## What changed
- Make `Session::spawn` return a `BoxFuture` while preserving lazy session
initialization.
- Prevent the startup wrapper from being inlined so callers do not expose its
concrete future.
- Await the returned future directly at session creation call sites.
GitOrigin-RevId: 5432745aae1c934b16202716a8d49a8553de2018
## Why
Context compaction can replace original conversation items that Guardian still needs to review approval-sensitive actions.
## What changed
- Retain a bounded, chronological review history independently of the model's compacted history.
- Keep separate limits for user messages and other transcript items so tool traffic cannot evict user instructions.
- Use the retained evidence for synchronous and asynchronous Guardian transcripts, while resetting it after rollback or history reconstruction.
## Testing
- Cover retention through compaction and eviction, reset behavior after rollback, independent size limits, and asynchronous transcript construction.
GitOrigin-RevId: c5851e992820504fc18de8cfec7f6f9f1b69f316
## What changed
- Add model metadata flags for enabling token budgeting and its history-notes extension.
- Apply those activation defaults once at thread startup when the user has not explicitly configured token budgeting, while respecting managed feature requirements.
- Continue resolving model-owned prompts for the active model so model switches can use their own token-budget guidance.
## Testing
- Update the token-budget integration test to verify that model metadata activates the feature without manual configuration.
GitOrigin-RevId: 73a32623eaeabb9243e240c43da7a5e7e93e9a19
## What changed
- Default `tools.update_plan.enabled` to `false`; users can explicitly enable it to expose `update_plan`.
- Remove bundled `update_plan` guidance from model, collaboration-mode, multi-agent, compaction, prewarm, and goal-continuation prompts when the tool is disabled.
- Preserve custom base instructions, model catalog instructions, collaboration policies, and user goal text even when they mention planning or `update_plan`.
## Testing
- Cover default and explicitly enabled tool registration, prompt consistency across request paths, and preservation of custom instructions.
GitOrigin-RevId: a53964e6e72f98e2557dd3090fdb7caccb956527
## What changed
- Set `history_ingest_requested` to `true` in Responses turn metadata when the
history-notes token-budget extension is enabled.
- Reserve the metadata key so caller-provided metadata cannot override the
core-owned value, and omit it when the extension is disabled.
GitOrigin-RevId: e750bcb729292c407e1d443857cdb584a88aac97
## Why
Compaction and host-injected context can rewrite the model-visible conversation
without changing what the user authorized. Treating those updates as authorization
changes prevents Guardian from reusing an otherwise valid review.
## What changed
- Track a host-owned user-message revision separately from the conversation history
generation.
- Advance the revision for genuine user messages and history resets, while preserving
it across compaction and internal context injection.
- Use message content-kind metadata to distinguish host context from user input,
conservatively treating unknown or incomplete metadata as user authorization.
## Testing
Added coverage that cached Guardian authorization survives compaction and internal
context, but is invalidated by user input and rollback.
GitOrigin-RevId: c7960fa182d1686042c23764d2abc58acbc4b882
## What changed
- Cover the default, explicitly enabled, and explicitly disabled states of
`tools.update_plan.enabled`.
- Verify prompt tool lists remain consistent across requests, including when
custom base and developer instructions are configured.
- Explicitly enable `update_plan` in existing fixtures that depend on the tool.
GitOrigin-RevId: 3be279cb5195dfc5892c2c8ff6a1461fac879309
## Why
Resuming a thread without an explicit `cwd` should restore that thread's latest
retained setting. Forked history can contain settings copied from another
thread, while compaction can move the latest setting outside the replay window.
## What changed
- Record the owning thread ID on new settings snapshots and only use snapshots
owned by the resumed thread when restoring `cwd`. Older snapshots without an
owner remain readable but do not override the startup `cwd`.
- Checkpoint the current settings after compaction and serialize checkpoints
with settings updates so the retained history contains an accepted, current
snapshot.
- Cover resume behavior across compaction, forks, reverts, legacy histories,
and concurrent settings updates.
