## What changed
- Carry each contextual fragment's `ContentItemKind` through initial context assembly and world-state updates.
- Attach the classifications to merged response messages through `content_item_kinds`, preserving their order alongside the message content.
- Convert extension prompt fragments into annotated developer fragments and classify built-in developer and guardian instructions.
## Testing
- Add an integration test covering roles and content annotations in the first request.
- Update world-state and persisted-history tests to account for annotation metadata.
GitOrigin-RevId: 71c48dad5b17bffd47ba03263e3eb6d5e74fbf4b
## What changed
- Require each `ContextualUserFragment` to provide a stable `<feature>.<name>`
`ContentItemKind`.
- Add `AnnotatedContent` and `RenderedFragment` so rendered text, its role, and
its classification can travel together to API boundaries.
- Derive extension-owned world-state classifications from the extension ID and
keep the skills catalog classification with its fragment implementation.
## Testing
- Verify that an extension-owned world-state section renders with an
`<extension-id>.instructions` content kind.
GitOrigin-RevId: e46b74a0bb41e0b6112667c9d36bc9e7f2714451
## What changed
- Require each `PromptFragment` to include a producer-owned `ContentItemKind`.
- Re-export `ContentItemKind` from `codex_extension_api` and assign stable kinds to skills, memories, examples, and test contributors.
- Limit prompt slots to developer policy and developer capabilities, and combine contributed fragments into a single developer update.
## Testing
- Update extension registry tests to verify fragment ordering with the new content-kind metadata.
GitOrigin-RevId: ef8624a9b4ae9dce025bca9c0954d91dd8075e85
## What changed
- Mark Guardian classifier requests with `thread_source: guardian_classifier` in turn metadata.
- Remove the classifier-specific `request_kind` and `is_guardian_mode` fields.
- Update sampler and extension tests to expect the new metadata shape.
GitOrigin-RevId: 1654389ea538b7bd879045634816cbace4bd8f5c
## What changed
- Emit a structured log event for each completed Guardian V2 classification with its thread, turn, tool call, risk score, review threshold, sample time, and whether the score was accepted or superseded.
- Verify that accepted async classification events appear in the reviewed thread's log export with the expected context and decision fields.
GitOrigin-RevId: a027d5991b12a1968b9327da57e2993f0a529ca1
## What changed
- Retain bounded evidence from completed synchronous Guardian allow and deny reviews and supply it to subsequent Guardian v2 async classifier samples as trusted developer context.
- Keep review evidence isolated from the conversation transcript, escape and truncate its fields, and ignore failed or incomplete reviews.
- Invalidate retained evidence after conversation history rewrites or new user messages, including authorization changes in a worker's root thread.
## Testing
- Cover approved, denied, malformed, and forged review inputs, plus root rollback and authorization-change scenarios.
GitOrigin-RevId: 27817e1fde9a136de727048c4220d148fcf72f42
## What changed
- Add an extension API for spawning host-owned internal sessions and a
`ThreadReadyInput` lifecycle callback that runs after thread registration.
- Add Guardian reviewer session scaffolding that records the parent thread and
effective model, plus the under-development `guardian_ext` feature flag.
- Start internal sessions with fresh history while preserving parent lineage,
shared session controls, and internal-thread visibility rules.
- Scope internal-session prompt cache keys to their source and parent thread,
and expose `guardian` as an internal session source.
## Testing
- Cover internal-session spawning, parent metadata, history isolation, prompt
cache keys, and extension spawner argument forwarding.
GitOrigin-RevId: 682dae80397d62bb36247796b447042e760ca364
## Why
Legacy classifier prompts without a `{{ tenant_policy_config }}` placeholder
append the security policy during rendering. Truncating the prompt before and
after that step can unnecessarily truncate the final classifier instructions
twice.
## What changed
Keep configured classifier instructions intact until the policy has been
rendered, then apply `max_classifier_instruction_tokens` once to the complete
prompt.
