## Why
Process lifecycle handlers could overwrite each other's updates by reading
and replacing the entire state. For example, a rejected stdin write could
restore an older snapshot after the output collector recorded an exit code.
## What changed
Apply lifecycle transitions in place under the watch channel's write lock.
Rejected stdin writes and confirmed termination only mark the process as
exited, preserving any exit code already reported by the executor.
## Testing
Regression tests cover both closed stdin and unknown-process responses.
They verify that subsequent lifecycle updates preserve the exit code and
sandbox denial recovered from the final read, and that repeated calls to
`fail_and_terminate` preserve the first failure message.
GitOrigin-RevId: f090c0860c86b6e668c31d122d986ea4f4a119b6
## Why
Local stdio discovery needs to accommodate legacy servers that return an
empty response to `server/discover`, take longer to initialize than the
discovery watchdog allows, or require `initialize` as their first request.
## What changed
Mark `mcp_connection_sharing` stable and enable it by default. Treat an
empty discovery response as an unsupported method and let legacy
initialization use the configured startup timeout. Give legacy restarts
a fresh timeout and keep modern version retries within the original
deadline.
## Testing
Cover empty discovery responses, slow initialization on the original
connection, and restarts after ignored or rejected initialization.
Explicitly select legacy mode in tests that depend on legacy behavior.
GitOrigin-RevId: 04d87ddfe02fac8491f879ce420bf294dc725128
Pass the host's existing `&[UserInput]` into `TurnInputContext` to avoid
cloning the input when building extension turn items. Use the borrowed
slice directly for skill mention detection and shadow selection.
GitOrigin-RevId: f4983b584830e481d914f0c9aa1e39e78b965a19
## Why
Automatic mailbox wakes can start new turns while the app-server is
draining or has paused turn admission. Native child results still need
to reach the parent during drain so they can survive a daemon restart.
## What changed
- Require a turn admission permit for ordinary automatic mailbox wakes
and hold it through turn registration. Retry queued work when a
temporary pause ends. Explicit follow-ups and native child results can
still wake the parent during drain.
- Let `thread/inject_items` queue plain-text agent messages for native
child threads when native tools are enabled. Injected mail follows the
ordinary wake rules.
- Extract the remote app-server client, shared event and request types,
and WebSocket compression into `codex-app-server-remote-client` so
callers can use them without linking the app-server runtime.
## Testing
Cover delivery of queued mail when admission reopens, injection into
native child threads, and recovery of a child result after a daemon
restart.
GitOrigin-RevId: 4c5ceeca8829c2612cc2b6b74c03d0c7677a9244
## Why
Namespace-wide guidance can repeat on every entry in flat Code Mode listings. The existing prefix limit covers only reachable `functions` tools, leaving indirect namespace guidance outside the budget.
## What changed
- Add `builtin_tool_namespaces` to put guidance on namespace headers in Code Mode prompts and tool search responses, including headers shared with client tools. The selector applies to namespaces with a registered built-in tool and is omitted from flat listings. Overlapping selectors must agree after trimming.
- Validate all four prefix maps together against the selected model's `description_prefix_budget_bytes`, which defaults to 512. Count each trimmed value once in UTF-8 bytes, including unused selectors. Exclude keys, paragraph separators, and rendered copies.
## Testing
Cover shared headers, flat listings, and conflicting selectors. Exercise the default budget boundary with multibyte text and verify that a model can raise the limit.
GitOrigin-RevId: 1570c23ea1dbad6041cc8e48d0dc2b77f92596c7
Take the payload from the successful tool result and pass it directly
to the post-tool-use hooks.
GitOrigin-RevId: bb7dbe82bbc7b565257b94a0473476326e60db12
## Why
Clients need to distinguish a caller who must enroll in Daybreak from a
model that does not support it. The list of accepted access programs
alone cannot express that distinction.
## What changed
Carry the server-resolved status for each caller and model from the
model catalog into `model/list` as `availableAccessPrograms.daybreakAvailability`.
Update the JSON schemas, TypeScript bindings, and Python SDK to expose
the field.
Rust decoders use `default` for missing or unrecognized statuses. The
Python SDK also defaults missing statuses and preserves new strings
from newer hosts.
