Commit Graph
265 Commits
Author SHA1 Message Date
Ruslan Nigmatullin 336b19ab26 Move JSON byte counting into codex-utils-string (#61465)
Export `serialized_json_bytes` from `codex-utils-string` and update core callers to use the shared helper. Preserve byte counting without retaining the serialized output.

GitOrigin-RevId: 9c75b05592d3171b9a0d54517f8f3a4281fde987
2026-11-14 05:00:21 +00:00
Hung 5158e9780e Add opt-in file-backed tool audio (#61394)
## What changed

- Add the disabled-by-default `file_backed_tool_audio` feature to prepare fresh function and custom tool audio through the attachment store before persistence. Store returned file IDs in history and preserve them on replay without uploading again.
- Validate and budget mixed text and audio output before uploading retained clips, preserving audio token charges and producer limits. Resolve tool audio durations before continuation compaction checks.
- Keep inline audio on backend upload failures and use text placeholders for invalid or rejected clips. Replace unavailable tool audio files with placeholders in prompts while retaining the original history, so restored files can be used again.
- Treat file-backed tool audio as a competing source when attributing audio or unknown-file errors to user uploads.

## Testing

Add coverage for direct tool and Code Mode audio across follow-up turns, file expiration, restoration, and resume in both legacy and paginated history. Extend tests for truncation before upload, upload failure handling, bounded error placeholders, and error attribution.

GitOrigin-RevId: 66c1e7699c7d92f98c3a2d0f5cecbbd63147eb6c
2026-11-13 23:02:16 +00:00
Vivian Fang e40b0e62e2 Update the harness provenance comment in context history (#61381)
GitOrigin-RevId: f53e93d355bb97737859b0f4c50862bda961872b
2026-11-13 21:51:23 +00:00
goodman-oai 83a23c0936 Support model catalog Apps usage instructions (#61213)
## What changed

- Add the disabled-by-default `catalog_apps_instructions` feature to use the selected model's `model_messages.apps.usage_instructions`, subject to existing Apps availability and model gates.
- Preserve valid catalog text literally, allow an empty string to clear guidance, and fall back to bundled instructions for missing, whitespace-only, or oversized values. Overrides are limited to 800 UTF-8 bytes.
- Reconcile guidance with the latest retained Apps fragment, emitting replacement or removal notices when needed and avoiding repeated updates across resumes. Preserve compatibility with bundled guidance and legacy checkpoints.
- Recognize catalog Apps fragments as contextual developer messages in event mapping and rollout rollback.

## Testing

Add resolver, world-state, client, and scenario coverage for literal text and byte limits, feature and availability gates, model switches, clearing guidance, resumes, and legacy history recovery.

GitOrigin-RevId: 6240d407ada8f22fa101753ff3ddb338632ae283
2026-11-13 00:29:24 +00:00
Ruslan Nigmatullin 9e6d1c5278 Preserve retained assistant revisions across rollout replay (#60839)
## Why

Replaying delivered assistant messages allocates new retained-source revisions instead of preserving those captured in live history.

## What changed

- Capture retained sources when recording assistant deliveries, persist them in rollout metadata, and reuse their revisions during replay.
- Add optional `next_item_ordinal` compaction metadata and a `next_model_id_ordinals` helper to recover item and window allocation positions from checkpoints and subsequent records. Ignore allocation metadata that refers to other threads.

## Testing

Extend tests to compare live and replayed delivery sources, preserve revisions through an older reader's rewrite, and verify ordinal recovery from checkpoints and subsequent local items.

GitOrigin-RevId: cfdeebe8e13b1700b34b9c98696880a90a99ace5
2026-11-11 18:56:47 +00:00
Hung 9018f3a3b0 Preserve file-backed audio in tool outputs (#60735)
## Why

Converting message content to tool output previously reduced file-backed audio to a text file ID, losing its audio type.

## What changed

- Allow `input_audio` tool output items to carry either `audio_url` or `file_id`, preserving both forms during content conversion.
- Account for file-backed tool audio in history estimates and output truncation, and prepare only inline audio data.
- Reject uploaded audio in raw function and custom tool output injection, extending the existing requirement to submit uploaded audio through typed turn input.
- Regenerate app-server schemas and TypeScript/Python types while preserving the Python inline audio class name.

