Commit Graph
71 Commits
Author SHA1 Message Date
felixxia-oai b3e0c49dfb Extract guardian transcript selection into guardian-context (#45417)
## What changed

Move full/delta transcript selection into the shared `TranscriptMode::select` API and use it when building guardian prompts. Export `TranscriptCursor`, `TranscriptMode`, and `TranscriptSelection` from `codex-guardian-context`.

Preserve full-transcript fallback when the history version changes or the saved cursor exceeds the collected entry count. Select entries before profile retention, preserving their numbering and returning a proposed cursor that counts all collected entries. Hosts remain responsible for committing and invalidating cursors.

## Testing

Add a regression test verifying that sliding-window retention preserves the collected-entry cursor and that an appended entry is selected and numbered correctly in the next delta.

GitOrigin-RevId: a2192c08e23c18302b0105ba47aeb2780eb4a015
2026-09-14 12:52:16 +00:00
felixxia-oai 9c4879f3a5 Preserve complete actions in Guardian approval reviews (#44569)
## 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
2026-09-10 16:01:02 +00:00
jif 4e6d5c0a96 Move Guardian reporting and denial accounting into the extension (#44544)
## What changed

- Move assessment event construction, metrics, and analytics tracking into `codex-guardian-reviewer` through `ReviewReport`.
- Store denial accounting in thread extension data through `ReviewDenials`, with core retaining turn interruption and lifecycle cleanup.
- Move failed-review record selection and bounded serialization into the extension, with core supplying captured review context.

## Testing

Add coverage that denial accounting clears on turn completion and interruption. Move the oversized-record test into the extension and exercise the new API.

GitOrigin-RevId: 6e5b3d9d4128b356f99b1d0e74a68111e40261ab
2026-09-10 14:13:57 +00:00
felixxia-oai 9688359977 Bound MCP descriptions separately from Guardian action JSON (#44493)
## Why

MCP tool and connector descriptions were included in required action JSON, consuming review input budget even though they are optional metadata.

## What changed

Move `tool_description` and `connector_description` into an optional, explicitly untrusted `guardian_tool_descriptions` fragment. Limit each description to 400 estimated tokens, escape closing tags, and allow budget enforcement to omit the fragment while retaining the required action JSON. Preserve nested arguments such as `arguments.description` and mention tool descriptions in the budget omission notice.

## Testing

Add regression coverage for oversized descriptions, escaped closing markers, and budget eviction without changing action arguments. Update MCP approval and elicitation tests to verify descriptions appear separately from the action JSON.

GitOrigin-RevId: 30734567c2ebc4b180110276c83d79ddb315ceab
2026-09-10 10:38:10 +00:00
felixxia-oai fcd90d8f07 Enforce complete request budgets for Guardian reviews (#44281)
## Why

Review evidence can fit its own limits while exceeding the reviewer context window once history, tools, output format, and reminders are included. Tool continuations can also push an existing review over budget.

## What changed

- Select evidence after tools and turn context are resolved, reserving request overhead and a 256-token margin. Omit optional evidence while preserving required sections, and reject required evidence that cannot fit.
- Keep incoming evidence pending until budget selection succeeds so cancellation and compaction cannot record unselected input in reusable history.
- Check the assembled prompt before sampling, including retries and tool continuations. Retry budget failures with summary compaction once per model step, and retire reviewers whose budgets remain exhausted. Avoid history-resetting token-budget compaction for this recovery.
- Record section costs after budget selection.

## Testing

Add unit and integration coverage for required-evidence rejection, optional-evidence omission, reminder overhead, cancellation safety, incoming-review and tool-continuation compaction, and reviewer retirement after ineffective compaction.

GitOrigin-RevId: 2f7ef191b3f46032d295ad363eb7debf9d7a9684
2026-09-09 19:21:08 +00:00
jif 2617ed2e1c Move synchronous Guardian orchestration into the reviewer extension (#44252)
## What changed

Move the synchronous review loop, outcome mapping, deadline helpers, and session pool into `codex-guardian-reviewer`. Core supplies adapters for evidence capture, authorization checks, session creation, and event publication through the new host interfaces.

Have `guardian-v2` initialize the reviewer pool and host separately, with the pool managing prewarming, session reuse, concurrent forks, invalidation, and shutdown.