## Testing
Cover catalog decoding, app-server response serialization, and Python
SDK round trips with known, missing, and future status values.
GitOrigin-RevId: c1e5ddef68fd19b25e583bb6996a93fd78ab9d18
## Why
A thread can start while managed Guardian policy is being refreshed. Startup
and refresh need to coordinate so the thread does not register with policy
that the refresh has already superseded.
## What changed
Let hosts provide a `GuardianPolicyLoader` to `ThreadManager` for startup
and on-demand refresh. Pause thread registration while a refresh loads and
publishes policy. If publication overtakes a thread's startup configuration,
reload Guardian policy before registering the thread.
Apply the published policy to live threads concurrently. Once published, the
updates continue even if the refresh caller is canceled. Return an error if
any live thread rejects the update.
## Testing
Add regression coverage for an idle thread receiving the update while
another thread holds its session lock, and for reporting a live thread's
rejection of invalid Guardian configuration.
GitOrigin-RevId: ca0cfa2a30f0d45c4ac5ded3400e0e18177ede1d
## Why
Attached reviews used fixed evidence limits even when elastic budgeting
was configured. This could reject a pending tool call whose exact
arguments would fit within the review request's input budget.
## What changed
Use the configured budgeting policy to prepare the pending action and
permissions for attached reviews. With elastic budgeting enabled, check
whether the exact call fits when enforcing the final request budget.
## Testing
Add a regression test for a tool call that exceeds the fixed evidence
limit. Elastic budgeting rejects it when the request budget is too small
and preserves the original call when it fits.
GitOrigin-RevId: 42cc4a8682a98badfe6be95434703500403e7118
## Why
Boxing the thread startup future at each call site makes each caller
compile its own copy of the future.
## What changed
Return a `BoxFuture` from `spawn_thread` and move the startup logic into
`spawn_thread_inner`. Thread start, resume, and fork paths now await
`spawn_thread` directly.
GitOrigin-RevId: 5004651423551e173b1cc98b185c6a1dc2bc097e
## What changed
Remove the legacy framed stdio protocol and its client and server.
Local Codex launches the standalone host with `--listen grpc+stdio://`,
which is also the host's default. The `stdio` and `stdio://` aliases
now select gRPC over stdio. The `features.code_mode_host_grpc` setting
is deprecated and ignored.
Carry execution trace context into gRPC tool callbacks and
notifications. Reject oversized execute requests before sending them
so an encoding failure does not appear as a connection reset.
## Testing
Exercise gRPC callback cancellation, trace context across overlapping
yielded cells, and isolation of tool preload failures between sessions.
Check that both values of the deprecated feature setting are ignored.
GitOrigin-RevId: f5faff3901dc60958627f677f50cfad834291b69
## Why
HTTP sampling can attach private connector result bodies, but Responses
WebSocket sampling leaves them out. WebSocket requests need the same
destination and account checks before sending those bodies.
## What changed
Apply the existing HTTP eligibility checks to each WebSocket attempt and
attach only result bodies that belong to the current account and match
the input being sent, including when that input is an incremental update.
Keep prewarm requests free of private result bodies. Update
`tungstenite` and `tokio-tungstenite` to keep sensitive payload bytes
out of frame diagnostics.
## Testing
Add regression coverage for connector results reaching the server while
staying out of request observers and client logs. Cover custom-provider
requests and retries across an account change with results from both
Code Mode `exec` and `wait`.
GitOrigin-RevId: df5ec8608e59bb3a0444fafd5268279e669ccb55
## Why
A turn interrupted during setup can be discarded before it runs. Starting
the worktree scan during setup can spend time scanning for a turn that
will never send a model request.
## What changed
Collect `HEAD` and remote URLs in the background after turn setup, and
start the worktree status scan only after the turn is admitted to run.
Model requests use the latest available Git metadata while collection
continues, and later requests can include the worktree status once it
is available.
Use shared background Git root discovery for the memory extraction
change probe and pass the resolved root into the probe.
## Testing
Regression tests cover requests proceeding while status is blocked,
later requests carrying the updated `HEAD`, remote URL, and dirty
status, and an interrupted wake skipping the worktree scan.
