Commit Graph
268 Commits
Author SHA1 Message Date
Zahan Malkani ad5dd4a110 Persist realtime handoffs before accepting turn input (#62288)
## Why

Realtime history needs to record which transcript frames reached the
coding agent and which turn received them. If a handoff fails or the
conversation reconnects, Codex needs to distinguish speech that was
already used from speech that is still available for the next input.

## What changed

- Route handoffs through normal turn admission to start or steer a turn.
  When realtime frame history is enabled, persist the transcript frames
  and the receiving turn's input before publishing acceptance or exposing
  the input to the model. Attribute only transcript frames represented
  in the rendered input.
- Reserve transcript frames across overlapping handoffs and release them
  on definite rejection. Retry an uncertain write with the same input
  and require history recovery if the write still cannot be confirmed.
- On V3 reconnect, restore persisted speech that has not been consumed
  and put it before newly captured speech.

## Testing

Add regression tests for attribution after transcript truncation,
overlapping handoffs, V3 reconnect ordering, and preservation of typed
input when a handoff in the same turn fails before model sampling.

GitOrigin-RevId: 432dd51971f7ced20344ebc7f0c7b007bb849f54
2026-11-18 14:26:53 +00:00
ioanbudea-OAI a05444bbeb Refresh managed Guardian policy independently of MCP (#62276)
## Why

Managed approval requirements can change while an MCP runtime is still in use.
Later turns need to honor those changes without restarting MCP, while in-flight
MCP actions need to retain the mandatory review requirements of their issuing
turn.

## What changed

- Add `CodexThread::refresh_guardian_policy` to apply managed policy updates to
  an existing thread. MCP configuration refreshes preserve the active Guardian
  policy, including when a policy update lands while MCP configuration is loading.
- Resolve MCP tool and elicitation reviewers against the issuing turn's policy.
  Live requirements can still require Guardian review, and a live policy change
  cannot waive mandatory review captured by the turn.

## Testing

Extend coverage for overlapping Guardian and MCP refreshes and for an existing
MCP runtime that must prompt the user on the next turn after a policy update
allows only the `user` reviewer.

GitOrigin-RevId: 5a1649f917313f6fb152a206bc89986fe6345481
2026-11-18 13:08:51 +00:00
jif 58627e6fea Fix turn interrupt reply ordering and shutdown handling (#62066)
## Why

A delayed terminal event from an earlier turn could acknowledge interrupts for a later turn. Interrupts could also remain pending after the thread listener shut down.

## What changed

- Associate pending `turn/interrupt` requests with their target turn and reply only when that turn completes or aborts.
- Track the current Core turn through stop hooks and resolve review delegate IDs to the owning turn when validating interrupts.
- Settle pending interrupts with an error on listener shutdown and reject further interrupts against the retired runtime.

## Testing

Add regression tests for delayed terminal events, queued turn starts, review delegates, and interrupt handling during thread revert and runtime replacement.

GitOrigin-RevId: 9695fc3f718baa6b792a14d54136bb87fc003100
2026-11-17 15:33:21 +00:00
jif cff061fca9 Add opt-in item-boundary interrupts for subagents (#61909)
## What changed

Add the multi-agent v2 setting `interrupt_at_item_boundary`, disabled by default. When enabled, `interrupt_agent` lets a subagent finish and retain its current output item while cancelling tools immediately. Later output items are discarded, and retained tool calls that have not run receive an aborted result.

Wait for the interrupted turn's terminal event before starting follow-up input, and prevent idle contributors from restarting the interrupted child.

## Testing

Add streaming tests for retaining an assistant message or tool call, discarding later output, and preserving history and follow-up input when the child resumes. Update configuration and interrupt-routing tests.

GitOrigin-RevId: 8b2b9b1383bfb4eefea7369a40012bebc5fff07b
2026-11-16 23:26:07 +00:00
tongzhou wang ce82e38b16 Gate new rollout work on shared budget exhaustion (#61582)
## Why

A budget can be exhausted while another thread waits to start work. New requests, tools, and automatic turns must observe that stop without discarding an already-consumed response whose cost exhausted the budget.