GitOrigin-RevId: 1efc9cf55472d6b88c465c4efe44672a232dbf64
## Why
Automatic goal continuations should remain attributable to the turn that created the goal. External input, hook context, or goal edits can make that attribution ambiguous and must not leave stale lineage metadata behind.
## What changed
- Carry the trusted root and previous parent turn through successive automatic goal continuations.
- Invalidate stored lineage when external context reaches an active turn or when a goal is edited or cleared.
- Preserve lineage for async hook results only when they belong to the receiving turn.
## Testing
- Cover goal continuations across intervening user turns and external objective edits.
- Cover active, buffered, and externally injected hook or response context.
GitOrigin-RevId: 22ccc07961cd0ed8f4b17912b7efa416b342bf50
## Why
The `core` and `tui` Bazel targets treated nearly their entire crate trees as
compile-time data, mixing embedded runtime resources with source files and test
fixtures.
## What changed
- Move embedded resources for `core` and `tui` into dedicated `assets/`
directories and update their `include_str!` paths.
- Restrict Bazel compile data to explicit asset and frame directories.
- Move test-only prompts into `tests/fixtures`, resolve the core prompt through
`find_resource!`, and reuse the integration-test file glob when assembling
Bazel test data.
GitOrigin-RevId: eee9b0e41df77ea2604215a7f7333a6fe8317bfa
## What changed
- Defer legacy sandbox policy projection until a working-directory change may require rebinding project-root writes.
- Keep client name and version updates from resolving filesystem paths or altering the existing permission snapshot.
## Testing
- Add regression coverage for updating client metadata with a managed, restricted filesystem policy that includes both a writable path and a nested denied path.
GitOrigin-RevId: d84dbcd09b9d1961a11b2889d79479e30eb54149
## What changed
- Add model message metadata for built-in tool descriptions.
- Use the active step model's catalog description for `send_user_message_async`, including after a mid-turn model change.
- Fall back to the built-in description when catalog metadata is absent, while preserving an explicitly empty description.
## Testing
- Cover catalog serialization, fallback and empty-value behavior, and model changes within a turn.
GitOrigin-RevId: 72b953214ab9708931065321debf12133f8d2d40
## What changed
- Add an optional `proactive` multi-agent mode message to model metadata.
- Use the catalog's proactive message for `Ultra` reasoning when no general mode hint is configured. A missing value falls back to the built-in proactive instructions, while an empty value suppresses the mode message.
- Keep explicit mode behavior for other reasoning efforts and refresh the applicable catalog message when the model changes.
## Testing
- Cover proactive overrides, built-in fallback, empty-value suppression, mode-hint precedence, non-`Ultra` behavior, and model switches.
GitOrigin-RevId: da0a9ebd9b58cf04ef13a703c210e5da2eed0884
## What changed
- Advertise `openai/elicitation` form support when the client declares an
object-valued `form` capability, without deriving it from the legacy
`openai/form` capability.
- Handle `openai/elicitation/create` requests in `form` mode and forward their
metadata and opaque schema through app-server as `openaiForm`, preserving
`x-openai-*` annotations.
- Keep legacy form handling independent and have the TUI automatically decline
the new form requests it cannot render.
## Testing
- Add an app-server round-trip test for capability negotiation, request
forwarding, and accepted responses.
- Extend MCP capability filtering coverage for `openai/elicitation.form`.
GitOrigin-RevId: 88f39257bffafe4ee98a9c910e38507843cd7eeb
## What changed
- Store the most recently captured execution `StepContext` in the active turn state.
- Keep speculative model-fallback captures from replacing that context, and update it only when the fallback is selected for remote compaction.
- Preserve the retained context when a turn completes or is interrupted.
## Testing
- Add coverage for completed and interrupted turns, successful primary compaction, fallback compaction, and interruption while a fallback is running.
GitOrigin-RevId: c7dee36ec8e88ba2f4fb4a9c0322e5c86a0caf0e
## What changed
- Treat a surviving full world-state snapshot as a context baseline when a fork removes the associated user message.
- Restore the previous turn settings and reference context from that baseline, without treating the segment as a user turn for rollback.
