## Why
Asynchronous score publication updated the risk score, authorization, and tool-call coverage separately. Approval checks could therefore read inconsistent evidence when deciding whether to reuse a cached score.
## What changed
Keep Guardian's cached score and observation state under one lock, and use consistent snapshots for approval checks. Publish successful scores together with their authorization and classified tool-call index. Preserve timestamp ordering, fail-closed precedence on timestamp ties, and per-call oversized-action tracking.
## Testing
Add regression coverage for rejected delayed results and timestamp ties. Add a gated integration test showing that a delayed score becomes stale after intervening tool calls, while a sufficiently fresh score permits cached approval.
GitOrigin-RevId: 8df6fafd5e0ee59aa0f78f541eac562efce3d4d0
## Why
Filesystem reads previously required a sandbox whenever writes were restricted, even with full-disk read permission. This made permitted reads depend on sandbox availability.
## What changed
- Select sandboxing independently for reads and writes, allowing full-disk reads directly while keeping restricted operations sandboxed.
- Use the executor's path convention when evaluating full-disk access, including `:slash_tmp` denials.
- Apply read-specific checks to capability discovery and skill resource reads, allowing unrestricted reads on executors without sandboxed discovery support.
## Testing
Add regression tests for direct read APIs, restricted writes and reads, executor-specific permission rules, and capability discovery without sandbox support. Update Windows tests to verify sandbox enforcement through writes.
GitOrigin-RevId: a03844bd1f0ea583bb54326683cddbfdd050119f
## Why
Removing the selected working directory can prevent filesystem sandbox helpers from launching, even when the requested absolute paths remain accessible. Permission rules must stay anchored to the selected directory while those operations continue.
## What changed
- Require a policy `cwd` in `FileSystemSandboxContext` and launch filesystem helpers from the filesystem root while preserving the policy directory and workspace roots.
- Carry explicit `policyContext` in filesystem RPCs, preserving legacy wire fields and resolving omitted directories from older clients at executor ingress.
- Keep permission paths as executor file URIs and validate host compatibility where they are enforced.
- Bind Windows relative denial globs to the policy directory before changing the helper's launch directory, preserving home-relative patterns.
## Testing
Add regression coverage for `apply_patch` after working-directory removal, legacy RPC directory fallbacks, cross-platform permission URI transport, and Windows relative read denials. The patch regression verifies that an allowed file is updated while an explicitly denied file remains unreadable and unchanged.
GitOrigin-RevId: b0f4db722b27cb72ec129fc297c85732afac11f7
## Why
File images contributed zero to history and compaction token estimates and were dropped from Guardian context. This undercounted context usage and omitted image evidence from reviews.
## What changed
- Estimate file images using the fixed resized-image cost, or 10,000 patch tokens for `original` detail because file references do not expose dimensions.
- Charge file images against remote compaction and Guardian request budgets.
- Preserve file images from user messages and tool outputs through Guardian selection and delivery, and deduplicate review images by file ID.
- Apply transcript image count and byte limits to both reference types, counting file ID bytes rather than referenced file contents.
## Testing
Add coverage for file-image estimates, selection limits, delivery, and compaction retention. Extend Guardian tests to verify file-image evidence reaches reviews and uploaded original-detail images trigger reviewer-history compaction.
GitOrigin-RevId: ee024204dcbe16e277fea9347c6a4d7fc120e18d
## What changed
Stop emitting a warning when skill descriptions are shortened to fit the skills context budget. Continue warning when skills are omitted from the model-visible list.
Update the app-server warning test to set `skills.max_context_tokens` to `1_000` and expect the omitted-skills warning. Remove assertions and tests for the description-shortening warning.
GitOrigin-RevId: feac10b3d159d73648a786120d477b7fe4e082d7
## 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
## What changed
- Move approval routing tests into `guardian-reviewer`, covering both reviewer choices across `UnlessTrusted`, `OnRequest`, `Granular`, and `Never` without host session setup.
- Combine reviewer configuration checks into one test for isolation from parent customizations.
