## Why
Some hosts assign a worker per connection. Idle connection eviction can prevent a quiet session from reaching the worker holding its lease, while retaining a connection after the last session closes unnecessarily keeps that worker allocated.
## What changed
- Disable the HTTP connection pool idle timeout for gRPC code-mode connections.
- Cache URL-based clients weakly and let live sessions own the shared client, releasing their ownership on closure or failure even while the provider remains alive.
- Acquire clients for deferred RPCs only when needed and return the session error if the client has already been released.
## Testing
Add regression tests that verify connection reuse and stored state after 120 seconds of inactivity, and connection closure after the last shared session fails while the provider and session handles remain alive.
GitOrigin-RevId: bd1ff00c4003c03ed173c23dd2ed373ed354ab65
## Why
Shared hosts serve independent chats and their subagents. Larger callback workloads need room for unrelated work and control messages, while gRPC completion bursts can exhaust HTTP/2's small-DATA-frame budget and disconnect shared sessions.
## What changed
- Raise the active-cell limit from 128 to 2,048 and the pending delegate callback limit from 1,024 to 4,096. Expand request and transport queues to leave room for completion bursts and control traffic.
- Negotiate larger callback budgets through the stdio `delegate-capacity-4096` capability and gRPC `max_pending_delegate_calls` field. Preserve the 1,024-callback budget for legacy clients and enforce both session and shared host limits on gRPC.
- Limit concurrent gRPC completion RPCs to 64 across a provider's sessions, allowing queued completions to be canceled without waiting for delivery capacity.
## Testing
Add stdio and gRPC integration tests covering 256 cells and 4,096 callbacks across shared sessions while unrelated work completes. Add tests for callback budget isolation and reuse, rejection of zero capacity, and cancellation of queued completions. Include an ignored resource-intensive test for full host capacity.
GitOrigin-RevId: e85bfa24cc46cc8a735b464d80adf8066a6a6d3a
## Why
Host-side truncation can bound Code Mode output before it crosses the transport, including script errors that would otherwise exceed the frame limit.
## What changed
- Negotiate server-side output truncation for stdio and gRPC sessions, with client-side fallback for older hosts.
- Forward output budgets for execution, waiting, and termination when server-side truncation is selected. Avoid truncating host-formatted output or appending script errors a second time.
- Preserve the negotiated mode across reconnects and reject hosts that cannot maintain it.
## Testing
Add transport and integration coverage for output budgets, oversized script errors, reconnect compatibility, and preservation of output-limit notices with a zero-token budget.
GitOrigin-RevId: 2075101e4845eca64f45491c27a3b326788645c8
## What changed
- Add `GetProcessMetrics` and provider sampling APIs for process ID, start time, cumulative CPU time, and resident memory, exposed through `ThreadManager` without creating a thread or session.
- Distinguish unsupported collection, a locally absent host, and sampling failures. Process-owned providers never start or replace a host when sampling.
- Bound gRPC sampling to one second without retries or explicit connection resets. Run optional collectors on a detached worker with at most one collection in flight, so stalled collection cannot block the HTTP/2 driver or runtime shutdown.
## Testing
Add coverage for snapshots, collector errors and panics, deadlines, overload, and shutdown with a stuck collector. Extend lifecycle tests to check that sampling preserves session state and keepalive behavior without starting or restarting local hosts.
GitOrigin-RevId: eeafe275891d0464636f28a29345a87aed1a12dc
## Why
Automatic gRPC reconnection could spawn a replacement host while dispatching an RPC for an existing session. Host loss must remain visible so recovery can report lost state without replaying execution or completed external effects.
## What changed
- Bind each local host channel to one child process and allow only `OpenSession` to replace a dead host.
- Add caller-owned transport support to `GrpcCodeModeSessionProvider` and map local transport failures to `Unavailable` so initial connection failures use the existing bounded admission retries.
## Testing
Add regression coverage for child exit during initial connection and between channel selection and dispatch, concurrent session recovery, and child reaping after teardown. An integration test kills the host after a command completes and verifies that recovery reports the reset, clears stored state, and leaves the command's external effect unchanged.
Increase the remote MCP stalled-handshake test's initialization timeout and mock response delay to let requests reach the server on loaded CI workers.
GitOrigin-RevId: 4ffd810e4cfcfb7a56e606816aa867bc206a09b6
## Why
Reloading an evicted sub-agent disabled its rollout tracing, leaving later turns out of the trace. Replacement Code Mode sessions could also reuse cell IDs from earlier sessions.