## Testing

Add coverage for tool audio wire-format round trips, repeated file audio sharing text and audio output budgets, and rejection of raw uploaded audio injection across messages and tool outputs. Extend Python SDK coverage for the inline tool audio shape.

GitOrigin-RevId: f4704466e6151662d2c8b8dc3f5ca2ab2f5710a2
2026-11-11 06:41:15 +00:00
Eric Traut 39b42aaacd Handle unfinished tool calls in restored history (#60519)
## Why

Forked and recovered sessions can restore a tool call before its output has been recorded. Missing outputs for these calls should be treated as interrupted work rather than an error.

## What changed

Track the restored history's sequence cutoff so custom tool and local shell calls in that history can receive synthetic `aborted` outputs without reporting a missing-output error. Keep the existing checks for newly recorded calls and clamp the cutoff when history is replaced.

## Testing

Extend the active-turn fork test with a pending `apply_patch` call and verify that the fork's model input includes its `aborted` output.

GitOrigin-RevId: d989fa0991db72f0f8fa8a6d5de403fb307c40c7
2026-11-10 04:11:29 +00:00
ioanbudea-OAI 2889d8e1bf Normalize turn context permissions to a required profile (#59963)
## What changed

Make `TurnContextItem.permission_profile` required and remove its legacy `sandbox_policy` and `file_system_sandbox_policy` fields. Normalize older histories into a permission profile during deserialization, preserving the restrictive fallback for malformed filesystem rules.

Keep legacy compatibility fields in rollout serialization so older readers remain usable. Update turn recording, metadata extraction, and permission checks to consume the profile directly.

## Testing

Add coverage for legacy normalization, profile-only serialization, missing permissions, and malformed filesystem rules. Extend rollout round-trip fixtures to cover deny rules and filesystem roots from another platform.

GitOrigin-RevId: 27515fec60cccbac5806b98fb71f54e311a09701
2026-11-06 19:29:06 +00:00
jif bbf57cec15 Avoid full image decoding for original-detail context estimates (#59908)
Read inline image dimensions with `ImageReader::into_dimensions` instead of
loading the full image into memory. Keep the existing patch-based estimate
and fallback when dimensions cannot be read.

GitOrigin-RevId: fd45eb748ce03134261c867397e1fe46c6023f2d
2026-11-06 13:18:18 +00:00
jif 1e9e8bdb00 Resample Guardian denials with the original request (#59659)
## What changed

When `features.guardian_denial_reconsideration` is enabled, retry the first terminal denial once using the unchanged inference request. Buffer provisional output so only the selected response enters reviewer history, while allowing tool calls to proceed normally.

Preserve the original denial if the retry fails or times out, and continue to honor cancellation. Account for token usage from both attempts, including failed reviews.

Remove `auto_review.experimental_denial_reconsideration_prompt` and warn when it remains in configuration.

## Testing

Update the Guardian scenario and snapshot to verify identical retry inputs, retained tool evidence, and reuse of the selected assessment in a subsequent review, with asynchronous compaction enabled.

GitOrigin-RevId: cc4e303d40a02acc824c8b4b954852325e64ad52
2026-11-05 13:38:12 +00:00
rhan-oai f3074e35c0 Render host-owned V2 subagent rosters in environment context (#59531)
## What changed

- Expose `AgentSessionReader` and `AgentRef` for read-only access to host-owned agent session membership.
- Read direct children when rendering V2 environment context, using the existing roster size limits. Refresh membership when rebuilding context after compaction, and omit the roster if the provider is absent or the read fails.
- Preserve unrecorded initial input across suspension or interruption while waiting for a roster read. Resolve the roster before advancing the compaction window.

## Testing

Add coverage for roster refreshes after compaction, unchanged environments avoiding extra reads, unavailable rosters, recoverable input during blocked reads, and stale compaction results rejected after a concurrent history replacement.

GitOrigin-RevId: 3c7e3822af09e9295d405e53dabb87b020c8f00c
2026-11-05 00:39:39 +00:00
Henry Levy d78f68f8d1 Allow per-turn overrides for direct-only tool namespaces (#59481)
## What changed

- Add `direct_only_tool_namespaces` to `TurnStartOptions`. Omission inherits the thread configuration; an empty list clears direct-only routing for that turn. Steering ignores the option.
- Persist explicit overrides in turn context and restore them when recovering an interrupted turn, including turns started by standalone tool output.
- Keep subsequent turns and spawned or resumed child agents on the thread's routing defaults.