GitOrigin-RevId: 00f465a5621d672d8b6d9063c384fb62e493c79f
## What changed
Add `ToolLifecycleContributor::mcp_tool_request_meta` so an extension can
use thread and turn data and the captured tool registration to attach an
entry to the outgoing request's `_meta`. Codex rejects duplicate keys.
Expose the registering plugin's package ID in `McpToolContext` so
contributors can scope metadata to that plugin.
Export `PERSISTED_MCP_RESULT_MAX_BYTES`, the existing 64 KiB limit for
untruncated MCP results in durable history.
## Testing
Extend the multi-environment MCP scenario to check that only the call
for the selected plugin receives the metadata and that the result
callback sees the same plugin context.
GitOrigin-RevId: 25a2afc5402a5c9323de184313d614aede3692fb
## What changed
Pass the resolved result-body destination into sampling, where its owner
is used to select retained bodies for the turn. Check that the capture
owner is current under the retention lock before converting an MCP
result and again before retaining its body.
## Testing
Extend the request-budget integration test for direct calls and Code Mode
to check when executed calls carry result metadata. Cover ungranted and
custom endpoints, independent model credentials, hosted Apps OAuth and
header overrides, and disabled result-body passthrough. Check that the
shared configuration preserves metadata on a subsequent turn.
GitOrigin-RevId: 8e5ac88ba323d10e0f3e696d4ae356d52bdf22a3
## Why
Result body capture covered direct Apps calls, leaving eligible calls made inside Code Mode without their result bodies in follow-up sampling metadata.
## What changed
Capture eligible nested MCP results and bind them to the originating `exec` or `wait` output for the current turn. Send `tool_result` and `tool_result_call_id` together when both fit the request budget. Keep the bodies out of conversation history and reclaim capture capacity when the output binding is lost.
Require analytics to be enabled and the provider to allow request metadata before starting turn-level capture. Check the call's credentials at dispatch and again when recording the result.
## Testing
Add coverage for yielded and finished outputs, late results, retries with replacement output items, capture budget recovery, and the request size boundary for sending the body and call ID together.
GitOrigin-RevId: 0983261ac39021099d14efa5c9a9acac47ac9269
## Why
The Responses client checks the gateway URL and request headers before
allowing connector result bodies to be attached. Provider query parameters
can also change the request while leaving those checks satisfied.
## What changed
Require an empty provider query parameter map, or no map, when selecting
a connector result destination. Check the ChatGPT credential variants
directly and retain the current auth owner for thread grant verification.
## Testing
Extend the client setup test to check that a custom provider URL is
ineligible even with refreshed ChatGPT credentials. Extend the enterprise
MCP lifecycle test to check that the configured provider receives the
account bearer token and executed tool-call metadata. Allow HTTP client
tests to cache a direct system proxy decision as well as a proxied route.
GitOrigin-RevId: 1e7e77ec6e65263eaff658ba39943ec51230d332
## Why
Callers of `CodexThread::inject_response_items` each compile a copy of
its concrete future, duplicating compilation work.
## What changed
Return a `BoxFuture` from `CodexThread::inject_response_items` so the
concrete future stays in the core crate. Raise the core integration
test crate's recursion limit to 256.
GitOrigin-RevId: 2ccb280c7887278fc8011c37df35df84c47b3e35
## Why
Remote runners can block bubblewrap from creating a nested network
namespace, preventing the shell from starting and reaching the
filesystem checks.
## What changed
Enable network access in `denied_read_in_second_workspace_root` so the
test can exercise denied reads across both workspace roots without
requiring a nested network namespace.
GitOrigin-RevId: 26eada1e5efde4a31190628f4696c6beb400454d
## What changed
Add the off-by-default `features.connector_tool_result_passthrough` flag. For direct ChatGPT Apps calls, capture eligible text and structured content from the original MCP result before tool output processing. Attach the captured body to the matching executed-tool record on follow-up HTTP Responses requests to the ChatGPT Codex endpoint when the request uses the same authenticated account and turn.
Bound captured bodies in memory for the originating turn. Enabling passthrough alone scopes executed-tool metadata to calls with a captured result ID.
## Testing
Add coverage for same-turn HTTP delivery, omission on later turns or after an account change, and omission of shell-call metadata on HTTP and WebSocket requests when only passthrough is enabled.