- Ignore partial snapshots and full snapshots superseded by compaction.
## Testing
- Cover resumed rollout reconstruction with removed task messages, partial snapshots, and compaction.
- Verify nested agents inherit developer instructions and environment context exactly once across history modes and compacted histories.
GitOrigin-RevId: 9f9f6992d7ea8cff9ac82a7fcdf0c8fb764db81c
## 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
## What changed
- Add an optional `timeoutMs` parameter to `thread/shellCommand` and propagate it to user shell execution.
- Preserve the one-hour default when the parameter is omitted or `null`, allow longer deadlines, and treat `0` as an immediate timeout.
- Reject negative or otherwise invalid values before execution. Timing out an auxiliary shell command leaves its active turn running.
## Testing
- Cover default, extended, immediate, invalid, standalone, active-turn, and interruption behavior.
GitOrigin-RevId: d72f67793d9766b7383f28d30140270a4a76a495
## Why
Fetching plugin recommendations while building the first-turn context delays that
context when the recommendation catalog is not already cached.
## What changed
- Start fetching recommendations during session initialization, in parallel with
MCP and plugin setup, and reuse the in-flight request when building context.
- Centralize the feature checks that enable plugin recommendations.
- Prevent an invalidated recommendation request from repopulating the cache after
a newer request has completed.
## Testing
Added coverage for startup feature gating, reuse of the in-flight fetch, legacy
behavior when recommendations are unavailable, and cache invalidation races.
GitOrigin-RevId: 708774b9bfc472c9847973ae8b2774b66e7321e4
## What changed
- Share the root thread's current service tier across its agent tree.
- Apply that tier to each subagent request when the selected child model supports it, including remote compaction requests.
- Propagate root service-tier updates to existing and newly spawned subagents without rewriting child-owned settings.
- Remove per-spawn and role-level service-tier overrides so routing remains controlled by the root thread.
## Testing
- Cover supported and unsupported child models, full-history forks, live tier changes, remote compaction, new subagents, and reloaded subagents.
GitOrigin-RevId: 6c299e224e90b41a558788fb7aa4d5f336fa383a
## What changed
- Emit turn-scoped authentication recovery started and completed events when a model provider refreshes expired credentials.
- Add stable `modelProvider/authRecoveryStarted` and `modelProvider/authRecoveryCompleted` app-server notifications with the thread, turn, provider, and user-facing message.
- Show recovery progress in the TUI and `codex exec`, including Amazon Bedrock session reauthentication.
## Testing
- Cover provider recovery success and failure events, app-server routing, client rendering, and Amazon Bedrock credential refresh.
GitOrigin-RevId: 3010c38d0676f18bced27761cf86dd38344d09f9
## What changed
- Add the under-development `powershell_shell_version` feature flag.
- When enabled for a single local PowerShell environment, query the selected
shell executable and include its major/minor version in
`<environment_context>`.
- Cache version lookups, bound command execution and output, and report when a
previously visible version becomes unavailable.
## Testing
- Cover environment-context diffs when the shell version appears or disappears.
- Verify on Windows that the version is model-visible only when the feature is
enabled.
GitOrigin-RevId: 3ec8e80425ec3c193134dc1b69f61ef09ede375e
## What changed
Resolve the review model's token budget from the parent turn's original
configuration and model-default setting when deciding whether a Guardian
follow-up review needs context rollover. This preserves explicit token-budget
preferences instead of unconditionally applying the review model's defaults.
GitOrigin-RevId: cb3adaea7dad22f4975d84c6ee62dfc2bfbddbb9
## Why
Long-lived Guardian sessions can exhaust the review model's context window. After
a rollover, a transcript delta is insufficient because the new window no longer
contains the original transcript or persistent review instructions.
## What changed
- Compact an exhausted Guardian session before submitting the next review while
respecting the review model's token-budget configuration.
- Detect whether the full transcript remains in the reviewer history. If it does
not, reset transcript and Node REPL evidence cursors so the next review replays
the complete transcript and associated image evidence.
