Commit Graph
540 Commits
Author SHA1 Message Date
Adam Perry @ OpenAI 1f52d40704 Bind permission grants to the originating turn (#49880)
## Why

Background code-mode cells can continue running after a new turn starts. Looking up grants and strict review requirements through the active turn can make those cells lose their approved access or use another turn's permissions.

## What changed

- Keep turn-scoped grants and strict review state in `TurnContext`, retained by captured steps and background cells.
- Record permission responses against the requesting turn, including requests made after a later turn starts.
- Share live session-scoped grants across turns while keeping turn-scoped grants isolated.
- Use the originating turn's permissions and review requirements for tool execution, patches, MCP calls, file uploads, and terminal input.

## Testing

Add regression coverage for background cells retaining grants across turn changes, turn versus session scope, permission requests after the next turn starts, and strict review enabled while `write_stdin` is queued.

GitOrigin-RevId: 641502fbfd43553c2c42424d4aaa53ae8020c850
2026-10-01 05:33:36 +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
sayan-oai 408f48ce0c Capture host-supplied extension data for each turn (#49846)
## What changed

Add `turn_extension_init` to thread startup and settings, and expose `WithTurnExtensionData` for turn input, recovery, and settings requests. Accept or reject data updates together with the request that supplies them.

Each new turn captures the configured data, including automatically started turns. Accepted steering updates the data for future turns while preserving the running turn's capture. Plain requests preserve the configured data; an empty initializer clears it. The data is not persisted and must be supplied again on cold resume.

Expose `current_turn_extension_data` to retrieve typed data only when the requested turn is still running and has not been cancelled.

## Testing

Add an integration test covering initial capture, accepted and rejected steering, inheritance by an automatic turn, clearing future data without changing the active capture, and rejection of lookups for an earlier turn.

GitOrigin-RevId: dd81eda3e32c7b62d783c772221858f4c0e68264
2026-10-01 03:31:48 +00:00
Owen Lin 8e44ad94b0 Add coordinated shutdown for local agent trees (#49814)
## What changed

Expose `ThreadManager::request_agent_tree_shutdown` and an `AgentTreeShutdown::wait` handle. Requests reject new starts in the shared runtime, cancel pending startup, and signal every session in that tree, including delegates outside the thread registry. Waiting covers admitted operations and session teardown, and reports cleanup or persistence writer failures. Other root trees remain unaffected; registry removal and persistence handoff remain caller responsibilities.

Retain tree membership through startup, cancelled spawn cleanup, residency eviction, and completion watchers. Release persistence resources before returning startup failures, and stop completion watchers during tree shutdown without reporting shutdown as a child result.

## Testing

Add coverage for tree isolation, children removed from the registry, delegate teardown, cancellation during required MCP initialization, aborted teardown, persistence writer failures, and successful retry after a failed resume releases its writer.

GitOrigin-RevId: a109e25adf23d20baf7595891099947050c6b6b5
2026-10-01 01:57:56 +00:00
felixxia-oai de02016798 Persist explicit user goal edits in model history (#49598)
## Why

User goal edits and clears need to reach the model as user instructions, including when a thread is unloaded or history is being compacted. Automatic goal lifecycle updates must not supply user authorization.

## What changed

- Add optional `origin` values of `user` and `automatic` to `thread/goal/set` and `thread/goal/clear`, and mark TUI goal requests as user actions. Omitted origins do not supply user authorization.
- Persist explicit user objective, status, and clear instructions before changing goal state. Recording failures leave goal state unchanged.
- Coordinate goal edits with thread resume, reload stored history under writer ownership, and preserve edits accepted after the compaction input snapshot.
- Update generated protocol schemas and TypeScript and Python bindings.

## Testing

Add coverage for user edits and clears on loaded and unloaded threads, concurrent path-based resume, exclusion of automatic goals from user instructions, preservation of edits during compaction, and recording failures that leave goal state unchanged.

GitOrigin-RevId: d779d7c678a0c32a6ef76975dbe5b1400d1df245
2026-09-30 11:58:03 +00:00
ningyi-oai 87d3e06847 Preserve live tool-call metadata across request windows (#49401)
## Why

Persistence and aggregate metadata budgets can omit tool-call observations that are still relevant to the current request, including tool result metadata needed during compaction.

## What changed

- Retain live direct-call observations separately from bounded persisted outputs, keyed by response item ID so reused call IDs cannot mix results.
- Restore those observations for sampling and compaction, pruning them only when sampling confirms they have left the history window. Preserve observations across shortened compaction retries.
- Remove aggregate prompt metadata caps while retaining per-call argument limits and the outgoing message size limit. Invalidate cell completeness when argument normalization truncates a recorded call.