GitOrigin-RevId: 410c81559bfc15022b896cf006cb631f3fd9454e
## Why
The telemetry-only integrity checks duplicate policy preparation already
performed by sandbox enforcement.
## What changed
Disable integrity checks at the shared entry point for exec-server and legacy
local execution. Command launches now skip integrity dependency discovery,
policy preparation, and telemetry.
## Testing
Ignore the command-launch integration cases for checker errors and timeouts
while runtime checks are paused. Keep the isolated checker tests.
GitOrigin-RevId: e29d3538e5c8fc943c1a265d75508ee904bfc166
## What changed
Allow callers to attach eligible Apps tool results as `tool_result` on
selected executed tool calls in Responses HTTP and WebSocket requests.
Keep text content and optional object-valued `structuredContent`, with
a 1 MiB per-result limit and a 15 MiB budget for the enriched message.
Attach results after request observers receive the original JSON. Keep
result IDs local and suppress WebSocket payload tracing for enriched
frames. Preserve the enriched HTTP body on retries to the original
endpoint and retry with the original JSON on HTTP 307 and 308 redirects.
Codex's default request construction supplies no result bodies.
## Testing
Add coverage for result eligibility and size limits, HTTP retry and
redirect behavior, and TLS fallback without forwarding the enriched
body to a redirect destination.
GitOrigin-RevId: c3f2f86f31349e40fcb699d48637c3cbbbe38837
## Why
Custom status lines previously depended on a feature flag in the TUI
client's configuration. When the TUI connects to a separate app server,
that flag can differ from the setting in the server's active profile.
## What changed
Add **Custom status line** to `/experimental` as an opt-in feature. Save
the toggle on the connected app server, restart the TUI, then use
`/statusline` to choose a provider.
Read `tui_extensions_statusline` from the app server's effective
configuration at startup. Keep that value when switching conversations
or reloading local status line preferences, so toggling the feature
takes effect on the next TUI launch.
## Testing
Add a menu snapshot and regression coverage for saving the toggle in a
server profile while preserving the active provider across a thread
change.
GitOrigin-RevId: 4025c9771e560d6d7b1e6bd97236bffc44201556
## Why
Proxy settings are bound to the command's execution environment. Preparing
them again at launch can overwrite the settings already bound to the
command, including credentials prepared for its shell.
## What changed
Have `SandboxManager` prepare the proxy environment and sandbox network
context together when the caller hasn't supplied a prepared route. Pass
the prepared environment through to the child and use the prepared
context for sandbox policy, including macOS Seatbelt's proxy ports and
Unix socket rules.
## Testing
Add regression coverage for preserving shell-specific credentials and
environment-bound proxy settings, reporting proxy setup errors for the
requested environment, and deriving Seatbelt rules from the prepared
route.
GitOrigin-RevId: 7a9b5822d8d53f6c93e6318a19222bf06abf0da0
## Why
The sandbox denial check can spend 20 ms waiting for more output even
after the stream has closed and all denial context is available.
## What changed
Classify the buffered output immediately if the stream is already closed.
While the stream is open, wait up to 20 ms for new output or closure.
Register for the closure notification before rechecking the stream state
so closure during setup wakes the check.
## Testing
Paused-clock tests cover output closed before or during the check,
preservation of the timeout for open streams, and classification of
sandboxed and unsandboxed failures. The remote process-event suite adds
a direct sandbox denial case with a `process/closed` event.
GitOrigin-RevId: 1f3029055d70ff4cd63939564872fe5d37240a33
## Why
The denial reconsideration feature only applied to standalone reviews, so
attached Guardian reviews could deny an action after a single verdict even
when the feature was enabled.
## What changed
Move denial reconsideration into the shared verdict runner so attached
reviews can also take one additional sample on a denial. Keep the first
denial and its rationale if the second sample runs past the review deadline.
## Testing
Extend the app-server integration test to verify that attached reviews
sample denials exactly twice with identical input, can approve on the
second sample, and retain the first denial and its rationale on timeout.