## Testing

Extend the app-server reviewer lifecycle test to cover interrupted concurrent reviews as well as completed reviews, including resuming a reviewer after parent shutdown. Adapt core reviewer tests to use the pool and factory interfaces.

GitOrigin-RevId: 2d69bd20c169b20534764b98cbdb63f38564a530
2026-09-09 17:59:10 +00:00
jif e8e7103cb9 Extract Guardian review policy into a dedicated crate (#44227)
## What changed

Move assessment parsing and schema, model selection, review outcomes, retry handling, and rejection circuit breakers into `codex-guardian-reviewer`. Update core to use the extracted APIs while retaining session execution and decision enforcement in the host.

## Testing

Move existing assessment, circuit breaker, and retry tests into the new crate, including coverage for transient errors, cancellation, and deadlines.

GitOrigin-RevId: 96ec9989a0066acb012bf3bec0b9f7d8bf11a4ef
2026-09-09 17:31:37 +00:00
felixxia-oai 2bba3a29a0 Use explicit histogram buckets for Guardian context metrics (#44181)
## Why

Guardian request and section cost distributions need shared bucket boundaries across synchronous and asynchronous reviewers so their measurements align.

## What changed

- Add `histogram_with_boundaries` to session telemetry and extension metrics while preserving session attribution.
- Use shared request-token buckets up to 2,000,000 tokens and section-cost buckets up to 16,777,216 for both review paths.

## Testing

Extend telemetry tests to verify explicit bucket boundaries and sample counts, and Guardian integration coverage to check exported request and section metric bounds for both synchronous and asynchronous reviews.

GitOrigin-RevId: e34f6972b3418bac3b061939b62eeeccaec5a299
2026-09-09 15:48:39 +00:00
felixxia-oai 17e64839eb Add aggregate budget enforcement for Guardian context (#44166)
## What changed

Add `ComposedContext::enforce_budget` to fit evidence within the input token allowance after reserving existing context. Preserve required content and message boundaries, reserve an omission notice, and return an error if required evidence cannot fit.

Carry retention policies through transcript rendering and composition. Protect user messages, protected messages, and the newest five tool entries. Remove oversized optional items first, then evict commentary, older tool evidence, and images in priority order. Add image admission support and record omitted content in truncation observations.

## Testing

Add tests for existing-context reservations, required-message preservation, image omission, framing costs, and protection of the newest five tool entries in both context profiles.

GitOrigin-RevId: 8a0f245bad82f562323b4533442385ddbcb6e541
2026-09-09 14:55:36 +00:00
felixxia-oai d3ffbbed5a Add Guardian context cost and request token telemetry (#44164)
## What changed

- Record per-section text bytes, estimated text tokens, image bytes, and image counts for synchronous reviews and asynchronous scoring without logging evidence payloads.
- Emit estimated request tokens through `codex.guardian.context.request_tokens`. Synchronous estimates include assembled history, instructions, tool definitions, and output format, and measure the full logical request before WebSocket delta generation. Asynchronous estimates cover the assembled input.
- Add shared context budgeting helpers, including conservative image token reservations independent of encoded payload size and model-aware input limit calculation.

## Testing

Add coverage for separate text and image accounting, image estimates independent of encoded size, and section estimates that bound delivered messages. Extend asynchronous scorer and app-server tests to verify cost metric emission.

GitOrigin-RevId: aed45ecd9c23706f88caa51f2a4f2c77872d3bdb
2026-09-09 14:48:08 +00:00
felixxia-oai 0337192dfd Centralize Guardian transcript policy in context profiles (#43806)
## What changed

Add `ContextProfile` to `codex-guardian-context` and route synchronous and asynchronous Guardian transcript rendering through it. Move default limits, retention, formatting, and async chunked eviction into the shared crate while preserving the distinct retention policies and host-managed full/delta cursors.

Carry transcript truncation observations through `RenderedTranscript` into context composition, and derive async image collection flags from the resolved profile.

## Testing

Add a profile regression test covering distinct sync/async retention priorities, original entry numbering, omission notes, and async truncation observations. Adapt existing transcript tests to the updated input API.

