## Why
Full-transcript fallback can select encrypted agent messages already present in the reviewer's history. These duplicates should not consume the input budget or displace new evidence.
## What changed
- Reconcile native agent messages with the effective reviewer history and earlier messages in the pending input before budgeting, including after compaction.
- Deduplicate messages with nonempty IDs while preserving messages without IDs. Allow delivery again when the prior copy is absent from the current history.
- Reject conflicting authors, recipients, or content under the same ID without modifying the pending context.
## Testing
Add coverage for duplicate admission, conflicting payloads, redelivery after history removal, and retained reviewer requests. Extend the history eviction regression to verify that full-transcript fallback reuses the reviewer and preserves a single encrypted agent message.
GitOrigin-RevId: 97e90f6359a7cf5a75d1c5967fc98cd885afb4e4
## Why
Guardian request counts grouped by input size should reflect admitted sampling inputs without counting synchronous transport retries more than once.
## What changed
- Add a shared `token_bucket` label to admitted total-input observations in `codex.guardian.context.request_tokens` for synchronous and asynchronous reviews. Use inclusive upper bounds in 20k-token steps through 240k, followed by 256k and overflow buckets.
- Record synchronous request tokens after budget admission, once per sampling step, including the complete input. A fresh step counts again even when its input is identical.
- Preserve asynchronous component histograms for oversized classifications, but omit `token_bucket` so rejected inputs do not contribute to counts grouped by that label.
## Testing
Add coverage for bucket boundaries, synchronous retry deduplication and budget rejection, asynchronous retained-history reuse and reset, and admitted versus oversized classification metrics.
GitOrigin-RevId: fac2f102a4c48c5fd21d56bfa701b95a1ff852d0
## Why
Contextual fragments and extension-owned world-state sections currently return strings, limiting them to text content.
## What changed
- Return `ContentItem` from fragment bodies and rendering, preserving classification and attribution. Apply text markers only to `InputText`; pass other content through unchanged.
- Allow extension prompt and world-state contributions to carry typed content, and preserve media in context-window hints.
- Hash non-text world-state fragments by their serialized content and role while retaining the existing text hashing behavior.
- Add shared content-to-tool-output conversion that preserves image references, image detail, and audio.
## Testing
Add coverage for typed world-state hashing and preservation of image content and attribution. Expand code-mode output tests to cover input text, output text, file-backed and inline images, and audio.
GitOrigin-RevId: 2ed00d466a3f169b151202f3eef71e667ea64be1
## What changed
Replace duplicated content-kind strings with constants shared by fragment producers and consumers, including authorization checks, forked-history filtering, and memory instructions. Share the compaction summary kind between core and Guardian's transcript filter.
## Testing
Add a regression test verifying that Guardian excludes compaction summaries while retaining surrounding user messages, including a user message with identical text.
GitOrigin-RevId: 36dbdf0f82c61b52d2571ac1c392fb151c98151b
## Why
Local compaction stores generated summaries with the `user` wire role, but those summaries are not user authorization. Guardian transcripts need to preserve that distinction when reviewing compacted history.
## What changed
Classify messages marked with `compaction.summary` as assistant entries in Guardian transcripts, while preserving the attribution of actual user messages.
## Testing
Add transcript collection coverage and regression tests for both line and JSON transcripts that verify generated summaries are attributed to the assistant and original user instructions remain attributed to the user.
GitOrigin-RevId: 66183028ce70a2500d47d3347b110f95471ba868
## What changed
- Add `ComposedContext::retain_new_images` and use it in Guardian review sessions and the async scorer. Compare host-prepared image references against admitted history, ignoring image detail, and skip scanning history when no images need checking.
- Restore `ConversationState` directly from a checkpoint, retaining committed history for rollback or transferring it with `take_history` for ephemeral forks.
## Testing
Extend tests for prepared-image matching across differing detail levels, image-free continuations that avoid scanning history, and checkpoint restoration and history transfer.
GitOrigin-RevId: bafeb6de9babc721182be43d0eb1a653955a0ebb
## Why
Unsupported evidence currently shares a single Decisions comparison reason, making it difficult to distinguish why a comparison was skipped.
## What changed
Report separate reasons for unsupported developer messages, other non-user messages, trusted skills, images, and audio. Keep these outcomes classified as `skipped` and retain the generic reason for other unsupported evidence. Expose `TrustedSkills::KIND` for identifying trusted skill fragments.
## Testing
Expand adapter tests for the distinct rejection reasons and add a lifecycle case verifying that trusted skill evidence skips the Decisions request, reports `trusted_skills`, and preserves the baseline score.