GitOrigin-RevId: 06cc730d08dc755863f715362328aa48d15678f0
## What changed
Set `features.guardianv3.message_board_in_memory = true` to give each
new or cleared root thread an in-memory SQLite board. The housekeeper
and reviewer share the board until the last handle closes. The reviewer
can read evidence and update its own subscriptions. On-disk storage
remains the default.
With in-memory storage enabled, the attached reviewer is unavailable
on resumed roots because evidence intake starts at the end of history
and a fresh board has no earlier evidence.
## Testing
Exercise evidence catch-up and later post delivery with both storage
modes. Check that only the on-disk mode creates a database file, the
reviewer cannot update channels, and a cold resume skips evidence intake
in memory mode.
GitOrigin-RevId: 338f44720fca510b2baf40fd4b3a4b30245d4fc2
## Why
Guardian V3 uses fixed guidance for grouping evidence into board channels
and deciding when to retrieve more evidence. Longer routing instructions
also leave less room for evidence in the model's context window.
## What changed
Under `features.guardianv3`, add `housekeeper.routing_instructions` and
`reviewer.retrieval_instructions` to customize those strategies while
preserving the routing format and the reviewer's board and authorization
rules. Blank or omitted instructions use the built-in guidance.
Add an optional `max_instruction_tokens` cap that validates each custom
prompt separately using an estimated token count. Ignore both prompts and
the cap in project config. The `housekeeper.batch_timeout_ms` setting
controls the per-batch routing timeout and defaults to 90 seconds.
Size routing batches to fit the model's input window after accounting for
the prompt and response schema. Route the largest fitting prefix of whole
evidence records within the tool-output budget and leave the rest pending.
## Testing
Cover prompt validation, project config filtering, feature toggle merging,
and default retrieval guidance. Exercise custom prompts in model requests
and verify that routing reduces a batch under context pressure or stops
without advancing the cursor when even one record cannot fit.
GitOrigin-RevId: 83b3fcd23c680e90d2b2aea82a4a290633729658
## What changed
Mark `mcp_lazy_stdio_startup` stable and enable it by default. When a
tool catalog is already cached, Codex can expose tools from an optional
stdio MCP server and defer starting the server until the first tool
call. Required servers still establish a live connection at startup.
## Testing
Exercise lazy startup with the default configuration in the tool-cache
test and explicitly disable the feature for the eager-start control.
Update the code mode and working-directory tests to proceed without
waiting for an optional server when its catalog is cached.
GitOrigin-RevId: a65d83fd1a4c141dba3c10e68b138539aa0a44f0
## Why
A connected account or app can choose an approval reviewer, but that
setting applies across its tools. Users may want to review calendar
event creation themselves while keeping automatic review for other tools.
## What changed
Set `approvals_reviewer` on an individual app tool in `config.toml`:
```toml
[apps.calendar.tools.calendar_create_event]
approvals_reviewer = "user"
```
Codex matches the tool by raw name first and falls back to its title.
The tool's reviewer takes precedence over the connected account, app,
app defaults, and session settings, subject to managed requirements.
This applies to tool-call approval prompts and MCP approval requests
that match a verified app tool invocation. The app-server
`config/read` response also exposes the per-tool setting.
## Testing
Extend coverage for reviewer precedence, tool title fallback, managed
review requirements, tool-call approval routing, strict MCP elicitation,
and app-server config reads.
GitOrigin-RevId: c92bd39b9e708dada9d314c19a446b4c580edf91
## What changed
Reserve `[permissions.credentials]` in `requirements.toml` for shared
credential policy. The table currently accepts no settings and should be
omitted. Empty tables and legacy profile definitions with profile fields
produce a migration error that tells administrators to rename the managed
profile and update references. Profiles named `credentials` remain valid
in `config.toml`.
Accept an empty `[permissions.<profile>.credentials]` table in named
profiles in both files and expose it in the `config.toml` schema.
## Testing
Add regression coverage for rejecting direct and inherited selection of a
user profile through an empty managed credentials table, preserving MCP
configuration reads, and enforcing network policy inheritance checks for
legacy managed profiles.
GitOrigin-RevId: 33dcca30b37d3f07ce670967342199f5302e1d98
## Why
The `incomplete` status alone gives no indication of which timing dependency was missing or which operation prevented capture.