## Testing

Add coverage for routing changes across turns in Responses and Responses Lite, recovery without carrying overrides into subsequent turns, and child-agent inheritance.

GitOrigin-RevId: cea1a997424572d4cd372cd15b098626904c6796
2026-11-04 19:58:49 +00:00
ioanbudea-OAI 4d62359072 Add opt-in elastic budgeting for synchronous Guardian reviews (#59444)
## What changed

- Add `features.guardianv2.sync_elastic_budgeting` and `sync_elastic_sections` to configure section minimums, per-entry token caps, and optional section maximums. Require and validate section limits when enabled.
- Use the selected budgeting policy consistently for history capture, resume, prompt preparation, and review input budgeting. Keep fixed budgeting as the default.

## Testing

Add configuration validation tests, capture-and-resume coverage for both budgeting modes, and a request-history snapshot verifying retained instructions and transcript section limits.

GitOrigin-RevId: c7b0bdfd3e0e008d75c4d8a204959bcc58eb1bf8
2026-11-04 16:33:22 +00:00
jif 48524f337a Preserve pending agent messages across compaction (#59404)
## Why

Agent messages that arrive after a model request starts must survive compaction until a normal inference request includes them. Compactor output alone does not establish that those messages were delivered.

## What changed

- Persist history sequences and sampling checkpoints in harness metadata, capturing the input prefix for each normal inference request.
- Retain pending agent messages within existing size budgets during local and remote compaction, placing the summary before pending messages and subsequent caller input.
- Preserve sampling progress across history replacement and resume, and drain queued mail after interrupted sampling even when compaction leaves context above the token limit.

## Testing

Add regression coverage for request-boundary sampling checkpoints, pending messages across local compaction and repeated remote compaction with resume, and injected messages that interrupt sampling before mid-turn compaction.

GitOrigin-RevId: 26c04ab96b5a1c3cc1d6625ccc74a0f00201d14c
2026-11-04 13:03:01 +00:00
ioanbudea-OAI fc79b7f53a Honor section limits in elastic Guardian context budgeting (#59401)
## Why

Elastic allocation could shrink section minimums under budget pressure and bypass section maxima when the complete request fit. Instruction deduplication also needs to account for which transcript entries actually survive allocation.

## What changed

- Fund demand-capped section minimums before scaling headroom up to section maxima. Reject requests whose minimum reservations cannot fit after duplicate credit.
- Prioritize required records over optional records within each section, reconsider selections when redistributing packing gaps, and reserve global space for protected content and framing.
- Deduplicate retained instructions only against complete selected transcript sources, keep that proof selected during refills, and reuse the released budget.
- Retain elastic transcript history by byte limits instead of fixed item counts. On non-user byte overflow, evict whole older entries toward half the byte allowance.

## Testing

Add regression coverage for section minimums and maxima, priority selection, shared framing, duplicate-proof refills, and elastic history retention and replay.

GitOrigin-RevId: d73b40b88ba1bdab97a619014577f4a297171413
2026-11-04 13:00:58 +00:00
Ruslan Nigmatullin d84c872464 Route response item IDs through a shared generator (#59312)
## What changed

Introduce `ItemIdGenerator` and pass the runtime-owned generator through sessions, model clients, history, hooks, tool outputs, and memory requests. Keep UUIDv7 generation as the default and replace direct calls to `ResponseItemId::new` with the shared callable.

Preserve existing item IDs and backfill missing or empty IDs through the same generator, including forked history before reconstruction and persistence.

## Testing

Extend coverage for UUIDv7 IDs on user messages, agent messages, and additional tools, and assert that forked model context matches the backfilled items persisted to history.

GitOrigin-RevId: 0dd9ccf0f46f0f1005a6acfc740133b03ce82259
2026-11-04 04:02:00 +00:00
Hung 1685f2d621 Use measured durations to budget uploaded audio (#59268)
## Why

Uploaded audio always reserved 12,000 tokens per file, even for short clips. This could discard recordings during compaction that would fit within the retained-message budget.