## What changed

- Share a work gate across the rollout tree while preserving native and host accounting ownership.
- Recheck the budget before HTTP and WebSocket requests, hook execution, tool dispatch, and compaction. Prevent automatic wakeups and late input from restarting stopped work, leaving pending input queued.
- Record response costs without rejecting the consumed response solely because its bill exhausts the budget, and stop new work when enforced accounting overflows.
- Expose a thread-registration receipt that remains committed after unload.

## Testing

Extend rollout-budget coverage to verify that accepted response text remains in history after exhaustion and that waiting tools observe the shared stop while completed tools retain their results.

GitOrigin-RevId: e583d657c6f29e374d374cdb014bf23fd6c28614
2026-11-15 00:03:23 +00:00
Dylan Hurd 347143f140 Preserve active additional context across compaction (#61390)
## Why

Compaction can drop active client context while deduplication prevents unchanged values from being sent again on subsequent turns.

## What changed

Track additional context by history item ID and restore active items when installing complete or prefix compaction results. Preserve their history order, refresh retained items with their original bounded payloads, and leave queued updates to arrive with their input.

## Testing

Add regression coverage for local compaction with message context and remote compaction with tool-output context. Verify that current context survives exactly once, stale values are absent, and restored context precedes the preserved user prompt.

GitOrigin-RevId: a3114f387c6a4ae1a2d1651993d0f7a101be7dd5
2026-11-13 22:19:12 +00:00
Dmitriy Kovalenko 76dcb8199b Honor permission grant scopes in executor capability discovery (#60461)
## Why

Approved filesystem access should make executor skills discoverable and readable for the lifetime of the grant. Turn-scoped access must not authorize persistent plugin servers or hooks, or bypass disabled-plugin policy.

## What changed

- Apply environment-specific session grants to capability discovery and refresh MCP after session grants change.
- Use a separate discovery inventory for skills with turn-scoped grants, keeping plugin servers and hooks on the session inventory. Reuse discovery results across unchanged model steps.
- Preserve disabled-plugin checks for unreadable manifests and symlinked manifest files or directories, including ancestor markers, while retaining standalone skill roots.

## Testing

Add coverage for V1 and V2 skill discovery before approval, after approval, and on the next turn; disabled-plugin ownership checks; discovery cache reuse; and session hooks surviving turn-scoped discovery failure.

GitOrigin-RevId: 929a7f8cb01f4bcfb4bee2c50ae02cd762066b2d
2026-11-09 22:20:07 +00:00
Dylan Hurd 8703d32bd0 Treat goal objectives as untrusted context (#60342)
## What changed

- Pass goal objectives as untrusted `goal_objective` context instead of embedding them in continuation instructions. Direct the model to use `get_goal` if the objective is missing or truncated.
- Inject newly created objectives into their originating active turn without dropping unrelated client context. Render them as tool outputs when `UntrustedAsToolOutput` is enabled.
- Queue additional context together with automatic turn input so both reach the first model request.

## Testing

Extend app-server coverage to verify that model-written objectives appear as tool outputs at creation and continuation, without duplicating them in user messages. Update steering and Python SDK tests to check objective preservation and delivery alongside continuation instructions.

GitOrigin-RevId: 6f44f6d6c895bb4dc378c9cf84d6470ccc885c67
2026-11-09 03:58:05 +00:00
rhan-oai f5ac7a757e Support asynchronous host service-tier reads and acceptance (#60075)
## What changed

Make `AgentServiceTier` reads asynchronous and have updates return a receipt for local acceptance. Await acceptance during root startup and settings updates outside state and registry locks, sharing the receipt with explicit same-tier retries.

Await shared-tier reads during child preparation, startup prewarm, and step capture. Allow cancellation while capturing a step, and keep host propagation independent of the acceptance wait.

## Testing

Extend the host-child WebSocket test to cover delayed acceptance, same-tier retries without duplicate updates, accessible settings while acceptance is pending, and tier inheritance during prewarm and subsequent steps.