## Testing
Add an extension test that verifies the bounded developer message sent for a
legacy prompt with an appended policy.
GitOrigin-RevId: 3d776a159340f4dffb6e3c2d7a10b33392fa2240
## Why
Models listed in `auto_review.required_on_models` must always use the full
automatic review path, regardless of any cached Guardian v2 risk score.
## What changed
- Skip Guardian v2 risk classification for models that require automatic
review and clear any cached `SecurityRiskScore` before review routing.
- Count thread lookup failures as failed scoring attempts so stale scores
cannot continue approving later tool calls.
## Testing
- Verify required-review models do not start a classifier and always run full
reviews.
- Verify failed thread lookups advance score lag and fall back to strict review.
GitOrigin-RevId: 048d9a80ac2a282e05437a3abb0c46ec21391be8
## Why
MultiAgent V2 workers can be reviewed after the root conversation has received
new user authorization that is absent from the worker's own transcript. Guardian
reviews need that genuine user context without treating forwarded or
assistant-authored claims as authorization.
## What changed
- Add the bounded root conversation to Guardian prompts for worker reviews.
- Preserve user and assistant roles, and explicitly treat only root user messages
as authorization evidence.
- Exclude summaries, review artifacts, and assistant commentary from the root
conversation evidence.
- Apply the same evidence to both core Guardian prompts and Guardian V2
classification.
## Testing
Add an integration test covering late root-user authorization, forged role text,
forwarded agent claims, and filtered synthetic authorization.
GitOrigin-RevId: 8846de69ba214a88f58f70675798d324e213427c
## What changed
Rename the token-budget configuration key and its Rust field from
`use_history_notes_history` to `use_history_notes_extension`. Update the config
schema, extension gating, explicit-setting detection, and tests to use the new
name consistently.
GitOrigin-RevId: b6a514e23f02095fcd1d35dc42eeb7394b800c77
## Why
Token-budget sessions need a way to recover prior conversation context and preserve working state across context-window transitions.
## What changed
- Add direct-model `history` tools for listing windows and items, reading items, and searching conversation contents.
- Add direct-model `notes` tools for listing, reading, searching, appending, and writing persistent notes.
- Route tool calls through the configured Codex backend with trusted session and agent context, bounded request arguments, and truncation-aware output handling.
- Expose the extension when `features.token_budget.use_history_notes_history` is enabled with an OpenAI provider and Codex backend authentication.
## Testing
- Cover tool registration, configuration changes, provider and authentication requirements, backend request context, encrypted output preservation, and request and response limits.
GitOrigin-RevId: 43b259f01014ba3f30803dd7cd5634942407a9bb
## Why
Guardian v2 applied an implicit token limit to classifier instructions even when
no limit was configured, which could truncate the rendered policy.
## What changed
- Leave classifier instructions unbounded by default.
- Continue honoring `max_classifier_instruction_tokens` from local or model
configuration, with local configuration taking precedence.
## Testing
- Cover full policy rendering without a configured cap and truncation when an
explicit cap is present.
GitOrigin-RevId: 70eb42d43858e0656129dd438b10b60940fd97c6
## What changed
- Treat standalone `function_call_output` items without a `call_id` as external context, including items injected into new and forked threads, and mark the thread memory mode polluted when `memories.disable_on_external_context` is enabled.
- Include these outputs in guardian transcripts with their namespaced tool name and a placeholder for non-text content.
- Allow image generation to reuse images from standalone and otherwise unpaired function or custom-tool outputs.
## Testing
- Cover injected thread items, external-context detection, guardian transcript rendering, and recent-image selection.
GitOrigin-RevId: ca4e8c16a4b731e1256a5956908d00f785b4feb1
## Why
External tool events may need to enter thread history without a preceding function call and therefore do not have a `call_id`.
## What changed
- Allow `function_call_output` items to omit `call_id` and carry optional `name` and `namespace` fields.