## What changed

- Add optional audio duration metadata to attachments and cache measured durations per session, resolving metadata for new input and replayed history.
- Use measured durations for history estimates and compaction budgets at 10 tokens per second, rounded up and capped at 12,000 tokens. Keep the full budget for missing or unusable metadata.
- Bound metadata lookups to 250 milliseconds per batch with up to eight concurrent requests, and record accepted input before lookup so interruption preserves the batch.

## Testing

Add coverage for duration rounding and fallback estimates, mixed measured and unknown audio, retention through compaction and cold resume, and interruption during metadata lookup.

GitOrigin-RevId: 26fe320c7024944d9d4c371d5fdcda84442b5cc5
2026-11-04 00:19:12 +00:00
Alex Kira a47a518928 Honor Guardian's additional-context setting for tool outputs (#59239)
## What changed

Tag untrusted additional context rendered as tool outputs with history metadata and apply `guardian_include_additional_context` when building Guardian review history. Preserve this metadata across session resume and apply the current session's setting to restored review checkpoints.

Keep ordinary tool outputs in review history and preserve the acting model's additional-context payload.

## Testing

Add coverage for inclusion and exclusion after restart, restored checkpoints after compaction, and asynchronous Guardian review. Verify both line and JSON transcripts and check that metadata does not alter the acting model's request.

GitOrigin-RevId: 8ba270a6aac16398c88db3282e9d666702f462fe
2026-11-03 21:54:52 +00:00
Hung 3780a5f750 Budget retained file audio during remote compaction (#59237)
## Why

Remote compaction did not charge retained file audio against its token budget, allowing audio-heavy history to exceed the intended limit.

## What changed

- Charge file audio using `FILE_AUDIO_TOKENS_ESTIMATE`, independently of image budgeting.
- Retain the newest media content that fits, keeping images and audio together with their adjacent labels and preserving content annotations. Stop at oversized media instead of backfilling older content.
- Share media-label detection with user-input event mapping.
- Add download timeouts and a Debian mirror fallback for the musl build's `libcap` source archive.

## Testing

Add unit coverage for audio budgeting in both image-budget modes, atomic media labels, and truncation boundaries. Add integration coverage showing that compaction retains the newest five of seven file-audio items within the retained-message budget, including a mixed-media case.

GitOrigin-RevId: 4b7cd69f53966461fdeec14096712671d480187b
2026-11-03 21:53:26 +00:00
Hung f1b0721c88 Enable uploaded audio with durable rejection recovery (#59234)
## Why

Uploaded audio was rejected even for audio-capable models. Invalid or expired recordings also need to be excluded from later requests so users can continue the conversation.

## What changed

- Accept uploaded audio when the selected model supports audio; otherwise, preserve valid sibling input and report the unsupported attachment before inference.
- Track successful audio decoding by model and provider, and attribute provider rejections against the exact submitted request. Preserve attachments when attribution is ambiguous.
- Persist confirmed rejected file IDs across compaction and resume, omitting them from future requests without rewriting retained user history.
- Check audio availability before remote compaction and retry once without an unambiguously rejected recording. Include guidance to remove or replace affected attachments in audio errors.

## Testing

Add unit and scenario coverage for uploaded audio across follow-up turns, compaction, cold and paginated resume, model switching, ambiguous rejections, expired files, and queued or interrupted input.

GitOrigin-RevId: e701901e7b1c0865c9f2a5918197215f549b06c0
2026-11-03 21:24:16 +00:00
rhan-oai 40cf168567 Allow hosts to restrict a thread's root-only capabilities (#58984)
## Why

Hosts need to apply subagent-style capability restrictions without changing a thread's session attribution, agent lineage, or permission profile.

## What changed

Add the public `RestrictRootCapabilities` marker to thread startup extension data. Capture the restriction once at startup and apply it alongside the restrictions native subagents already receive. Hosts must supply the marker again when cold-resuming a restricted thread.

Use the restriction to gate user-input and plugin-install requests, async user messaging, legacy MCP tool approvals, and Full Access MCP form input. Exclude inherited user messages from Guardian's retained authorization context for restricted threads, including after cold resume.

## Testing

Add coverage for host-restricted tool registration, MCP approval rejection, Full Access form-input handling, and inherited instruction authorization after cold resume. Update handler tests to exercise the shared restriction flag.