GitOrigin-RevId: 0e2e1899834c37684e904c4a36e66344a0c6c7f3
2026-11-07 06:51:10 +00:00
rhan-oai 94ca55a93c Add shared service-tier support for host-owned agents (#59770)
## What changed

- Expose an optional `AgentServiceTier` provider so host-owned children, including detached children, inherit the root's committed tier during spawn preparation, startup prewarm, and each inference step.
- Reject per-turn tier overrides for children using the shared provider. Root updates apply to subsequent requests without changing requests already in flight.
- Preserve independent configured tiers and role overrides when no provider is installed, including across eviction and reload. Spawned children inherit the parent's thread preference rather than its per-turn override.
- Publish the latest committed root tier only after successful runtime registration, preventing failed duplicate starts from changing shared settings.

## Testing

Add coverage for detached-child WebSocket prewarm and between-step tier changes, shared-child override rejection, independent-tier inheritance and reload, and settings updates before registration.

GitOrigin-RevId: f0f3676813a1c59bd77baf95121e2d13e0163a8e
2026-11-06 00:24:27 +00:00
rhan-oai f69c660a83 Add host reporting for terminal agent turn outcomes (#59725)
## What changed

Expose `AgentTurnOutcomeReporter` for hosts to install through `StartThreadOptions::thread_extension_init`. Await its callback after emitting the terminal turn event, independently of native parent notifications, so hosts can handle result delivery and retries.

Track outstanding outcome handoffs and wait for them during graceful shutdown before running thread-stop hooks. Prevent new turns from starting once handoff registration closes.

## Testing

Add integration coverage for completed and errored outcomes, turn lineage, queued follow-up turns, and shutdown waiting for handoffs. Extend Guardian tests to verify that strict safety stops reach the reporter while default interruptions remain silent.

GitOrigin-RevId: 0d825f6208f0250969734488881f32fd1dcd7509
2026-11-05 19:20:56 +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
Ruslan Nigmatullin 71a4c07036 Share context window ID allocation through the agent runtime (#59354)
## What changed

Store a shared `ContextWindowIdGenerator` in `LocalAgentRuntime` and pass it into each session. Use it for both initial context window IDs and subsequent window advances, retaining UUID v7 allocation by default.

Make `AutoCompactWindow` accept caller-provided IDs instead of allocating them directly.

## Testing

Update the window boundary test to verify that advancing installs the supplied ID while preserving the first and previous window IDs.

GitOrigin-RevId: a9da4232fe6ab9a53b75e239cb412ceae0fe3216
2026-11-04 08:13:41 +00:00
Ruslan Nigmatullin 19438adcb6 Introduce a dedicated context window ID type (#59328)
## What changed

Add `ContextWindowId` as a string-backed protocol type with UUIDv7 generation and parsing. Use it throughout compaction state, rollout reconstruction, token budget context, and request metadata in place of raw UUIDs, preserving the serialized string format and UUIDv7 validation.

GitOrigin-RevId: 786b74fcc331707ca5d6765f429b6e5e322d9328
2026-11-04 06:05:16 +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
cooper-oai 079e17a422 Wire shared runtime authentication through threads and API clients (#59177)
## What changed

- Share a lazily registered runtime across threads, model discovery, and extension clients when `runtime_agent_instances` is enabled, while honoring each thread's authentication mode and credential owner.
- Bind model catalogs to the effective provider and authentication context while preserving cache policy and live API-key discovery settings.
- Resolve credentials, routing, and transport policy together for each request. Preserve proxy and TLS settings, validate runtime account selectors, and prevent pooled connections from crossing runtime identities.
- Accept supported ChatGPT subdomains when selecting the registration environment.

## Testing

Add integration coverage for shared runtime registration, dual-header and assertion-only modes, disabled and third-party providers, credential replacement, subagents, and delegated review. Add coverage for catalog cache policy and runtime-aware connection reuse.

GitOrigin-RevId: 99dd9d987c29c93597a648f6586f98ac6a932c3c
2026-11-03 15:02:57 +00:00
rhan-oai d893b14475 Add host-owned turn admission to the agent runtime (#58963)
## What changed

Add `AgentRuntimeOwnership::Host` and an `AgentTurnAdmission` provider so hosts can acquire execution capacity before runtime tasks run, including root turns. Require the provider at startup and retain its execution guard for the task lifetime. Controller ownership remains the default; host ownership skips controller rollout-budget accounting, budget reminders, and subagent roster context.

Keep pending admission interruptible without blocking input submission. Report admission failures through the turn lifecycle, record accepted input before exposing the error, and unwind task setup such as review mode. Preserve recoverable input during suspension. On interruption, skip prompts that still require a `UserPromptSubmit` hook rather than recording them without inspection.