- Cover filtering of ordinary developer messages while preserving manual approvals in `guardian-context`, removing the duplicate async scorer test.
GitOrigin-RevId: b39ef21475e2f5d99d758f2d27615449b252dc65
## What changed
Remove standalone tests across Guardian review sessions, configuration, assessment schemas, transcripts, and trusted skill paths.
Strengthen the existing request layout snapshot test to check exact transcript separators, the action header, and parent session ID placement before snapshot whitespace normalization. Concatenate user input text without inserting newlines so the assertions validate the actual request formatting.
GitOrigin-RevId: 06d262a1a85b1a8401796958f300c894146d2e1b
## Why
Orchestrator skill caches were tied to the entire MCP connection set, so runtime updates could discard cached catalogs and contents even when the Apps connection was reused. Plugin changes still need to refresh those resources without requiring a reconnect.
## What changed
Key orchestrator skill caches by the Apps server connection and an explicit invalidation generation. Preserve cached resources while that identity stays unchanged, and invalidate them on MCP runtime refresh requests or configuration refreshes. Replacing the connection also changes the cache identity.
## Testing
Extend the production-turn integration test to verify cache reuse across ordinary turns and approval-policy changes, refreshed skill contents after plugin invalidation without reconnecting, and fresh resource requests after a forced reconnect.
GitOrigin-RevId: 43ce059972369c4c735742e8224b60a692244b02
## What changed
- Add `ResolvedModelMessages` to resolve catalog text and bundled defaults while preserving explicit empty overrides and their source.
- Move base-instruction rendering, Guardian prompt composition, multi-agent role rendering, and `update_plan` guidance filtering into `codex-prompts`; migrate consumers to the shared APIs.
- Separate permission-profile resolution from prompt composition, and annotate Guardian policy and classifier instructions with content kinds.
## Testing
Add and update coverage for missing versus empty templates, literal overrides, multi-agent role composition, Guardian policy substitution and truncation, and preservation of permission path spellings and order.
GitOrigin-RevId: 52335bb7acec0f432d5c57acb2accd5f0276056e
## What changed
Add an optional `protocol_mode` to `McpServerContribution::HostedApps` and apply it when registering the server. Preserve the hosted Apps default when no override is provided.
## Testing
Add an integration test that selects MCP `2026-07-28` with the global feature disabled and completes a native user verification flow, including resuming the tool call with the verification response. Extend capability tests to confirm that the override preserves hosted Apps user verification without granting it to ordinary extension registrations.
GitOrigin-RevId: e8662c45f964c00b0bb4503bbcfd8432d7e2fa70
## Why
Executor skill discovery enabled every discovered skill, with no provider option for callers to disable selected skills.
## What changed
Add `ExecutorSkillProvider::with_disabled_skill_paths` to accept `SKILL.md` paths grouped by environment ID. Mark matching catalog entries disabled in both discovery routes so they are omitted from model context. Skills remain enabled by default, and paths configured for another environment have no effect.
## Testing
Extend catalog tests to cover default enablement, matching paths, and environment isolation across both discovery routes. Add a model-context snapshot test confirming that an active skill appears while a disabled skill is omitted.
GitOrigin-RevId: 7426eb94566db2bcf7e49a5dee28673ab743ad25
## What changed
Add `codex_history::CompactionCheckpoint` to keep a checkpoint and its recorded producer model hash together, with shared checks for usability and reviewer compatibility.
Replace `ConversationHistorySnapshot::latest_compaction_model_hash()` with `latest_compaction()` and use the shared abstraction in Guardian context selection, review sessions, and parent compaction reuse. Preserve selection of the latest checkpoint even when unusable, and leave missing producer metadata unknown.
## Testing
Add coverage showing that the latest checkpoint retains its own provenance even with missing or empty encrypted content, and that compatibility requires matching producer and reviewer hashes. Adapt existing history, migration, and parent compaction tests to the new API.
GitOrigin-RevId: 98e794644edf979c98fcc8294baa48ff82ea0ed0
## What changed
Move duplicated action preparation into `codex_guardian_context::action_for_review`, shared by approval rendering and cached-evidence size checks. Preserve the existing behavior: omit top-level `tool_description` and `connector_description` metadata from MCP tool calls while keeping tool arguments intact.