## Testing

Add regression coverage for exhausted persistence budgets, reused call IDs and reversed completion order, shortened compaction retries, and completeness after argument truncation. Extend integration coverage for large Code Mode metadata, provider metadata support, and fork/resume behavior.

GitOrigin-RevId: 004082b69bff9cc8b05f36d1052057f163c67f07
2026-09-29 23:24:37 +00:00
Celia Chen a44afa527c Track credential storage outcomes and redact sensitive errors (#49384)
## Why

Successful file fallback can hide a secure-store failure. Storage diagnostics need to distinguish that fallback from explicit file storage while keeping credential values out of error output.

## What changed

- Record credential load, save, delete, and fallback-file cleanup outcomes, including secure-store failures and fallback reasons.
- Preserve the initiating client's attribution across app-server requests, login tasks, sessions, turns, and blocking token persistence.
- Preserve storage error context and typed causes while redacting credential JSON data errors and payload-bearing keyring errors.
- Include sanitized fallback completion warnings in the default `codex login` file logs.

## Testing

Add coverage for storage policy and metric attribution, credential redaction, login diagnostics, concurrent app-server clients, device-code login, and token refresh during a turn.

GitOrigin-RevId: 6d5651c1718d184016fa77befa02aa528750b2ae
2026-09-29 22:24:09 +00:00
jif 63475131ce Notify parent agents when Guardian stops a subagent (#49312)
## Why

A subagent stopped by repeated Guardian denials remained interrupted without notifying its parent. The parent needs guidance to report the stop and wait for explicit user confirmation before continuing the blocked work.

## What changed

Forward `TooManyDenials` interruptions from `MultiAgentV2` children to their parents with Guardian-specific guidance to inform the user and avoid resuming or retrying until the user explicitly confirms. Preserve the child's `Interrupted` status and keep ordinary interruptions silent.

## Testing

Add integration coverage for parent notifications under `CircuitBreakAction::Strict`, silence under `CircuitBreakAction::Default`, and preservation of the child's interrupted status. Add a unit test ensuring long denial messages stay within the notification token limit while retaining the guidance.

GitOrigin-RevId: c9bb90827393068d9c92edaf0443a52e6ff57f9f
2026-09-29 16:53:11 +00:00
Nick Steele 0b1b78a4f1 Preserve thread overrides and cloud policy validity during config reloads (#49269)
## Why

Reloads can reject a valid thread configuration when the host configuration references a missing model provider that thread overrides replace. Concurrent cloud refreshes can also leave older policy snapshots or staged cache writes eligible for publication.

## What changed

- Compose preserved session layers before materializing refreshed configuration, retaining source-aware validation errors.
- Carry cloud policy revision bindings through config loading and role overrides. Reject stale MCP refreshes and disable enterprise MCP authorization when the policy is suspended. Retire the old policy when replacing or clearing its loader.
- Stage cloud cache writes in temporary files and publish them atomically only while their policy revision remains current.
- Route file reloads through the shared runtime refresh path, resolve paths relative to each config file, and preserve session settings. Restore host defaults for `secret_auth_storage` and `mcp_oauth_refresh_coordination` when explicit overrides are removed.

## Testing

Add regressions for thread-provided model providers, managed plugin and feature requirement refreshes, stale and suspended policy snapshots, superseded and retired cache writes, and file reloads preserving runtime settings and restoring feature defaults.

GitOrigin-RevId: 3280a2445eeb7074a142f720690bb0e4f01d174b
2026-09-29 13:58:09 +00:00
Nick Steele af0d68a236 Restrict enterprise MCP auth and fail closed on config refresh (#49260)
## Why

Configuration reloads must honor current managed restrictions without granting new enterprise MCP authority to existing sessions or overwriting newer configuration with stale snapshots.

## What changed

- Allow enterprise managed authentication only for configured MCP servers. Remove plugin `ema_auth` overlays from the schema and disable servers that still use them.
- Refresh ordinary MCP settings while preserving session settings and initially admitted enterprise registrations. Disable enterprise MCP when its registration changes, policy revokes access, or configuration loading or resolution fails; require a new session to restore enterprise authority.
- Publish refreshes only when their captured configuration is still current, with bounded retries in app-server reloads. Preserve updated managed restrictions even when a refresh is rejected.
- Finish processing other thread refreshes before `config/mcpServer/reload` reports a rejection, and document that an error can accompany partially applied changes.

## Testing

Add regression coverage for stale refreshes, bounded retries, managed-policy revocation, persistent enterprise disablement, ordinary MCP transport changes, and rejection of plugin enterprise authentication. Verify that denied enterprise servers receive no requests.

GitOrigin-RevId: 34429cd339c1ec152a2ee8697cb29f0cff5d9e5f
2026-09-29 13:08:37 +00:00
andrewgu-oai 458f7046a5 Treat explicit provider model catalogs as authoritative (#49135)
## Why

Providers with `model_catalog_url` could advertise bundled models absent from their catalog or reuse stale models after a failed refresh. Model metadata could also match by prefix or namespaced suffix instead of the provider's exact model ID.