- Restore the follow-up reminder and Node REPL policy when compaction removes
them.
## Testing
- Cover rollover both when the reviewer window is already exhausted and when a
follow-up reminder crosses the limit.
- Verify that multimodal reviews preserve the fallback buffer and replay browser
evidence after rollover.
GitOrigin-RevId: f2a02a84e1f25d15c6e2b8eddb93e2d1805449f5
## What changed
- Add the optional `platformOs` field to exec-server environment metadata, with
local exec servers reporting `std::env::consts::OS` and legacy servers
remaining compatible when the field is absent.
- Preserve the reported OS in each selected turn environment, including the
local environment.
## Testing
- Extend protocol serialization and environment-selection coverage for remote
and local platform values.
GitOrigin-RevId: 0f325f6cb83cd8f8c7ec1ce2a4041f45a29e8f38
## What changed
- Count remote executor MCP discovery attempts, including refreshes, with `success` and `error` outcomes.
- Count each discovered server by name and report whether it was found or unavailable before policy and startup filtering.
GitOrigin-RevId: b912aa64b606e6243cce79df0fe85d485a2c3547
## What changed
- Report the executor's user home in environment metadata.
- Cache the reported home on selected turn environments and include it in
filesystem sandbox contexts, including `apply_patch` requests.
- Preserve compatibility with executors that do not report a home directory.
## Testing
- Cover local, remote, inherited, serialized, and `apply_patch` sandbox
contexts.
GitOrigin-RevId: ea65e68e33a80bc9bcfc7c53b1486ab944699470
## What changed
- Add `mcp_optional_startup_grace_ms` with a default of 1,000 ms to control how long tool catalog capture waits for optional MCP servers.
- Treat a value of `0` as disabling the shared grace, so optional servers use their configured `startup_timeout_sec` instead.
- Apply updated grace values during runtime and MCP configuration refreshes, and reset cached startup deadlines when the configured duration changes.
## Testing
- Cover custom, disabled, refreshed, and shared startup grace behavior in MCP connection and core integration tests.
GitOrigin-RevId: 936196215ed6e0a29c9ec72cfee17663978df8ef
## Why
Planning tools for a candidate model must not change the selected model or its
Responses Lite tool inventory. Otherwise, preparing a fallback can overwrite
metadata before the current model's request is sent.
## What changed
- Build tool registries and visible specs from an explicit `ModelInfo`.
- Store the finalized tool mode, code-mode names, namespace inventory, and
terminal and child-management capabilities on `ToolRouter`.
- Publish namespace inventory only after selecting the step's router, and use
that router's mode when dispatching and recording tool calls.
## Testing
Added coverage for candidate-model isolation, fallback compaction metadata,
code-mode worker selection, normalized tool names, and router capabilities.
GitOrigin-RevId: 56df2de8f42c7e93a4c4dcd117720e40a0cca3db
## Why
Model settings can change between steps in the same turn. Token-budget defaults,
context-window limits, and model-visible guidance need to stay aligned with the
model captured for each step.
## What changed
- Preserve the turn's original token-budget preferences and resolve model-owned
defaults when each `StepContext` is created.
- Use the step's model to determine context-window availability and token-budget
guidance, while keeping explicit user settings unchanged.
- Emit one-time replacement or removal notices when context-window guidance
changes, including compatibility with previously persisted world-state
snapshots.
- Add `ModelInfo::usable_context_window()` to distinguish reserved-headroom
capacity from the resolved context window and auto-compaction limit.
## Testing
- Cover mid-turn model switches with model defaults, explicit settings, config
reloads, missing context windows, and missing destination guidance.
- Cover unchanged, replaced, removed, blank, and legacy guidance transitions.
GitOrigin-RevId: 91100a9f3625bef8a79ecd158eb028051c309eee
## What changed
- Recognize `codex_sensitive_action: true` on MCP approval elicitations.
- Route marked actions to the synchronous Guardian reviewer instead of extension fast approval. Absent or false values preserve the existing path.
- Cover ordinary and strict auto-review configurations, including marked, unmarked, and explicitly false metadata.