## Testing
Add unit tests verifying that MCP arguments retain description fields, including nested fields, and that other action types remain unchanged.
GitOrigin-RevId: 8e890fb263556596f695177f917750377c04d43c
## What changed
Introduce `ReviewTurnResult` and `ReviewSessionResult` to carry review outcomes, session disposition, and completion or analytics data. Propagate `SessionDisposition` directly through review execution and pooling, replacing boolean reuse flags and positional tuple access while preserving existing behavior.
Update existing review-session tests to assert the named fields and explicit session dispositions.
GitOrigin-RevId: 93eea8b482cd592167956444042becc1ef5f2eb5
## What changed
- Extend `GuardianModelPolicy` with controls for uncategorized tools, unscored actions, the initial computer-use call allowance, and sandboxed command coverage.
- Add `GuardianPolicyLoader` in `codex-config` to translate legacy settings, preserve catalog policy precedence, and enforce reviewer requirements. Apply live model review requirements through `ConfigRequirements::constrain_guardian_policy`.
- Use the shared model policy for Guardian scoring and approval, replacing the extension-local policy wrapper while retaining legacy defaults.
## Testing
Add configuration tests for catalog precedence, legacy scope fallback, required-model constraints, and legacy computer-use opt-in and feature gating. Adapt existing extension tests to consume the shared policy.
GitOrigin-RevId: 6fed1c3a831964ad28ea7de5cb19e74e323c1204
## Why
Code Mode wrappers should preserve cached scores for nested actions so adaptive review can reuse them across cells.
## What changed
Remove `code_mode` from `GuardianModelPolicy` and its approval scopes. Skip scoring and cache invalidation for direct Code Mode `exec` wrappers under model policies, leaving nested tools governed by their own categories. Preserve legacy wrapper scoring behavior.
## Testing
Expand app-server coverage for score reuse across cells, synchronous nested reviews, required model policies, and legacy configurations. Retain checks that ordinary tools named `exec` or MCP tools named `wait` invalidate cached scores.
GitOrigin-RevId: 7a0692e5330f3d4b4bda988ecbe7061bc8b2318c
## Why
HTTP callers currently cannot bound response bodies. Callers accepting provider-controlled model catalogs need a size limit before decoding the response.
## What changed
- Add per-request `response_body_limit_bytes` for buffered, streaming, and error responses, leaving requests unbounded by default. Reject oversized declared lengths early and count observed bytes across chunks.
- Return a non-retryable `ResponseTooLarge` error that reports only the byte limit. Preserve HTTP status and headers when a bounded error body fails to read without exceeding the limit.
- Expose `ModelsClient::list_models_raw` to fetch bytes and an optional ETag using provider authentication and retries, with an optional body limit. Keep existing `list_models` decoding behavior.
## Testing
Add HTTP fixture tests covering size boundaries, chunked and missing-length responses, early rejection, stream termination, interrupted bodies, error text decoding, and request isolation. Add a models-client test verifying that limits survive authentication retries without affecting subsequent ordinary requests.
GitOrigin-RevId: 61b8940bc8a549588cee6865a3b5d1cff789074d
## What changed
- Accept `fileId` alongside the existing `url` form for app-server image inputs, and forward file references to the Responses API as `file_id`. Update generated schemas and client types.
- Preserve file references, image detail hints, and mixed inline/file image ordering through user-message events, thread history, and rollout migration. Retain file images when truncating tool output.
- Pass file references through image preparation without resolving them, while keeping resize-notice numbering correct. Omit them from unsupported TUI display and Guardian image context.
- Reject image-edit requests whose recent-image window includes a file reference, preventing selection of an older inline image instead.
## Testing
Add coverage for serialization, request and rollout preservation, mixed-image history ordering, incomplete ordering metadata, tool-output truncation, and rejection of unsupported image-edit selections.