GitOrigin-RevId: 98f076d1aa6771d0209e7f5910eed0dcc40c6263
## Why
Encrypted compaction payloads can include sampled-token replay caches that increase their size roughly threefold. Budgeting them as ordinary text can reject checkpoints that fit the model context budget.
## What changed
Use a shared estimate of one token per twelve ciphertext bytes, rounded up, for encrypted `Compaction` and `ContextCompaction` items in history and Guardian context budgeting. Guardian continues to count framing and metadata separately.
Check parent compactions against the configured token budget and a separate hard wire-size limit. Preserve the complete encrypted checkpoint when it fits. The token estimate remains a heuristic and can underestimate legacy checkpoints without replay caches.
## Testing
Add coverage for token-estimate rounding, parent compaction token and wire-size limits, and reviewer continuation with a cached compaction preserved intact.
GitOrigin-RevId: cc18ca29802491865167eef992830efcf93d97da
## What changed
- Add the disabled-by-default `output_token_replay` feature to request `output.encrypted_content` from OpenAI providers.
- Preserve encrypted message and tool-call output, message and function-call status, and output-text `annotations` and `logprobs` through history and replay. Update the generated protocol schemas and SDK types.
- Strip output ciphertext from requests to other providers while retaining it in stored history, including after replay is disabled. Remove encrypted content from memory extraction inputs.
## Testing
Add integration coverage for the feature/provider combinations, metadata replay across follow-up requests, and history preservation when resuming with replay disabled.
GitOrigin-RevId: 77729a021b6771933402aebe74ece8eb39c057e2
## What changed
Add `preserve_fork_prefix` to the multi-agent v2 configuration, defaulting to `false`. When enabled, full forks retain the parent's model-visible history, including commentary and tool records, and append the child's instructions. Legacy compaction without a reusable context baseline still rebuilds the child context.
Mark copied history as inherited, complete unfinished tool calls in the fork snapshot, and exclude inherited manual approvals from authorization review while retaining them in model context.
## Testing
Add coverage for preserved request prefixes, child instructions, legacy and paginated histories, Code Mode, missing inherited tool outputs, compaction fallback, and approval isolation across checkpoint replay. Update the noisy helper failure test to read its response before starting cleanup.
GitOrigin-RevId: ce9235dfa111c88d6ce66cb93d564fb63b326489
## What changed
Add the disabled-by-default `guardian_trust_orchestrator_connectors` feature flag. When enabled, Guardian V2 accepts connectors with nonblank IDs from the host-owned plugin service orchestrator as a trusted source without requiring a matching user configuration declaration.
Distinguish orchestrator provenance from user configuration in trusted-tool context. Trust remains scoped to the exact server or connector identity and source; descriptions, outputs, other tools, and other connectors remain untrusted.
## Testing
Add coverage for host-owned connector context, rejection of extension-owned Apps lookalikes and blank connector IDs, source-specific trust metadata, and context token limits. Snapshot the connector approval request history.
GitOrigin-RevId: fbd87dd71110ca7b2b92d50eef02e1e739a74a02
## What changed
Update retained instruction splitting and empty assistant-context detection to match `SectionContent::Other`, and convert assistant framing markers into `SectionContent`.
## Testing
Extend the snapshot-splitting regression test to use `prepare_transcript` with both line and JSON formats, with and without source message IDs. It checks evidence preservation, deduplication, and a stable instruction prefix as assistant context and reviews change.
GitOrigin-RevId: e2205eaf9b31a2b8291dc6c0c22a899596b10206
## Why
Retained assistant context changes as the conversation grows. Keeping it separate from retained user instructions lets independent Guardian snapshots preserve a reusable instruction and transcript prefix.
## What changed
Move retained user instructions before the transcript and place assistant originals and their omission notice in a separate `RETAINED ASSISTANT CONTEXT` section afterward. Preserve evidence, delivery roles, source metadata, and deduplication across both retained sections. Clear omission delivery metadata when a split assistant notice cannot be attributed to a single section.
## Testing
Add regression coverage for prefix stability as assistant context and previous reviews change, evidence preservation, repeated preparation, and deduplication. Update approval-routing and scorer tests to check the new section order and retention of assistant text outside a smaller transcript budget.
GitOrigin-RevId: 15cda59d24bee0106f6a66c067f6a790fa91df69
## What changed
Introduce `TranscriptRecord` and `PreparedTranscript` to retain attributed transcript text through preparation and composition. Cache rendered content, byte counts, and token estimates, refreshing them when text or framing changes.
