Commit Graph
41 Commits
Author SHA1 Message Date
jif 98a5cb46b1 Manage synchronous Guardian reviewers through the thread manager (#43570)
## What changed

- Store the shared `GuardianReviewSessionManager` in thread extension data and let the Guardian extension spawn and fork reviewers through `ThreadManager`, independently of async scorer startup.
- Preserve reviewer reuse, concurrent forks, cancellation, inherited environments, and user instructions. Gate managed reviewer startup on parent registration and remove terminated reviewers from the manager.
- Retain the `guardian` subagent identity in saved conversations. Reject resume of live Guardian reviewers and reject client archive or delete of live owner-managed workers with JSON-RPC error `-32600`; allow normal access after the owner releases them.
- Keep standalone reviewer support for inline delegates and remove duplicated extension prompt and configuration code.

## Testing

Add app-server tests covering reviewer reuse and concurrent forks, inherited instructions and prompt cache keys, live resume and removal rejection, saved reviewer discovery and resume after parent shutdown, and strict Guardian assessment inside an inline review.

GitOrigin-RevId: 4f9893c1ade151eacb4c11ffeae3d9e9ec019187
2026-09-07 19:43:10 +00:00
jif ce5c4133bd Route MCP elicitations through the shared approval decision path (#43447)
## What changed

- Use `decide_approval` for MCP elicitations, carrying the effective approval policy, reviewer, and synchronous-review requirement into the decision.
- Preserve unsupported form and URL elicitations for user review.
- Use `computer_use_review_required()` for review evidence, session policy, and turn metadata so model policy takes precedence over the legacy flag.
- Derive strict-review notifications from each assessment's review reason.

## Testing

Add coverage for model policy precedence, user and full-access modes, sensitive elicitations, independent code-mode and nested-tool policies, cached score reuse, and user review of unsupported elicitations.

GitOrigin-RevId: d0a9c40572dca7b3dbf876230d2162b72728084d
2026-09-07 12:04:27 +00:00
jif 7769bccbb2 Avoid WebSocket connection waits in Guardian v2 classification (#43408)
## Why

When no healthy pooled WebSocket is available, classification waits for a new connection to open and can stall on its handshake.

## What changed

- Use HTTP streaming when no healthy idle WebSocket is ready, and replenish the pool in the background with a five-minute cooldown after connection timeouts.
- Share concurrency limits and the sampling retry budget across both transports, and bound the wait for HTTP response headers.
- Cancel superseded requests while awaiting response headers and stop SSE polling when the consumer closes.

## Testing

Add coverage for HTTP fallback and recovery after cooldown, shared transport capacity, stalled response headers, and cancellation before headers or during body draining. Update sampler and extension tests to cover fallback after socket expiry, authentication changes, and transient failures.

GitOrigin-RevId: 0b23ccc1bfb2dfdbc8d2791d097706cd4dd44f39
2026-09-07 09:24:03 +00:00
felixxia-oai e4ce83419b Move Guardian thread context into guardianv2 configuration (#43104)
## What changed

Replace `features.guardian_thread_context` with `features.guardianv2.thread_context` and update the configuration schema. The setting controls thread-owned context for synchronous and asynchronous Guardian, defaults to `false`, and remains independent of `features.guardianv2.enabled`.

## Testing

Update configuration tests to cover the nested setting, profile overrides, and enabling thread context while Guardian v2 is disabled. Migrate Guardian approval, history, and async scorer tests to the new configuration key.

GitOrigin-RevId: 54a91555eb68001e0adedc8a07bdbc5ea2273f9b
2026-09-05 21:28:11 +00:00
Benjamin Carlsson ddf04ad267 Wait for turn analytics before shutting down the Guardian v2 test (#42933)
In `guardian_v2_routes_scoped_tool_approvals`, wait for the reviewed
thread's `codex_turn_event` before shutting down the app server. Use the
returned event for assertions instead of looking it up in the
post-shutdown analytics snapshot.