## What changed
- Reattach reloaded children to the root rollout trace and emit `ThreadResumed` without duplicating `ThreadStarted`.
- Reopen existing threads during replay, preserving their identity and history while clearing stale runtime cell and terminal process mappings.
- Allocate Code Mode cell generations across sessions for both gRPC and stdio hosts to prevent ID reuse after reload.
- Record an aborted trace when session startup fails before the session loop begins.
## Testing
Add regression coverage for child eviction and reload, including a failed reload, continued inference tracing, distinct cell IDs, and preserved mail and follow-up context. Extend replay tests for history preservation and reused runtime handles, and host tests for cell IDs across session replacement and reconnection.
GitOrigin-RevId: 6c146c04b5cda9295afb7d973c0d5474b363a822
## What changed
Add session options for client- or server-owned output truncation and optional `max_output_tokens` budgets for wait and terminate observations. Introduce `TerminateRequest` and update session implementations and callers to use it.
Extend the stdio and gRPC schemas with optional truncation settings and budgets, and define the `bounded-output-v1` capability. Existing sessions retain client-side truncation; providers reject server mode by default. This change defines the contracts without enabling server-side truncation.
## Testing
Extend protocol serialization coverage for `outputTruncation: "server"` while preserving existing wire shapes when optional fields are absent.
GitOrigin-RevId: 22145897ea16d385c5ba9e7549109f7a669fceb4
## What changed
- Configure the code-mode host to send zstd-compressed responses when clients advertise support, and enable both client transports to accept them.
- Preserve uncompressed responses for clients that do not advertise zstd support.
- Patch `tonic` in Bazel builds to use zstd compression level 1, favoring response latency over compression ratio.
## Testing
Extend the TCP session integration test to check response encoding and successful session opening for both zstd-capable and uncompressed-only clients.
GitOrigin-RevId: b6cf0da51dd30fc2ac9a1b0af11bec26555c559d
## What changed
- Add `CodeModeCellTiming` notifications for yielded executions, terminal waits, and unavailable cells, with thread, call, and cell identifiers.
- Measure how long completed results remain buffered before retrieval and carry `result_buffered_for` through the runtime, stdio protocol, and gRPC response metadata. Immediate delivery reports zero; waits without a completed result report no buffering measurement.
- Default missing buffering data to zero for compatibility with older hosts.
## Testing
Extend runtime tests to verify buffering duration with a paused clock and retrieval through termination. Add gRPC conversion tests for nanosecond precision, optional metadata, and the distinction between immediate delivery and no completed result.
GitOrigin-RevId: 60c0c5a3affa69ac3779bf1812fca50f5748af76
## Why
Replacing a gRPC code-mode runtime loses stored values and running cells. Silently continuing after reconnection hides that loss, while prior external tool operations may already have completed.
## What changed
- Capture the runtime generation when `execute` and `wait` calls are created. If recovery replaces it, reject those calls with an explicit reset error, including concurrent calls and calls first polled after recovery.
- Explain the lost state and advise inspecting prior tool results before retrying. Allow a fresh `exec` call to continue on the replacement runtime.
- Keep `terminate` from reopening the runtime so cleanup cannot hide a reset from the next request.
## Testing
Extend gRPC recovery tests to cover reset errors after host restarts, missing sessions, and route loss, rejection of queued calls, and successful fresh execution with cleared stored state.
GitOrigin-RevId: 914cab3bb45e30a3f3b2afacb5631e42b18e21b0
## Why
Sessions sharing a transport can be admitted to different hosts. Each session needs its own routing token so subsequent RPCs reach the host holding its state.
## What changed
- Capture the optional `x-code-mode-route` metadata from successful `OpenSession` responses and attach it to subsequent session RPCs. Reject tokens longer than 256 bytes and keep new admission attempts free of existing session routes.
- Invalidate a routed session when an `Unavailable` response explicitly sets `x-code-mode-route-lost: true`, allowing a later operation to reopen it without replaying the failed execution. Ordinary `Unavailable` responses preserve the session.
- Allow up to five admission attempts for `Unavailable` and `ResourceExhausted`, with an immediate first retry followed by jittered exponential backoff.
- Order session failure with shutdown so late callback errors cannot cancel an in-flight `CloseSession` RPC.
## Testing
Add regression coverage for independent routing on a shared provider, route-loss recovery without execution replay, preservation of unaffected session state, and callback failures racing shutdown. Extend admission tests to verify the five-attempt limit and recovery tests to distinguish explicit route loss from ordinary unavailability.
