Commit Graph
141 Commits
Author SHA1 Message Date
Charlie Marsh fb2b06a829 Surface MCP server startup errors when preparing cached tool calls (#60038)
## Why

A cached MCP tool can remain advertised after its server fails to start. Calling it previously returned a generic unavailable-tool error, hiding the startup failure from the model.

## What changed

Propagate preparation errors through MCP tool calls and recorded approval preparation, preserving the error chain and adding the server name to startup failures. Keep the unavailable-tool error for tools that are missing, filtered, or no longer visible to the model.

## Testing

Update the executor capability downgrade test to require the credential resolution startup error in the model's tool output and assert that the host token is absent.

GitOrigin-RevId: 434074eb748165b2fe19de3a40046835f26992df
2026-11-07 02:28:22 +00:00
Charlie Marsh 7ccf78dec4 Preserve MCP sharing discovery when deferring cached stdio startup (#60037)
## Why

Subagents reusing cached MCP tool definitions were forced to use the legacy protocol, preventing connection sharing from being negotiated on first use.

## What changed

Keep stdio protocol discovery enabled during deferred startup while allowing dormant subagents to advertise cached legacy tools. Track modern discovery separately from catalog updates so discovery clients reject older legacy catalogs while legacy clients can still reuse them. Keep modern tools out of the legacy catalog when joining pooled connections.

## Testing

Add coverage for deferred subagents negotiating sharing on first use and reusing the same server process, while a legacy sibling retains cached definitions without starting its server. Extend cache coverage to ensure older legacy fetches cannot undo modern discovery.

GitOrigin-RevId: 5f8f4deb22cfb81a08c05ff3fcd7ba5c9199551c
2026-11-07 02:26:13 +00:00
fahad-OAI 8fcabd75a0 Preserve app invocation context across MCP tool refreshes (#59991)
## What changed

Keep the advertised `_codex_apps.invocation_context` when an app's tool metadata refreshes on the same server connection, while using refreshed values for other metadata. If the invocation context changes on a different connection, skip the call as unavailable.

## Testing

Extend the code-mode compaction test to refresh app tools before resuming a cell and verify that calls retain the original invocation context while picking up the refreshed `resource_uri`.

GitOrigin-RevId: 1b6072faed74623994ea82035a6ed977451984a4
2026-11-06 21:53:04 +00:00
cooper-oai 219ab4043e Use runtime authentication for OpenAI MCP and app requests (#59749)
## What changed

- Select fresh authentication for each HTTP operation on eligible OpenAI MCP connections while preserving the connection's credential source and account network policy. Stop redirects when requests carry runtime assertions.
- Pass the owning thread's authentication and runtime context through app discovery, connector metadata, and trusted-access requests. Include the runtime authentication mode in MCP connection and tool-cache identities.
- Give resumed actors analytics clients bound to their own authentication context, and serialize stored-credential reloads with token refresh while preserving host-provided credentials.

## Testing

Add coverage for fresh request assertions, revoked account policies, redirect handling, resumed-actor isolation, and MCP uploads in dual-header and assertion-only modes.

GitOrigin-RevId: f06900a6db61cd6cf53c3aa8075afeec4d5706d9
2026-11-05 22:03:29 +00:00
sayan-oai 0585dcb5db Preserve sandbox context in prepared MCP calls (#59020)
## Why

MCP permission profiles need the directory context used to interpret them. Prepared calls must retain that context alongside their captured permissions across runtime refreshes.

## What changed

- Store each server's permission source as thread permissions or a captured environment `FileSystemSandboxContext`.
- Capture environment working directories, home and temporary directories, and materialized workspace roots. Retain the host working directory for thread permissions.
- Expose the captured source through `PreparedMcpCall::permission_source()`, while keeping hosted Apps on thread permissions even when a local environment is attached.

## Testing

Extend binding tests to verify that prepared calls retain environment sandbox paths across refreshes and expose the host working directory for thread permissions. Update elicitation coverage to include hosted Apps with an attached local environment.

GitOrigin-RevId: 87e94605daece324786eebc07b62a507d35e5ced
2026-11-03 01:08:08 +00:00
victor-openai 62bf900680 Skip attribution for direct MCP tool calls (#58938)
## Why

`CodexThread::call_mcp_tool` returns only the tool result, so computing attribution and creating a cached runtime binding for these calls is unnecessary.

## What changed

Add `McpRuntime::latest_call_tool_without_attribution` and use it for direct thread calls. Validate the call, wait for server startup, and dispatch through the latest connections without capturing an attribution binding.

## Testing

Extend the runtime binding cache test to verify that a successful call without attribution leaves the cached binding empty. Update the test server to return a successful tool result.

GitOrigin-RevId: 898e655d9a0174ddaf71c390757d6fc9a5a61780
2026-11-02 17:55:39 +00:00
bkotsopoulos 9630b2e019 Allow Apps catalog discovery before the initial tool inventory completes (#58352)
## Why

Plugin and skill catalog reads currently wait for the Apps server's initial `tools/list` to finish, even after its MCP transport is initialized.

## What changed

Add the disabled-by-default `features.apps_early_resource_discovery` flag. When enabled, host-owned Apps requests for `mcp/plugin` and `mcp/skill` resources can use the initialized transport while the initial tool inventory is pending. Server readiness still waits for the inventory to complete, and retries use normal readiness.

Bound early client acquisition by the configured startup timeout, propagate startup failures and cancellation, and reject pages if the connection, authentication scope, or resource cache generation changes during discovery.