## Testing

Add integration tests for missing-provider startup cleanup, denial and immediate retry, steering during admission, cancellation, execution guard lifetime, submit-hook handling, suspension recovery, and review-mode cleanup. Add a context snapshot verifying retained input reaches the next admitted turn.

GitOrigin-RevId: 9ee85fad2ecfbfb38735bac8ef88888fb179a967
2026-11-02 20:10:37 +00:00
Alex Kira 1a604d58da Refine additional-context byte histogram buckets (#58939)
## Why

The previous histogram placed post-truncation values just over 4,000 bytes in a bucket ending at 8,000 bytes, obscuring their size near the truncation cap.

## What changed

Use logarithmic boundaries through 4 MiB for `codex.additional_context.value_bytes`, with an explicit 4 KiB boundary near the truncation cap.

## Testing

Add a regression test verifying that post-truncation samples between 4,024 and 4,043 bytes land in buckets ending at or below 4 KiB while preserving their sum and count.

GitOrigin-RevId: af268fc0de3d5ceb648ce51fced3918317f2457a
2026-11-02 18:05:13 +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
Charlie Marsh 4969675f01 Resolve network approval prompts when their requests end (#58350)
## Why

A network request can disappear while its approval prompt is still pending. Leaving that prompt active allows a late cancellation to interrupt the turn even though the request no longer needs approval.

## What changed

- Track network prompt lifetimes in core and notify the app server when a prompt is resolved or withdrawn.
- Cancel the corresponding server request, emit `serverRequest/resolved`, and release its permission guard. Resolve outstanding network prompts during thread shutdown as well.
- Claim a live pending approval and validate its originating turn before applying a decision, so stale responses cannot interrupt a turn or persist policy changes.
- Distinguish network approvals that require the originating turn from proxy approvals that can be serviced by a later active turn.

## Testing

Extend regression coverage for disconnected network requests, late cancellation responses, turn completion during review, shutdown cleanup, mismatched turn IDs, and command approvals from background code running during a later turn.

GitOrigin-RevId: 78f656d92607ec21ddd0c5d781ed0947588cb2eb
2026-10-30 14:40:32 +00:00
Alex Kira 2ca71f1683 Record additional context counts, sizes, and truncation metrics (#58144)
## What changed

Add `codex.additional_context` to count submitted entries, including unchanged entries, and `codex.additional_context.value_bytes` to measure UTF-8 value sizes before and after the existing truncation limit. Sizes exclude message wrappers and include the truncation marker after truncation.

Tag metrics with context kind, truncation status, and an explicitly supplied name when present; never use opaque source keys as tags. Record metrics through the current turn's telemetry.

## Testing

Add integration coverage for message and tool-output formats, repeated submissions, truncation boundaries, empty and multibyte values, optional name tags, and turn-specific model attribution.

GitOrigin-RevId: 463f257af27be58c372970fac853b85fde495dd5
2026-10-29 18:24:34 +00:00
jif 9cc14e7988 Add managed command checkpoints requiring fresh human approval (#58045)
## Why

Managed command policies need to require a human decision independently of automatic review, hook allowances, and saved approvals.

## What changed

- Support `decision = "prompt-user"` in managed `requirements.toml` prefix rules, requiring a fresh, one-time human decision for matching commands.
- Add experimental `humanCommand` approval requests for clients advertising both `experimentalApi` and `requiredHumanApproval`. Accept only `accept`, `decline`, or `cancel`; reject commands when no capable client is available.
- Restrict responses to capable clients that received the request, and replay pending prompts to compatible clients on reconnect.
- Enable the capability in the TUI and update protocol schemas and client bindings.

## Testing

Add coverage for fresh decisions overriding saved allowances, hook allowances, and automatic review, plus child-thread approval routing and rejection of responses from incompatible or unrelated clients.

GitOrigin-RevId: 0b95ff6aab7f8a04b7298c8a4329d1fddf6a603e
2026-10-29 11:23:51 +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
Steve Coffey f088e0c4ff Allow replacing dynamic tools between model steps (#57553)
## What changed

Add `CodexThread::replace_dynamic_tools` to replace the tool catalog without starting a turn. Updates apply to subsequent model steps and future turns while preserving in-flight calls; an empty list clears the catalog.

