collect_mcp_snapshot_from_manager_with_detail() still returns a map
keyed by qualified callable name.
mcpServerStatus/list no longer uses that; it uses
McpServerStatusSnapshot, which is raw-inventory shaped.