## Testing

Add coverage for gated discovery, concurrent plugin and skill reads, startup timeouts and failures, cancellation, cache invalidation, and account switches. Verify that early discovery reuses a single handshake and tool inventory, including with the newer MCP protocol and connection sharing enabled.

GitOrigin-RevId: 955b1f6d43b733bdf9763c1caf2ce6913c686c81
2026-10-30 14:47:56 +00:00
khong bd2e9bf0f4 Skip resource template discovery for hosted Apps (#57861)
## Why

Hosted Apps returns no resource templates, so querying `resources/templates/list` is unnecessary.

## What changed

Return an empty template list for host-owned Apps in single-server binding requests and skip it during aggregate template discovery. Use server registration metadata so extension-provided servers named `codex_apps` continue to expose their templates.

## Testing

Add coverage for hosted Apps versus extension registrations and for full MCP server status with standalone and thread connections. Verify that status requests still call `resources/list` but omit `resources/templates/list`.

GitOrigin-RevId: 3182b619391aefbb2a47ca7b8d9611402846bfe3
2026-10-28 23:21:04 +00:00
bkotsopoulos d03a6e1a6a Support Apps tool catalogs without shared snapshots (#57572)
## What changed

- Add `ConnectorRuntimeManager::without_shared_snapshot()` to keep connector catalogs local to each client and disable disk caching.
- Use the thread's own Apps discovery when snapshot sharing is disabled, preserving its catalog across turns.
- Expose `apps_tools_from_mcp_client()` to read a ready, unshared Apps client's tools without starting or waiting for discovery.

## Testing

Extend coverage for shared-cache eligibility and verify that an unshared thread waits for its own Apps discovery and uses its own tools on successive model requests, even when a peer finishes discovery first.

GitOrigin-RevId: d57ddacaede047e67c09d378c86c1aa3577ca56f
2026-10-27 22:01:13 +00:00
Charlie Marsh 5e3496e0ad Share eligible local stdio MCP connections across chats (#57211)
## What changed

Extend `mcp_connection_sharing` to pool compatible local stdio connections that negotiate MCP `2026-07-28`. Keep legacy servers and servers that request input during startup private.

Preserve each chat's tool catalog and route input requests to the calling chat. Release connections independently for each chat, shutting down the shared server after the final release, and retire pooled connections on forced refresh without interrupting other chats.

## Testing

Add pool, runtime, app-server, and core scenario tests covering protocol gating, startup cancellation, refresh, final-release shutdown, and separate tool visibility and simultaneous input responses across chats sharing one process.

GitOrigin-RevId: 896b107267a20ec1bffccdefca6c40fe2945a022
2026-10-26 19:37:32 +00:00
Charlie Marsh 0bc73bbeb1 Add opt-in protocol discovery for local stdio MCP servers (#57210)
## Why

Stdio servers without `CODEX_MCP_PROTOCOL_VERSION` use the legacy handshake, even when they support MCP `2026-07-28`.

## What changed

- Add the disabled-by-default `mcp_connection_sharing` feature to enable discovery for eligible local stdio servers.
- Recover from failed or late discovery with a fresh handshake when needed, waiting for the previous helper to release resources before restarting.
- Retain a working legacy connection when an upgrade attempt fails for the same account, except on authentication errors.
- Keep tool catalogs separate by protocol and invalidate legacy catalogs when discovery negotiates the modern protocol.

## Testing

Add regression coverage for failed and late discovery, legacy fallback, protocol-specific catalog reuse, message limits, and Windows descendant cleanup.

GitOrigin-RevId: 36c675ef3b81d32fccf06ccbddda2312cfe602c9
2026-10-26 19:37:06 +00:00
viyatb-oai d124fe9de1 Add negotiated network policy metadata support for MCP calls (#56278)
## What changed

Add an optional `NetworkPolicyMetadataProvider` to the MCP runtime. Advertise `codex/network-policy-meta` only to trusted ChatGPT origins when a provider is available, and attach policy metadata only when the sending session also advertises support.

Strip caller-supplied policy metadata and derive restrictions from the call's captured permission profile and server environment. Resolve metadata against each session's capabilities, including after recovery, while retaining one policy snapshot per call. Fail the call before sending if policy resolution fails. Existing runtime callers leave the provider unset.

## Testing

Add tests for capability negotiation, trusted-origin restrictions, forged metadata removal, both tool dispatch paths, and policy snapshot reuse. Add HTTP session-recovery coverage for changing capabilities and policy-resolution failures while preserving caller metadata and read-only restrictions.

GitOrigin-RevId: 4b978f986abc0475fc67659be98fc01a2d6ce7cd
2026-10-22 22:33:07 +00:00
jif c419fd01a7 Fix enterprise MCP credential selection and reauthentication (#55859)
## What changed

- Prefer an eligible account grant over saved enterprise credentials, with a fallback to saved enterprise sign-in during initial connection if the account grant is rejected.
- Remember rejected account grants per enterprise MCP registration across threads and report authentication failures as requiring reauthentication. Keep administrator configuration errors distinct from sign-in failures.
- Compare account grants independently of access and ID token rotation, recheck the latest identity eligibility after exchange, and retry once when the refresh grant changes.
- Make MCP authentication status reflect usable account grants, saved enterprise credentials, and completed connection failures.

## Testing

Extend regression coverage for enterprise sign-in failures and recovery, MCP authentication status, account grant identity across token refreshes, and retries after transient credential failures.