GitOrigin-RevId: 5d60ef1afa56beb8c70ba7ab27865128225b0d3d
## Why
A host can reject a request for an unknown session while its lease streams remain open. Keeping that session binding prevents subsequent execution from recovering.
## What changed
Invalidate the session when `Execute` (including its response stream), `Wait`, or `Terminate` returns `NOT_FOUND`, allowing the next execution to open a fresh session. Return the failed call's error without replaying execution.
Document that these RPCs reserve `NOT_FOUND` for unknown sessions, while `CompleteToolCall` can use it for a missing invocation.
## Testing
Add regression coverage for recovery with lease streams still open, no replay of failed execution, and preservation of session state after `INVALID_ARGUMENT`. Add a cancellation-race test confirming that a late tool-completion `NOT_FOUND` leaves the session open.
GitOrigin-RevId: 85b38fdbd6d59268447b6aa6a85ea7c60ec869a7
## Why
Transient admission errors or connection resets can prevent a code-mode session from opening before any execution starts.
## What changed
- Retry `OpenSession` and the initial lease message on `Unavailable` or `ResourceExhausted`, with at most three attempts and 100 ms and 200 ms delays within a single startup deadline.
- Classify HTTP connection and request failures as `Unavailable` so admission can recover from connection resets, while preserving error sources.
- Keep retries limited to initial admission so established executions are never replayed.
## Testing
Add gRPC integration tests covering transient status recovery, the retry limit, immediate failure for non-retryable statuses, and successful session execution after a connection reset.
GitOrigin-RevId: 3b654bcd377a5addf7c3ed58bb1c72a449b2f7f7
## What changed
Add an optional `preempt` signal to `CodeModeSession::execute` and `CodeModeSession::wait` so callers can end an observation early while the cell continues running and remains available for later waits.
Support the signal in the in-process runtime and across both remote transports: `operation/yield` with the negotiated `yield-observation` capability for stdio, and `YieldObservation` for gRPC. Preserve signals received before gRPC observation registration, and retain normal timeout behavior for older hosts that lack support.
## Testing
Add runtime and transport tests covering early execute and wait yields followed by successful cell completion, yields before observation registration, and suppression of yield frames for older stdio hosts.
GitOrigin-RevId: 3f8049d8f0582be6b8f0832cc03f2d974a0b8f18
## 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
## What changed
- Load `application.network` requirements at startup and explicit config/account reloads, applying local policy before authentication and cloud configuration bootstrap.
- Block traffic when policy loading fails, cancel requests that a new policy no longer permits, and reject configuration loads based on superseded policy snapshots. Local requirements edits take effect on the next explicit reload or restart.
- Bind authenticated clients to the current account and revoke them when account ownership changes. Carry policy enforcement through backend requests, realtime connections, workload identity exchanges, and code-mode gRPC streams while preserving gRPC trailers.
- Treat denied credential refreshes as policy errors without retrying them or invalidating stored credentials.
## Testing
Add regression coverage for startup restrictions, reload ordering, cancellation after restrictive or malformed requirements edits, account revocation, gRPC destination denial, workload identity cancellation, and credential preservation after denied OAuth refresh.
GitOrigin-RevId: 0cb4c720a906397e9df06675f01ecf60d741101f
## Why
A delegate bound at session creation cannot provide different callbacks for cells sharing that session. Each execution needs to retain its own delegate across yields and release it when the cell is cleaned up.
## What changed
- Pass `CodeModeSessionDelegate` to `execute` instead of session creation.
- Route tool calls, notifications, and cell closure callbacks through the execution's delegate in the in-process runtime and the gRPC and stdio transports.
- Retain delegates with pending executions and live cells, releasing them through closure and cancellation cleanup.
## Testing
Add coverage for distinct delegates across yielded cells, gRPC callbacks before cell admission, and delegate release after completion or abandoned execution cleanup. Update transport tests to verify callbacks reach the owning cell's delegate.
GitOrigin-RevId: 7469f52104b993e790a40c65fb800ad02b7fd606
## Why
Code mode wall time should measure the host operation itself, without including
client-side response delays or idle time between requests.
## What changed
- Measure each execute, wait, and terminate request in the code mode host.
- Carry the duration through the stdio and gRPC protocols and use it for
model-visible wall time.
- Emit a structured `codex.code_mode.host_timing` event correlated with the
conversation, turn, tool call, and cell.