GitOrigin-RevId: 3ba4d37838852842c6caf7701d3abf7c89166a98
2026-09-05 05:31:48 +00:00
felixxia-oai 4636819a35 Harden Guardian reviews after context compaction (#42852)
## Why

Guardian reviews must not lose user authorization constraints or reuse an
unreadable or incompatible parent checkpoint after compaction.

## What changed

- Retain bounded excerpts of oversized root user messages across compaction,
  resume, and rollback, and recover excerpts from legacy Guardian history when
  available.
- Require thread-owned Guardian review sessions and asynchronous scoring to
  reuse only non-empty checkpoints produced by a compatible review model.
- Fail closed instead of reviewing without the required checkpoint when reuse
  is disabled, compatibility is unknown, or checkpoint content is unusable.

## Testing

Expanded retained-context, review-session, asynchronous-scoring, and app-server
coverage for oversized instructions, legacy recovery, model incompatibility,
missing content, resume, rollback, and disabled checkpoint reuse.

GitOrigin-RevId: f30d837197e0d838e656a632109f490c38208f5e
2026-09-04 21:07:14 +00:00
felixxia-oai 87628df77a Preserve root authorization context in Guardian reviews (#42832)
## Why

Guardian reviews for delegated workers need the current root instructions and verified answers even after the parent context is compacted. Approvals must also become stale when that root authorization changes.

## What changed

- Build bounded root review evidence from retained context, preserving source order and answer scope while prioritizing user instructions over optional assistant context.
- Recover retained instructions from Guardian history after compaction and mark authorization incomplete when required instructions or answers are unavailable.
- Version root authorization in synchronous and reusable review sessions so an allow result is cancelled when its evidence changes.
- Strip parent-only Guardian approvals when forking worker history in retained-context mode.

## Testing

- Cover retained and legacy context modes, oversized evidence, message limits, parent compaction, and authorization changes during review.

GitOrigin-RevId: 658219b7cee08f2752adcea9966268fd21727976
2026-09-04 19:19:04 +00:00
jif 4e48cd02da Honor model-provided Guardian review policies (#42744)
## What changed

- Add optional per-model Guardian modes for computer use, shell, code mode,
  file changes, MCP, network, and permission requests.
- Apply disabled, synchronous, or adaptive review by action category while
  preserving legacy configuration when model policy is absent.
- Bind cached risk scores to the active model policy, thread settings, and
  environment selections so outdated scores cannot approve later actions.
- Keep unknown review modes on the synchronous path and ignore unknown policy
  fields for forward compatibility.

## Testing

- Cover policy serialization, legacy fallback, scoped tool classification,
  disabled categories, and score invalidation after policy changes.

GitOrigin-RevId: f334a5f4fffb198da66cde3a400233e0df650516
2026-09-04 11:59:18 +00:00
jif 38ba8cdceb Honor model requirements in Guardian computer-use scoring (#42422)
## Why

Computer-use-only Guardian scoring should follow the active model's REPL
auto-review requirement, including when the model changes within a live thread.

## What changed

- Run computer-use scoring and fast approval decisions only when the active
  model sets `node_repl_auto_review_required`.
- Invalidate prior or in-flight scores when a model switch skips scoring, so
  switching back to a reviewed model cannot revive a stale decision.

## Testing

- Cover model switches for both `node_repl` and `cua_repl` MCP servers.
- Verify skipped scoring and stale-score rejection across requirement changes.

GitOrigin-RevId: 46aec4d017bea8f135b435bcd769b87369b8ce95
2026-09-03 02:27:43 +00:00
jif a0dcfe2ada Skip Guardian scoring in User approval mode (#42256)
## What changed

- Skip Guardian prewarming and asynchronous scoring when `approvalsReviewer` is
  `"user"`, including when the reviewer changes during an active turn.
- Automatically accept ordinary `node_repl.js` execution confirmations in this
  mode while continuing to surface sensitive-action checks and requests for user
  input.
- Cover transitions between User approval, automatic review, and Full Access in
  the app-server integration tests.

GitOrigin-RevId: 599c8ebb672728b9bf3d64c34377ddec230325f8
2026-09-02 10:32:22 +00:00
jif e576993911 Skip Guardian reviews in Full Access (#42147)
## Why

Full Access already combines `approvalPolicy: "never"` with unrestricted
permissions, so confirmation-only actions do not require a model review.