GitOrigin-RevId: eeac65f668d5d2d55e9d8bf4547fbcbdd6a6c42d
2026-10-21 10:40:59 +00:00
jif feadd9441e Preserve enterprise MCP authorization across runtime refreshes (#55696)
## Why

Configuration reloads must cancel stale enterprise MCP startups while preserving access for running threads that still pass validation. Replacing a runtime during `thread/revert` must also leave the replacement able to use enterprise tools.

## What changed

- Separate pending startup cancellation from published thread lifetimes, with independent startup state for forks and replacement runtimes.
- Retain eligible enterprise MCP connections across refreshes and explicit reconnects while applying updated tool restrictions and approvals. Do not create replacement enterprise connections after initial startup.
- Revoke enterprise authorization when a thread is removed, drain requests when access is disabled, and revoke a connection's authority when token renewal requires authentication.

## Testing

Extend regression coverage to verify enterprise tool calls survive `thread/revert`, inherited fork configurations are cancelled after a reload, and enterprise calls fail after logout while ordinary MCP calls remain available.

GitOrigin-RevId: 2d634be3011854b57f73c741bd1ec6b5aed94ee5
2026-10-20 23:03:56 +00:00
cooper-oai f29d55ff06 Propagate authentication header failures before sending requests (#55607)
## Why

Synchronous auth header construction could silently return no credentials when managed authentication became unavailable. Callers need an error to avoid sending requests with missing or partial authentication.

## What changed

- Make `AuthProvider::add_auth_headers`, `to_auth_headers`, and auth header telemetry fallible, and propagate failures through request and WebSocket setup.
- Return a transient error when managed authentication is unavailable, and preserve network policy denials through auth error conversion.
- Retry transient MCP authentication errors while treating build errors and policy denials as terminal.
- Cancel local MCP event streams before resolving authentication so cleanup still works when credentials become unavailable.

## Testing

Add regression coverage for rejected partial authentication, async header resolution, terminal auth errors without sending requests, managed account changes, MCP retry classification, and local stream cancellation after authentication becomes unavailable.

GitOrigin-RevId: 39d981e232cba1e302de502a2fa6239101114ed1
2026-10-20 15:43:19 +00:00
jif 1569abc5dc Enable enterprise-managed authorization for MCP connections (#55334)
## What changed

- Connect directly configured MCP servers using account-scoped enterprise grants stored in the keyring, with token exchange and authenticated HTTP requests.
- Retain authorization for the thread across MCP refreshes and revoke it when the account owner changes, cloud configuration becomes invalid, or an enterprise login replaces the grant.
- Negotiate the enterprise-managed authorization extension with the `2026-07-28` protocol, bypass shared tool catalog caching for enterprise servers, and report credential status with enterprise sign-in guidance.

GitOrigin-RevId: c0f96d37194212758a211ae6a4107205df96df26
2026-10-19 17:57:35 +00:00
victor-openai 0e67502123 Preserve MCP tool-list metadata in server status (#54686)
## Why

MCP tool discovery discarded response-level `_meta`, leaving server-status clients unable to inspect metadata associated with the tool catalog.

## What changed

- Preserve opaque `_meta` from the first `tools/list` page through catalog refreshes and caches.
- Expose it as `toolListMeta` in `McpServerStatus`, with `null` when unavailable, and update generated schemas and bindings.
- Scope cached Apps catalog metadata to discovery settings, including client capabilities and extensions, while retaining tools from older caches without metadata.
- Restrict server capabilities in status snapshots to the selected servers.

## Testing

Add regression coverage for listing legacy cached Apps tools during failed startup, and update protocol serialization expectations for nullable `toolListMeta`.

GitOrigin-RevId: 62335e2fd8c482067a7a78a7612b58c3eaa9fb55
2026-10-16 19:26:46 +00:00
jif 0df4aa6778 Honor MCP tool-list changes within a turn (#54656)
## Why

MCP servers can expose new tools during discovery. Refreshing the catalog after a tool-list notification lets Codex discover and call those tools in the same turn.

## What changed

- Add the disabled-by-default `mcp_tool_list_changed` feature for ordinary MCP servers.
- Track notifications in transport order and refresh catalogs before subsequent discovery or tool binding. Coalesce concurrent refreshes and retry listings invalidated by notifications, including during pagination.
- Retain the previous catalog on refresh failure, apply a retry cooldown, and invalidate affected startup caches. Allow notification refreshes to proceed while admitted calls finish.

## Testing

Add coverage for same-turn discovery and invocation in direct and Code Mode, notifications during startup and pagination, disabled behavior, cache invalidation, and concurrent refreshes with timeout recovery.

GitOrigin-RevId: 9afee8deba00dd9851b5f32625f61f96965aa00c
2026-10-16 16:40:17 +00:00
lsingh-oai 160b2b1ef6 Preserve ChatGPT authentication for Gov MCP servers (#54170)
## Why

MCP servers at `https://gov.chatgpt.com` were treated as untrusted for ChatGPT authentication and switched to OAuth, preventing local Gov Apps from retaining the ChatGPT auth provider.

## What changed

Trust the `https://gov.chatgpt.com` origin alongside the existing ChatGPT origin when resolving MCP server authentication. Origin matching preserves scheme, host, and port restrictions.

## Testing

Add regression coverage for Gov Apps URL construction, trusted origin normalization, rejection of lookalike hosts and nonstandard schemes or ports, and retention of the local ChatGPT auth provider. Verify that explicit OAuth servers and servers in remote environments do not receive the host provider.