## What changed
Emit deduplicated `critical_path_incomplete_reasons` in turn analytics for incomplete captures. Reasons cover unsupported Code Mode and agent operations, unattributed reviews and user waits, missing turn starts, and forced teardown, with an `unclassified` fallback. External tools that share a built-in name can still invalidate capture without being classified as the built-in operation.
During interruption, use the task's join result to distinguish forced teardown from normal completion so a task that already finished can still report a complete path.
## Testing
Add coverage for analytics serialization, concurrent reason deduplication, tool name collisions, and interrupted tasks that finish normally or require forced teardown.
GitOrigin-RevId: fab9aa1ff881997a34dcf75adc5638133cedc7fe
## Why
If a provider keeps an SSE connection open after reporting a failure, Codex
can hit the idle timeout and lose the original error and its retry advice.
## What changed
End the SSE stream immediately on `response.failed` or a recognized
`error` event. Classify statusless quota, plan usage, invalid prompt, and
misalignment errors consistently across both event forms. Preserve
`Retry-After` for overload and capacity errors, preferring error-level
headers over event-level headers when both provide valid advice.
## Testing
Add SSE regression tests that hold the connection open and verify immediate
error delivery, terminal error handling, and retry delays. Extend WebSocket
coverage for statusless errors received before the close handshake.
GitOrigin-RevId: 69afd1d3b18ca3139e9b7ad98ae1481a432179f0
## Why
When the Responses API sends a policy rejection in an `error` event
without an HTTP status, the SSE client can wait until its idle timeout
to report the failure.
## What changed
Recognize `bio_policy` and `cyber_policy` in the shared terminal error
parser so both SSE and WebSocket clients surface the corresponding
policy error. Preserve the provider's message and use a default when
it is missing or blank. End the SSE stream as soon as the error arrives.
## Testing
Add SSE tests for both error codes, including when the stream stays
open, and WebSocket cases that verify both errors surface without a
close handshake.
GitOrigin-RevId: 1b46b6557a39a61a1bb10bdd0d8ede48ecd6fa13
## Why
A parent turn can acknowledge a child agent's final reply while a model
request is in flight. If the parent is interrupted before consuming the
queued reply, turn cleanup can discard it.
## What changed
Record queued agent messages in conversation history during interruption
and flush the rollout before emitting `TurnAborted`, even when the
interruption marker is disabled.
## Testing
Add a regression test that interrupts a parent while a child's final
reply is queued. The test checks that the reply is persisted when
`TurnAborted` arrives and appears exactly once in the next model request.
GitOrigin-RevId: 5bd828c4101942fa4fe7ba5040355a8244f02018
## Why
The MCP tool call future reserves space for the large config reload
future even on ordinary calls that never reload configuration.
## What changed
Pin the reload future on the heap after successfully persisting a tool
approval. Ordinary MCP calls no longer carry the inline storage for it.
GitOrigin-RevId: a883f7619f520e730dda139d967c6a0f945b5df2
## Why
Generic quota errors hide whether the project or organization spend limit
blocked a request. Reporting the scope tells users which limit to check.
## What changed
Recognize API key, project, and organization spend limit errors across
HTTP, SSE, and WebSocket responses. Show a message that names the scope and
treat the error as non-retryable, even when the provider sends
`Retry-After`. Apply the existing usage-limit handling to goals and
compaction.
Expose distinct `CodexErrorInfo` values in the app-server protocol and
generated bindings so clients can identify the limit in live errors and
failed turn history.
## Testing
Extend the transport tests to cover all three scopes and verify that they
stop retries. Extend the app-server test to check that the error survives
`thread/read` and that a subsequent turn can succeed in the same thread.
GitOrigin-RevId: 03341f7a78daa7540e7baa519f1a29ec07e34509
Record how long initial context assembly waits for prompt-only plugin
recommendations in `codex.initial_context.plugin_recommendations_wait_ms`.
Use named info-level spans for loading recommendations, reading the
tooling change marker, and fetching the recommendation catalog.
GitOrigin-RevId: 67b26f79252f8cba6e770f43df252ec1d73a569f
## Why
Remote Noise executors previously connected outbound to a rendezvous relay. An inbound listener lets harnesses that can reach the executor connect directly while retaining registry-backed authorization.