Keep JSON records structured during budget recovery so truncation replaces only `text`, preserving metadata, field order, and framing without reparsing serialized JSON. Preserve whole-entry truncation for line transcripts and update synchronous and asynchronous consumers to use the preparation API.
## Testing
Expand coverage for cached estimates and delivery with escaped text, Unicode, and large payloads; metadata and framing preservation during recovery; and stable asynchronous history prefixes in both line and JSON formats.
GitOrigin-RevId: 8b3b5231ca4a6ee47d17dcc788ea68c56dd1a9bc
## Why
Transcript content can contain forged role headers, transcript boundaries, or claims of user approval. Guardian needs to distinguish that content from host-assigned authorship metadata.
## What changed
- Add `features.guardianv2.transcript_mode` with `line` (the default) and `json` options for synchronous review and asynchronous scoring.
- Render JSON transcript entries with host-assigned metadata and escaped `text`, and prepend matching provenance instructions to reviewer and classifier prompts.
- Preserve JSON metadata when budget recovery shortens historical text, and keep provenance instructions outside the configured classifier prompt limit.
## Testing
Add coverage for forged roles and approvals remaining within their original entries, JSON metadata surviving budget recovery, and provenance instructions surviving prompt limits. Extend review and scoring integration tests to verify JSON evidence and matching instructions.
GitOrigin-RevId: 81a3580390d3c4d1a664e554976cda5f1adc9a2c
## Why
Partial answers contain answer fragments while a turn is still in progress. Treating them as final answers defers mailbox delivery and can count empty fragments as empty final responses; filtering them from forked histories and guardian transcripts loses answer content.
## What changed
- Allow mailbox preemption after `partial_answer` messages and keep mailbox delivery open.
- Preserve partial answers in forked agent histories and include them as protected assistant entries in guardian transcripts.
- Exclude empty partial answers from empty-final-response accounting for goals.
## Testing
Add regression coverage for partial-answer phases in live events, sampling continuations, follow-up requests, and forked child requests, plus mailbox preemption, goal accounting, and guardian transcript inclusion.
GitOrigin-RevId: 208c8933e211b42b6c2d0bec6952527b8c9c120e
## Why
The Decisions adapter rejected developer messages, including host-annotated trusted-tool context. Requests containing that context could not be classified.
## What changed
Append developer text annotated with `TrustedTool::KIND` to the classification rubric, preserving its authority while keeping user evidence in the input messages. Accept only a single text part with a single matching annotation; continue rejecting other developer messages and unsupported evidence.
## Testing
Add a regression test covering trusted-tool context alongside untrusted user text, plus rejection of incorrect roles, missing or mismatched annotations, multiple annotations or content parts, and image content.
GitOrigin-RevId: 65b6301b2ee1ac54d67275978ccb5e37dcac7c16
## Why
World-state sections could return only a single optional text fragment, with message separation controlled by the fragment itself. They could not contribute explicit `ResponseItem` values alongside instructions.
## What changed
- Let sections return ordered `WorldStateUpdate` values containing fragments or response items, with `Prefix`, `Standalone`, or `Mergeable` placement.
- Preserve explicit item boundaries when grouping context updates, and place prefix updates first when assembling initial context.
- Move standalone-message decisions from fragments to world-state updates and retain snapshot handling.
## Testing
Add tests for mixed fragment/item ordering, role and standalone boundaries, initial-context prefix placement, and snapshot restoration without repeated updates.
GitOrigin-RevId: c2514c3c98538da83f5140c10a9d89f531de6549
## Why
The rolling retained-context window can change as assistant messages are added or evicted. Placing it before the transcript invalidates the reusable history prefix.
## What changed
Move retained user instructions and context after the transcript and action-specific attestations, immediately before the planned action in asynchronous Guardian requests. Keep synchronous section ordering unchanged.
## Testing
Extend the cache-prefix regression test to cover retained assistant message growth and eviction alongside changing attestations. Update app-server and async scorer assertions to verify the new request ordering.
GitOrigin-RevId: 59953d582d4b174ce3e083606626c7532ce0776d
## Why
Guardian review continuations can repeat retained-instruction omission notices even when availability has not changed. When retained context becomes complete, earlier notices also need to be superseded explicitly.
## What changed
Track delivered user-instruction and assistant-context omission states in host metadata scoped to the retained-instructions section. Suppress redundant notice-only sections when the state is unchanged, and emit an availability-change notice when an earlier omission clears, including when no retained originals remain.