GitOrigin-RevId: 7ceaf7c524d963005d9f23204489bbde269f0dd5
2026-10-15 03:07:27 +00:00
Anthony Tafoya 12fd621f62 Preserve specific MCP failure codes in metrics and spans (#54094)
## Why

MCP request failures and tool errors reported only through `codex/telemetry` metadata previously received an `unknown` error code, obscuring the failure category.

## What changed

- Classify typed request failures into bounded codes such as `request_timeout`, `transport_closed`, and `protocol_error`, and carry them into MCP call metrics and spans before errors become strings.
- Preserve typed timeouts when a prepared tool call exhausts its deadline during trusted access lookup.
- Accept allowlisted `error_code` values from `codex/telemetry` as a fallback after structured content and connector authentication metadata. Keep model output unchanged and ignore unrecognized metadata codes.

## Testing

Add coverage for typed errors wrapped in context, prepared-call timeout classification, metadata precedence and rejection, unchanged model output, and exported span error attributes.

GitOrigin-RevId: d7a440a6c9119b8a07a7f1fad697eb25af506f00
2026-10-14 20:53:59 +00:00
jif 4c1cd8bbcc Keep discovered Apps tools available during startup recovery (#53672)
## Why

A transient Apps MCP startup failure can remove cached tools that the model has already discovered, preventing a later Code Mode invocation from using them.

## What changed

Keep cached Apps tools in the model's tool catalog while startup recovery runs. Let invocation readiness start or join a shared retry before acquiring the tool execution lock, while tool discovery remains nonblocking. Keep recovery running if a waiting invocation is cancelled.

## Testing

Extend connection-manager tests to cover cached tool bindings and shared recovery. Add a Code Mode regression test that discovers a cached Calendar tool, invokes it after startup fails, and verifies that it succeeds after recovery while unrelated tools remain able to execute.

GitOrigin-RevId: e888501a1050884a50a473e71b893cf6672fe256
2026-10-13 13:59:41 +00:00
alishobeiri-oai de8fab6d7a Upgrade rmcp to 3.5.1 and harden lifecycle tests (#52972)
## What changed

- Update the pinned `rmcp` dependency and Cargo/Bazel lockfiles, and migrate MCP handlers and fixtures to `ClientConfig` and `ServerConfig`.
- Stop requiring responses for cancelled elicitation requests in cancellation tests while retaining checks for cleanup and continued connection usability.
- Handle client disconnects during WebSocket writes in the TUI test server, use an asynchronous HTTP mock for loopback proxy testing, and improve remote executor test diagnostics and the Windows subprocess fixture.
- Use `invalid_prompt` for the memory worker failure fixture.

## Testing

Add a regression test covering transport EOF during both form elicitation and user verification. Keep the client alive while checking that pending response channels close, the timeout pause clears, and the service exits with `QuitReason::Closed`.

GitOrigin-RevId: c0166b83fab3b59d84a5207c6630cb6adcf3b230
2026-10-10 19:43:24 +00:00
Owen Lin cef2a64531 Share MCP binding tool metadata with prepared calls (#51823)
## Why

Prepared MCP calls cloned `ToolInfo` already stored in the binding's model-visible catalog, duplicating tool metadata.

## What changed

Store the binding catalog in an `Arc<[ToolInfo]>` and let prepared calls reference its entries by index. Keep separately owned metadata for tools hidden from the model and calls prepared outside the binding catalog.

## Testing

Extend the binding test to assert that advertised and prepared tool metadata share the same address, while preserving ready-client catalog selection and hidden-tool behavior when tool names collide.

GitOrigin-RevId: 24b92f7e1dbb421c9ebce875b9b1024f7d697c5e
2026-10-07 20:47:19 +00:00
jif cc31e374fc Follow MCP tool pagination in legacy protocol mode (#50035)
## Why

Legacy MCP tool discovery discarded pagination cursors, leaving tools on later pages unavailable.

## What changed

Pass `tools/list` pagination cursors to the existing collector in both legacy and modern protocol modes. Legacy discovery now follows subsequent pages and applies the same pagination limits and cursor validation as modern discovery.

## Testing

Extend integration coverage to both protocol modes, checking that tools from later pages reach the model and can be called successfully, while catalogs with oversized cursors are rejected. Update the legacy discovery test to expect an error for a repeated cursor.

GitOrigin-RevId: a6daa8034d2a403c1f2a5846e385789d8a620af7
2026-10-01 17:10:03 +00:00
victor-openai ea64727556 Add single-server MCP status discovery with thread connection reuse (#48783)
## Why

Inspecting one MCP server should not require full-inventory discovery or a separate connection when a thread already has one.

## What changed

- Add optional `serverName` to `mcpServerStatus/list` and update generated schemas and bindings.
- With `threadId`, apply any pending runtime refresh and reuse the thread's current connection and tool catalog, waiting only for the selected server during discovery.
- Without `threadId`, create a discovery connection for only the selected server. Unknown names return an empty page; omitting `serverName` preserves full-inventory discovery.

## Testing

Add a regression test covering both status detail modes, checking that tool metadata is preserved and that status reads do not repeat `initialize` or `tools/list` requests.

GitOrigin-RevId: 1b361b67a993ab45cae493e88c146f50cfaf753a
2026-09-27 21:07:54 +00:00
rreichel3-oai c7e80f873f Preserve executor MCP credential boundaries across reconnects (#48143)
## Why

Cached executor MCP declarations must not gain access to host environment credentials when reconnecting to an executor that lacks environment credential resolution.