## What changed
Use `--listen ws://IP:PORT` with `--remote` in Noise mode to accept inbound WebSockets. The listener serves RPCs over encrypted Noise streams after the registry authorizes the harness. It limits concurrent connections and handshake attempts, closes sockets that miss the first authorization deadline, and drains connections on shutdown.
Noise clients can attach optional admission headers from the connection bundle to `wss://` requests.
## Testing
The inbound integration test exercises a registry-authorized file read, rejection of a registry-denied harness, the absence of a plaintext RPC response, and connection closure on shutdown. Unit tests cover handshake limits and rejection of admission headers before dialing a plaintext WebSocket.
GitOrigin-RevId: b388672e49a3fe806e38f6db38825435ba25d27a
## What changed
Have `SandboxManager` handle backend selection and shared preparation, then
delegate command wrapping and session setup to the Linux, macOS Seatbelt,
Windows token, and Windows MXC backends.
Centralize Windows managed proxy routing in `codex-sandboxing` so the exec
server and other callers use the same rule for selecting dedicated MXC
listeners.
GitOrigin-RevId: a75886809aefebf6c267bb39dbc104f76bda3409
## Why
A prepared agent fork needs its ancestor's conversation for context.
Treating that conversation as the child's own authorization evidence
can attribute the ancestor's messages and approvals to the child.
## What changed
Allow host-started prepared forks that reference stored history to
inherit the root session ID. Keep ancestor messages as conversation
context and exclude ancestor authorization evidence when starting or
resuming the fork and when reconstructing sender evidence from storage.
Preserve the inherited disabled-plugin settings and the child's own
authorization evidence.
## Testing
Cover prepared forks across startup and resume, including inherited
context, plugin settings, and child-local authorization. Verify that
live and stored sender evidence both contain only the child's messages.
GitOrigin-RevId: 5c5a6547d31fe78dfad3742efb8a41df539173c6
## Why
An attached reviewer needs to catch up on evidence before issuing a verdict. If the approval context changes during that preparation, the first verdict must still be checked against the authorization captured before preparation began.
## What changed
Bind the attached reviewer to the pending action and reserve a review session after evidence preparation. Submit the action and its permissions for a synchronous verdict, checking authorization before and after the verdict. Retries reserve a new session under the original review deadline.
## Testing
Extend the review tests to check that preparation and verdict execution share a deadline. Cover when optional synchronous review falls back to user approval after local input exhaustion and when context-limit or server failures result in denial.
GitOrigin-RevId: 0ea835ba897069d19d045605d97b58f47406a4b4
## What changed
- Extract synchronous verdict execution into the Guardian review session.
It starts the review turn, tracks evidence admission, and clears
pending routing and context when the attempt ends.
- Add a context builder for reviewers that already have the evidence.
It presents the exact pending action and any applicable permission
restrictions without repeating the transcript.
## Testing
Add a composition test that checks that an encrypted tool call is
replayed unchanged with its matching evidence-only output.
GitOrigin-RevId: a0c120936c80459b0928550d23ba1137e125013c
## What changed
In opt-in local MCP mode, deliver board notifications through
`inject_response_items` when the recipient uses Multi-Agent V2 with native
tools enabled. Other deliveries continue to use `Op::InterAgentCommunication`.
## Testing
Extend the HTTP board scenario to check that the parent receives exactly one
`CHANNEL_POST` agent message with the worker as author, the parent as
recipient, and the expected post preview text.
GitOrigin-RevId: deaa1451619a94d9c70c771327892069325c5e60
## Why
A realtime call can end with completed speech that has yet to reach a
Codex turn. The next turn needs that context even if the call's transport
has already disconnected.
## What changed
- For V3 calls with realtime frame history enabled, save pending speech
as context for the next turn. Restore it on thread reload and exclude
it from the turn count used by rollback.
- Drain already received transcript events and settle pending handoffs
before saving the call's closing marker. Report recovery required if
shutdown cannot confirm that history was saved.
- Keep transport detach separate from ending a call. Reattaching restores
saved pending speech, and an open call can be ended from persisted
history after a reload. Use the call ID for TUI attachments so successive
calls on one thread have distinct history.
## Testing
Add coverage for queued speech at hangup, reconnect and stop after
reload, retries after an uncertain write, and rollback across saved
session-end context.