GitOrigin-RevId: 7f7dc249629d8017e9547085f6617f815ae43a1b
2026-09-08 12:17:28 +00:00
felixxia-oai 0034ef93a7 Centralize Guardian context composition (#43805)
## Why

Synchronous approval reviews and asynchronous scoring assemble Guardian evidence separately, duplicating transcript framing and section placement.

## What changed

- Move evidence ordering, framing, and delivery into `codex-guardian-context`, while keeping transcript retention, budgets, and delta cursors with their callers.
- Pass composed messages directly to the asynchronous sampler, preserving roles, annotations, trusted message boundaries, and image placement.
- Reject unsupported synchronous delivery with `SectionError::UnsupportedDelivery` instead of silently dropping content.

## Testing

Add composition coverage for preserving separate messages and rejecting them for synchronous delivery. Extend Guardian integration coverage to check trusted tool and skill messages, mixed text and image evidence ordering, and full and delta review framing.

GitOrigin-RevId: efc06695ee4e2b7400b123b77ef2a21c8bf40646
2026-09-08 12:06:44 +00:00
felixxia-oai d75ed505d7 Move Guardian REPL evidence rendering into the shared context registry (#43602)
## What changed

Move completed REPL response rendering into `codex-guardian-context` and register a synchronous-only `NodeReplEvidence` section. Core supplies a borrowed view of its evidence snapshot and retains responsibility for capture, storage eviction, and the delivery cursor.

Preserve the existing text and multimodal layouts, rendering bounds, omission markers, image ordering, and deduplication. Exclude evidence payloads from the rendered section's debug output.

## Testing

Adapt the existing evidence tests to exercise the shared renderer. Extend registry coverage to verify synchronous-only inclusion, rendered content, and omission of evidence payloads from debug output.

GitOrigin-RevId: eb181870558a7f2f77270dee7c17d36700f3e148
2026-09-07 21:54:56 +00:00
felixxia-oai b4373e53ab Move Guardian image selection into shared context sections (#43601)
## What changed

Move bounded transcript image collection into `codex-guardian-context` and expose it through an optional `TranscriptImages` section. Update the async scorer to consume this section, preserving source filtering, image detail, omission accounting, and the existing limits of four images and 8 MiB of image URL data.

## Testing

Add coverage for source order, tool-output filtering, disabled image collection, oldest-image eviction, oversized images, and omitted-byte accounting.

GitOrigin-RevId: 097799ea1be6d73b347088f959de1fd23f02ec2c
2026-09-07 21:46:23 +00:00
felixxia-oai f5331dc237 Move trusted skill evidence into the Guardian context registry (#43599)
## What changed

Move trusted skill rendering into `codex-guardian-context` as `TrustedSkills` and collect it through an async-only context section. Pass the collected section to the sampler, preserving the separate developer message and bounded JSON rendering. Omit skill paths from `Debug` output, exposing only their count.

## Testing

Move rendering tests alongside the shared type, covering multiple paths and escaped UTF-8 within the size budget. Extend registry tests to cover async-only inclusion, omission for empty input, section ordering, and path redaction in debug output.

GitOrigin-RevId: c0778067f19d31e693ca954aee1a7d7be8f9739a
2026-09-07 21:38:25 +00:00
felixxia-oai 2554239561 Move trusted tool metadata into shared Guardian context (#43597)
## What changed

Represent host-verified tool metadata as `TrustedTool` in `codex-guardian-context` and collect it through an async-only context section. Route Guardian v2 sampling through that section while preserving separate developer-message delivery and the 512-token rendering budget. Omit tool metadata from debug output.

## Testing

Move the rendering-budget test into the shared crate, extend registry coverage for async-only collection and debug redaction, and narrow the scoped approval test to messages tagged `guardian.trusted_tool`.

GitOrigin-RevId: ee7cb4e9b9ceced70438cb58060d1a2973e1dc6a
2026-09-07 21:29:31 +00:00
felixxia-oai 0b9b5ecff3 Centralize bounded Guardian review evidence in guardian-context (#43595)
## What changed

Move prior-review rendering and developer-message construction into `codex_guardian_context`, and route async scorer evidence through a dedicated `PreviousReviews` context section.