## What changed

- Retain credential policy through MCP registration, catalog materialization, and runtime setup. Derive selected HTTP plugin policy from the plugin's source environment.
- Reject host bearer-token fallback and transports requiring host environment headers or helpers for executor-owned declarations. Preserve host-configured credential resolution.
- Include credential policy in connection and tool catalog cache identities, and avoid reading host environment values for executor-only identities.

## Testing

Add regression coverage for legacy executors, cached declarations after a capability downgrade, policy preservation through catalog rebuilds, and connection and tool cache isolation.

GitOrigin-RevId: 2d5b85e4e98668c637c538cbe528914c936bd831
2026-09-25 17:15:29 +00:00
Vivian Fang 339e981ba7 Make MCP and Code Mode input schema budgets configurable (#47936)
## 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
2026-09-24 20:39:37 +00:00
Henry Levy b62fcd3c4b Allow cached catalogs to satisfy MCP startup readiness (#47935)
## Why

Required MCP servers can block startup while establishing a live connection even when a usable tool catalog is already cached.

## What changed

- Add per-server `startup_readiness = "catalog"` to expose cached tools while the connection starts, retaining `"connection"` as the default.
- Allow an eligible cached catalog to satisfy required-server startup validation and model binding capture. Explicit server or plugin requirements still wait for a live connection, and tool execution still requires the current connection.
- Check tool filters and model visibility when accepting cached catalogs, and preserve cached catalog revisions for binding reuse.
- Serialize the new setting, expose it in the configuration schema, and trace startup readiness at validation and model binding gates.

GitOrigin-RevId: a1cbdf1757ebb5824cffe2cc56408be49711c422
2026-09-24 20:38:50 +00:00
willwang-openai 83b56bc5ad Support client secrets for pre-registered MCP OAuth clients (#47891)
## Why

MCP OAuth configuration supported a pre-registered client ID but could not supply a client secret for token exchange or refresh.

## What changed

- Add `oauth.client_secret` and `codex mcp add --oauth-client-secret`, requiring a nonempty secret and client ID. Persist the secret in server configuration and accept `clientSecret` in plugin configuration.
- Pass configured credentials through CLI, app-server, and plugin login flows and token refresh. Require a new login when a configured confidential client's ID differs from stored credentials.
- Redact secrets in debug output and keep them out of authorization URLs and persisted OAuth token records.
- Invalidate cached OAuth connections when the configured client ID or secret changes.

## Testing

Add coverage for configuration validation and round trips, CLI argument redaction and subsequent login, app-server login, connection invalidation, and refresh with `client_secret_basic` and `client_secret_post` in both refresh modes. Verify secret exclusion from token records and reject mismatched client IDs before contacting the provider.

GitOrigin-RevId: 85bb0e410e9e68e6cd68eb6359badfc7eac78147
2026-09-24 17:26:18 +00:00
Rennie c45ea25ffb Allow subagents to request MCP elicitation input (#46877)
## Why

MCP requests requiring user input previously failed in subagents with guidance to ask the parent agent, blocking browser sign-in, form input, and interactive tool approval in the child thread.

## What changed

Remove the root-only MCP elicitation guard and `allow_user_interaction` plumbing so subagents can surface prompts and wait for responses under the existing approval policy and automatic review rules. Connector authentication can now return retry guidance after acceptance or preserve the original failure when declined.

Keep legacy tool approvals through `request_user_input` restricted to the root thread.

## Testing

Update unit and integration coverage for subagent prompt resolution, form responses, connector authentication acceptance and decline, automatic approval, strict review, and approval-policy updates on reused connections. Update the browser sign-in scenario snapshot to show the subagent continuing after an accepted response.

GitOrigin-RevId: c226fc6dfde724ed22b91dc04dd32bbc15245bee
2026-09-20 19:42:18 +00:00
Rennie 40584fad87 Keep MCP user interaction on the root thread (#46066)
## Why

MCP requests that need human input, including browser sign-in, must be handled by the root thread. Subagents need guidance to hand these blockers to their parent without prompting the user or automatically accepting requests that require input.

## What changed

- Reject interactive MCP elicitations and tool approval prompts in subagents with guidance to ask the parent and wait before retrying.
- Recognize browser authentication and `codex_requires_user_input` metadata even when the form schema is empty, while preserving automatic permission approvals and review decisions.
- Carry user-interaction eligibility through MCP runtime creation and connection reuse, and guard prompt registration.
- Preserve connector authentication diagnostics alongside handoff guidance, subject to normal tool-output limits.

## Testing

Add unit and integration coverage for blocked subagent prompts, root browser authentication, automatic approval and review, connection reuse, and authentication diagnostic preservation and truncation. Add a request-history snapshot for browser-auth handoff guidance.

GitOrigin-RevId: 83e146b6e5c1a2a22a34681ad45615f504b3c96c
2026-09-17 01:18:37 +00:00
Rennie b97abdbe30 Add read-only policy support to MCP tool requests (#46042)
## Why

Read-only MCP discovery and invocation need a consistent policy across requests, connection reuse, and tool catalogs. Reusing an unrestricted connection or cached catalog can bypass filtered discovery.