- Preserve named standalone outputs during history normalization and agent forks while retaining existing pairing behavior for outputs with a `call_id`.
- Accept, persist, and forward these outputs through `thread/inject_items`, and update the app-server schemas and documentation.
## Testing
- Cover paired and standalone JSON round trips, history normalization, agent forks, and injected thread history.
GitOrigin-RevId: a3258163a7dc93777c7c3023116fe204819bdbb0
## What changed
- Add `otel.tool_result.max_bytes`, defaulting to 2048 bytes, to control the
text included in `codex.tool_result` logs independently of model-visible
output limits.
- Apply truncation centrally when emitting telemetry, preserve UTF-8
boundaries, and report whether the logged output was truncated.
- Record typed tool names and namespaces, agent names, and process-local result
sequence numbers while keeping tool arguments and output out of trace events.
## Testing
- Cover configuration propagation, UTF-8-safe truncation, raw tool-output
logging, and log/trace event routing and fields.
GitOrigin-RevId: 89ca76e55f1a09b729e7770ce01f2e4d98f9a22d
## What changed
- Remove the legacy `shell_command` handler and runtime, leaving `exec_command`
and `write_stdin` as the shell execution tools.
- Treat legacy `shell_command` model metadata as `unified_exec`, and normalize
legacy user opt-outs so they do not disable command execution. Managed feature
requirements and `shell_tool` can still disable it.
- Preserve shell approvals, sandboxing, zsh-fork support, and output truncation
through the unified execution path.
## Testing
- Cover legacy configuration and model-metadata compatibility.
- Exercise unified shell execution, approvals, truncation, and `apply_patch`
serialization across the app-server and core test suites.
GitOrigin-RevId: 5c2fd6164fc3519cdae4944cb9db276b8467311c
## What changed
- Make `ExecutorFileSystem::walk` a required backend operation instead of
providing a fallback built from directory reads and metadata requests.
- Implement bounded local walks on a blocking task with cancellation, symlink
cycle detection, deterministic ordering, error collection, and response-size
limits.
- Have remote filesystems use the server's walk operation directly.
## Testing
- Cover local and remote handling of invalid roots and limits, directory
symlinks, non-UTF-8 names, cancellation, sandbox contexts, and response
budgets.
GitOrigin-RevId: 7499bf05080c3f9965a5eb7ffd593de604d62c2a
## What changed
- Add `max_tool_call_lag`, `reuse_parent_compaction`, and transcript
`include_images` to the Guardian model configuration.
- Inherit these settings from model defaults while preserving explicit local
overrides.
- Enable image capture for Node REPL review evidence when transcript images are
included.
## Testing
- Cover inheritance and local override precedence for the new settings.
- Verify that enabling transcript images initializes review-evidence capture.
GitOrigin-RevId: 7c1965c2ec00a78a0436fa916b630be0004dbb51
## Why
An `apply_patch` path can be replaced with a symlink after verification, allowing an unsandboxed patch operation to reach a different file than the one that was approved.
## What changed
- Add `follow_symlinks` options to executor filesystem reads, writes, metadata lookups, directory creation, and removal, including the corresponding `followSymlinks` protocol fields.
- Implement no-follow filesystem operations on Unix and Windows that reject links in any path component and restrict file access to regular files.
- Run `apply_patch` with symlink traversal disabled when an otherwise-required sandbox is bypassed, while retaining the existing follow-symlink default for standalone callers.
## Testing
- Cover leaf and ancestor symlinks across patch add, update, delete, and move operations, including a path swap after verification.
- Exercise local and remote no-follow filesystem behavior, concurrent directory creation, special-file rejection, and Windows reparse points.
GitOrigin-RevId: 43fd479084891493ce13564fbd894b98f329c6dd
## What changed
- Exclude sandboxed `exec_command` and `shell_command` calls from Guardian v2 classification by default while continuing to classify calls that request `require_escalated` permissions.