GitOrigin-RevId: dce13d1696568ca81c8d2b08a7ef0d5f13e155bf
## What changed
- Parse optional `usage_metadata.amount` values from Responses API completion events without converting their string representation.
- Propagate the metadata through SSE and WebSocket streams, regular turns, and remote compaction.
- Add `usageMetadata` to the app-server `rawResponse/completed` notification and its generated schemas.
## Testing
- Cover missing, null, zero, and high-precision amounts, including completion events emitted during compaction.
GitOrigin-RevId: b16d81e8350eab0dc7ff613d0dc9dbed73041713
## Why
Executed-tool metadata for a Code Mode cell can span its `exec` and `wait`
outputs. Consumers need to know whether that metadata contains the cell's full
tool call inventory or only a partial record.
## What changed
- Associate recorded tool calls with the originating Code Mode cell across
`exec` and `wait` outputs.
- Add a host-owned `tool_calls_complete` marker when recording finishes without
dropped or truncated calls. The marker describes inventory completeness, not
tool success.
- Preserve the cell metadata through replay and pruning, and clear the
completeness marker when prompt budgeting makes the record incomplete.
## Testing
Add coverage for multi-wait cells, retries, interruptions, recording limits,
argument truncation, prompt budgeting, and rejection of input-forged metadata.
GitOrigin-RevId: 377014edec06aab42aadab78b8b5dccc0a1c83ba
## What changed
- Add bundled proactivity and follow-up guidance when `ReasoningEffort::Persistent` is selected.
- Allow model metadata to override or disable the guidance with `persistent_instructions`.
- Track the instructions as world state so effort and model changes replace or retire stale guidance without duplicating unchanged context.
- Tailor approval guidance to `send_user_message_async` availability, rebuild it for forked agents and rolled-back history, and exclude Guardian sessions.
## Testing
- Cover default, overridden, disabled, replaced, and retired instructions.
- Cover model changes, history transitions, agent forks, Guardian sessions, and asynchronous-message tool availability.
GitOrigin-RevId: 78e05d0335a70b385bd069b45edcd4ba570cef91
## Why
Trusted Guardian reviewer sessions can emit events without an app-server listener, leaving their turns and tool usage absent from analytics.
## What changed
- Translate Guardian session lifecycle and canonical tool item events into analytics notifications while excluding messages, reasoning, and review content.
- Preserve Guardian lineage during prewarming even when inherited client metadata is not yet available, without emitting an incomplete thread initialization event.
- Limit direct session tracking to trusted Guardian children so resumed threads that have an app-server listener do not report duplicate events.
## Testing
- Cover completed, failed, and interrupted Guardian turns, tool counts and timing, lineage, content exclusion, prewarmed lineage, and resumed-thread deduplication.
GitOrigin-RevId: 70ace5f7ed30a7e3b6a9d6097c8b2a6a326dc4f7
## What changed
- Add callback lifetimes to extension `ToolCall`, `ToolEnvironment`, turn-input context, and skill-read request types.
- Require extension tool executors to handle calls for any invocation lifetime and tie their returned futures to that lifetime.
- Update built-in extensions, adapters, and tests to use the lifetime-scoped APIs without retaining host-owned environment capabilities.
GitOrigin-RevId: aca9ce3a1b6870df550bc84ec9308cc65533f161
## Why
Guardian treated all skill instructions as untrusted, so it could not use the
invocation of a user-owned skill as authorization evidence.
## What changed
- Record explicit and implicit skill invocations through extension contributors.
- Send Guardian a bounded, deduplicated list of canonical skill paths only when
they resolve under the user's Codex or `.agents` skill directories.
- Keep repository skills and symlinks that escape trusted roots untrusted, and
send only verified paths rather than skill contents to the classifier.
- Avoid recording host skill prompts that were superseded by core skill
injection.
## Testing
- Added unit coverage for trusted roots, symlink escapes, deduplication, and
size limits.
- Added integration coverage showing that Guardian trusts invoked user skills
while rejecting a forged repository skill.