Require host opt-in through `allow_persistent_dynamic_tool_updates` for persistent threads. Persist replacements before publishing them, restore the latest catalog on resume, and preserve replacements and explicit clears through compaction. Keep explicit resume overrides separate from the persisted catalog.

## Testing

Add coverage for pending calls during replacement, replacement and clearing across restart, compaction recovery, resume overrides, submission ordering after caller cancellation, host opt-in, and persistence failures.

GitOrigin-RevId: fd2634906895a9b72ae7892ddd9435dced45eccb
2026-10-27 19:50:41 +00:00
Alex Kira 6d8f311b00 Render untrusted additional context as named tool outputs (#57279)
## What changed

When `features.untrusted_as_tool_output` is enabled, emit untrusted additional context as `function_call_output` items in the `agent` namespace before user input. Use the explicit context name when supplied; otherwise derive a bounded name from the source key, falling back to `untrusted_content`.

Deduplicate these outputs by name and truncated value so name changes are emitted while changes confined to truncated content are ignored. Application context remains in developer messages, and disabling the feature preserves the existing message format.

## Testing

Extend core and app-server tests and add request-history snapshots covering both formats, turn start and steering, derived names, truncation, and deduplication.

GitOrigin-RevId: cd5e2f4b3121714acb7f4baf5607f3c27d408cae
2026-10-26 21:55:00 +00:00
Alex Kira 8ebd272491 Add optional names to untrusted additional context (#57181)
## What changed

- Add an optional `name` to `AdditionalContextEntry` across the protocol, app-server schemas, and generated clients.
- Accept names matching `[a-z][a-z0-9_]{0,63}` only for `untrusted` context. Reject invalid names and names on `application` context before admitting turn input, returning invalid-request errors from `turn/start` and `turn/steer`.
- Preserve existing model-input rendering and deduplicate context when only its name changes.

## Testing

Add coverage for valid name boundaries, callers that omit `name`, invalid-name rejection without admitting input, and deduplication across name changes.

GitOrigin-RevId: 2caffa6fddd40ddb178733b335f520f317239edd
2026-10-26 18:23:07 +00:00
eknight-oai b337fe06b7 Defer agent mail after empty final answers (#56133)
## Why

An explicit `final_answer` with empty or whitespace-only text could allow queued agent mail to restart the turn instead of leaving it queued for the next turn.

## What changed

Treat explicit `final_answer` messages as terminal regardless of visible text in both completed-item handling and turn-item finalization. Untagged messages still require visible text to defer mailbox delivery; `commentary` and `partial_answer` messages do not defer it.

## Testing

Extend regression coverage for empty and whitespace-only final answers, delivery of queued agent mail on the next turn, and message board notifications arriving before or after an empty final answer.

GitOrigin-RevId: 466f1a9c613e07d67ce45458858ba32d5d23f11d
2026-10-22 10:03:05 +00:00
Eric Traut 430a16799d Preserve thread model settings across resume and revert (#55923)
## Why

Resuming a thread can replace saved model settings with global defaults. A selected service tier can also be lost when the current model does not support it.

## What changed

- Restore saved service tier, reasoning summary, and personality on resume unless the caller explicitly overrides them, respecting thread ownership of settings snapshots.
- Retain the selected service tier for persistence, settings notifications, and resume comparisons while filtering the effective tier against the current model's capabilities.
- Carry the runtime model, service tier, and personality into thread revert reconstruction.
- Persist initial thread settings before user input during MCP startup so an interrupted first turn retains its settings checkpoint.

## Testing

Extend regression coverage for cold resume and explicit overrides, unsupported service tiers, settings preservation during revert, and settings checkpoints before user input when MCP startup is interrupted.

GitOrigin-RevId: e5ec012a5678048c5eedd443e524f38c142bf9fd
2026-10-21 15:34:43 +00:00
rka-oai 1a69ad3237 Add infrastructure for speculative remote V2 compaction (#53614)
## What changed

- Capture history, prompt metadata, routing state, authentication ownership, and context-window identity for background compaction requests.
- Send background requests over HTTP while preserving foreground routing, and retain the latest candidate in session memory.
- Add feature and threshold eligibility checks and deduplicate attempts by history version, model, compaction hash, and authentication ownership.
- Cancel and drain background tasks during session shutdown, preserving completed usage accounting.