- Add `guardianv2.review_scope.sandboxed_exec_commands` to opt sandboxed shell commands back into classification.
- Keep other tools and namespaced shell tools in scope, and advance tool-call progress when a call is skipped.
## Testing
- Cover the default and configured review scopes, tool namespaces, permission modes, and skipped-call progress tracking.
GitOrigin-RevId: 32fb540c69959b9a82569f0f2fc76b5517496e6b
## What changed
- Use `ExecutorFileSystem::read_file_stream` for capability files and executor skill resources regardless of whether filesystem sandboxing is active.
- Enforce per-file and bundle size limits incrementally as chunks arrive.
- Preserve the existing error for Windows resources when the required filesystem sandbox is unavailable.
GitOrigin-RevId: 1a40602a8f913588ec9a6ad7edd2d62eb82436e1
## What changed
- Move the Guardian thread lifecycle contributor and subagent-spawn context into `codex-guardian-v2`.
- Install the lifecycle contributor and asynchronous risk scorer through a single extension entry point.
- Remove the redundant `codex-guardian` crate and group the existing scorer implementation under `async_scorer`.
GitOrigin-RevId: 97141190c843e588b8c09c9ccd3fe00e1fae39a4
## What changed
- Add environment-provided MCP restrictions for configured and plugin-provided servers.
- Disable attachment-scoped servers while their environment configuration is pending, failed, or unselected, while preserving selected-plugin access and the controller-owned Apps server.
- Apply the resolved policy to runtime startup, model tool exposure, telemetry, OAuth flows, and skill dependency installation.
## Testing
- Cover policy filtering for configured and plugin servers, environment state transitions, Apps ownership, and managed OAuth rejection.
GitOrigin-RevId: 7ca5d4dda14068e758497f88835d5cba685e159e
## Why
Selecting only the newest entries changes the retained transcript prefix whenever a new entry arrives, reducing cache stability.
## What changed
- Replay non-user transcript entries through a bounded buffer and evict half of the applicable pool when an entry or token limit overflows.
- Keep message and tool token budgets independent, and reject entries that cannot fit without evicting retained evidence.
- Preserve user-message selection and emit retained entries in transcript order.
## Testing
Added coverage for stable tool prefixes, message-budget overflow, rejected messages, and chunked entry-count eviction.
GitOrigin-RevId: e6181f5a229a6de06c7ef5db21a5c74980a53e97
## What changed
- Carry selected plugin identities from MCP discovery into per-turn extension data, limited to capability roots that are ready for the turn.
- Annotate executor skill catalog entries with the matching plugin ID and `user` scope.
- Include that attribution in `skill_invocation` analytics for both explicit and implicit executor skill invocations.
## Testing
- Extend the executor skills app-server test to verify plugin ID, skill scope, and invocation type for explicit and implicit analytics events.
GitOrigin-RevId: 9345600ae29bcb8eb894fffd2d6f7b05f22ebf74
## What changed
- Treat configuration, action serialization, thread lookup, and classification errors as elevated risk instead of retaining a prior low-risk result.
- Track asynchronous scoring failures separately from completed scores so approval review falls back to strict review when the latest tool call could not be scored.
- Preserve newer classification results when recording a fail-closed score.
## Testing
- Add coverage for each failure path and for ordering fail-closed scores with concurrent classifications.
GitOrigin-RevId: 7012f078a24031848f2943354a206866286ad9f4
## What changed
- Stop writing Guardian v2 security risk scores to rollout history.
- Treat resumed and forked threads as having no previous Guardian score, so their
first tool approval is classified and reviewed normally.
- Remove the now-unused `codex-history` dependency from the Guardian v2
extension.
## Testing
- Update app-server coverage for new, resumed, and forked threads to wait for
classification completion and verify the expected approval reviews.