Treat missing delivery metadata as unknown. Matching notice text in coalesced tool output cannot establish host delivery, and incomplete history entries discard omission delivery proof.
## Testing
Add coverage for repeated notices, independent histories, availability recovery, and matching notice text in multimodal REPL output. Extend app-server review-continuation tests to assert that each effective classifier context contains the retained user-instruction omission notice exactly once.
GitOrigin-RevId: c6d52dd1dafb284af34b46174aecc7c8cc9b8404
## Why
Retained Guardian classifier conversations already contain previously delivered synchronous reviews. Appending those reviews again repeats evidence and consumes input budget.
## What changed
Track each review completion with a stable identity in host-only history metadata. Filter already delivered reviews before composing and budgeting continuation input, while preserving distinct completions even when their text is identical.
Fresh conversations receive the review evidence again, and shortened messages no longer count as complete delivery. Delivery metadata stays out of model requests.
## Testing
Add coverage for delivering retained reviews once, appending newly completed reviews, restoring evidence after failed streams, and retaining reviews when history lacks complete delivery metadata.
GitOrigin-RevId: b8d58ffdd7d9c484676af707c65fa50020aae73c
## What changed
- Add `features.guardianv2.async_classifier_mode` with `snapshot` and `conversation` modes. Local configuration overrides model defaults; snapshots remain the fallback.
- Retain completed classifier requests and responses in conversation mode, appending transcript deltas without repeating previously delivered user instructions.
- Rebuild retained history when authorization, classifier instructions, or parent compaction changes, after an incomplete response, or when the next request exceeds the configured token limit. Add `async_classifier_conversation_token_limit`, defaulting to 100,000 tokens, while preserving the model's input limit for fresh requests.
- Serialize conversation requests in observation order, bound outstanding work, and reject stale work after authorization changes or thread shutdown. Record retained and new input token costs.
## Testing
Add tests for configuration precedence, history reuse and resets, ordered admission, queue overflow, and stale or stopped requests. Exercise the approval matrix in both modes, including resume, fork, mandatory review, authorization changes, and stream failure after an early score.
GitOrigin-RevId: 35b9800104648cde4f467086e678dee12458c3e5
## What changed
- Add request preparation that preserves committed classifier history and appends transcript deltas, filtering instructions and images already delivered. Account for the complete input budget before transferring history; rejected requests leave the checkpoint and cursor intact.
- Add retained sampling that publishes the early score while collecting completed assistant output and encrypted reasoning for reuse. Preserve item IDs and input metadata, and discard unusable history without retracting an already published score.
- Keep the existing snapshot sampling path and share input preparation and execution with the new retained path. Retained sampling is not yet wired into the active backend.
## Testing
Add coverage for transcript and image deduplication, immutable history prefixes, budget boundaries, rejected preparation, WebSocket output retention, and retention failures that preserve the early score.
GitOrigin-RevId: 0fa33db8b8ea4483eb64d75e207690f1d349123b
## Why
Guardian transcripts omitted encrypted agent messages, leaving reviews without evidence such as encrypted parent replies.
## What changed
- Carry encrypted `agent_message` items through synchronous reviews and asynchronous scoring, preserving their order, author, recipient, and ciphertext.
- Account for native messages in context budgets and omit them whole with an explicit notice when limits require it.
- Support mixed text and native messages during review admission while retaining user-content annotations.
## Testing
Add coverage for transcript ordering and budget eviction in both reviewer profiles, plus a mocked scenario verifying that an encrypted parent reply appears exactly once in each of two incremental Guardian reviews.
GitOrigin-RevId: abc2f1ce5accd83beb06f799ef9ed9d7fa2cd90a
## Why
Guardian reviews need to retain original evidence across parent compaction when checkpoint reuse is disabled, including after resume and rollback.
## What changed
- Make disabling `guardian_reuse_parent_compaction` select a bounded, independent review transcript while preserving thread-owned authorization.
- Keep synchronous reviews and asynchronous scoring independent of parent checkpoints, and allow reviewer sessions to continue using transcript deltas across parent compaction.
- Persist transcript entries with rollback provenance while remaining compatible with older checkpoint readers and metadata-free checkpoints.
- Exclude compaction output from the transcript and synthetic summaries from rollback turn boundaries. Use acceptance ordering to remove rolled-back evidence even when persistence order differs.
## Testing
Add regression coverage for reviewer continuity, evidence retention across compaction and resume, rollback after local and remote compaction, and checkpoint serialization compatibility.