GitOrigin-RevId: 6ca20a8577155cc934b720803c3b7b3bffdf972a
## Why
A checkpoint produced by a previous model may be incompatible with the selected Guardian reviewer. Legacy review must retain user restrictions and verified answers across compaction and restart until a compatible checkpoint is available.
## What changed
- Expose retained user evidence independently of whether Guardian reviews the legacy transcript or parent context.
- Resolve the selected reviewer's compaction compatibility hash during resume and remote compaction. Preserve the legacy transcript when checkpoint compatibility is unknown or mismatched.
- Activate parent-context review immediately after compatible compaction and invalidate pending reviews bound to the previous evidence policy. Ordinary compaction preserves pending reviews.
- Keep strict compatibility checks for previously migrated checkpoints whose complete legacy transcript is no longer available.
## Testing
Add coverage for mismatched checkpoint hashes, retained evidence during legacy review, and pending-review cancellation on migration. Add a request-history scenario covering a model switch, incompatible automatic compaction, restart, and compatible manual compaction with preserved user restrictions and verified answers.
GitOrigin-RevId: b52bfd7c21ce7ebed482174c8d7c327e50b1ada8
## Why
Older compaction checkpoints can lack a producer model hash. Enabling thread-owned Guardian context must accommodate those checkpoints without losing user instructions or verified answers.
## What changed
- Select review policy from each history snapshot independently of the session's evidence capture policy.
- Keep legacy review for checkpoints with missing or empty producer hashes when the transcript preserves the evidence. Preserve strict compatibility checks when review depends on retained-only evidence.
- Continue capturing retained instructions and answers during migration, while supplying legacy reviews with runtime answers and subagent reviews with root authorization evidence.
- Record producer model hashes on new compactions regardless of context mode.
## Testing
Add coverage for checkpoint migration across compaction and resume in both history storage modes, snapshot policy stability, retained-evidence preservation, and authorization revalidation for owning sessions and subagents.
GitOrigin-RevId: ec72b4a51d40729173e24ac1b53b7d516ddaa825
## What changed
- Replace dedicated Guardian endpoints with `/responses`, sending `x-codex-guardian: reviewer` or `x-codex-guardian: classifier` for eligible Codex backend requests over HTTP and WebSocket.
- Add model-scoped thread headers, recheck backend authentication on each request attempt, and reconnect WebSockets when the applicable headers change.
- Retain `features.guardianv2.free_guardian` for configuration compatibility while removing its routing gate; the backend now controls Guardian billing.
## Testing
Update tests for reviewer and classifier headers, model and authentication scoping, HTTP fallback, WebSocket reuse, and parent-response metadata across retries.
GitOrigin-RevId: 1c2b3c458ab77fb40aae4ee6784f6d827c60e76c
## Why
MXC is a sandbox implementation, not a restricted-token sandbox level. Executor requests need to represent that choice separately from `WindowsSandboxLevel`.
## What changed
- Introduce `WindowsSandboxSelection` for executor sandbox contexts and remove `Mxc` from `WindowsSandboxLevel`.
- Preserve the `windowsSandboxLevel` wire field and its serialized values for compatibility.
- Share sandbox selection between executor process launches and filesystem helpers, and use the new selection in capability discovery and skill reads.
- Disable Windows sandbox selection for executor paths that do not use Windows path conventions.
## Testing
Extend coverage for MXC wire serialization, Windows skill-read sandbox checks, and capability discovery with distinct permissions. Exercise remote filesystem write restrictions with both restricted-token and MXC sandboxes, including rejection when native MXC is unavailable.
GitOrigin-RevId: 266211377bcb138a0dc75861e9ff2225fa37a53d
## What changed
Move concrete reviewer settings into `guardian-v2` and replace `ReviewerConfigOverrides` with a `ReviewerConfig` callback stored in thread extension data. Apply the callback to each captured parent configuration before preparing context and checking session reuse.
Keep model selection, policy prompt construction, and live network state in the core adapter. Preserve the existing read-only reviewer settings and share `reviewer_permission_profile` with inherited environment configuration.
## Testing
Update existing Guardian configuration, session reuse, and prewarming tests to use the extension's actual configuration builder, including compiling the shared source in core's test host.