GitOrigin-RevId: 74d9caa903a1ff9d38e577b4697b5f8970f11ebb
## What changed
- Pass session-attributed extension metrics to the Guardian v2 sampler.
- Record `codex.guardian_v2.classification.token_usage` histograms for total, input, cached input, cache-write input, non-cached input, output, and reasoning output tokens.
- Preserve token accounting when streamed JSON produces an early score by recording usage from the terminal response in the background.
## Testing
- Cover token metrics through the Guardian v2 extension and the sampler's early-score path.
GitOrigin-RevId: 04fadc07f3b9cbe83f9fa3737536860b5cf5b4be
## What changed
- Replace the minimal default classifier prompt with predictive security-review guidance covering authorization, action risk, untrusted evidence, and browser or computer use.
- Trigger full Guardian review at an `action_risk` score of `0.5` or higher for the new default prompt, while retaining the `0.8` calibration for custom and model-provided prompts unless they specify a threshold.
- Render security policy configuration into prompt templates before enforcing the classifier instruction token limit. Prompts without the policy placeholder retain the existing append-and-truncate behavior.
## Testing
- Cover default, custom, and model-provided prompt thresholds and precedence.
- Verify policy substitution, truncation, sampling requests, and full review at the threshold boundary.
GitOrigin-RevId: 7aad9e2d8c9fbea453dc52147a3d0a46bf191aec
## Why
Guardian sampling uses a pool of WebSockets, while the bridge routes each
connection by thread ID. Each socket therefore needs its own identity instead
of reusing the source thread ID across the pool.
## What changed
- Generate a classifier thread ID for every pooled connection and use it for
request, window, and subagent attribution.
- Preserve the originating thread ID in turn metadata alongside the Guardian
classifier request kind.
- Build request metadata after leasing a connection so retries use the identity
of the socket that carries the request.
## Testing
- Cover unique identities across pooled and replacement connections.
- Verify classifier attribution metadata for extension requests and retries.
GitOrigin-RevId: 1c36207125157ff05cab9487a2fa37bf85e5fdd8
## What changed
- Emit a `codex.guardian_v2.classification` counter for each classification.
- Record classification latency in `codex.guardian_v2.classification.duration_ms`.
- Tag both metrics with `success`, `failure`, or `superseded` outcomes.
## Testing
- Verify metrics for successful classification and fail-closed handling.
GitOrigin-RevId: 021702f3b26565d6da5708d2b8f1f8881942a441
## Why
`node_repl` screenshots are nested tool results and do not appear in the parent
conversation history, so Guardian v2 transcript image collection could miss
them.
## What changed
- Allow Guardian v2 to enable bounded, thread-scoped `node_repl` image capture
when transcript images are configured.
- Add retained `node_repl` screenshots to review requests when transcript
sources include tool outputs, while deduplicating images and keeping them out
of the parent turn input.
## Testing
- Add an integration test covering screenshot capture without the synchronous
Guardian transcript feature flags and verifying that screenshots stay out of
the parent request.
GitOrigin-RevId: 5f70c5f2e80389aa8b2a286c3b7f9b1104865486
## What changed
- Pass metrics bound to the active turn's model to approval-review contributors.
- Add counter support to `ExtensionMetrics` while preserving host session attribution.
- Record Guardian v2 tool-call score lag and count stale-score fallbacks with the `score_lag` reason.
## Testing
- Cover counter forwarding and session metadata tags.
- Verify Guardian v2 metric samples for current and stale scores.
GitOrigin-RevId: 0ec34736d29d2e37ed217a27e773b4f6230ae38a
## Why
Guardian sampling WebSockets authenticate when the connection is opened. Reusing
a pooled connection after credentials change can therefore keep using the old
authorization.
## What changed
- Track authentication changes on each pooled sampling connection.
- Discard stale connections after an auth change and reject connections whose
authentication changes while the handshake is in progress.
## Testing
- Verify the sampler reconnects with the refreshed bearer token instead of
reusing its existing connection.