## What changed

- Suppress bundled model fallback and merging for explicit provider catalogs, including when discovery is disabled.
- Require unique, non-empty catalog slugs and match metadata by exact model ID.
- Serialize explicit catalog refreshes and invalidate in-memory and persisted catalog entries after fetch failures, preventing stale cache reuse.
- Report a clear configuration error when startup has no available model, while allowing an explicitly configured `model`.

## Testing

Add regression coverage for exact model matching, hidden catalog entries, cache invalidation across restart and recovery, invalid slugs, and empty-catalog startup with and without an explicit model. Update app-server model-list expectations and verify the TUI configuration error.

GitOrigin-RevId: ac323a94178e3c48df8d4a159a86fdbac20b03a2
2026-09-29 02:33:55 +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
pakrym-oai 1bf73324ca Add opt-in code-mode yielding on new user input (#48135)
## Why

User input queued during a long-running code-mode `exec` or `wait` can remain pending until the call returns. Allow these calls to yield early so Codex can receive the new input while the cell continues running.

## What changed

- Add the `instant_interrupt` feature flag, disabled by default.
- Watch queued user input for each sampling request and pass a shared preemption signal to code-mode `exec` and `wait` calls, including calls received later in the same response.
- Yield running cell IDs without stopping the cells, allowing subsequent `wait` calls to collect their results.

## Testing

Add integration coverage for enabled and disabled behavior, repeated steering during `exec` and `wait`, input deferred during compaction, and later calls in the same response. Verify that direct tool results and queued user messages are preserved, and add a scenario snapshot showing continuation after a cell yields.

GitOrigin-RevId: e33466a980d98e113bfb6c6a45c59de0cc7b3c1d
2026-09-25 16:28:36 +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
sayan-oai e8098eb406 Separate selected plugin identities from MCP contributions (#47939)
## Why

Plugin identity and skill ownership need to be available independently of MCP data, including for plugins without servers or connectors. Hosted plugins have no executor root and should not own executor skills.

## What changed

- Add `McpServerContributor::selected_plugins` to declare executor plugin identities with deferred MCP data. Let the host assign server attribution and precedence across contributors.
- Represent hosted connectors with `HostedPluginConnectors` and make `SelectedPluginIdentity::selected_root_id` optional instead of assigning synthetic roots to hosted plugins.
- Resolve executor contributions before reading hosted connectors, and clear stale hosted state on account changes before executor loading can wait.
- Preserve connector declarations from disabled plugins so Apps can hide those connectors.

## Testing

Add coverage for account changes while executor loading is pending. Update tests to verify hosted identities have no executor root and disabled plugins hide their connectors even when plugin IDs differ from root IDs.

GitOrigin-RevId: e7f511264c5874c1b56d5f3e4a46820724046db0
2026-09-24 21:03:50 +00:00
felixxia-oai e7f119819a Preserve human overrides across repeated heartbeat instructions (#47851)
## 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
2026-09-24 14:48:41 +00:00
vkg-oai d838c2346d Allow idle threads to prewarm and repair WebSocket connections (#47701)
## Why

A previously used WebSocket can close while a thread is idle. Hosts need a way to prepare the connection before the next turn without repeating warmup on a healthy connection.

## What changed

- Expose `CodexThread::prewarm()` using the existing startup warmup and next-turn handoff.
- Reuse healthy connections and pending warmups, retry failed warmups, and repair closed connections. Treat a finished WebSocket reader task as a closed connection.
- Skip scheduling during an active turn or shutdown, and stop warmup admission before shutdown takes the pending task.
- Expose `active_turn_environment_selections()` so hosts can inspect the environments captured by the active turn, including those still starting or failed, when authorizing steering.

## Testing

Add regression coverage for repeated idle prewarm calls, healthy connection reuse with preserved response continuation, reconnection with `generate: false`, and warmup rejection after shutdown. Update delayed MCP discovery coverage to verify prewarm repairs an eagerly closed socket and the first turn reuses it.

GitOrigin-RevId: aa164b91a399de2375eda3bf28397725439fce93
2026-09-24 00:37:56 +00:00
felixxia-oai 7b66c21f58 Remove obsolete Guardian context capture mode branches (#47690)
## What changed

Remove the session-level capture mode and make thread-owned retained context the default throughout history management. Simplify input ordering, compaction identity preservation, fork handling, and approval freshness checks by removing legacy capture branches.

Keep checkpoint compatibility review separate: unknown or incompatible checkpoints can still use legacy review while retaining user evidence. Set the already-removed `guardianv2.thread_context` feature's default to false and remove redundant opt-out test variants.