## Testing
- Cover successful and failed execution timing, delayed response reads,
repeated waits, termination, and missing cells across stdio and gRPC.
- Verify timing survives protocol serialization and is reflected in app-server
model output and structured telemetry.
GitOrigin-RevId: d24af30c3fc5820521b4beba1f9970714dad6482
## What changed
- Inject W3C `traceparent` metadata into code-mode session and execution requests.
- Carry each execution's trace context on streamed tool callbacks so callback and nested-tool spans remain connected across the gRPC boundary.
- Preserve the relevant spans when execution, callback, and dispatch work moves into spawned tasks.
GitOrigin-RevId: 6482963cb9b2e2f0ea27f1345bba208b70a80d0b
## What changed
- Forward notification text to the session delegate without applying the previous 1,024-byte limit or appending a truncation suffix.
- Update the gRPC host integration test to verify that oversized multibyte notification text is delivered unchanged.
GitOrigin-RevId: 9a9e24b359a07540f70ec4e98b28524db3f7a4a0
## Why
Code-mode tool failure messages larger than 64 KiB were truncated before they
reached the host.
## What changed
- Remove the tool error size limit from the gRPC protocol and host validation.
- Forward failed tool completion messages without truncation.
## Testing
- Verify that a multibyte error larger than 64 KiB is preserved exactly.
GitOrigin-RevId: 264ae4ba4adea5c19b669e4f41ccfd41a0c30fb5
## What changed
- Accept root `http://` and `https://` URLs in `--code-mode-host` and use the
shared gRPC session provider for those endpoints.
- Keep `ws://` and `wss://` URLs on the existing WebSocket transport.
- Reject paths, queries, fragments, and credentials where unsupported, without
exposing gRPC URL credentials in command-line validation errors.
## Testing
- Cover argument parsing and transport selection for both remote protocols.
- Exercise a gRPC host shared across app-server threads and verify credential
rejection does not disclose usernames or passwords.
GitOrigin-RevId: b6516a85cf76db5c4cea620f89ef866d8af30cf0
## What changed
- Reopen a cached code-mode session when its gRPC host stops, while
serializing concurrent reconnection attempts and coordinating shutdown.
- Scope cell IDs to the new host generation so callbacks remain consistent
and stale `wait` or `terminate` requests are rejected.
- Accept both `unix://` and `unix:` endpoints for gRPC hosts on Unix systems.
## Testing
- Cover host restart recovery, concurrent execution after reconnection,
generation-aware callbacks and cell operations, stale cell rejection, and
Unix socket execution.
GitOrigin-RevId: 548e168fdcef7f7d54bd32262e614886bf7bdd32
## What changed
- Build URL-based gRPC code-mode connections with `HttpClientFactory` so they support the application's outbound proxy and custom CA configuration.
- Accept `http` and `https` origins while rejecting endpoints with unsupported schemes, paths, queries, or fragments.
- Preserve custom tonic channel injection and gRPC frame-size limits through the new transport adapter.
GitOrigin-RevId: 142f0b572b3ab0154e8fe860cb304752a9af5784
## What changed
- Subscribe each gRPC code-mode session to nested tool calls and forward tool and notification callbacks to its delegate.
- Complete tool calls through the host while bounding oversized results and errors.
- Track callback ownership and cancellation so completed cells drain notifications, terminated cells cancel them, and shutdown revokes outstanding work.
- Validate callback identifiers, cell ownership, enabled tools, and pending callback limits without serializing independent callbacks or sessions.
## Testing
- Add integration and state tests for callback forwarding, completion ordering, cancellation, malformed callbacks, delegate panics, oversized results, and concurrent work.
GitOrigin-RevId: 005afbb90eea0eb77d746b930a1a96ca6dfcd4e7
## What changed
- Add `GrpcCodeModeSessionProvider` for opening code-mode sessions over HTTP/2 or an existing `tonic` channel.
- Support execution, waiting, termination, per-session limits, cell-closure callbacks, and graceful shutdown over the gRPC protocol.
- Bound transport waits and error messages, validate host identifiers and responses, and clean up abandoned executions and observers.
## Testing
- Add end-to-end TCP tests covering session persistence, cancellation, concurrent waits, shutdown, cell cleanup, and independent yield limits.
- Add unit coverage for protocol conversion, deadlines, and session lifecycle state.
GitOrigin-RevId: d4729ce608ad4b42a99744b07e1f230e46cb24ec
Alex Kotliarskyi