GitOrigin-RevId: 1cc0be230f7e6649110201c23c3de8f26447ff36
## What changed
- Fetch ChatGPT account access and attach `cyber_trusted_access` under
`openai/entitlementContext` when an installed or selected plugin explicitly
requests it from a local, read-only stdio tool with no arguments.
- Reject caller-supplied entitlement context on direct MCP calls, and fall back
to an `unknown` status when access cannot be verified or the account changes
during the lookup.
- Count entitlement lookup time against the MCP tool-call timeout.
## Testing
- Cover eligible calls plus unowned, remote, HTTP, mutable, argument-bearing,
API-key, undeclared, direct, and account-switch cases in core and app-server
integration tests.
GitOrigin-RevId: 303b61585ef4530276c373d914a2d0ccef1f4ac2
## What changed
- Allow `start_or_steer_turn` to accept named `function_call_output` items without a `call_id`, starting a turn when idle or steering the active turn.
- Preserve standalone outputs as pending input, assign response item IDs, and record them in conversation history without treating them as user prompts.
## Testing
- Cover starting a turn with a standalone output and steering an active turn with one.
- Verify passive response items do not signal pending steering activity.
GitOrigin-RevId: e32b292106b3c5a5f03577b8ebf67d2d7dee4e5d
## What changed
- Add the zero-based `window_number` to Responses turn metadata and include it in prewarm requests.
- Persist `forked_from_ordinal_exclusive` separately from the physical history base so fork lineage remains accurate after reverts and cold resumes.
- Report context-inheriting subagents through `parent_thread_id` without also emitting fork lineage.
- Reserve the new metadata keys while accepting and filtering previously valid configured values so Codex-owned values take precedence.
## Testing
- Cover window numbering across compaction, resume, and fork operations.
- Cover fork cutoff persistence across direct forks, legacy rollouts, reverts into inherited history, and cold resumes.
GitOrigin-RevId: 288f6cd632254814c1f166668b33828082d799a1
## What changed
- Add the `write_stdin_approval` feature flag, disabled by default, to require a fresh approval before sending non-empty input to an escalated unified-exec terminal.
- Route these reviews through hooks, Guardian, app-server, and the TUI as `writeStdin` approvals while keeping the parent command item unchanged.
- Preserve the terminal environment, launch directory, and escalation state across turns, and revalidate the process identity after approval before writing input.
## Testing
- Cover approval routing, policy decisions, process reuse, terminal identity preservation, app-server events, analytics, and TUI rendering.
GitOrigin-RevId: 4ba29eaae6208e934737ba078c96e589c7c2164d
## Why
App-server clients need the explanation and steering instruction returned with a
misalignment policy violation to offer a user-confirmed continuation. Missing or
incomplete details must continue to behave as a terminal block.
## What changed
- Parse optional misalignment classification, explanation, and steering details
from streamed and HTTP Responses errors and propagate them through core errors.
- Include the details in live app-server `error` and `turn/completed` payloads and
export the corresponding protocol schemas and TypeScript types.
- Keep explanations and steering messages out of serialized rollout events and
redact them from debug output.
- Document how clients can resume with `turn/start` after user confirmation.
## Testing
- Cover streamed, HTTP, and WebSocket-wrapped errors, including malformed and
classification-only details.
- Verify live app-server notifications expose resumable details without writing
the explanation or steering message to the rollout.
GitOrigin-RevId: 329258a444c2cd91d0c57ab1720830b33ddcfac5
## What changed
- Default the current-time reminder and interruptible `clock.sleep` tool on for
turns using persistent reasoning effort.
- Preserve explicit `features.current_time_reminder` settings and managed
requirements, and gate reminder recording on the resolved feature state.
## Testing
- Cover configured, persistent, ordinary-effort, explicitly disabled, and
managed-disabled clock setups.
- Verify review turns do not expose clock tools or current-time reminders.
GitOrigin-RevId: 4ff8805807ceeef374abd962b1d9e7c37f1fbd83
## What changed
- Add effective workspace roots to `EnvironmentConfig` and use them when
materializing permission profiles and filesystem context.