GitOrigin-RevId: fea0852cc5537372854f4a34bacb6bfda6d13e88
## Why
Guardian needs the assistant's delivered question to interpret a user's reply. Messages sent through nested Code Mode tools must remain available as review context, and approvals based on earlier context must become stale when a new delivery is confirmed. Assistant messages do not themselves grant authorization.
## What changed
- Capture successful hosted User Messaging sends after input rewriting and before result callbacks or post-tool hooks.
- Retain bounded delivery text in acceptance order, persist it through cancellation and shutdown, and preserve it across compaction and resume with an older-client-compatible rollout encoding.
- Track assistant review context separately from user authorization, invalidating pending and cached approvals for both local and worker reviews when that context changes.
- Associate retained deliveries with the correct instruction or communication boundary during rollback.
## Testing
Add coverage for nested messaging review context, fast replies and communication ordering, shutdown persistence, stale local and worker approvals, rollout compatibility, and rollback retention.
GitOrigin-RevId: 1bc1ec9e8947560b99f4b39a42ddb9f2099f519b
## Why
Complete assistant messages with empty text can produce misleading omitted-context notices in Guardian reviews.
## What changed
- Skip complete, empty assistant messages when recording and rendering retained context, while preserving omission notices for incomplete messages.
- Separate retained instruction fragments with blank lines and keep blank lines free of role prefixes, while labeling every nonempty line with its original role.
- Update source-order guidance detection and empty-section deduplication for the added spacing.
## Testing
Extend retained-context tests to cover empty completed messages across resume, restored empty messages, incomplete-message notices, blank-line spacing, and role labeling. Update Guardian prompt snapshots.
GitOrigin-RevId: e945717b18689fa4538d9b617b89a644858ab1c8
## Why
Async Guardian requests include retained user instructions even when complete copies are already present in the same request's transcript.
## What changed
- Include retained source metadata and order labels in async user transcript entries.
- Omit retained instructions when the current transcript contains the same complete source revision. Keep ordering guidance in every async sample and never rely on previous samples for deduplication.
- Preserve retained originals when transcript copies are incomplete or lack source metadata.
## Testing
Extend retained-instruction tests to cover async deduplication, preservation of originals without complete source proof, and budget protection. Update async scorer expectations for source order labels and deduplicated instructions.
GitOrigin-RevId: 7029731b296d2342eea4ad74ffa9d3447578d83e
## Why
Synchronous Guardian reviews can include the same user instruction in both the retained instructions section and the conversation transcript. Avoid this duplication while preserving complete authorization evidence and its original acceptance order.
## What changed
- Carry retained source metadata into transcript rendering. Omit a retained instruction when the transcript delivers its complete matching source revision.
- Label qualifying transcript entries with retained source order and protect them from budget truncation. Keep separate retained instructions when source metadata is missing or incomplete.
- Track delivery of source-order guidance so it remains available even when all retained instructions are delivered through the transcript. Invalidate that delivery proof when messages are shortened.
## Testing
Add coverage for complete-source deduplication, missing or incomplete metadata, budget protection, guidance redelivery, and source metadata preservation across resume. Update Guardian review assertions and snapshots to verify instructions are delivered once.
GitOrigin-RevId: 20ad0304501fa49f1596c4cfaf3ef89d64815753
## Why
User messages could fill the root context limit before assistant questions were considered, while commentary could displace relevant conversational context. Guardian needs that context to interpret user authorization when reviewing subagent actions.
## What changed
- Select recent user and assistant evidence together under a shared message limit, discarding assistants with unknown ordering first.
- Exclude explicit assistant commentary and retain message phases across compaction. For older retained records, check available source messages for commentary phases.
- Mark missing instructions or assistant context when the shared limit removes corresponding evidence, and use remaining space for legacy instructions.
## Testing
Update Guardian authorization tests to cover shared message limits, commentary filtering across compaction and resume, and preservation of ordered replies when legacy assistant context has unknown ordering.
GitOrigin-RevId: c369a2760917cba9d7dfca37fb5e85a7ec49012f
## Why
Reused Guardian sessions already contain previously delivered instructions. Follow-up reviews can avoid repeating that evidence, but must deliver it again if it changes or compaction removes it.
## What changed
- Track retained source identity, revision, and completeness in host-owned history metadata, preserving revisions during replay.
- Omit retained instructions from synchronous review inputs only when the same complete revision remains in reviewer history.
- Restore missing evidence after reviewer compaction, enforcing the request budget before persisting additions. Classify restored evidence as context rather than new user authorization.