## What changed

- Detect Full Access consistently across the thread and every selected
  environment. Pending, failed, or restricted environments are not treated as
  Full Access.
- Approve confirmation-only Guardian and MCP requests without synchronous
  review, sampler prewarming, or background scoring. Cancellation, explicit
  denials, and forms that require user input keep their existing behavior.
- Re-evaluate the active permission state on each turn so an existing thread
  can enter or leave Full Access safely.

## Testing

Added coverage for switching approval modes, strict sensitive MCP
confirmations, failed and pending environments, and suppression of Guardian
requests and background-scoring connections.

GitOrigin-RevId: 089ab4296dde473b8e33ab8324be79c5446c46f6
2026-09-01 22:16:54 +00:00
jif 86b7d12742 Add Guardian V2 analytics events (#42144)
## What changed

- Emit `codex_guardian_v2_classification` events with the outcome, risk
  level, duration, model, and thread attribution.
- Emit `codex_guardian_v2_fast_decision` events when Guardian V2 approves an
  action without a full review.
- Add `guardian_v2_enabled` to turn analytics, accounting for review scope
  and model requirements.

## Testing

- Extend the Guardian V2 app-server coverage to verify event payloads,
  attribution, enablement state, and scoped event emission.

GitOrigin-RevId: 0805975f8e8f9d954aa1058ea8c60545a090ae74
2026-09-01 22:02:42 +00:00
felixxia-oai 2350823caa Unify Guardian context section collection (#42076)
## What changed

- Add typed context sections for root-conversation evidence, trusted user answers, and conversation transcripts to `codex-guardian-context`.
- Use the shared section registry when building core Guardian prompts and asynchronous Guardian V2 scoring context.
- Preserve source-role labeling while delivering authorization evidence as user-role context, and fail closed when context collection fails.

## Testing

- Cover shared section ordering and role preservation for synchronous and asynchronous targets.
- Verify trusted user answers appear exactly once as user-role input in Guardian V2 approval flows.

GitOrigin-RevId: 1d5d793ffbf166077411a66e85e039596907fb7e
2026-09-01 15:22:21 +00:00
felixxia-oai 9f97cb79eb Preserve Guardian review evidence across compaction (#41879)
## Why

Guardian reviews need the original conversation evidence even after the parent context has been compacted, while rollback must still remove evidence from reverted turns.

## What changed

- Build synchronous and asynchronous Guardian transcripts from the retained review-history snapshot.
- Use the shared Guardian transcript collector for consistent entry formatting, tool attribution, and per-entry truncation while preserving the existing transcript retention budgets.

## Testing

- Add app-server coverage for retaining user and tool evidence after compaction and discarding it after rollback.
- Extend transcript tests for standalone tool outputs, shell calls, attribution, truncation, and retention limits.

GitOrigin-RevId: a18b1176f14a4d006c263aa82d4e593afee66bc3
2026-08-31 17:10:38 +00:00
jif cefa060695 Approve the first Node REPL execution without a Guardian wait (#41666)
## Why

The first REPL execution should proceed while its initial asynchronous Guardian
classification is still pending.

## What changed

- Fast-approve the first `js` execution from a Node REPL-backed server while
  continuing its asynchronous classification.
- Track `js` executions separately so setup and reset tools do not consume the
  first-execution allowance.
- Apply the normal Guardian review policy to subsequent executions.

## Testing

Add coverage for browser and computer-use startup, reset, and module-directory
setup sequences, verifying that only the first `js` execution skips the wait.

GitOrigin-RevId: 7297b35411a6317bcf9e7058c08c6db3e3310ac8
2026-08-30 13:02:11 +00:00
jif 31d338a1ea Isolate required-model Guardian approval coverage (#41322)
## What changed

Update the Guardian v2 required-model-switch test to continue the thread with
empty input. This keeps authorization unchanged, so the test specifically
verifies that the required-model check prevents reuse of a cached approval for
the sandboxed command.