## Testing

Update checkpoint tests to exercise legacy review through restored checkpoint data. Adapt existing compaction, resume, rollback, and fork tests to use unconditional context capture.

GitOrigin-RevId: ab23a0d7aca71d5a103458cff8b0c68137dd9708
2026-09-23 22:51:13 +00:00
eknight-oai f2654ea6c9 Support model-specific descriptions for agent message board tools (#47670)
## Why

Channel tools use bundled descriptions even when the model catalog supplies tool messages. Let their descriptions follow the active model, including model changes within a turn.

## What changed

- Add channel tool entries to `MultiAgentToolMessages` and pass them to the message board extension for each step.
- Apply catalog descriptions while retaining bundled text for missing or null values and honoring empty strings. Ignore parameter overrides with a warning and keep bundled schemas.
- Clarify bundled descriptions for channel discovery, thread reading, posting, and subscription notifications.

## Testing

Add coverage for full and partial overrides in namespaced and Code Mode tools, preserving Code Mode declarations and parameter schemas. Extend the mid-turn model-change test to verify that the `post` description follows the active model.

GitOrigin-RevId: fca20bed1203d5214713f82cb16a91eeebb38f56
2026-09-23 21:34:11 +00:00
Ben Romano ce7df36237 Overlap startup WebSocket preconnect with tool discovery (#47635)
## Why

Startup prewarm waited for tool discovery before opening its WebSocket connection. Slow MCP startup therefore delayed connection setup even though the handshake could proceed independently.

## What changed

- Run the WebSocket handshake concurrently with tool capture, then send `generate=false` with the discovered tools and finalized request metadata.
- Include model and effective service-tier routing hints in preconnect, including the inherited tier for spawned subagents. Omit routing hints for guardian reviewer requests.
- Recover authentication failures during preconnect and switch to HTTP fallback on `426 Upgrade Required`.

## Testing

Add and extend regression coverage for handshake overlap with delayed MCP discovery, cancellation before warmup, finalized Responses Lite tool metadata, socket reuse and reconnection, service-tier changes and inheritance, and authentication recovery during preconnect.

GitOrigin-RevId: ecb53540765df8227e55e79910a13a05260a4c8d
2026-09-23 19:16:20 +00:00
jif 851d9e95f1 Bind Guardian reviews to the action's target environment (#47630)
## 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
2026-09-23 19:03:12 +00:00
pmccrary-oai fdeaa6c1ec Allow history and notes without experimental context capability (#47625)
## Why

Enabling `features.token_budget.use_history_notes_extension`, explicitly or through model defaults, rejected session startup when the model lacked `supports_experimental_context`, even with experimental mode disabled.

## What changed

Remove that startup restriction so token-budget history and notes can be used independently of the model's experimental context capability.

## Testing

Extend core tests to cover explicit configuration and model defaults on startup and cold resume, checking `new_context`, token-budget contexts, and `history_ingest_requested` metadata. Add app-server coverage for native history and notes tools and context-window hints without experimental context capability.

GitOrigin-RevId: 6157479aeb75ffd785b8d4b2f48b11c6f46e8f35
2026-09-23 18:54:02 +00:00
reia-oai c114f232f9 Refresh realtime context for each model request (#47596)
## Why

A realtime call can start or end while a text turn is still running. Using call activity captured at turn creation leaves subsequent model requests with stale realtime context.

## What changed

Capture realtime activity and mode instructions together in a `StepContext` snapshot for each model request. Use that snapshot to build realtime instructions and persist `realtime_active` in the request's turn context. Store the conversation and mode instructions under one mutex so they are installed together when a call starts.

## Testing

Extend the realtime handoff integration test to start and close a call during one text turn, verifying that successive requests include the appropriate start and end instructions.

GitOrigin-RevId: 4c5c2b47b7c0c06ae28167013218dc2e3cd72c42
2026-09-23 17:34:58 +00:00
felixxia-oai 876ce590e1 Preserve streamed assistant message order in retained context (#47584)
## Why

A user can reply to a streamed question before its completed message is recorded. Assigning the assistant message's order at completion can place the reply before the question in retained context.

## What changed

Reserve an assistant message's order when its source stream starts in thread-owned Guardian context, then reuse that reservation when recording the message. Reserve completion-only messages before publishing their text, and give text and plan output from the same source message a single order.

## Testing

Add streaming regression coverage that accepts a reply before the question completes and verifies question-before-reply ordering both live and after resume. Cover server-provided and generated message IDs, plan output, and text combined with a plan.