Validate a maximum of eight review fragments and a per-fragment byte budget equivalent to 1,000 approximate tokens. Reject oversized evidence without rewriting or dropping records, and fail closed when validation fails. Preserve authorization-version filtering and the existing developer-role delivery. Keep review contents out of debug output.

## Testing

Add coverage for count and UTF-8 byte-size boundaries, preservation of accepted fragments, async-only section delivery, and omission of review contents from debug output.

GitOrigin-RevId: 05e9f53a6405f490b2424b2491a834cd90ab24f0
2026-09-07 21:20:31 +00:00
felixxia-oai 81f23bc186 Move Guardian permission context into the shared section registry (#43538)
## What changed

Represent denied read paths and globs with `PermissionContext` and render them through a sync-only section in `codex-guardian-context`. Keep policy resolution in the host and preserve the existing permission warning in Guardian review prompts.

## Testing

Extend registry tests to verify permission sections appear only for synchronous reviews. Extend the session reuse test to check that resolved denied paths and globs appear in both the initial review request and the next review's delta.

GitOrigin-RevId: a966a14b367f1eedb5a577f7ed74a869559ae646
2026-09-07 18:10:05 +00:00
felixxia-oai 93ac341410 Preserve Guardian context sections and share planned-action rendering (#43534)
## What changed

- Replace `ComposedContext` with ordered `ContextSection` values that retain section identities through transcript selection and rendering.
- Add a shared `PlannedAction` section and renderer for core synchronous approval prompts and the Guardian V2 asynchronous scorer, preserving full, delta, terminal-input, and network-request framing.
- Keep action JSON and retry reasons out of `PlannedAction` debug output.

## Testing

Update registry and transcript tests to assert section identities and source roles. Extend registry coverage to verify planned-action collection for both consumers and omission of action JSON and reasons from debug output.

GitOrigin-RevId: 675b841a2cad06afce03f414615b4d7aa0553b13
2026-09-07 18:02:06 +00:00
jif d70044072c Expose shared Guardian reviewer helpers through guardian_review (#43490)
## What changed

Add a public `codex_core::guardian_review` module exposing `GuardianAssessment`, the assessment parser and output schema, and the review session configuration builder for reuse by the Guardian extension.

Extract assessment handling and reviewer configuration into dedicated modules, preserving the existing parsing defaults, policy prompt, and read-only reviewer settings.

## Testing

Move the existing embedded-JSON, bare allow/deny, and output-schema tests alongside the assessment implementation.

GitOrigin-RevId: cb2aba3ccdb597e2876015e718e923d4f8f36802
2026-09-07 15:01:16 +00:00
felixxia-oai 5b85aea979 Keep Guardian review evidence consistent and reject stale approvals (#43442)
## Why

Concurrent parent compaction can remove evidence between Guardian checkpoint selection and prompt construction. New user instructions or answers received during a review can also invalidate an approval based on earlier authorization.

## What changed

- With `GuardianThreadContext` enabled, use the same parent history snapshot for checkpoint selection and prompt evidence.
- Cancel completed allow decisions if the owning session's user-message revision or root authorization version changed during the review, even without cacheable review evidence.

## Testing

Add regression coverage for evidence preservation during concurrent compaction and aborted approvals after new user instructions or verified answers. Extend reviewer context rollover coverage to exercise `GuardianThreadContext`.

GitOrigin-RevId: 1c2c82197a27c0d3d45b6fe08d200a835d80cc30
2026-09-07 11:51:20 +00:00
felixxia-oai 9c4253ffc1 Retain user instructions in Guardian context (#42844)
## Why

Guardian reviews need the original user instructions even when compaction or
transcript selection removes them from the active conversation window.

## What changed

- Capture complete user messages in host-owned retained context and include them
  as ordered authorization evidence alongside verified answers.
- Omit oversized instructions atomically and emit an explicit incomplete-evidence
  notice instead of exposing a potentially misleading partial permission.
- Mark user messages inherited by forked agents so they remain model-visible
  without becoming child-local authorization.

## Testing

- Cover retained instructions across compaction, resume, rollback, and agent
  forks, including oversized evidence and legacy context behavior.