GitOrigin-RevId: dfa1dae9f8d4469f2e4f6e64ab842cc662970898
## What changed
- Add `AllowedTools` to the extension API, captured once at thread startup to restrict which tools can be advertised or executed. An empty list disables all tools; an absent value preserves ordinary tool setup. Callers must supply the list again when resuming a thread.
- Filter tool registration and hosted tool specifications before Code Mode and discovery. Match names with their namespaces and require generated tools such as `exec` and `wait` to be explicitly allowed.
- Move Guardian reviewer tool selection into the allowlist while retaining feature and sandbox restrictions, with a fallback for older reviewer sessions.
## Testing
Add coverage for selected, empty, and absent allowlists across tool sources and Code Mode. Add managed-thread tests verifying advertised tools, rejection of excluded calls, and that replacing extension state after startup cannot change the captured allowlist.
GitOrigin-RevId: 33c1722c4241f368963d8385fa0415112e16a72d
## What changed
Introduce `ReviewRequest` in `codex-guardian-reviewer` to own contributor routing, cached approvals, synchronous fallback, and review cancellation. Keep action validation and session-specific preparation in the host.
Track the complete approval operation through parent shutdown, driving cancellation through reporting and reviewer cleanup before releasing it. Preserve fresh-review requirements and validate cached approvals before recording their outcome.
## Testing
Extend coverage for cancellation before routing and during cached approval, parent shutdown cleanup, and required fresh review overriding a cached allow result with matching assessment events and a denial warning.
GitOrigin-RevId: 8dcf26cc3660e91ef94e28e0bee7327ebe34ae6c
## What changed
Make `parse_guardian_assessment`, the rejection circuit breaker types and methods, and `AUTO_REVIEW_DENIAL_WINDOW_SIZE` crate-visible with `pub(crate)`. Update their re-exports in `guardian-reviewer` to match.
GitOrigin-RevId: 39521c7859936ebc9af570d7f1ad7b2348f3673d
## What changed
- Mark `guardian_ext` as a removed compatibility flag.
- Remove `InternalSessionSpawner`, `InternalSessionSpawnFuture`, `ApprovalReviewInput`, and `ApprovalReviewError` from the extension API, along with the spawner test.
- Remove `NodeReplReviewEvidence::review_inputs` and make the core re-exports of `NodeReplReviewEvidenceMode` crate-private.
GitOrigin-RevId: 1c532bc57264644cf521ef8077285a9b8ed67391
## What changed
- Make `SynchronousReview` own assessment events, telemetry, warnings, evidence-recording decisions, and consecutive-denial accounting.
- Keep action preparation, stale-approval validation, evidence storage, and event publication in the host adapter.
- Extract `Session::interrupt_turn_with_warning` to apply extension-requested interruptions to the selected active turn and emit the interrupted thread-idle lifecycle event.
## Testing
Adapt session tests to verify that extension interruptions emit the thread-idle lifecycle event and survive shutdown of the calling runtime.
GitOrigin-RevId: b234ad86206fedc2434f48fc44dc57679443c69f
## What changed
- Use cancellation guards to tie reusable reviewers and temporary forks to their lifetimes, including when a review future is dropped. Replace cancelled reusable reviewers before reuse.
- Share `ReviewerTasks` between the pool and `ThreadManager`, and wait for tracked cleanup during pool shutdown. Remove the separate session shutdown protocol.
- Move denial cleanup from core task handling into Guardian's turn start, stop, and abort hooks. Expose `install_reviewer` to register both thread and turn lifecycle contributors together.
## Testing
Add an integration test that exercises stale denial cleanup at turn start, completion, and interruption, including starting another turn after interruption.
GitOrigin-RevId: 85badb235309c1e1576547d1746f7ee9fc66d545
Update the image generation tool guidelines to avoid printing full results
or base64 image data with `text()` or `notify()`. Recommend printing only
small metadata when needed.