GitOrigin-RevId: 657491f1d7e446a1a48e22189758810c8bdd82cb
## Why
Raw response injection bypasses the native mailbox wake policy, so
injected agent messages cannot wake a sleeping thread through that
policy.
## What changed
In native Multi-Agent v2, route batches consisting entirely of
single-text agent messages with valid agent paths through the mailbox
as queue-only mail. Hold a residency pin across the batch so capacity
eviction cannot interrupt delivery between submissions. Mixed batches,
multipart or encrypted messages, and configurations with native tools
disabled continue to use raw response injection.
Only preempt model sampling for pending mail when the matching turn
still accepts mailbox delivery.
## Testing
Extend coverage for waking live and reloaded sleeping threads,
delivering older queued mail before a new task, and retaining mail in
order when a batch exceeds the submission channel capacity and
eviction is attempted.
GitOrigin-RevId: 3df60d2a4a9e7b193ef298e01f31c565ad8a2785
## Why
Realtime history needs to record which transcript frames reached the
coding agent and which turn received them. If a handoff fails or the
conversation reconnects, Codex needs to distinguish speech that was
already used from speech that is still available for the next input.
## What changed
- Route handoffs through normal turn admission to start or steer a turn.
When realtime frame history is enabled, persist the transcript frames
and the receiving turn's input before publishing acceptance or exposing
the input to the model. Attribute only transcript frames represented
in the rendered input.
- Reserve transcript frames across overlapping handoffs and release them
on definite rejection. Retry an uncertain write with the same input
and require history recovery if the write still cannot be confirmed.
- On V3 reconnect, restore persisted speech that has not been consumed
and put it before newly captured speech.
## Testing
Add regression tests for attribution after transcript truncation,
overlapping handoffs, V3 reconnect ordering, and preservation of typed
input when a handoff in the same turn fails before model sampling.
GitOrigin-RevId: 432dd51971f7ced20344ebc7f0c7b007bb849f54
## Why
The in-memory MCP backend depended on the runner's clock, so a runner
clock failure could prevent posting. The standalone SQLite backend
already uses the system clock.
## What changed
Read UTC time from the MCP backend's system clock after validating
caller membership. Apply this to both tree and agent-specific bindings.
## Testing
Update the HTTP integration scenario to make the runner clock fail and
verify that a child agent's post timestamp falls within the posting
window.
GitOrigin-RevId: bad5dec5a9be7246cf627b017897fbfe2c960cd1
## Why
Managed approval requirements can change while an MCP runtime is still in use.
Later turns need to honor those changes without restarting MCP, while in-flight
MCP actions need to retain the mandatory review requirements of their issuing
turn.
## What changed
- Add `CodexThread::refresh_guardian_policy` to apply managed policy updates to
an existing thread. MCP configuration refreshes preserve the active Guardian
policy, including when a policy update lands while MCP configuration is loading.
- Resolve MCP tool and elicitation reviewers against the issuing turn's policy.
Live requirements can still require Guardian review, and a live policy change
cannot waive mandatory review captured by the turn.
## Testing
Extend coverage for overlapping Guardian and MCP refreshes and for an existing
MCP runtime that must prompt the user on the next turn after a policy update
allows only the `user` reviewer.
GitOrigin-RevId: 5a1649f917313f6fb152a206bc89986fe6345481
## Why
Legacy rollout migration rewrites rolled-back history and can disrupt
the realtime stream. Enroll local threads once they use paginated history
so the stream stays intact.
## What changed
- Add the off-by-default `features.realtime_frames_local` flag for
paginated local threads with a state database. Accept runtime updates
through the app server and keep threads enrolled once they have a
persisted canonical session start.
- Add thread-store APIs to recover a realtime session and look up the
latest session ID from durable history. Route frame completion through
a store seal so completion acknowledges durable writes.
- Add a local seal for accepted model input that writes source frames
before the input and lets callers retry an uncertain flush without
appending the input twice.
## Testing
Cover enrollment for legacy and paginated threads, continued enrollment
after disabling the flag, and retrying a local input seal after a failed
directory sync.
GitOrigin-RevId: 2012feeca3857da90f9b026f077b31a26d15191b