GitOrigin-RevId: 97122509d818c90d2f87c6a1b29c99c05ad0316e
2026-09-04 20:27:18 +00:00
felixxia-oai 1d74c3ba1e Persist verified user answers in Guardian thread context (#42579)
## Why

Guardian authorization reviews need host-verified `request_user_input` answers to
remain aligned with thread history across compaction, resume, and rollback.

## What changed

- Record verified question-and-answer pairs in retained thread context when
  `GuardianThreadContext` is enabled, and source both local and delegated Guardian
  review evidence from that context.
- Preserve complete answer records within the evidence budget. Mark authorization
  context incomplete when records are unavailable or oversized, and defer cached
  fast approvals in that case.
- Keep the existing runtime-only answer path for threads that do not enable the
  thread-owned context.

## Testing

- Cover answer retention through compaction and resume, removal after rollback,
  source-call eviction, oversized answers, delegated review evidence, and legacy
  behavior.

GitOrigin-RevId: 01f6f4541515434a0d00fdde318ff59f2f8bd2eb
2026-09-03 16:41:11 +00:00
felixxia-oai 6127478086 Centralize Guardian context composition (#42085)
## What changed

- Add `SectionRegistry::compose` to group collected authorization fragments and transcript entries while preserving fragment boundaries and registration order.
- Use the shared composed context in the core Guardian prompt and Guardian V2 synchronous and asynchronous reviewers.
- Keep transcript rendering, retention, and delta handling with each caller while removing duplicated section grouping and authorization rendering.

## Testing

- Cover composition for synchronous and asynchronous targets, including root conversation role handling, trusted user answers, transcript entries, empty context, and contributor failures.
- Verify the synchronous reviewer preserves authorization fragment boundaries in prompt items.

GitOrigin-RevId: 9268343631fd6417bb05a3c0abce883559a37685
2026-09-01 15:54:18 +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 8209978616 Share Guardian user-message retention logic (#42031)
## What changed

- Add a shared `select_user_messages` helper in `guardian-context` that always
  anchors the first user message and fills the remaining token budget from
  newest to oldest.
- Use the helper for both core Guardian prompt rendering and the Guardian V2
  async scorer so they apply the same retention policy.

## Testing

- Cover empty and single-message transcripts, over-budget anchors, and
  newest-first selection across multiple token budgets.

GitOrigin-RevId: 7db246013875263ceb5806d2a6a19b4e2a3db1c4
2026-09-01 10:40:33 +00:00
felixxia-oai 032d15cba7 Use shared transcript collection for Guardian reviews (#41870)
## What changed

- Build Guardian review transcripts with `codex-guardian-context` while preserving Guardian-specific filtering and transcript budgets.
- Apply per-entry truncation during collection, including the larger limit for Node REPL output.
- Retain standalone function and custom tool outputs even when their matching calls are unavailable.

## Testing

- Update Guardian transcript tests to cover shared entry types, Node REPL truncation limits, standalone tool outputs, and recent-tool retention.

GitOrigin-RevId: ec73e80c09fe63c3cf100684cb15574d5d6c2c95
2026-08-31 15:56:32 +00:00
jif 1c1e17782a Preserve Guardian review evidence across compaction (#41846)
## Why

Context compaction can replace original conversation items that Guardian still needs to review approval-sensitive actions.

## What changed

- Retain a bounded, chronological review history independently of the model's compacted history.
- Keep separate limits for user messages and other transcript items so tool traffic cannot evict user instructions.
- Use the retained evidence for synchronous and asynchronous Guardian transcripts, while resetting it after rollback or history reconstruction.

## Testing

- Cover retention through compaction and eviction, reset behavior after rollback, independent size limits, and asynchronous transcript construction.

GitOrigin-RevId: c5851e992820504fc18de8cfec7f6f9f1b69f316
2026-08-31 13:57:59 +00:00
Tamir Duberstein 6478a751fd Organize bundled Rust resources under asset directories (#41477)
## Why

The `core` and `tui` Bazel targets treated nearly their entire crate trees as
compile-time data, mixing embedded runtime resources with source files and test
fixtures.

## What changed

- Move embedded resources for `core` and `tui` into dedicated `assets/`
  directories and update their `include_str!` paths.
- Restrict Bazel compile data to explicit asset and frame directories.
- Move test-only prompts into `tests/fixtures`, resolve the core prompt through
  `find_resource!`, and reuse the integration-test file glob when assembling
  Bazel test data.