GitOrigin-RevId: f62b67b384dd39a1cc1adad0956b6fcf69d09c72
## What changed
Represent images in `ContentItem` and `FunctionCallOutputContentItem` with `ImageReference::Inline`, flattened to preserve the existing `image_url` wire format. Update image producers and consumers and regenerate app-server schemas and SDK artifacts.
Preserve the Python SDK's `InputImageContentItem` and `InputImageFunctionCallOutputContentItem` class names during generation.
## Testing
Add a regression test for stable Python image class names and adapt existing image tests to the shared representation.
GitOrigin-RevId: c38a780ac3314c2ac2deb3afc1b93b94b6f93fec
## Why
Guardian reviews must stop when their parent shuts down or their history is reset. Shutdown must also finish reviewer cleanup before closing the parent's persistent history, including when a review is waiting to retry after a rate limit.
## What changed
- Let the Guardian extension own reviewer startup, prewarming, and shutdown through `ThreadManager`, tracking and joining outstanding work during teardown.
- Cancel reviews on history reset or parent shutdown and reject decisions returned after cancellation, including cached extension decisions.
- Keep background prewarm previews from overwriting the active turn's model metadata.
## Testing
Extend regression coverage to verify reviewer cleanup after parent shutdown, prompt shutdown during a 60-second Guardian retry with the network request denied, and preservation of active model review requirements during prewarming.
GitOrigin-RevId: 05efc369b9a1642a10365eac8a09da77f3f6e28c
## What changed
Replace `ReviewerSessionFactory` with a startup callback installed through
`ReviewerPool::new`. Review requests supply shared setup data and reuse context,
while the pool uses its callback to create both reusable and forked reviewers.
Update Guardian session setup, prewarming, and existing test fixtures to use the
new pool API.
GitOrigin-RevId: fa171503afcee8bdbdf6822573bea8ded50ce117
## Why
Inline delegates have no entry in the thread registry, so reviewer creation cannot depend on looking up the parent or waiting for its thread-ready notification.
## What changed
- Capture parent identity, authentication, shared agent control, originator, and inherited instructions in `StartThreadOptions` so `ThreadManager` can start a child without a registered parent.
- Route Guardian reviewer creation through this path, remove the standalone fallback and readiness gate, and require a Guardian extension host.
- Install explicit reviewer hosts in unit tests and the Guardian reviewer extension in the integration test harness, using `ExtensionRegistry::to_builder()` to preserve existing contributors.
## Testing
Extend the thread-manager regression test to remove the parent from the registry before starting a child, then verify inherited lineage, originator, session identity, and authentication, and exclusion from the public thread list.
GitOrigin-RevId: 468ded6fdce7520cb39d55c1a884dcfa5aaea2d9
## What changed
Restrict `run_before_review_deadline_with_cancel` and its re-export to `codex-guardian-reviewer`. Move its timeout, abort, and successful-completion tests from core into the reviewer's deadline module, and remove the standalone `run_before_review_deadline` tests from core.
GitOrigin-RevId: dd9f1ed571a40a4bd66b08c88f3ee2be071f4870
## What changed
Extract tool observation and evidence capture into `observation.rs`, background classification into `classification.rs`, and score tracking and failure handling into `score.rs`. Keep lifecycle hooks in `extension.rs` and pass captured evidence through a `Classification` struct, preserving the existing snapshot and background task boundaries.
## Testing
Move the fail-closed score-ordering test into `score_tests.rs` and extend it to verify that a failed sample replaces an equally dated score while preserving newer scores.
GitOrigin-RevId: b25b9e828cce78fb2be522d7209346b63403d824
## What changed
- Remove `AgentSpawner` and `AgentSpawnFuture` from the extension API, along with the Guardian wrapper, thread lifecycle context, and app-server injection plumbing.
- Define `InternalSessionSpawnFuture` directly as a boxed future instead of aliasing `AgentSpawnFuture`.
- Raise the workspace `rustls` minimum version to `0.23.45`.
GitOrigin-RevId: b7319dee41bfb869479afeb7555a6f050c4d00a5
## Why
Command execution and plugin measurement events lack model and reasoning-effort labels. Attribution needs to reflect the step that invoked the command, even when model settings change before a background process finishes.