- Verify the installed Guardian extension reconnects after an external auth
refresh.
GitOrigin-RevId: 6d2e7df776fd21c78be0928f71162f5419a8b8f0
## What changed
- Add the experimental `autoApprovalReview/strictReviewRequired` app-server notification with the review's `threadId`, `turnId`, and `startedAtMs`.
- Emit it when Guardian v2 leaves an approval review in progress because the action risk is elevated or its score is stale.
- Export the notification through the Rust, TypeScript, and JSON schema surfaces and route it to the associated thread.
## Testing
- Cover strict-review reason tracking for elevated and stale risk scores.
- Verify that high-risk approval routing emits one notification whose identifiers and timestamp match the review-started event.
GitOrigin-RevId: e65352894520bfb756c75b2c95e2e149f9767bb2
## What changed
- Give `PluginsManager` a shared `AuthManager` instead of a separately mutable authentication-mode snapshot.
- Read the current authentication mode and credentials from that shared manager for plugin discovery, startup tasks, CLI commands, MCP setup, and external-agent migration.
- Update test helpers and coverage to exercise plugin projections and curated marketplace selection as authentication changes.
GitOrigin-RevId: 600c94de5130eda2da5727e1a0b4d39083fefc56
## What changed
- Resolve MCP server policy directly from the effective user configuration for selected executor-plugin roots.
- Apply server enablement, tool allow/deny lists, and default and per-tool approval modes while preserving stricter plugin-declared restrictions.
- Cover policy merging and opaque selected-root IDs in the executor-plugin MCP integration tests.
GitOrigin-RevId: dce3021969ba71e642de52312449cab6277515c3
## Why
Image generation did not save its result when the extension host did not provide a host-local save root.
## What changed
- Save generated images under the turn environment's `generated_images` directory through its sandboxed filesystem executor when no host save root is configured.
- Limit executor-written images to 32 MiB, reject symlinked output directories and existing destinations, and continue returning the image if persistence fails.
- Include the saved workspace path in the tool output hint while preserving host-managed persistence when a save root is available.
## Testing
- Extend the granted-turn-permissions integration test to verify the image and path hint are written to the turn workspace without creating a host-local artifact.
GitOrigin-RevId: 3d24381a75fe4595641b5250e19fac8ed473b2cd
## What changed
- Stop parsing and exposing the `model` field from skill frontmatter.
- Remove the skill model delegation types and instruction generation.
GitOrigin-RevId: 01198c68c095da5062e8abec417cc3539099d9d5
## Why
Durable thread queues can be updated through another SQLite connection, but loaded idle threads need to notice those updates before they can dispatch the new messages.
## What changed
- Track a durable revision for each thread queue, including backfilling existing queues and updating revisions on inserts, updates, and deletes.
- Poll SQLite's data version and query revisions to find changed queues for loaded threads.
- Wake idle threads with pending external messages, discover queued work when threads are loaded or resumed, and retry each thread independently so one blocked queue does not stall others.
## Testing
- Cover cross-runtime queue writes, edits, independent dispatch, wake retries, and resumed threads.
- Cover migration backfills and revision tracking for queue updates and deletes.
GitOrigin-RevId: 906d902bbf6a760be206a1b2c08fd71427b63c46
## Why
Short continuation prompts such as `continue` do not contain enough context for
skill selection on their own.
## What changed
- Add the `task_context_fusion_v1` shadow selector, combining the current request
with up to two prior substantive requests and recently relevant skills.
- Record explicit skill intent and successful skill invocations for future turns,
while excluding same-turn observations from predictions.
- Bound retained requests, augmented queries, and skill history, including safe
truncation at UTF-8 character boundaries.
## Testing
Add unit and extension tests for continuation prompts, explicit intent, turn
isolation, cold thread state, bounded history, and unchanged control selectors.