Continue asserting that a Guardian review starts and that the command skips a
second classification.

GitOrigin-RevId: 5ce914cb6e66f0527d9ec154331cc356104bfb72
2026-08-28 11:22:20 +00:00
jif f98649cde9 Honor required reviews when reusing Guardian scores (#41309)
## Why

After switching a thread to a model that requires automatic review, a
standard-scope approval could reuse a cached low-risk score for an action that
skipped classification. This allowed the action to bypass the required full
review.

## What changed

Check the thread's current model requirements before making a fast Guardian
decision. Defer required models to the full review path instead of approving
from cached scoring state.

## Testing

Add coverage for switching to a required model and reviewing a sandboxed
command without running another classification.

GitOrigin-RevId: e79e6a5df19c46acd5c4176f522a859793177639
2026-08-28 10:03:19 +00:00
rka-oai 035295b46e Improve sandboxing, MCP errors, and cached approvals (#41196)
Expand Windows deny-read globs robustly across protected directories while preserving filesystem enforcement. Preserve structured MCP tool and resource errors through app-server responses. Bind cached Guardian classifications to current local and root authorization state so stale scores cannot approve actions.

Add regression coverage for the sandbox resolver, structured protocol errors, and authorization changes.

GitOrigin-RevId: 4b80ed724d869afeca79204222d8465fa99d3a24
2026-08-27 20:05:08 +00:00
felixxia-oai 694edc23b2 Propagate trusted root skills to delegated workers (#41118)
## Why

Guardian reviews for delegated workers need to recognize skills explicitly
invoked by the user on the root turn.

## What changed

- Record verified user-owned skill paths as bounded, turn-matched Guardian
  evidence.
- Include trusted paths from the current worker and its root snapshot when
  building review context.
- Exclude skill trust from later root turns so stale authorization is not
  inherited.

## Testing

- Add an app-server lifecycle test covering inheritance of a root-invoked
  skill by a delegated worker and its removal on the next root turn.

GitOrigin-RevId: aa28f485edb7a447e51d51de8dc89b959fd63d40
2026-08-27 11:31:40 +00:00
Ankush Gupta 89650c66f2 Require synchronous review for sensitive MCP actions (#41094)
## 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
2026-08-27 08:44:54 +00:00
felixxia-oai b68acc4d4b Trust invoked user skills in Guardian reviews (#41006)
## 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
2026-08-27 00:07:26 +00:00
jif 102ae5e2e6 Prewarm Guardian WebSockets without blocking thread startup (#40985)
## Why

Opening Guardian's initial WebSocket connections can be delayed, but thread startup and resume do not need to wait for those connections.

## What changed

- Install the Guardian sampler and related thread state before opening its initial connections.
- Prewarm the sampler's WebSocket pool in a background task while retaining the existing on-demand connection behavior.

## Testing

- Verify extension startup returns before a delayed WebSocket handshake completes, then warms the full initial connection pool.
- Verify resuming a thread likewise returns before Guardian's delayed handshake completes.

GitOrigin-RevId: 58c91cf045b223f917c67d7e8dc82eac529db4ce
2026-08-26 22:10:47 +00:00
felixxia-oai d61ba72f2f Give Guardian trusted context for configured MCP tools (#40982)
## What changed

- Add a bounded developer context fragment identifying the MCP server or connector and the user-owned configuration that declared it.
- Emit the fragment only when the effective entry matches the user configuration or an active plugin declaration resolves inside the Codex home directory.
- Keep tool descriptions, outputs, and unrelated tools untrusted, and reject unsupported sources or paths that escape through symlinks.

## Testing

- Cover user-configured servers and connectors, plugin-provided capabilities, token truncation, symlink escapes, and app-server request integration.

GitOrigin-RevId: 0bfe2a2f3a48334d1d5faad692b5d36452febb68
2026-08-26 21:48:05 +00:00
jif daa3eaf10f Allow Guardian scoring for required computer-use models (#40967)
## Why

Models that require automatic review previously skipped Guardian v2 risk scoring
entirely, even when Guardian was limited to computer-use tools. That scope cannot
use a low-risk result to approve unrelated tools, so the blanket skip is
unnecessary.