GitOrigin-RevId: eee9b0e41df77ea2604215a7f7333a6fe8317bfa
2026-08-29 06:05:28 +00:00
felixxia-oai bae69125df Roll over Guardian context before follow-up reviews (#41215)
## Why

Long-lived Guardian sessions can exhaust the review model's context window. After
a rollover, a transcript delta is insufficient because the new window no longer
contains the original transcript or persistent review instructions.

## What changed

- Compact an exhausted Guardian session before submitting the next review while
  respecting the review model's token-budget configuration.
- Detect whether the full transcript remains in the reviewer history. If it does
  not, reset transcript and Node REPL evidence cursors so the next review replays
  the complete transcript and associated image evidence.
- Restore the follow-up reminder and Node REPL policy when compaction removes
  them.

## Testing

- Cover rollover both when the reviewer window is already exhausted and when a
  follow-up reminder crosses the limit.
- Verify that multimodal reviews preserve the fallback buffer and replay browser
  evidence after rollover.

GitOrigin-RevId: f2a02a84e1f25d15c6e2b8eddb93e2d1805449f5
2026-08-27 22:28:46 +00:00
Dylan Hurd a57b398351 Require approval for input to escalated terminals (#40978)
## What changed

- Add the `write_stdin_approval` feature flag, disabled by default, to require a fresh approval before sending non-empty input to an escalated unified-exec terminal.
- Route these reviews through hooks, Guardian, app-server, and the TUI as `writeStdin` approvals while keeping the parent command item unchanged.
- Preserve the terminal environment, launch directory, and escalation state across turns, and revalidate the process identity after approval before writing input.

## Testing

- Cover approval routing, policy decisions, process reuse, terminal identity preservation, app-server events, analytics, and TUI rendering.

GitOrigin-RevId: 4ba29eaae6208e934737ba078c96e589c7c2164d
2026-08-26 21:39:05 +00:00
joeflorencio-openai a10c8127f7 Add live turn settings updates (#40653)
## What changed

- Add a feature-gated `TurnSettings` operation that can update the model,
  reasoning effort, reasoning summary, and service tier for subsequent steps of
  a named running turn without changing future thread settings.
- Report whether an update was applied, rejected, or lost its live target.
- Revalidate managed constraints and preserve the turn's admitted approval and
  Guardian safety properties before publishing an updated settings snapshot.

## Testing

- Cover sparse updates, step capture and ordering, target replacement, managed
  policy changes, model metadata safety, and separation from future settings.

GitOrigin-RevId: ecaaaa95b4fd4ec9d62265bef551582dddb36d78
2026-08-25 17:44:50 +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
Ankush Gupta 2161ec272a Support cua_repl as a Node REPL-backed MCP server (#40257)
## What changed

- Recognize `cua_repl` alongside `node_repl` when collecting Guardian review evidence, applying computer-use policy, and capturing transcript images.
- Render `cua_repl.js` results with the compact REPL history and transcript views.
- Preserve failed MCP call status and result content when replaying app-server history.

## Testing

- Cover both REPL server names across Guardian policy, evidence, image capture, history rendering, and failed-call replay tests.

GitOrigin-RevId: 8c13a721f718add403b07e727c1634b59203c51a
2026-08-23 17:02:21 +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
jif d12a7f3fd8 Preserve root user authorization in subagent Guardian reviews (#39975)
## 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
2026-08-21 17:12:42 +00:00
rka-oai aead844f64 Handle standalone tool outputs as external context (#39791)
## 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
2026-08-20 19:46:29 +00:00
rka-oai 763787d061 Support standalone named function call outputs (#39782)
## 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
2026-08-20 19:19:41 +00:00
sayan-oai d75c85f651 Separate thread settings from environment configuration (#39597)
## Why

Environment-owned permission profiles and workspace roots must remain effective
without being persisted or restored as thread-owned settings.

## What changed

- Keep thread settings snapshots and restore data separate from the effective
  configuration supplied by the primary environment.
- Use resolved environment permissions, active profiles, workspace roots, and
  network policy when building turn context and session events.
- Preserve environment selections when agents are evicted, resumed, or spawned,
  and validate owner configuration when it becomes ready.
- Start memory initialization only after the primary environment is configured.