GitOrigin-RevId: cb67e7459c804df7beb0841908244202448c655a
2026-09-23 16:16:29 +00:00
felixxia-oai 29aa1df62e Retain assistant context for Guardian authorization reviews (#47582)
## 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
2026-09-23 16:15:55 +00:00
jif 27969c0ae9 Route session agent operations through AgentControl (#47520)
## What changed

- Store `Arc<dyn AgentControl>` in session services and route agent tools, turn admission, usage accounting, completion notifications, and shared settings through the trait.
- Add `AgentControlInit` so delegated and guardian sessions inherit the parent's controller and local runtime.
- Move local registry and context helpers to `LocalAgentRuntime`, keep resume, close, inspection, and status subscriptions on the local controller, and simplify synchronous and infallible trait methods.

GitOrigin-RevId: ad3f213eb483ac524bde56d769f20a78f20989a0
2026-09-23 09:39:30 +00:00
Henry Levy 0a2eb4696c Parallelize instruction refresh and tool preparation (#47458)
## What changed

Run `AGENTS.md` and thread-instruction refresh concurrently with capability discovery, MCP setup, and tool construction. Box both preparation futures to keep them off caller stacks while they are live together.

Finish delivering returned instruction warnings even when tool preparation fails or is cancelled. Preserve instruction-error precedence and publish the selected-plugin snapshot only after both branches succeed and cancellation is checked.

## Testing

Add tests covering concurrent preparation, warning delivery on tool failure and cancellation, instruction-error precedence, and plugin snapshot publication. Verify that cancellation does not repeat consumed warnings on the next capture.

GitOrigin-RevId: 6b2d7b091aae2dba67ad154a19bdb0ed0f96f6f7
2026-09-23 05:22:16 +00:00
sayan-oai 370cd7cba8 Allow environment selection updates during an active turn (#47435)
## What changed

- Add `environments` to `TurnSettingsUpdate` to replace the selection for subsequent steps without changing future turns.
- Capture model settings and environments together so combined updates apply or reject atomically, while existing steps retain their captured selection.
- Validate environment IDs, working directories, configuration ownership, and permission requirements against the running turn's selections.
- Preserve executor startup behavior when switching environments and use the step's local working directory for tool hooks.

## Testing

Add coverage for workspace switching and restoration on the next turn, atomic model and environment updates, permission validation, executor connection waits, and replacement of pending environment selections.

GitOrigin-RevId: 02b5c57a78fd594b2aed84fc09d9e99eb97be28c
2026-09-23 03:12:33 +00:00
sayan-oai 8a3c4ea3b5 Apply inherited environment settings at turn boundaries (#47428)
## Why

Thread permission updates could affect later steps of a running turn. Keep inherited environment settings consistent throughout the turn and apply saved updates when the next turn starts.

## What changed

Store active thread defaults alongside selected environments and refresh them when activating a new turn, even when its environment selections are unchanged. Resolve new selections from those active defaults while preserving owner-provided configuration, selection workspace roots, and the local Windows sandbox type.

## Testing

Add a regression test that changes permissions during a paused turn and verifies that both steps retain restricted filesystem access while the next turn receives unrestricted access. Update session and MCP elicitation tests to check that permission changes take effect on the next turn.

GitOrigin-RevId: 37ee44198fa40c138d140974dc90c4177d8c1f4c
2026-09-23 02:18:40 +00:00
guinness-oai 90e653014a Add opt-in reasoning status for realtime V3 delegations (#47377)
## What changed

Add `backend_reasoning_status`, disabled by default, to relay public reasoning summaries during active, server-managed realtime V3 delegations.

Use the last non-empty summary from each completed reasoning item, truncate it to 256 bytes at a character boundary, and append it as `[STATUS]` context on the `commentary` channel. Discard queued updates if their delegation is no longer active.

## Testing

Add an integration test covering enabled and disabled behavior, selection of the latest non-empty summary, exclusion of raw reasoning and partial summaries, and delivery without triggering `response.create` or changing the model's reasoning-summary setting.

GitOrigin-RevId: 75ce87dbbbab15a6d9d0ad0b23737738837192fd
2026-09-22 22:20:00 +00:00
iceweasel-oai 418199f6ad Add an opt-in preference for the local MXC sandbox (#47375)
## What changed

Add the default-off `features.prefer_mxc` flag to select MXC for local Windows execution when native support is available and effective network settings do not explicitly forbid local binding. Otherwise, retain the configured backend and legacy setup behavior.

Resolve the preference during config loading and use the effective local backend for execution, network proxy setup, sandbox diagnostics, and `windowsSandbox/readiness`. Preserve the configured backend for remote executors and `config/read`. Explicit `windows.sandbox = "mxc"` remains strict, and command failures do not trigger backend fallback.

## Testing

Add coverage for preference resolution, startup readiness without rewriting configuration, local and remote backend selection across environment updates and child threads, and workspace-write permission context in an agent turn.