## What changed

- Thread `requires_read_only_mcp_tools` through MCP configuration and client setup, leaving it disabled by default.
- When enabled, set `openai/readOnly` to `true` in `tools/list` and `tools/call` request metadata, overriding a caller-supplied `false` while preserving other metadata, pagination, and arguments.
- Include the policy in connection identity and bypass shared Apps and persistent tool catalog caches for read-only connections.

## Testing

Add coverage for metadata propagation in legacy and modern protocol modes, preservation through session recovery, and isolation from unrestricted connections and shared Apps catalogs.

GitOrigin-RevId: f4a3a4021a54e302bffd53f08092b080c056246b
2026-09-16 22:25:54 +00:00
Charlie Marsh f8bed26f7b Share Apps tool catalogs without retaining unused snapshots (#45440)
## Why

Idle Apps clients and cached MCP bindings can retain replaced tool definitions. Shared catalog updates can also invalidate prepared calls even when the current definitions match the captured catalog, including after tools are restored while a call awaits approval.

## What changed

- Share immutable tool arrays across equivalent live discovery contexts and reuse storage for equal results. Let unused providers and cached bindings expire.
- Include requested capabilities and initialization results in the sharing scope, and detach servers that disable catalog caching from live sharing.
- Capture catalog snapshots for prepared calls. Accept equivalent shared catalogs regardless of tool-list order, while rejecting changed definitions and calls captured before an explicit refresh on that client.

## Testing

Add regression coverage for shared storage and scope isolation, release of replaced tools and unused bindings, equivalent catalog restoration, explicit refresh invalidation, and an Apps call completing after catalog restoration while awaiting approval.

GitOrigin-RevId: a0516186e4286d0ff13405fafbab6fcf3f1f3773
2026-09-14 15:06:28 +00:00
Gabriel Peal 7a6f469dcf Expose advertised MCP server capabilities in status responses (#44826)
## What changed

Add `serverCapabilities` to `mcpServerStatus/list` responses in both `full` and `toolsAndAuthOnly` modes, including thread-scoped reads. Return the initialized server's capabilities object, including its `extensions` map, or `null` when unavailable.

Capture capabilities during initialization so they remain available when tool discovery fails. Clear them on each new connection attempt and keep them independent of shared tool caches. Update the protocol schemas, generated TypeScript and Python types, and documentation.

## Testing

Add coverage for extension capabilities surviving tool-discovery failures in both detail modes. Extend existing tests to check advertised extensions, absent capabilities after initialization failure, and serialization of unavailable capabilities as `null`.

GitOrigin-RevId: 8e9d67c10fac3a54e82bef8be5d99caaf82597d1
2026-09-11 13:55:24 +00:00
Matthew Zeng 8570091e14 Honor thread-level plugin exclusions across runtime capabilities (#44655)
## What changed

- Apply `disabled_plugin_ids` to plugin skills, recommendations, hooks, and MCP servers without modifying shared plugin state.
- Activate selection changes when the next task starts, keeping pending settings from changing the active runtime.
- Hide app tools contributed only by disabled plugins from model tool catalogs and search. Preserve connectors shared with enabled plugins and existing direct Apps RPC behavior.
- Include plugin identity in MCP approval keys and use that identity when persisting plugin tool approvals.

## Testing

Add regression coverage for disabling and re-enabling plugin skills and tools, deferred activation of settings, shared connector ownership, selected plugin servers, executor hooks, and plugin-specific approval keys and persistence.

GitOrigin-RevId: 1153ff6aa7ab6953d50e838da05203904ecd2956
2026-09-10 21:51:34 +00:00
thomas 6baa076eb6 Allow extensions to select MCP protocol mode per HTTP server (#44571)
## Why

Extension-owned HTTP MCP servers previously inherited the default protocol mode. Extensions need to select a mode for their own server independently of other HTTP servers.

## What changed

- Add `McpServerContribution::SetWithProtocolMode` and re-export `McpProtocolMode` through the extension API.
- Carry the winning registration's protocol override through catalog resolution and materialization, and apply it to Streamable HTTP connections.
- Preserve existing defaults when no override is present. Selecting a protocol mode does not grant host-owned Apps cache access or environment authority.

## Testing

Add coverage for registration precedence and materialization, and extend cache isolation tests to cover explicit protocol selection. Add an integration test verifying that extension servers can select either the legacy or newer protocol while other HTTP servers retain the default mode.

GitOrigin-RevId: 606cecc03094a93258ffc433849f575020b81473
2026-09-10 16:02:59 +00:00
Andrei Eternal d996b4f02a Report OAuth authentication failures in MCP status snapshots (#44359)
## Why

Status-only discovery could retain an `OAuth` status based on stored credentials even after the connection attempt failed authentication. This discovery path has no event channel to report the failure.

## What changed

Update the snapshot to report `NotLoggedIn` for servers previously marked `OAuth` whose connection attempts failed authentication. Preserve other authentication statuses and OAuth statuses for failures unrelated to authentication.

## Testing

Add a regression test covering OAuth authentication failure, successful OAuth startup, an unrelated provider error, and a bearer-token authentication failure.

GitOrigin-RevId: 9df68fb1f818e40acab0c71bb4fda118f03a7611
2026-09-10 01:05:56 +00:00
thomas eb680c0558 Give hosted Codex Apps an independent MCP protocol opt-in (#44318)
## Why

Hosted Codex Apps shared the `mcp_2026_07_28` protocol setting with other MCP servers, preventing independent protocol selection.