## What changed
- Add `model_slug` and `reasoning_effort` to command execution and plugin measurement analytics.
- Capture model context from resolved step settings and carry it through execution, approval, Guardian review, and plugin metrics collection.
- Preserve the first command-start model context when subsequent start notifications arrive.
- Keep the carried context out of serialized protocol items and generated schemas.
## Testing
Extend analytics tests to cover model switches before invocation and during background execution, default reasoning effort, Guardian-denied commands, and repeated start notifications retaining the original model context.
GitOrigin-RevId: af90e1c0d39bab625f2e89786085b61a9b96c0ce
## Why
MCP tool parsing eagerly cloned structured output schemas and built full call-result envelopes. Cloning tool definitions also copied their output-schema JSON, even before a consumer needed it.
## What changed
- Introduce `ToolOutputSchema` with immutable `Arc` storage so tool definitions share output schemas when cloned.
- Retain MCP structured output schemas and materialize the call-result envelope only when JSON is requested.
- Update code-mode consumers and schema mutation sites to materialize JSON explicitly, reusing uniquely owned storage when possible.
- Move structured content into the MCP envelope without an extra clone, preserving property order.
## Testing
Add tests for JSON preservation, mutation isolation, equality between lazy and materialized schemas, reuse of uniquely owned storage, and equivalent code-mode definitions.
GitOrigin-RevId: e98ba4c2f0efedc99f7cbc7bba206cc63a3bd8f4
## What changed
Move request execution from `LunaSampler` into `SamplingExecution` in
`sampler/execution.rs`, keeping request preparation and active-request tracking
in the sampler. Preserve the existing retry, authentication recovery,
cancellation, streaming, connection reuse, and token accounting behavior.
GitOrigin-RevId: 5e50116459fe9cc196c30c58e42773fc32be3d02
## What changed
Add `ConversationState` and `ConversationCheckpoint` to `codex-guardian-reviewer` and use them in core review sessions to track transcript cursors, completed review counts, and committed snapshots. Keep history and admitted evidence host-owned.
Preserve the separation between live review progress and committed checkpoints so forks inherit the history, cursor, and review count from the last committed snapshot.
## Testing
Add a unit test verifying that forks retain committed history and progress after an uncommitted review, then advance when the next snapshot is committed.
GitOrigin-RevId: 9f92410b11beec6b8f413c4c922fabba65852399
## What changed
After resolving startup configuration and model defaults, reject `features.token_budget.use_history_notes_extension` when the starting model lacks `supports_experimental_context`. Return an error directing users to disable the option or select a compatible model.
## Testing
Add startup coverage for explicit configuration and model defaults, verifying rejection for unsupported models and successful activation for supported models and standalone token budgets. Update history-notes test fixtures to declare experimental context support.
GitOrigin-RevId: abf1a024efc5acf97cfc858fbb93821363769dc0
## Why
ChatGPT derives Responses cache affinity from the `session-id` header. Ephemeral forks need to reuse their parent's cache routing while retaining their own session and thread identities.
## What changed
- Inherit the parent session ID as the `prompt_cache_key` for ephemeral root forks.
- Use the prompt cache key for root-agent Responses `session-id` headers, including WebSocket handshakes. Preserve the actual session identity in turn metadata and leave non-root-agent routing unchanged.
- Keep enabled goal tool definitions visible on ephemeral threads, but reject execution with `Goal tools require a persistent thread.` and disable automatic goal continuation without persistent state.
## Testing
Add regression coverage for inherited cache routing with distinct fork identities, matching parent and fork tool definitions, WebSocket session headers, and rejection of ephemeral goal tool execution.
GitOrigin-RevId: d235e97630068b27f7ebd562dd23ac6266ebc1b9
## What changed
- Expose `ThreadInstructionsProvider` through `StartThreadOptions`. Load its snapshot at startup and model-request boundaries, composing it after global instructions and before repository instructions. Empty or blank output clears only the thread contribution.