GitOrigin-RevId: b51178f0371d38ebb58e41af54a068fca67fa316
2026-09-22 21:57:52 +00:00
Owen Lin e7bbc79f48 Resume model context from the latest compaction boundary (#47365)
## Why

Resuming a compacted thread should use the compaction and its newer records without depending on older turn contexts or restoring stale state from before the compaction.

## What changed

- Stop paginated model-context scans at the newest compaction with `replacement_history` and `window_number`. Fall back to full replay if the newest compaction lacks either field.
- Align reconstruction with that boundary and restore `last_started_turn_id`, previous turn settings, and the multi-agent version from compaction resume metadata. Allow newer completed turns to override the saved settings.
- Keep canonical session metadata alongside the bounded replay and use the same scan boundary for subagent rollout migration.

## Testing

Add coverage for cold resume after mid-turn compaction, bounded/full replay equivalence, missing companion state, settings precedence, and incomplete-compaction fallback.

GitOrigin-RevId: f836464bb5cadd9be4a4e05b5d1679223713fc87
2026-09-22 21:17:08 +00:00
Alex Daley 9124e23e2c Move plugin recommendations into developer context (#47349)
## What changed

Emit `<recommended_plugins>` with the `developer` role and append it to the shared developer message after other context. Retain recognition of legacy user-role recommendations in existing rollouts.

## Testing

Add a startup context snapshot and assertions that recommendations share the developer message with permissions instructions. Update core and app-server tests to check developer-role recommendations, including across successive model requests.

GitOrigin-RevId: 0326ff347e86ec38b948f93e39e90e6a88a8730c
2026-09-22 19:51:33 +00:00
Owen Lin 5121aeff13 Overlap subagent spawn persistence and clean up cancelled children (#47348)
## Why

Forked subagent startup serializes history persistence and spawn-edge writes. Cancellation during startup can also leave a provisional child behind.

## What changed

- Defer copied-history durability so it can run alongside spawn-edge persistence, while still awaiting both before submitting the child's initial input.
- Commit agent and residency reservations and notify clients only after initial input is accepted.
- Guard provisional children so cancellation or startup failure removes and shuts them down, discards their stored state, and closes their spawn edge after any pending edge write finishes.

## Testing

Add coverage for cancellation at the child-history durability barrier, checking child termination, removal, and edge closure. Extend fork-context tests in legacy and paginated history modes to verify inherited context can be read from storage after spawn acknowledgment without flushing the live child first.

GitOrigin-RevId: e02dda57efc32b7237bb7af0ce5e5844922890c3
2026-09-22 19:39:18 +00:00
Owen Lin 286d4ecf44 Persist resume metadata in compaction checkpoints (#47323)
## Why

Compaction companion records do not represent all resume state, including the last started turn and previous turn settings. The selected multi-agent runtime also needs a persisted value when no companion turn context is present.

## What changed

- Add optional `resume_metadata` to `CompactedItem`, recording the multi-agent version, last started turn ID, and previous turn settings. Preserve the distinction between absent metadata and explicitly empty values while retaining compatibility with older records.
- Capture this metadata under the settings persistence lock alongside the compaction checkpoint.
- Rewrite inherited metadata to use the child agent's runtime, and clear previous turn settings when the child must rebuild context.
- Clear pending continuation state before delivering acceptance of standalone settings updates, while holding the persistence lock.

## Testing

Add coverage for metadata round trips with empty values, legacy reader compatibility, checkpoint companion records, inherited fork metadata, and continuation invalidation while settings event delivery is blocked.

GitOrigin-RevId: f30b7feaa7e2fdb5b65fc3c53b37fd7a7fe27903
2026-09-22 16:35:30 +00:00
jif acc20df49f Prevent V2 agent eviction from racing with queued messages (#47240)
## Why

Accepted inter-agent messages can still be waiting in the submission queue when a worker appears idle. Evicting that worker can lose the queued work, while cancelling an eviction caller can leave shutdown unfinished and allow delivery to a closing runtime.

## What changed

- Pin V2 recipients until queued submissions are handled or dropped, and verify that delivery targets the registered runtime.
- Move `Submission` into core so it can carry the residency guard through queue handoffs.
- Give eviction exclusive access to each runtime and keep it counted against capacity until removal. Run teardown in a separate task so it completes even if its caller is cancelled, and remove registry and residency entries together.

## Testing

Add a regression test covering queued follow-up delivery, eviction refusal while mail is pending, cancelled eviction completing teardown, rejection of late delivery, and capacity reuse after removal.

GitOrigin-RevId: d6de3d68cb6b911942e9ddfc05c65b3741b796b2
2026-09-22 09:50:48 +00:00
Krish Chainani 4a8cfe6bb7 Pass thread IDs to attachment uploads (#47121)
## Why

Attachment stores need the receiving thread ID to support thread-based placement.

## What changed

Add `thread_id` to `UploadRequest` and pass the session's thread ID through image preparation for messages and tool outputs. Include the ID in debug output while continuing to redact attachment bytes.