## Testing

- Extend remote-environment and multi-agent tests to cover settings persistence,
  owner permission profiles, workspace roots, spawning, eviction, and resume.

GitOrigin-RevId: d82f126981cd0190233020bbdd61d96c70558868
2026-08-20 05:29:12 +00:00
olliem-oai 4a3e829c56 Inject Node REPL policy into Guardian review sessions (#39267)
## What changed

- Add a dedicated developer policy for Guardian reviews of Node REPL browser and computer-use actions.
- Inject the policy once before the first eligible Node REPL approval and retain it when the reviewer session is reused.
- Include Node REPL policy eligibility in the reviewer reuse key so sessions are not reused across incompatible model settings.
- Keep the approval request itself on the generic planned-action format.

## Testing

- Cover first-action injection, reuse after a shell review, repeated Node REPL reviews, and ineligible parent models through the production approval path.

GitOrigin-RevId: bfce190ff8599788aa85e388cc6a959d27fb489c
2026-08-18 19:22:31 +00:00
felixxia-oai 14a6813fa8 Apply Guardian policies to v2 risk classification (#38618)
## What changed

- Include the resolved Guardian security policy in the v2 classifier's developer instructions.
- Prefer an explicit `guardian_policy_config`, then the reviewer model's catalog policy, and finally the bundled policy.
- Limit the combined classifier instructions and policy to 10,000 tokens.

## Testing

- Cover configured and catalog policies, policy truncation, and the bundled-policy fallback.

GitOrigin-RevId: 1b84d70081dd924e214636b4b470d57a63d7cb88
2026-08-14 18:21:22 +00:00
Ankush Gupta 4d9f3021c8 Include node_repl images in Guardian review evidence (#38454)
## What changed

- Preserve valid image content from completed `node_repl` Code Mode responses alongside text evidence for Guardian reviews.
- Enable multimodal evidence for models that require automatic `node_repl` review, or when both `guardian_enhanced_node_repl_transcripts` and `guardian_node_repl_transcript_images` are enabled.
- Bound retained and rendered evidence, deduplicate images already present in review history, and fall back to text-only evidence when the reviewer lacks image support or prompt capacity.
- Keep review evidence out of the parent turn input.

## Testing

- Cover feature gating, mixed text and image ordering, truncation and eviction, unsupported reviewers, prompt pressure, image detail normalization, and reused review sessions.

GitOrigin-RevId: a3e482402f7b3484401b231dadc662e6e3504fdf
2026-08-13 23:29:52 +00:00
Ankush Gupta 93327c852a Gate Node REPL Guardian guidance on model metadata (#38432)
## What changed

- Use the specialized Node REPL approval guidance only when the parent turn's
  model sets `node_repl_auto_review_required`.
- Keep Node REPL JavaScript requests on the generic approval prompt otherwise.
- Cover both prompt paths in the Guardian tests.

GitOrigin-RevId: e3eb89a59cc74fe2b53bb1c3f29a30d450e173de
2026-08-13 20:43:13 +00:00
socrates-oai 507ef0b371 Add Guardian guidance for Node REPL tool calls (#38427)
## Why

JavaScript executed through the Node REPL can invoke connected MCP, browser,
or computer-use tools, so reviewing only the outer tool call does not capture
its immediate effects.

## What changed

- Add dedicated Guardian review guidance for `node_repl` `js` requests that
  evaluates nested calls, authorization boundaries, staged data, and concrete
  side effects.
- Keep the existing generic approval prompt for other MCP tools and shell
  requests.

## Testing

- Cover the dedicated Node REPL prompt, retry reasons, serialized request data,
  and the generic fallback behavior.

GitOrigin-RevId: bf8d49d7e3dc3d36aa41b6b56ff8a0eb02da3e53
2026-08-13 20:11:38 +00:00
Ankush Gupta 053dda6b89 Include Node REPL results in Guardian reviews (#38397)
## What changed

- Capture accepted, successful `node_repl` results from Code Mode and include them as bounded, untrusted evidence in Guardian review prompts.
- Enable the enhanced transcript through `guardian_enhanced_node_repl_transcripts` or when the selected model requires automatic Node REPL review.
- Keep this evidence out of the parent model history, exclude encrypted and failed results, and avoid resending admitted evidence when a Guardian session is reused.
- Clear retained evidence and invalidate review sessions when a thread is rolled back.