## What changed

- Retain Guardian v2 risk scoring for required-review models in
  `ComputerUseOnly` scope while continuing to skip it in standard scope.
- Preserve strict automatic approval review for Node REPL-backed tools, including
  when classification fails.

## Testing

Added coverage for low-risk, high-risk, and invalid classifications across the
Node REPL and computer-use REPL paths, plus an unrelated MCP tool.

GitOrigin-RevId: ee38fabf82e196df0ae13b4216b5892be9b799cf
2026-08-26 20:07:39 +00:00
jif 10d5a603ae Persist Guardian V2 risk scores without restoring them (#40884)
## 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
2026-08-26 13:54:36 +00:00
olliem-oai 9dea1f709f Default Guardian v2 to computer-use reviews with images (#40846)
## What changed

- Default Guardian v2's review scope to computer-use tools. Set
  `features.guardianv2.review_scope.computer_use_only = false` to retain the
  broader tool review scope.
- Include images in Guardian transcripts by default. Set
  `features.guardianv2.transcript.include_images = false` to disable them.

## Testing

- Update configuration and app-server coverage for the new defaults and
  explicit broader-scope overrides.

GitOrigin-RevId: 88c3cfe6dd2a96df052111794ad59c5c9ea39b78
2026-08-26 09:13:33 +00:00
Owen Lin 17c4c47dfa Default durable threads to paginated history (#40677)
## What changed

When `thread/start` omits `historyMode`, select `paginated` for durable
threads whose store supports `thread/turns/list` and `thread/items/list`.
Continue to default ephemeral threads and threads backed by stores without
those APIs to `legacy`.

Explicit `legacy` and `paginated` selections keep their existing behavior.

## Testing

- Verify durable threads use paginated history with the local store.
- Verify ephemeral threads and stores without history-list support use legacy
  history.

GitOrigin-RevId: c94b6a3b522049d541da33c41618c4142370f1a7
2026-08-25 19:10:35 +00:00
jif 4b81410a80 Treat user input answers as Guardian authorization changes (#40623)
## Why

Answers collected by `request_user_input` can change what the user authorizes, but post-tool hooks may replace or reject the tool output that Guardian normally sees.

## What changed

- Record bounded, host-observed answers before post-tool hooks run and provide them as trusted evidence to synchronous and asynchronous Guardian reviews.
- Include root-thread answers in worker review context.
- Count successful answers as authorization changes so earlier review evidence becomes stale, while ignoring empty or unrelated answers.

## Testing

Add coverage for allowed and denied stale reviews, empty and oversized answers, hook-modified output, and root-to-worker propagation.

GitOrigin-RevId: 970c52e40338ecd057c7cd2a69d04b06e97cd068
2026-08-25 15:03:46 +00:00
jif 219c203417 Add a computer-use-only Guardian v2 review scope (#40480)
## What changed

- Add `features.guardianv2.review_scope.computer_use_only` to restrict asynchronous classification and fast approvals to browser and computer-use REPL tools.
- Keep other tools on the synchronous approval path without letting them age or reuse scores from in-scope calls.
- Give this scope precedence over `sandboxed_exec_commands` when both options are enabled.

## Testing

- Cover configuration precedence, direct and code-mode tool detection, fast-approval isolation, and end-to-end approval routing for browser, computer-use, and unrelated MCP servers.

GitOrigin-RevId: bfd9ab6abdbd90634b265d354e038724b1e81d65
2026-08-24 21:33:32 +00:00
felixxia-oai c2904b9c5a Record Guardian classifier input truncation metrics (#40465)
## What changed

- Track truncation and omission across Guardian v2 actions, transcript entries and images, and synchronous review evidence.
- Emit `codex.guardian_v2.classification.truncation` counters and byte histograms for original, retained, and omitted content after successful classifications.
- Tag observations by input component and whether content was truncated or fully omitted.

## Testing

- Extend async scorer coverage for action truncation metrics.
- Verify the app-server Guardian flow records truncated review evidence.