## Testing

Extend the image rollout test to assert that a new image uploaded after resume carries the resumed session's thread ID. Update the debug-output test for the new field.

GitOrigin-RevId: f81e3bd761dca704064287177b8ecd7175662f07
2026-09-21 22:58:49 +00:00
peilin-openai ac98537678 Track cumulative MCP attribution across requests and thread history (#47081)
## What changed

- Record each MCP source and its first turn, including connector and plugin identifiers when available. Cover direct and Code Mode calls, including returned errors and results that are not printed.
- Persist cumulative attribution checkpoints and restore them across compaction, resume, and forks, including forks limited to recent turns.
- Attach fresh attribution to Responses API `client_metadata` under `mcp_attribution`, outside model-visible content. Restrict transmission to allowed OpenAI HTTPS destinations for both HTTP and WebSocket requests.
- Report `attribution_error` for legacy history without checkpoints, malformed or conflicting attribution, and serialized attribution exceeding 16 KiB.

## Testing

Add coverage for source deduplication, checkpoint persistence and restoration, recent-turn forks, direct and Code Mode calls, destination filtering, and the serialized size limit.

GitOrigin-RevId: fd285e152ca9e24bfa000debb20013709d67a86e
2026-09-21 18:19:07 +00:00
felixxia-oai f585f13a12 Preserve ordered assistant context in Guardian reviews (#47030)
## 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
2026-09-21 13:30:55 +00:00
vinh-gpt e24ac448b3 Allow thread instruction providers to share updates with subagents (#46577)
## Why

Subagents inherit applied instruction snapshots, so later provider updates do not reach running descendants. Hosts need an explicit way to share current thread instructions across an agent tree.

## What changed

- Add `ThreadInstructionsProvider::share_with_subagents()`, defaulting to `false`. Opted-in descendants refresh instructions at their next model-request boundary.
- Retain shared providers across root unloads and descendant reloads. Let a resumed root replace the provider for surviving descendants, or preserve their last shared instructions when it supplies a non-sharing provider.
- Keep providers scoped to their root tree and preserve snapshot-only instructions for root forks and isolated sessions, including Guardian reviews.

## Testing

Add coverage for child and grandchild instruction updates and clearing, provider replacement after root restart, lazy descendant reloads, default snapshot inheritance, and isolation between roots and Guardian sessions.

GitOrigin-RevId: 44c9fb7fe9ae1f96156bdf3c47ab6c4a81bd65e9
2026-09-19 02:15:20 +00:00
sayan-oai 5685183570 Use captured environment state for permissions and daemon recovery (#46568)
## Why

When a selected environment is unavailable, permission fallbacks can use the turn's initial environment and workspace roots. Daemon recovery can also miss a local environment that finishes starting after the turn begins.

## What changed

- Resolve fallback permissions from thread defaults using the captured selection's workspace root URIs, including when setup is starting or has failed.
- Apply this resolution to Guardian reviews, permission instructions, and sandbox context selection.
- Capture daemon recovery from the latest step inputs and refresh environment readiness before checking recovery eligibility.

## Testing

Add regression tests for Guardian resolving project-root restrictions against an unavailable captured environment's workspace and for daemon recovery including a local environment that finishes starting during a turn.

GitOrigin-RevId: c76d83d786fdc9c444b78ecbc335b02a20805c58
2026-09-19 01:56:57 +00:00
sayan-oai c7828dd010 Keep step settings and approval environments consistent (#46556)
## Why

Background processes can outlive their launching turn. Their approval reviews need the originating environment's filesystem restrictions, even when the active turn uses a different environment. Remote restrictions must also be resolved using executor paths rather than the local filesystem.

## What changed

- Capture settings and environment selections together in `StepInputs`. Preserve the captured selection during active model updates while allowing environment startup to finish.
- Carry environment IDs into Guardian network and `execve` requests, and use the owning process's environment with current review settings. Refresh environment readiness for network requests that fall back to the active turn.
- Resolve remote denied paths and globs using the executor's policy context. Decline automatic approval when the request's environment is unavailable or an explicitly denied temporary directory cannot be resolved.

## Testing

Add regression coverage for active model updates retaining their environment while the next turn adopts a new selection, background approvals using their owning environment, newly ready network environments, and remote denied paths and globs, including Windows paths and missing temporary-directory metadata.

GitOrigin-RevId: b620d4b030c597a4f430fe0500864fa5469db3d4
2026-09-19 01:27:26 +00:00
iceweasel-oai a633ebc124 Always use private desktops for legacy Windows sandboxes (#46554)
## What changed

- Remove the private-desktop opt-out from elevated and unelevated Windows sandbox launches.
- Remove `windows.sandbox_private_desktop` and its managed requirement and API fields. Warn users to remove the obsolete setting.
- Require a private desktop name when launching through the Windows sandbox wrapper and command runner.