GitOrigin-RevId: 72eca3f0d64620a0d24e95d5675f126ac982f8c7
## Why
Guardian stores a `SecurityRiskScore` before advancing its scored-tool-call
counter. Waiting only for the score can therefore let the test continue while
the asynchronous sampler is still updating its progress state.
## What changed
Wait for `latest_scored_tool_call` to advance before exercising Guardian v2
approval decisions, and reuse the same progress tracker for the subsequent
assertions.
GitOrigin-RevId: 3d92c9a3ccf52f4f43a2454aac60d2479258c1c8
## What changed
- Track the latest tool call and latest scored tool call for each Guardian v2 thread.
- Skip approval review when the score lags by more than `max_tool_call_lag` tool calls.
- Add `max_tool_call_lag` to the Guardian v2 configuration and schema, with a default of three.
## Testing
- Cover approval review at, above, and after recovering from the configured lag limit.
- Cover deserialization of `max_tool_call_lag`.
GitOrigin-RevId: fd870e143d95ee76491180e2aed742717248ff78
## Why
Guardian sampling can encounter temporary connection or service failures while
opening a WebSocket, starting a request, or consuming its response stream.
## What changed
- Retry transient transport, streaming, rate-limit, and server failures up to
two times across the full sampling flow.
- Recover from unauthorized responses through the configured auth manager and
discard pooled connections that may contain stale credentials.
- Allow sampler initialization to continue when its initial WebSocket prewarm
fails, so the first sample can establish a fresh connection.
## Testing
Added coverage for recovery after prewarm failures and `503` responses, and for
stopping once the transient retry budget is exhausted.
GitOrigin-RevId: af74c2df95c4241184ca641ef4418db6aa4ab695
## What changed
- Add shadow selectors that fuse recently invoked skills with character routing-card matches using reciprocal rank fusion.
- Add a second variant that also incorporates weighted lexical matches.
- Record both variants in the existing shadow-selection experiment, capped at 50 candidates.
## Testing
- Cover cold starts, duplicate removal, short exact matches, deterministic ranking, result limits, and truncation metadata.
- Extend the skills integration test to verify both selectors recover a skill invoked on an earlier turn.
GitOrigin-RevId: 4ff9e91da8cb49ecbce54ab8e760820616e1a1d3
## What changed
- Add optional `guardian_v2` model-message defaults for classifier instructions,
review thresholds, reasoning effort, transcript selection and limits, and
action and compaction token limits.
- Apply those defaults when Guardian v2 samples a tool call, while preserving
explicit `[features.guardianv2]` settings over catalog values.
- Preserve the new catalog configuration when applying model overrides.
## Testing
- Cover model-catalog serialization and model override behavior.
- Verify that sampling uses catalog defaults while retaining local overrides.
GitOrigin-RevId: 4ae9ec235c647acbca05a20b7c4b87a4e1331161
## What changed
- Add the `features.guardianv2.transcript.include_images` opt-in setting.
- Include recent images from user and assistant messages and configured tool outputs in Guardian v2 sampling requests.
- Bound image context to the four most recent images and 8 MiB total, and omit image detail hints when constructing the sampling request.
## Testing
- Cover configuration parsing and image forwarding from messages and tool outputs.
GitOrigin-RevId: b0ae6b6182535fd084ea103ed01dc3c803d7d1f7
## What changed
- Add `max_parent_compaction_tokens`, defaulting to 25,000 tokens, to the Guardian v2 feature configuration.
- Reuse the latest encrypted parent compaction only when its complete serialized item fits within the configured limit.
- Fail closed with the maximum action-risk score when the latest encrypted compaction is oversized, instead of sampling with older context.
## Testing
- Cover exact-boundary and oversized compaction items, including oversized passthrough metadata.
- Verify that an oversized latest compaction bypasses sampling and prevents approval.
GitOrigin-RevId: 0b308e7bec0f41d46433dfdf656948cee5cba667