GitOrigin-RevId: 70680897f75716770aa072f866d60bd37e171c1d
2026-08-24 19:47:48 +00:00
felixxia-oai 523519d974 Render Guardian review evidence with async scorer truncation (#40431)
## What changed

- Retain completed synchronous Guardian reviews as structured records until the async scorer builds its classification input.
- Render and bound review correlation, action, rationale, and full evidence body with the async scorer's transcript truncation marker.
- Extend the Guardian V2 integration test to verify oversized action and rationale content is truncated and the resulting review evidence remains bounded.

GitOrigin-RevId: 0000ba77b810748e34254dd00df4a69f8492e175
2026-08-24 16:41:55 +00:00
jif a9e7920da1 Classify Guardian V2 risk with a single token (#40393)
## What changed

- Replace structured numeric classifier output with a single `high` or `low`
  token and map those classifications to `1.0` and `0.0` risk scores.
- Require the classification contract in both default and overridden prompts,
  including when classifier instructions are truncated.
- Return the first streamed classification immediately while draining the
  remaining response for connection reuse and token accounting.

## Testing

- Cover prompt rendering and truncation, classification parsing, early streamed
  results, conflicting later output, WebSocket reuse, and approval routing.

GitOrigin-RevId: 31a57e4eb9bf8df2a161b84aa3d5b16ab8f2937b
2026-08-24 12:22:02 +00:00
Dylan Hurd 50ea8fd411 Log Guardian V2 classification results (#40028)
## 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
2026-08-21 22:24:28 +00:00
felixxia-oai 9949c9eafa Reuse Guardian reviews in async risk scoring (#40013)
## 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
2026-08-21 21:22:35 +00:00
felixxia-oai c517cc6d84 Bypass risk scoring for models that require automatic review (#39981)
## 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
2026-08-21 17:33:45 +00:00
Ankush Gupta 4e1a772a7d Let Guardian V2 satisfy required model reviews (#39658)
## What changed

- Allow the Guardian V2 approval monitor to handle reviews for models that require automatic review.
- Preserve full Guardian review when Guardian V2 is disabled by configuration or managed requirements.
- Cover required-model routing for low- and high-risk actions, plus both Guardian V2 disable paths.

GitOrigin-RevId: d1201cc3dbd7c4a07de633388b14f17ada9feb39
2026-08-20 08:07:06 +00:00
felixxia-oai e51a91b2f4 Keep Guardian v2 risk scores in memory (#39304)
## 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
2026-08-18 22:37:50 +00:00
olliem-oai 846a16852f Improve Guardian v2 risk classification (#39264)
## 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
2026-08-18 19:12:30 +00:00
Shijie Rao f5e9d66851 Notify clients when Guardian requires strict review (#39157)
## 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
2026-08-18 06:48:12 +00:00
felixxia-oai 9dd3d6a13e Restore Guardian risk scores across thread lifecycles (#39038)
## Why

Resumed and forked threads need their persisted Guardian risk state available when routing later approval requests.

## What changed

- Initialize a session's `SecurityRiskScore` from the latest score in its rollout history when no score was supplied explicitly.
- Preserve the inherited score when resuming or forking a thread so low-risk actions can skip redundant Guardian reviews.

## Testing

- Add app-server coverage for inheriting the latest Guardian score when resuming and forking threads.

GitOrigin-RevId: 95a6c803f8dae75516d1da9d66a722379a0a7ef2
2026-08-17 17:33:30 +00:00
felixxia-oai 71e5e1ec50 Add app-server coverage for Guardian V2 approval routing (#39035)
## What changed

- Add end-to-end app-server tests for MCP tool approvals with Guardian V2 and automatic review enabled.
- Verify that sampling receives the thread context, uses a thread-scoped prompt cache key, and persists the resulting action risk score.
- Cover both routing outcomes: low-risk actions skip later Guardian reviews, while high-risk actions continue through full review.

GitOrigin-RevId: 78ba2b0560caf701ad333785c4dda4f7d2be4f2c
2026-08-17 17:24:06 +00:00