## Testing

Add coverage for the obsolete-setting migration warning and update wrapper tests to verify a live private desktop is passed and a missing desktop name is rejected.

GitOrigin-RevId: c7135f8d211aac8d812180691c2c1e433d77cde4
2026-09-19 01:21:32 +00:00
jif 7a15548c6a Move child completion routing into the agent controller (#46552)
## What changed

Have `Session` capture terminal child state in an `AgentTurnOutcome` and delegate delivery to `LocalAgentControl`. Move parent notification and completion activity routing into `agent/control/completion.rs`, so routing no longer takes the child's live `TurnContext`.

Preserve best-effort, queue-only parent delivery, completion activity for the initiating agent, and result tracing only after successful delivery.

GitOrigin-RevId: e06871d92c9ca8660de29f50356cf7c37f7e5199
2026-09-19 01:19:55 +00:00
iceweasel-oai 74af2496c9 Default local binding to true for MXC managed networking (#46523)
## Why

MXC's native host-loopback access is bidirectional, so it cannot enforce `allow_local_binding = false`. Treating an omitted setting as `false` prevents managed networking from working with the default configuration.

## What changed

- Preserve an omitted `allow_local_binding` until the executor's sandbox policy is known. Default to `true` for Windows MXC and `false` elsewhere, including remote execution.
- Reject an effective `false` for MXC managed networking after applying policy restrictions, without enabling disabled networking.
- Use the executor's resolved value for remote network approval decisions while preserving explicit controller restrictions.
- Document that local binding permits local servers and direct host-loopback connections and skips additional private-network destination checks; proxy domain rules still apply.

## Testing

Add coverage for per-executor defaults, explicit values, MXC rejection of `false`, and remote network review cleanup with the resolved policy.

GitOrigin-RevId: 0fa7c1eaec68bebada2b8f7af45688315eea70a9
2026-09-19 00:00:01 +00:00
jif 7fb599de8c Rename AgentControl to LocalAgentControl (#46513)
GitOrigin-RevId: a8e8a5c818a1751c5d2667f5af87e481482fc7db
2026-09-18 23:39:30 +00:00
jif 0f7c8608bd Capture Guardian review checkpoints from live context (#46512)
## Why

Committing a Guardian review snapshot currently flushes and reloads the transcript from storage, even though the completed model context is already available in memory.

## What changed

- Build forkable checkpoints directly from live session state, preserving model history, retained Guardian context, world state, turn context, and token usage.
- Preserve compaction window metadata and MCP resource origins in the checkpoint.
- Recognize prior user turns inside compacted replacement history.

## Testing

Add round-trip checkpoint tests for legacy and thread-owned Guardian context without storage, including preservation after live history is mutated. Extend the review compaction test to assert that checkpoints do not load history from storage.

GitOrigin-RevId: cb2638ca93bba6ab4472de50c7e2913e1a2f25a8
2026-09-18 23:37:35 +00:00
jif 36430b3688 Flush completed Guardian reviews before delivering decisions (#46509)
## Why

The parent could receive a Guardian decision before the review's terminal event was flushed, allowing the reviewed action to start while completion was still being saved.

## What changed

Flush the review transcript and `TurnComplete` event together before delivering completion to the parent. Clear the reviewer's active turn before delivery so the parent can immediately request another review. Skip redundant completion flushes for Guardian reviewers while retaining the existing pre-completion flush on errors and cancellation.

## Testing

Add a gated thread-store regression test covering two consecutive reviews in the same reviewer thread. Verify that each save includes `TurnComplete` and that the approved command cannot start while the save is blocked.

GitOrigin-RevId: 9274b08ddba1ea93d4fb4709bbfdc2ccaea866ab
2026-09-18 23:28:13 +00:00
sayan-oai 7498521d28 Keep MCP policy evaluation consistent with turn environments (#46335)
## Why

Environment settings saved for the next turn must not change MCP tool availability during the active turn.

## What changed

- Use one captured environment snapshot for MCP policy evaluation and runtime publication, including selections that are still starting or have failed.
- Compare both captured selections and ready environment handles when deciding whether to refresh the runtime, and rebuild the startup configuration when resolved selections change.
- Preserve the configuration origin of failed selections so MCP authority evaluation retains whether configuration comes from the thread.
- Box the MCP refresh future to keep it off the sampling request's stack.

## Testing

Add a regression test that updates the environment MCP policy while a turn waits for user input, verifies the tool remains visible when that turn resumes, and verifies it disappears on the next turn. Extend snapshot tests to cover starting and failed selections.

GitOrigin-RevId: cc2744b5137be69f000d70b0c7009b14079f39ae
2026-09-18 00:59:10 +00:00