GitOrigin-RevId: 4d6d7102a20dc747431c7511a74a204c773e8420
2026-11-02 21:09:37 +00:00
Hung 2546265e3e Add audio file references with explicit runtime rejection (#58430)
## What changed

- Allow audio inputs to carry uploaded file references via `fileId` in the app-server API and `file_id` in model content, preserving existing inline audio shapes and Python class names.
- Preserve mixed inline and file-backed audio ordering through history reconstruction and migration, and budget file-backed audio conservatively when duration is unknown.
- Keep uploaded audio disabled at runtime: reject it before inference or raw history injection, preserve accompanying text, and prevent rejected attachments from reaching follow-up or resumed requests.

## Testing

Add coverage for wire-format round trips, SDK compatibility, mixed audio history ordering, and rejection during turn start, steering, task completion, and raw injection.

GitOrigin-RevId: 70895c47a2494c0c9936348ad1250eb1b7b892a6
2026-10-30 20:53:25 +00:00
Dmitriy Kovalenko 7a030b9769 Fix context token estimates for responses without usable usage (#58404)
## Why

Retained output from failed responses or responses without usage data can be omitted from context token estimates, delaying automatic compaction even as reasoning accumulates.

## What changed

Track the history boundary covered by the last usable usage report and estimate all items added after it. Spending-only updates preserve that boundary, while successful usage reports cover the request input even when the response contains no output items. Avoid counting unreported reasoning twice and align recomputed context counts and remote compaction estimates with the boundary.

## Testing

Add unit coverage for initial input, empty responses, spending-only updates, and reasoning accounting. Add regression coverage verifying that accumulated unreported reasoning triggers compaction for incomplete responses with or without usage and completed responses without usage.

GitOrigin-RevId: 80c0b800c7236ff40ae53fdd18320e0d06e026ab
2026-10-30 18:12:39 +00:00
andrewgu-oai c62e9df39f Preserve reasoning token accounting across resume and fork (#57839)
## Why

Resuming or forking a thread loses the session flag indicating that reported token usage already includes retained reasoning. Counting that reasoning again can trigger automatic compaction too early.

## What changed

Persist `last_usage_includes_reasoning` with `TokenUsageInfo` and use it when estimating context usage. Record the flag with completed sampling and compaction usage, and mark local estimates as including reasoning. Older records without the field retain the existing accounting behavior.

## Testing

Add coverage for saved inclusive, exclusive, and legacy token counts. Extend the encrypted-reasoning compaction test to verify the compaction threshold for live, forked, and cold-resumed threads.

GitOrigin-RevId: 9f22262a693111ec0406dce11804eb4844ad8983
2026-10-28 21:22:06 +00:00
Krish Chainani 55503b23d4 Limit concurrent image availability checks to eight (#57835)
Replace `join_all` with a buffered stream in `replace_unresolvable_image_files` so at most eight distinct image file IDs are checked concurrently. Collect unavailable IDs as checks complete before replacing their image content with placeholders.

GitOrigin-RevId: 2a9ed8742fbb6157c99c69a15f212df3ed88810c
2026-10-28 21:08:10 +00:00
goodman-oai e8e6a922c0 Support catalog headings for saved command-prefix notices (#57599)
## What changed

- Add `permissions.approved_command_prefix_saved` to model messages and use valid plain-text headings in saved-prefix notices, with or without full permissions instructions. Reject empty, oversized, control-character, and markup-like headings.
- Keep `Approved command prefix saved:` as the leading marker for older clients and recognize saved-prefix content kinds when trimming rolled-back context.
- Shut down analytics test exporters asynchronously and serialize workspace-switching tests that share authentication environment state.

## Testing

Extend coverage for heading validation, sparse wire values, single-notice emission, rollback with and without metadata, and saved-prefix behavior across model switches.

GitOrigin-RevId: a70175deca64a6339abac2fa69af970f470e3f4f
2026-10-27 23:53:34 +00:00
Krish Chainani cb66dbecf2 Replace unavailable image references in model requests (#57585)
## Why

Image files can become unavailable between turns. Model requests should reflect their current availability while preserving the original references in saved history.

## What changed

- Add opt-in availability checks through `AttachmentStore::is_resolvable`, retaining references by default and on backend or access errors.
- Check distinct file IDs concurrently for each prepared request and replace confirmed-missing images with `[image unavailable]` in messages and tool outputs.
- Apply the same handling to compaction requests, with background image checks running outside the session lock without delaying foreground turn completion.