## Testing

- Add unit coverage for evidence ordering, escaping, truncation, and empty responses.
- Add an integration test covering feature-disabled, feature-enabled, and model-required behavior, including filtering and review-session reuse.

GitOrigin-RevId: edc3dca5d5b88d472a492f35531ec46889a89d72
2026-08-13 15:50:48 +00:00
rka-oai 99915080b6 Store model history in response item envelopes (#38045)
## What changed

- Add `ResponseItemEnvelope` as the history-owned wrapper for a `ResponseItem`.
- Store envelopes in `ContextManager` while keeping raw-item iteration and consuming APIs for existing callers.
- Preserve envelopes when compaction and history normalization rewrite retained items, and wrap newly synthesized items.

## Testing

- Cover envelope access, mutation, and conversion back to raw response items.
- Cover annotated history replacement and annotated user-message collection for compaction.

GitOrigin-RevId: 3fa2e7bdd981b08cc7aec410c80deccaecdabab4
2026-08-11 17:58:11 +00:00
sayan-oai dd43a9967f Use step environments for Guardian approval reviews (#37618)
## Why

Deferred environments can become ready after a turn starts. Guardian approval
reviews must use the environment selected for the current step so their working
directory and permission context are not taken from the stale turn snapshot.

## What changed

- Carry `StepContext` through tool execution and approval handling.
- Add a Guardian review context that pairs the parent turn with the current
  step's environment snapshot.
- Use that snapshot for Guardian prompts, reviewer sessions, and delegated
  reviewer threads while retaining turn-level fallbacks for callers without a
  step context.

## Testing

Added an integration test that makes a deferred remote environment ready before
an escalated command and verifies Guardian receives that environment's working
directory and denied-read policy instead of the initial local environment's.

GitOrigin-RevId: 779096bc7d0e0c7453a009c02141029714b9b407
2026-08-08 20:07:27 +00:00
sayan-oai bac3ef1d8e Use turn environment permissions for tool execution (#37038)
## What changed

- Derive sandboxing, approval, and permission-profile environment variables from the selected turn environment for shell, unified exec, `apply_patch`, image viewing, and network approval flows.
- Materialize project-root permissions against every workspace root in that environment, allowing `apply_patch` to write to secondary workspace roots while preserving protected paths.
- Carry the selected environment's permission profile into spawned or resumed agents and Guardian permission context.

## Testing

- Add an integration test covering `apply_patch` writes in a secondary workspace root.
- Extend tool, Guardian, and multi-agent tests to distinguish environment permissions from thread-level permissions.

GitOrigin-RevId: e22da8974bcd3de1e724978b1a7790307a51e4bc
2026-08-05 06:38:57 +00:00
Dylan Hurd 2b1357c27c Include policy approval reasons in Guardian reviews (#36939)
## Why

Guardian reviews received sandbox retry reasons, but not the execution-policy reason that triggered an initial approval request.

## What changed

- Propagate approval and retry reasons through the tool approval flow and include the applicable reason in the Guardian prompt.
- Prefer a sandbox retry reason when both are available.
- Truncate approval context to 512 tokens while preserving the start and end of the reason.

## Testing

Added prompt and integration coverage for policy reasons, retry precedence, and truncation.

GitOrigin-RevId: 5b0f4e1c40b792b031bc8f7b31685a61d12b8d92
2026-08-04 18:01:04 +00:00
jif 90314a9207 Read turn permissions from the current configuration (#36930)
## What changed

- Remove the cached permission profile from `TurnContext` and derive the
  effective profile, filesystem policy, network policy, and legacy sandbox
  policy from `config.permissions`.
- Update sandbox consumers and tests to use the current turn configuration.
- Verify that role-based agent spawning reapplies runtime permissions and that
  cold-resumed agents retain their disabled permission profile.

GitOrigin-RevId: d67c5d2bcbe6dc76f15b56defb485b155fd1f138
2026-08-04 17:13:13 +00:00