## What changed

- Add `features.codex_apps_mcp_2026_07_28`, disabled by default, for the host-owned HTTP `codex_apps` server. Enabling it attempts 2026-07-28 discovery with Legacy fallback.
- Keep other servers, including third-party registrations with Apps-like names or URLs and local stdio servers, governed by `mcp_2026_07_28`.
- Support the new flag through `experimentalFeature/enablement/set` and refresh protocol settings in existing sessions. Track protocol mode per server when deciding whether to reuse connections.

## Testing

Add coverage for independent feature and protocol selection, hosted event-stream protocol selection, standalone Legacy defaults and discovery fallback, and runtime enablement refreshing an existing app-server thread.

GitOrigin-RevId: 640e8e251f0c45e5f4829cc2b73c3082cf42f715
2026-09-09 22:22:08 +00:00
jif 20f109eadb Reuse MCP bindings while cached servers remain dormant (#44121)
## Why

Dormant MCP servers with usable cached tool catalogs prevented binding reuse, causing each model step to capture a new binding even when the catalog was unchanged.

## What changed

Track dormant catalog revisions alongside ready client revisions so bindings can be reused until the catalog changes or the server starts. Explicit server and plugin requirements still trigger binding capture and startup. Check cached catalog availability without cloning tool definitions.

## Testing

Extend the lazy-startup integration test to verify that two model steps share one binding while the server stays dormant, and that starting the server invalidates that binding.

GitOrigin-RevId: ef0469a85000f84a928946c24350e3bda1522fb5
2026-09-09 10:51:24 +00:00
Matthew Zeng 6d377e96eb Propagate Apps tool refreshes to existing threads (#43900)
## Why

Refreshing installed Apps without a thread should update the tools available to existing threads on their next turn.

## What changed

- Publish live tool catalogs to clients with matching transport, auth, protocol, and listing settings within the same account and home directory. Ignore the thread attribution header when matching scopes.
- Adopt updated tools before catalog reads and new calls, preserving running calls and rejecting calls prepared against an outdated catalog revision.
- Keep the newest successful fetch per scope so older refreshes cannot overwrite newer tools, and exclude disk snapshots from live updates.

## Testing

Add regression coverage for refresh propagation to an existing thread without another tools listing, scope and account isolation, out-of-order refreshes, late client startup, and running versus stale prepared calls.

GitOrigin-RevId: 7a5ee34e23742ce374c6647dc8928b76ea622448
2026-09-08 19:56:25 +00:00
jif 0df39752cb Notify opted-in stdio MCP servers of auth changes (#43428)
## What changed

Advertise the experimental `codex/auth-change` capability for stdio MCP connections with an auth manager. When the server opts in, send `notifications/codex/authChanged` after initialization and on subsequent auth changes, with credential and owner generation counters and no credentials.

Track owner changes separately from credential refreshes so consumers can detect login, logout, or user, workspace, and auth-mode changes even when notifications coalesce. Treat credential changes with incomplete owner identity as owner changes as well.

Tie the notification watcher to the managed client's lifetime, limit each send to five seconds, and close the connection if a subsequent notification fails.

## Testing

Add tests for credential refreshes versus owner changes, coalesced logout and account switches, capability opt-in, notification payloads, and watcher cleanup.

GitOrigin-RevId: 35a99f5252bd48ca60a71fbdd1f708a8158e812c
2026-09-07 10:37:46 +00:00
riley-oai 555b82afa9 Add opt-in MCP user-verification transport (#43352)
## Why

Device-authenticated verification needs an app response outside automated approval and review. Previously, MCP user-verification requests were always cancelled.

## What changed

- Add typed `openai/userVerification` elicitations carrying a title, description, and challenge through core and app-server.
- Route verification to one connection enabled by trusted host activation. Restrict responses to that owner, exclude requests from replay, and cancel on disconnect or authentication changes, including account switches away and back.
- Return proofs in `content`, discard response metadata, and cancel malformed acceptances. Suppress verification response logging and analytics, and send only a generic verification notice to realtime.
- Keep the mode experimental and exclude it from stable schema exports. Capability advertisement remains disabled, and the TUI cancels verification requests.

## Testing

Add coverage for owner isolation, disconnect and authentication races, proof response handling, automated-review bypass, inactive-host cancellation, stable schema filtering, and realtime payload exclusion. Update the MCP integration test to verify that an activated request holds the tool until the client responds.

GitOrigin-RevId: c8c7b6691a3f7756682bbcca9def1df0d1feb906
2026-09-07 04:33:22 +00:00
Alex Daley 32351a7b1a Keep refreshed MCP tool catalogs with their clients (#43031)
## Why

Refreshed Apps catalogs need to survive connection reuse. Catalog revisions scoped to the connection set can lose refreshed tools during reconciliation and invalidate prepared calls for unrelated servers.

## What changed

- Store tools, revisions, and refresh coordination on each MCP client, and key cached bindings by client identity and revision.
- Preserve existing binding snapshots while rejecting stale calls before preparation. Wait for active calls to finish before publishing a refreshed catalog.
- Add `CodexThread::refresh_codex_apps_tools` to reconcile the runtime and return the refreshed client's raw Apps inventory alongside tool names eligible under that runtime's generic MCP policy. Keep hidden and disabled tools in the inventory for callers to apply app-specific policy.

## Testing

Add regression coverage for refreshed catalogs surviving client reuse, unrelated calls remaining valid, binding cache invalidation, and refresh snapshots retaining the client's own inventory when another client wins the shared-cache race. Update stale-call and preparation tests to use client-owned catalogs.