- Reject thread instructions exceeding 10,000 estimated tokens independently of the repository instruction budget. Allow host-provided instructions without a filesystem source and rename the shared future type to `LoadInstructionsFuture`.
- Retain the provider across warm resumes; require hosts to supply it again for cold resumes and offline forks. Live forks and subagents inherit applied snapshots without inheriting the source thread's provider, including when the parent is unloaded during setup or reload.
- Include thread instructions in guardian reviewer inheritance and reuse decisions.
## Testing
Add coverage for composition and clearing, refresh within an active turn, size rejection before sampling, cancellation-safe refresh, fork and resume behavior, parent eviction, and reviewer reuse invalidation.
GitOrigin-RevId: 7be9a523cbbfd67704067dfd526188dad89a3c88
## Why
Running root threads retained their startup global instructions, so edits to global `AGENTS.md` files did not take effect during an active session.
## What changed
- Reload global instructions when capturing model-request context, including after tools within the same turn. Apply changes without repeating unchanged instructions or rediscovering repository instructions when the environment and trust level are unchanged.
- Preserve the last successful global instructions on read failures, suppress recurring warnings until recovery, and clear instructions when their source is removed or blank.
- Serialize refreshes and allow cancellation without blocking subsequent requests.
- Give new subagents the parent's applied instruction snapshot and update Guardian reviewer reuse to account for refreshed instructions.
## Testing
Add regression coverage for live edits and removal, read failures and recovery, warning suppression, cancellation, subagent inheritance, and Guardian reviewer reuse. Update resume, fork, and compaction tests to verify refreshed instructions.
GitOrigin-RevId: f7e9b399740fa4f45e482c070f6e901ee7cf85cd
## Why
An unscored `exec_command` requesting additional permissions could reuse an earlier Guardian v2 approval score, even though that score did not cover the expanded permissions.
## What changed
When `UnscoredAction::AgeScore` applies, invalidate the cached score for default-namespace `exec_command` calls with `sandbox_permissions` set to `with_additional_permissions` by marking the call as a scoring failure.
## Testing
Extend the review-scope regression test to verify that an ordinary sandboxed command preserves cached approval, while a command requesting additional network permissions clears it and causes review to fail closed with `scoring_failure`.
GitOrigin-RevId: d6e3e1ec7d618dff2ef03ffcb92869249f0e547a
## 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
## Why
Transcript limits could shorten user instructions or omit later messages before Guardian evaluated the complete request budget, losing restrictions or prior approvals even when they would fit.
## What changed
- Keep user messages and manual approvals complete and in source order through transcript collection and retention.
- Preserve historical instructions while synchronous review can still compact history. If the final request still cannot fit, discard optional evidence before shortening older historical entries with truncation markers, preserving later restrictions where possible.
- Warn that shortened instructions and approvals are incomplete and that missing evidence does not authorize actions.
- Let asynchronous review defer to synchronous review when complete instructions exceed its budget.
## Testing
Add coverage for complete instruction retention, approval and restriction ordering, marked truncation of oversized Unicode text, compaction before truncation, and asynchronous fallback when instructions do not fit.
GitOrigin-RevId: e6159e876231347986b937ee7e07f944b83bfbe6
## Why
Truncating action arguments can leave approval reviewers evaluating incomplete actions. Large actions need complete review input and explicit handling when they exceed the review budget.
## What changed
- Remove action truncation and the fixed synchronous action byte limit. Admit complete actions against the whole-request budget, splitting long text losslessly into bounded transport parts and accounting for their framing.
- Route actions exceeding the asynchronous action budget to synchronous review. Prevent cached scores from covering oversized calls, including expanded approval arguments, while allowing later small actions to recover score reuse.
- Request user approval when optional review exhausts its local input budget. Keep required review and compaction service failures closed to approval, and retire exhausted review sessions.
## Testing
Add coverage for complete large-action delivery, optional user fallback, required-review denial, subsequent review recovery, async overflow through MCP approval routing, and lossless text splitting with budget accounting.
GitOrigin-RevId: 08f06e68a94b2f779480dd6b6cf5bc30b241f6dd