- Keep source-order labels stable across eviction and retain full delivery for legacy records without reliable provenance.
## Testing
Add coverage for consecutive-review deduplication, changed revisions, missing or incomplete metadata, and restoring compacted evidence exactly once. Extend compaction and resume coverage to verify retained revisions.
GitOrigin-RevId: e90fca16cde0a0970da36dd873af2ca1dbaaf4d3
## Why
Unchanged scheduled heartbeat instructions should not become fresh authorization evidence that supersedes a later human override or floods Guardian review context.
## What changed
- Capture heartbeat provenance per input, including when steering an active turn, and persist it in history and retained context.
- Coalesce unchanged instructions for each automation at their original acceptance order. Preserve changed instruction versions and ordinary human messages.
- Represent repeated heartbeats as references to earlier instructions in Guardian transcripts, while tracking the latest input turn separately for trusted skill selection.
- Limit coalescing to recognized, bounded heartbeat envelopes with host-provided provenance.
## Testing
Add regression coverage for human overrides across repeated and steered heartbeats, instruction changes, checkpoint restoration, transcript references, envelope parsing, and latest-turn trusted skills in delegated reviews.
GitOrigin-RevId: 31e444ddb685a1a618be26d3cc32fb562296946b
## Why
Asynchronous Guardian scoring omitted the target environment's denied-read restrictions, and cached score authorization did not include that permission context. A score from one environment could therefore be reused for an action in another environment with different restrictions.
## What changed
- Share permission resolution between synchronous review and asynchronous scoring, including resolving `write_stdin` targets from the owning terminal session.
- Include environment identity and denied-read paths/globs in async scoring context and cached score authorization.
- Fall back to synchronous review when permission evidence cannot be resolved or exceeds the async limit of 3,000 bytes, rather than omitting restrictions.
## Testing
Add regression coverage for rejecting cached scores with different environments or denied paths, passing the target environment's restrictions to the classifier, and rejecting oversized async permission evidence.
GitOrigin-RevId: 14b6f3454bbea458be037560eb1da47d22874df1
## What changed
- Deprecate and ignore `features.guardianv2.thread_context`, including profile overrides, with a migration notice. Ignore managed requirements for the removed flag with a warning.
- Preserve surviving legacy root instructions across compaction and resume when retained instructions are incomplete. Include unmatched instructions in worker authorization context with an explicit unknown-ordering notice, without replacing newer retained facts or expanding the reviewer's model window.
- Use surviving source calls to roll back verified answers from older checkpoints that lack acceptance-order metadata.
## Testing
Add regression coverage for legacy instruction recovery, compaction and resume, source matching, and answer rollback around same-turn user input. Update configuration and Guardian review tests to verify that the deprecated opt-out still uses thread-owned context and rejects incompatible compaction evidence.
GitOrigin-RevId: a65936c9e0b912e7667d1385a2ec53df726565de
## Why
Actions targeting a secondary environment could be reviewed with the primary environment's denied-read restrictions. Each review needs the permission context of the environment that owns the action, including when a Guardian session is reused across environments.
## What changed
- Carry `environment_id` through environment-scoped approval requests and their planned-action JSON.
- Select and label permission context using the target environment, explicitly stating when it has no denied-read paths or globs.
- Reject reviews for unavailable target environments and reject environment IDs longer than 256 bytes before Guardian inference, while preserving manual approval routing for oversized IDs.
## Testing
Add coverage for `exec_command`, `apply_patch`, and `request_permissions` reviews switching environments while preserving the existing review prefix. Extend tests for environment-specific restrictions, unavailable environments, action serialization, and the 256-byte ID boundary.
GitOrigin-RevId: ad3292edddb1c6f5b180142c6857cba6f1308328
## Why
Ordinary user replies can depend on preceding assistant messages. Losing those messages during compaction can leave Guardian without the context needed to interpret a reply's scope.
## What changed
- Retain original assistant messages, including commentary, with their recorded order across checkpoints and resume. Exclude compaction output from this evidence.
- Include assistant context in local and delegated reviews as untrusted context, with separate storage limits and lower budget priority than user instructions.
- Prefer assistant context preceding ordinary replies when selecting bounded root evidence, and report unavailable context without inferring what a reply refers to.
- Preserve inherited context scope and remove assistant evidence beyond rollback boundaries, including during rollout migration.
## Testing
Extend coverage for compaction and resume, checkpoint migration, queued-input ordering and rollback, role labeling, and evidence budgets that preserve user restrictions while dropping assistant context.