GitOrigin-RevId: f2308c81f7474eb31f3380c091c977030123cda2
2026-09-05 15:04:46 +00:00
Matthew Zeng 460b63e5f4 Honor explicit plugin mentions during MCP startup (#42406)
## Why

Selected plugin MCP servers may be skipped after the optional startup grace
period. When a user explicitly mentions a plugin or MCP server, the turn should
continue waiting so the requested tools are available to the model.

## What changed

- Track explicitly mentioned plugin IDs and MCP servers across later samples,
  steered input, and same-turn restarts.
- Require matching selected-plugin MCP servers to finish startup before
  capturing the tool catalog, while leaving unmentioned servers optional.
- Match `plugin://` identities independently of display names and ignore
  targeting query parameters such as `app` and `browserFamily`.

## Testing

Added coverage for linked and structured plugin mentions, direct and batched
capability discovery, delayed environment attachment, steering, and same-turn
restarts.

GitOrigin-RevId: c509ef8df45df1a7edb0da6afdb2682dc53a4de8
2026-09-03 01:33:32 +00:00
Matthew Zeng bd89ce67cc Prepare MCP connections for coordinated OAuth refresh (#42128)
## What changed

- Add the under-development `mcp_oauth_refresh_coordination` feature and pin its selected refresh mode to each MCP connection.
- Treat the refresh mode as part of OAuth connection identity so runtime and MCP configuration refreshes reconnect when the mode changes.
- Keep legacy refresh behavior in builds where coordination is unavailable and emit a warning when coordinated mode is selected.

## Testing

- Cover connection replacement when the feature is toggled through both runtime and MCP configuration refresh paths.

GitOrigin-RevId: 89596150f0338e6ede0e712badecea4450831766
2026-09-01 20:26:57 +00:00
Evan Fannin 53691bad98 Keep MCP event subscriptions alive after task unloading (#41899)
## Why

Event subscriptions may need to outlive the MCP runtime associated with a task.

## What changed

- Add `McpEventStreamOpener`, which retains the hosted Apps MCP connection settings and opens each event stream with an independent client.
- Keep subscription cancellation tied to account changes and removal of the task's event server, including when cancellation ownership is forwarded across runtime replacement.
- Expose the opener and cancellation forwarding through `McpResourceClient`.

## Testing

Add an HTTP transport test that shuts down the runtime, refreshes credentials for the same account, and verifies that a new event stream connects with the current authorization and receives notifications.

GitOrigin-RevId: e102aa382e63121661ff0567f7523515e459922d
2026-08-31 19:18:18 +00:00
Eric Traut 94cbbddafc Support package-style MCP server names (#41700)
## What changed

- Allow MCP server names to contain `:`, `@`, `/`, and `.`, enabling names such as `npm:@modelcontextprotocol/server-sequential.thinking`.
- Preserve these names across `mcp add`, `get`, `list`, and `remove`, runtime tool namespaces, and OAuth credential lookup.
- Quote non-bare server names in generated `config.toml` recovery hints and keep similarly escaped OAuth credential names isolated.

## Testing

- Add CLI and runtime round-trip coverage for npm-style names.
- Add snapshot coverage for quoted recovery hints and a regression test for OAuth credential-name collisions.

GitOrigin-RevId: 97e9353cdbc544db4eb2377ddbfa1d2d539b1e04
2026-08-30 16:45:34 +00:00
pakrym-oai f742dabc6f Support per-tool MCP output limits (#41421)
## What changed

- Add a positive `output_token_limit` setting to each entry under an MCP server's `tools` configuration.
- Apply the most restrictive limit when plugin and user policies overlap, while keeping approval policy independent.
- Carry the effective MCP output budget in conversation history so tool output, post-tool hook responses, and resumed sessions use the same truncation limit.

## Testing

- Cover configuration parsing, serialization, schema validation, and plugin policy merging.
- Cover MCP output below and above the configured limit, post-tool hook responses, and session resume.

GitOrigin-RevId: d0beb4fca9ba6055d9e1d31c137373b465d50d61
2026-08-28 21:40:23 +00:00
jif 92f887ec35 Use refreshed MCP tool caches during binding capture (#41344)
## Why

An MCP binding can wait for one server while another connection updates a
shared tool catalog. Preferring the catalog captured before that wait can omit
newly available tools from the first model request.

## What changed

- Recheck the current MCP tool catalog before using the captured fallback.
- Keep the captured catalog when it expires during the wait, while honoring a
  server that disables catalog caching.

## Testing

- Cover a shared Apps cache populated by a peer while binding capture waits.
- Cover cache opt-out while capture waits for another server.

GitOrigin-RevId: e4389c7ba23e4d8397dbae0930c9e36ba0a172fb
2026-08-28 14:53:53 +00:00
jif 7625343977 Preserve cached MCP tools during binding capture (#41336)
## Why

Binding capture can wait for multiple MCP servers concurrently. During that wait, a tool catalog may be published by another connection or expire after it has already allowed a server to remain dormant. Re-reading only the current cache could omit those tools or trigger unnecessary startup.

## What changed

- Retain the cached catalog that permits lazy startup for the duration of binding capture.
- Recheck the cache after the optional startup grace so catalogs published while waiting are included.
- Add regression coverage for catalogs published during startup and catalogs that expire while another server is resolving.

GitOrigin-RevId: b5c66831db8f9ebecc79ff25c5b4b0182ad68a2f
2026-08-28 13:42:37 +00:00