## Testing

Add unit tests and multi-turn scenarios covering placeholders, preserved history and annotations, concurrent deduplicated checks, backend errors, and nonblocking background compaction.

GitOrigin-RevId: ec98c7014c6e6b5063d87b943d2cd66a9ce7c1e3
2026-10-27 22:59:22 +00:00
ioanbudea-OAI 5c6ef3d1a6 Centralize Guardian root context preparation (#57425)
## Why

Verified answers omitted because of request-size limits must be marked as missing. Legacy checkpoint messages also need to match retained excerpts after truncation so workers receive only one copy.

## What changed

Move root message preparation into `codex-guardian-context::prepare_root_context`, including handoff selection, shared message limits, legacy matching, and incomplete-context markers. Pass original legacy text into preparation and mark verified answers incomplete when oversized records are excluded.

## Testing

Add a handoff regression scenario for an ID-less checkpoint with a retained excerpt and an oversized verified answer. Assert that the worker receives one capped user message, both incomplete-evidence markers, and an incomplete authorization version.

GitOrigin-RevId: bd3f5fe164ae2fddba440e9269fd63b52e8b12f8
2026-10-27 10:29:14 +00:00
Charlie Marsh 1cc7dc8f2a Share additional tool declarations across response item clones (#54945)
## Why

Cloning `ResponseItem::AdditionalTools` deep-copies its tool declarations for history and request copies.

## What changed

Store `tools` in an `Arc<[serde_json::Value]>` so clones share immutable declarations. Update constructors and existing test fixtures to use the shared representation.

GitOrigin-RevId: eb8831774afb45fe8431a796f34ae9d1243e884a
2026-10-17 23:17:51 +00:00
pakrym-oai ce0323f2f2 Support typed content in contextual prompt fragments (#54776)
## 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
2026-10-17 04:02:21 +00:00
rhan-oai cb03f28a40 Refresh model instructions in incremental-tool context windows (#53872)
## Why

Incremental-tool windows recorded base instructions only at setup, so a catalog refresh for the same model could leave active instructions stale. Instruction changes need to reach the model without rewriting the existing history prefix.

## What changed

- Track model identity and instruction content independently, appending a `model_instructions` developer message when either changes. Explicitly mark replacement or removal of base instructions, and avoid repeating unchanged updates.
- Start new windows after compaction with the latest model instructions while preserving custom base instructions.
- Preserve the recorded window base during local compaction retries, including empty instructions, and rebuild current model instructions when reconstructing child context.

## Testing

Add integration tests and request snapshots for catalog refreshes, instruction removal, model switches, compaction, restart, and forks. Extend regression coverage for legacy snapshots and local compaction retries.

GitOrigin-RevId: eeca69067b4d9cd91c817732d79b4d5c1faa18b9
2026-10-14 05:36:44 +00:00
tongzhou wang 1b6e0d8145 Track content provenance through prompt assembly and compaction (#53831)
## Why

Message roles and content classifications do not identify who supplied each context part. Source attribution needs to survive prompt assembly and history truncation, including messages that combine text from different producers.

## What changed

- Add per-part `ContentItemMetadata` for user input, developer instructions, `AGENTS.md`, skills, harness context, command hooks, and tool context with optional callable namespaces.
- Carry attribution alongside content and classifications through contextual fragments, extension prompts, hook output, and agent messages. Keep tool-provided context-window hints separate from harness text.
- Preserve the source of retained text slices during compaction and attribute generated truncation notices to the harness. Default missing historical attribution to unknown.
- Attribute host MCP calls using cached tool namespaces only when catalog revisions remain stable, without starting servers or refreshing catalogs for attribution.
- Send the new metadata to Responses Lite and strip it from requests to endpoints that do not support it.

## Testing

Extend compaction tests to cover mixed sources, missing or extra metadata entries, and discarded media. Update Responses Lite request assertions to verify endpoint-specific attribution handling.