GitOrigin-RevId: 886189b0be29ffc3d324b56f5d0b673d6b8d84da
## Why
Guardian's retained root context appended assistant messages after user evidence and omitted questions sent through User Messaging. This could leave subagent authorization reviews with missing or misplaced conversational context.
## What changed
- Record assistant messages and function calls in the same order sequence as accepted user inputs, then merge retained review context by recorded position while prioritizing user evidence within the message budget.
- Capture bounded User Messaging text after input hooks when a send succeeds, and match it to the original call for assistant context.
- Label assistant context without a recorded position explicitly and clear parent assistant and tool positions when forking agent history.
- Clarify that recorded order does not establish delivery order or pair ordinary replies with questions; assistant messages remain untrusted context.
## Testing
Add coverage for hook-rewritten messages, failed or unconfirmed sends, queued replies with delayed tool results, ordering across resume, and removal of inherited assistant positions from forked history.
GitOrigin-RevId: ac717cf0f5efe4b05c18f2d8370dd48453ac4e35
## Why
Changing previous review decisions or trusted tool and skill evidence should not invalidate the reusable conversation history prefix.
## What changed
Move previous reviews, trusted tool metadata, and trusted skills after the transcript and permission context in Guardian context composition. Keep them before the current action, with history remaining user-role evidence.
## Testing
Add a regression test that varies reviews, tools, skills, and actions while asserting an identical history prefix, both with and without retained context. Update the Guardian v2 integration test to verify the separate transcript and action messages.
GitOrigin-RevId: 8f5b570b17b6715beac526afff7d52139d13a3ad
## Why
A delegated task can omit restrictions from the sender's user messages. Guardian needs that context when reviewing actions in the receiving thread.
## What changed
- When thread-owned Guardian context is enabled, capture up to three recent local user messages when a `send_message_to_thread` delivery from `codex_app` or `codex_tui` is accepted.
- Provide a bounded snapshot to synchronous and asynchronous reviewers, with explicit notices for unavailable evidence and a reminder that historical context does not transfer permission. Keep the snapshot out of the worker prompt.
- Retain snapshots with their deliveries across replay, compaction, and rollback, and strip them from forked agent history.
## Testing
Add coverage for reviewer-only visibility, missing sender provenance, context-budget preservation, checkpoint replay and rollback, fork isolation, and input serialization that rejects harness metadata.
GitOrigin-RevId: 3870ba762aa81689713b56d4cb3b98566bffed88
## Why
File images contributed zero to history and compaction token estimates and were dropped from Guardian context. This undercounted context usage and omitted image evidence from reviews.
## What changed
- Estimate file images using the fixed resized-image cost, or 10,000 patch tokens for `original` detail because file references do not expose dimensions.
- Charge file images against remote compaction and Guardian request budgets.
- Preserve file images from user messages and tool outputs through Guardian selection and delivery, and deduplicate review images by file ID.
- Apply transcript image count and byte limits to both reference types, counting file ID bytes rather than referenced file contents.
## Testing
Add coverage for file-image estimates, selection limits, delivery, and compaction retention. Extend Guardian tests to verify file-image evidence reaches reviews and uploaded original-detail images trigger reviewer-history compaction.
GitOrigin-RevId: ee024204dcbe16e277fea9347c6a4d7fc120e18d
## What changed
- Move approval routing tests into `guardian-reviewer`, covering both reviewer choices across `UnlessTrusted`, `OnRequest`, `Granular`, and `Never` without host session setup.
- Combine reviewer configuration checks into one test for isolation from parent customizations.
- Cover filtering of ordinary developer messages while preserving manual approvals in `guardian-context`, removing the duplicate async scorer test.
GitOrigin-RevId: b39ef21475e2f5d99d758f2d27615449b252dc65
## What changed
Remove standalone tests across Guardian review sessions, configuration, assessment schemas, transcripts, and trusted skill paths.
Strengthen the existing request layout snapshot test to check exact transcript separators, the action header, and parent session ID placement before snapshot whitespace normalization. Concatenate user input text without inserting newlines so the assertions validate the actual request formatting.
GitOrigin-RevId: 06d262a1a85b1a8401796958f300c894146d2e1b
## What changed
Move duplicated action preparation into `codex_guardian_context::action_for_review`, shared by approval rendering and cached-evidence size checks. Preserve the existing behavior: omit top-level `tool_description` and `connector_description` metadata from MCP tool calls while keeping tool arguments intact.
## Testing
Add unit tests verifying that MCP arguments retain description fields, including nested fields, and that other action types remain unchanged.