- Preserve selection roots for thread-owned configurations while allowing a
ready environment attachment to supply its resolved roots.
- Propagate resolved roots to child environments and validate both command and
`apply_patch` writes in owner-provided secondary workspaces.
GitOrigin-RevId: 1832c168ce1fc37fa67efc2a472998ab4f448c70
## What changed
- Add optional `call_id` and `action` fields to `SecurityRiskScore` so a persisted score identifies the tool call and bounded action that produced it.
- Populate the fields after Guardian V2 classifies an action while leaving fail-closed and legacy scores without provenance.
## Testing
- Verify scores preserve action details for direct tool calls and nested code-mode calls.
- Verify rollout serialization retains the new fields.
GitOrigin-RevId: fce69af37f368cecf1d9bce42bfc38a969e969b7
## What changed
- Add the opt-in `features.guardianv2.free_guardian` setting to route eligible
Guardian reviews to `/guardian` and asynchronous classifications to
`/guardian-classifier`.
- Support endpoint selection for both HTTP and WebSocket Responses clients,
including endpoint-aware connection reuse.
- Keep `/responses` for API-key authentication, custom provider URLs,
non-Guardian models, and configurations where the option is disabled.
- Omit routing hints and `service_tier` from requests sent to the dedicated
Guardian endpoints.
## Testing
- Cover route selection across authentication modes, provider URLs, HTTP
fallback, WebSocket review sessions, and classifier sampling.
GitOrigin-RevId: 9c88673c710d009e0239559b9245a4e6c4f4befc
## What changed
- Append accepted Guardian V2 classification results to rollout history for non-ephemeral threads.
- Keep resumed and forked threads from restoring a persisted score into active Guardian state.
## Testing
- Verify that asynchronous scoring records the resulting `SecurityRiskScore` rollout item.
- Seed resume and fork tests with a persisted score and verify that approvals ignore it.
GitOrigin-RevId: 40d63a2ea7e7cf6398402c6aed7d5e5727dc9d68
## Why
Tool actions and approval retries can outlive the step that issued them. If approval handling reads newer turn settings, a delayed action can use the wrong approval policy or reviewer.
## What changed
- Capture the approval policy and reviewer in `GuardianReviewContext` from the issuing step.
- Use step settings for command, patch, permission, sandbox-retry, and remote-network approval decisions.
- Use the active task's published settings for background network approvals and long-lived Unix shell interception.
- Preserve explicit MCP reviewer overrides.
## Testing
Added regression coverage for permission requests, same-turn network policy changes, delayed commands after a new turn starts, sandbox retries, patch policy selection, and MCP reviewer overrides.
GitOrigin-RevId: b7e84a2e1e79a71502272b15b5ef5b0e5643f8b0
## Why
MCP tool calls can be issued by a model step whose settings differ from the
turn's initial settings. Request metadata must describe the issuing step.
## What changed
- Build MCP tool-call metadata from the captured `StepContext` model and its
effective reasoning effort.
- Populate `node_repl_disabled` from the issuing step while leaving turn-owned
Responses metadata unchanged.
- Apply the same step-scoped values to MCP requests emitted by stop hooks.
## Testing
Add coverage for configured and model-default reasoning effort, model changes,
Node REPL restrictions, and preservation of the original step and turn
metadata.
GitOrigin-RevId: f9cce52fbe559f15437bd19d699b657229c0d145
## Why
Tool execution can use an environment whose sandbox configuration differs from the turn-wide configuration. Sandbox selection and process setup need to follow the environment that owns the tool request.
## What changed
- Source Windows sandbox level, legacy Landlock mode, and Windows private desktop settings from the selected turn environment across tool orchestration, unified exec, patch application, and zsh fork escalation.
- Preserve Windows sandbox level overrides in restorable thread settings and invalidate MCP configuration when that level changes.
## Testing
- Add a resume test that verifies an elevated Windows sandbox override survives thread restoration.
GitOrigin-RevId: 9d8fc46404919f5ae97c9ed712c0932a1f095e36
Benjamin Carlsson