GitOrigin-RevId: b0ebc929b7c60f3ec49d1affbf897312d63edcc8
2026-10-14 02:01:49 +00:00
ioanbudea-OAI 979771476a Account for replay caches in encrypted compaction budgets (#53691)
## 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
2026-10-13 15:05:16 +00:00
pakrym-oai 5ef96ab278 Add opt-in output token replay for OpenAI requests (#52742)
## 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
2026-10-10 01:25:00 +00:00
jif a06545b311 Add opt-in history prefix preservation for full agent forks (#52535)
## 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
2026-10-09 11:08:24 +00:00
Adam Perry @ OpenAI 806d2d1cc1 Remove per-content source attribution metadata (#52329)
## What changed

- Remove `ContentItemMetadata`, provenance and namespace types, and the `content_item_metadata` protocol field while retaining `content_item_kinds` classifications.
- Simplify context fragments, extension contributions, and hook output to carry content without source attribution. Return MCP tool results directly without an attribution wrapper.
- Render context-window hints together in a single text fragment.

## Testing

Update Responses and Responses Lite request assertions to expect content-kind classifications without per-content attribution, and adapt context, hook, and extension tests to the simplified types.

GitOrigin-RevId: 201894cd4c93c6d4b4c41c845e7a9af1dd3c95ca
2026-10-08 23:57:36 +00:00
tongzhou wang ad54b75574 Track source attribution for individual context parts (#52195)
## What changed

- Add `ContentItemMetadata` to distinguish user input, developer instructions, `AGENTS.md`, skills, harness context, and tool-supplied context independently of message role and content classification.
- Keep attribution attached to content through prompt assembly, history processing, and hook-output spilling. Distinguish command hooks from MCP hooks and retain tool namespaces when known.
- Treat missing historical attribution as unknown and send per-content metadata only to Responses Lite.

## Testing

Update request and history assertions to cover attribution persistence and Responses Lite metadata handling.

GitOrigin-RevId: 24e56005f5bdb98bf82f2b3943f29c4c81f3d587
2026-10-08 17:56:54 +00:00
pakrym-oai 28cd3e2501 Remove obsolete fields from persisted turn context (#51492)
## What changed

Remove `workspace_roots`, `current_date`, `timezone`, `network`, `personality`, and `multi_agent_mode` from `TurnContextItem` and stop recording them in rollouts. Update the precomputed protocol schema accordingly.

Make `summary` optional when reading turn context and write `"none"` as a compatibility placeholder for older readers, rather than persisting the active reasoning-summary setting.

## Testing

Update protocol tests to cover omitted `summary` fields and compatibility-placeholder serialization. Extend compaction tests to verify that model requests retain the updated reasoning-summary setting after compaction while the rollout checkpoint records the placeholder.

GitOrigin-RevId: d49eabaa1b340be1d79f7f09115281a307ea8a64
2026-10-06 22:39:06 +00:00
pakrym-oai e79c498b5e Preserve tool declaration mode across resumed context windows (#51480)
## Why

Changing the incremental tools setting on resume should not move tool declarations and base instructions within an existing context window.

## What changed

- Determine the current window's tool declaration mode from its history, and apply the incremental tools setting when building a new window after compaction or a context reset.
- Use that mode consistently for sampling, remote compaction, startup prewarm, prompt debugging, and Guardian input budgeting.
- Preserve tool declarations alongside their comparison baseline in full-history agent forks, including compacted histories; remove them together when rebuilding context.

## Testing

Add request snapshots covering legacy resume, migration after remote compaction or a context reset, disabling incremental tools, and tool updates after restart. Extend agent fork and WebSocket startup prewarm tests to cover retained declarations and incremental tools.

GitOrigin-RevId: 7c46849a1788027a8435082e4d1f7e7f48d2700c
2026-10-06 21:28:13 +00:00
pakrym-oai 062439b2f8 Record base instructions in incremental tool history (#51188)
## What changed

- Record nonempty base instructions as a developer message after tool declarations at the start of each incremental tool context window, and omit the request-level `instructions` field.
- Preserve this prefix when rebuilding history after local or remote compaction, while keeping the compaction summary last.
- Count recorded base instructions once when estimating context usage and trimming tool outputs.

## Testing

Add regression coverage for instruction placement, omission of request-level instructions, compaction ordering, and token accounting. Exercise remote compaction trimming with both request-level and recorded base instructions.