GitOrigin-RevId: 8e890fb263556596f695177f917750377c04d43c
## What changed
- Accept `fileId` alongside the existing `url` form for app-server image inputs, and forward file references to the Responses API as `file_id`. Update generated schemas and client types.
- Preserve file references, image detail hints, and mixed inline/file image ordering through user-message events, thread history, and rollout migration. Retain file images when truncating tool output.
- Pass file references through image preparation without resolving them, while keeping resize-notice numbering correct. Omit them from unsupported TUI display and Guardian image context.
- Reject image-edit requests whose recent-image window includes a file reference, preventing selection of an older inline image instead.
## Testing
Add coverage for serialization, request and rollout preservation, mixed-image history ordering, incomplete ordering metadata, tool-output truncation, and rejection of unsupported image-edit selections.
GitOrigin-RevId: 6ca20a8577155cc934b720803c3b7b3bffdf972a
## What changed
Remove core guardian tests for transcript numbering and unpaired tool outputs. Extend `guardian-context` coverage to assert that named image-only tool outputs retain their source and use `[non-text output]` for both sync and async contexts, with tool calls included or excluded.
GitOrigin-RevId: ba483e8f6e335697106a987051687a12ad2fde01
## What changed
Represent images in `ContentItem` and `FunctionCallOutputContentItem` with `ImageReference::Inline`, flattened to preserve the existing `image_url` wire format. Update image producers and consumers and regenerate app-server schemas and SDK artifacts.
Preserve the Python SDK's `InputImageContentItem` and `InputImageFunctionCallOutputContentItem` class names during generation.
## Testing
Add a regression test for stable Python image class names and adapt existing image tests to the shared representation.
GitOrigin-RevId: c38a780ac3314c2ac2deb3afc1b93b94b6f93fec
## 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
## Why
Direct tool-call records need to stay associated with the invocation that produced each output, including when call IDs are reused. Completeness must describe the recorded call inventory, independently of tool success.
## What changed
- Attach direct-call records to outputs before they enter history, and set `tool_calls_complete` when the invocation's arguments are fully recorded.
- Bound pending recordings and retained metadata, release reservations on completion or cancellation, and invalidate pending records when capture is disabled.
- Apply request budgets to direct metadata and strip it from inference and compaction inputs when capture is disabled.
- Remove executed-call metadata from app-server raw response notifications and exclude its size from Guardian history retention budgets.
- Track call IDs that bypass dispatch so their reuse cannot incorrectly establish Code Mode completeness.
## Testing
Add regression coverage for direct-call attribution, malformed calls, metadata budgets, cancellation, configuration changes, compaction, notification filtering, and Guardian context isolation.
GitOrigin-RevId: 2ebd39c7f141d04788736491495109841656b4c0
## Why
Transcript limits could shorten user instructions or omit later messages before Guardian evaluated the complete request budget, losing restrictions or prior approvals even when they would fit.
## What changed
- Keep user messages and manual approvals complete and in source order through transcript collection and retention.
- Preserve historical instructions while synchronous review can still compact history. If the final request still cannot fit, discard optional evidence before shortening older historical entries with truncation markers, preserving later restrictions where possible.
- Warn that shortened instructions and approvals are incomplete and that missing evidence does not authorize actions.
- Let asynchronous review defer to synchronous review when complete instructions exceed its budget.
## Testing
Add coverage for complete instruction retention, approval and restriction ordering, marked truncation of oversized Unicode text, compaction before truncation, and asynchronous fallback when instructions do not fit.
GitOrigin-RevId: e6159e876231347986b937ee7e07f944b83bfbe6
## Why
Truncating action arguments can leave approval reviewers evaluating incomplete actions. Large actions need complete review input and explicit handling when they exceed the review budget.
## What changed
- Remove action truncation and the fixed synchronous action byte limit. Admit complete actions against the whole-request budget, splitting long text losslessly into bounded transport parts and accounting for their framing.
- Route actions exceeding the asynchronous action budget to synchronous review. Prevent cached scores from covering oversized calls, including expanded approval arguments, while allowing later small actions to recover score reuse.
- Request user approval when optional review exhausts its local input budget. Keep required review and compaction service failures closed to approval, and retire exhausted review sessions.
## Testing
Add coverage for complete large-action delivery, optional user fallback, required-review denial, subsequent review recovery, async overflow through MCP approval routing, and lossless text splitting with budget accounting.
GitOrigin-RevId: 08f06e68a94b2f779480dd6b6cf5bc30b241f6dd
jif