GitOrigin-RevId: 6a6fda92c5c6eb33852a77b38d34c028cf78b164
2026-10-05 22:20:38 +00:00
pakrym-oai 0df76892b1 Support ordered response items in world-state context updates (#50441)
## 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
2026-10-02 22:12:07 +00:00
pakrym-oai 08a7031b6a Return world-state snapshots and context updates together (#49894)
## What changed

Replace the separate `WorldStateSection::snapshot` method with a typed `SectionTransition` returned by `render_diff`. Snapshot updates and model-context fragments are independently optional; omitting a snapshot preserves the stored comparison state.

Update all section implementations to use this contract, skip snapshot updates for unchanged sections where applicable, and cache the multi-agent usage-hint fingerprint.

## Testing

Adapt world-state and context-history tests to the combined return value, including an assertion that unchanged permissions return neither a snapshot nor a context fragment.

GitOrigin-RevId: 79c1977a87481009ff193ea4cce4546be87458fe
2026-10-01 06:11:42 +00:00
pakrym-oai b44ca872b1 Persist world-state snapshots alongside rendered context (#49847)
## What changed

Return snapshots and model-visible fragments together from world-state rendering, and use those snapshots for history baselines, step updates, compaction, and new context windows.

Update `WorldStateSectionContribution` so its renderer returns an optional replacement snapshot and an optional fragment. A missing snapshot preserves stored section state, even when retained history requires a fresh render; a null snapshot removes the section. Derive step patches from the history baseline rather than a separately captured live world state.

Adapt skills and Git attribution contributors to the new contract.

## Testing

Add coverage for skipped snapshot updates, explicit section removal, changed-state patch replay, unchanged-state deduplication, and matching live and persisted checkpoints when starting a new context window.

GitOrigin-RevId: af71606a278358ae937853cb76e8a019a501408e
2026-10-01 03:38:50 +00:00
felixxia-oai 21eb35513d Preserve independent Guardian history across parent compaction (#48779)
## 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
2026-09-27 20:15:23 +00:00
Ankush Gupta 88235f881d Retain confirmed Code Mode messages for Guardian reviews (#48725)
## 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
2026-09-27 15:03:38 +00:00
vkg-oai e72da2b538 Stabilize skill catalogs across executor availability changes (#48353)
## Why

Changes in executor availability can alter cloud skill catalog rendering under a shared budget and repeat an unchanged executor catalog when it reconnects.

## What changed

- Cap cloud skills at three quarters of the metadata budget initially, leaving the remaining budget and unused cloud allowance for filesystem skills. Reduce the cloud cap only when doing so allows all skill entries to fit; description truncation alone does not trigger rebalancing.
- Persist the allocation across disconnects, compaction, and thread resume. Recompute it when the cloud inventory or total budget changes.
- Emit a short availability update when an unchanged selected-environment catalog returns and its full text remains in history. Reinject the full catalog when that text is missing.

## Testing

Add coverage for allocation stability, catalog and budget changes, omission handling, and full catalog reinjection after compaction and resume. Extend the selected capability stack test to verify reconnection avoids repeating the catalog.

GitOrigin-RevId: 839fe98024e00c4082a46e6a40675e15fd2c16c6
2026-09-26 04:19:46 +00:00
felixxia-oai d5cbfe1455 Deduplicate retained instructions against Guardian transcripts (#48109)
## 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
2026-09-25 14:54:51 +00:00
felixxia-oai b35a7afbe8 Preserve recent authorization context for Guardian reviews (#48098)
## 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
2026-09-25 14:06:57 +00:00
felixxia-oai 86be5320b0 Deduplicate retained instructions across Guardian reviews (#48060)
## 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
2026-09-25 10:49:50 +00:00
jif df06bb054c Initialize media estimates outside the global cache lock (#47856)
## Why

Image and audio estimate cache misses computed values while holding the cache mutex, blocking access to unrelated keys during initialization.

## What changed

Add `BlockingLruCache::get_or_init` using `Arc<OnceLock<V>>` and migrate the image and audio estimate caches to it. Initialization and same-key waits run in a Tokio blocking region after releasing the cache mutex. Callers share initialization while an entry remains cached; eviction can allow another initialization, so factories must be deterministic.

## Testing

Add a concurrency test covering access to other keys during initialization, eviction of an in-flight entry, and successful completion of both the evicted initializer and its replacement.

GitOrigin-RevId: 079226a58828b854eddc2b501079527bcbe5e603
2026-09-24 14:58:04 +00:00