The background launch helper is not yet connected to the turn loop.

GitOrigin-RevId: 0e5e6ab7dac4d98a836733668a775ac48b5a59f1
2026-10-13 08:44:23 +00:00
Eric Traut 7a0a8ac9c4 Close input admission while a turn finishes recording (#53452)
## Why

Inputs arriving during turn finalization could be queued after the final batch had already been drained, or recorded ahead of that batch.

## What changed

Mark active turns as finishing before draining pending input. Reject injection into finishing turns and make late client responses and notifications wait until the final batch is recorded before appending to history. Continue accepting input into turns reserved but not yet started.

## Testing

Extend the task-completion test to block final recording, verify that input admission is closed, and confirm that a late notification is recorded after queued input while preserving the notification truncation budget. Strengthen the retained-tool-output scenario to assert tool output names and ordering.

GitOrigin-RevId: a57e44bd974e60e134183dab989795e2c3538596
2026-10-12 18:54:22 +00:00
jif cd85a26cae Remove the patched zsh shell execution backend (#52160)
## What changed

- Use the standard shell executor and remove the zsh fork escalation runtime, `codex-execve-wrapper`, and `shell-escalation` crate.
- Ignore `shell_zsh_fork` and `unified_exec_zsh_fork`, emit migration notices for legacy configuration, and warn when ignoring managed requirements for these flags.
- Keep the active session's `apply_patch` alias directory readable under restricted filesystem permissions.

## Testing

Add coverage for legacy configuration startup and shell execution, migration notice rendering, active alias directory selection, and nested `apply_patch` execution while other files in `CODEX_HOME` remain unreadable.

GitOrigin-RevId: 43af461fea32b4980550ee7237effd2bc0646ed8
2026-10-08 16:19:09 +00:00
liame-oai ec8251c545 Add experimental prediction forks that inherit parent context (#51884)
## Why

Prediction forks need to preserve the parent's effective context and request settings to maximize prompt-cache reuse.

## What changed

- Add `experimentalPredictionMode` to `thread/fork` for ephemeral forks of loaded, persisted parents. Reject configuration overrides, `path`, `lastTurnId`, and `beforeTurnId` in this mode.
- Inherit the parent's effective configuration, tools, environments, collaboration settings, disabled plugins, and persisted history.
- Enable `features.reasoning_effort_override` on the fork while preserving the parent's request-effort baseline, including when overrides were disabled on the parent.
- Report the fork's effective model, service tier, and reasoning effort in the response.

## Testing

Add coverage for matching parent and fork requests and prompt-cache keys across legacy and paginated history, invalid fork parameters, and reasoning-effort inheritance. Add a context snapshot for identical parent and fork continuations.

GitOrigin-RevId: 46b41f2ac331562b441dadff1a0a9f926c94880d
2026-10-08 00:09:03 +00:00
Adam Perry @ OpenAI 9479e1fdb7 Bind capability roots to environment selections (#51493)
## Why

Capability discovery needs the roots captured with each environment snapshot. Threads sharing an executor can select different roots, and changing environment order must preserve executor skill aliases.

## What changed

- Attach the receiving thread's capability roots when constructing `TurnEnvironmentSelection` at startup and settings updates, including roots loaded from saved history.
- Resolve capability roots from environment snapshots while preserving their original order. Retain startup roots so deselecting and reselecting an environment restores its capabilities.
- Match inherited configuration by environment and workspace so differences in capability root metadata do not interrupt configuration propagation.

## Testing

Add regression coverage for root ordering across environment states and configuration changes, independent root selections on a shared executor, reselection with startup or saved roots, active turn updates, and pending configuration failures reaching child and grandchild threads.

GitOrigin-RevId: 4f2e4bb414d27195a6752406117a6050718a1198
2026-10-06 22:39:32 +00:00
Eddie Chen 551bd409eb Preserve turn attribution across recovery and compaction (#51402)
## Why

Recovering an interrupted turn replaced its original trigger with `retry` and restored only its root lineage from model context. Turns suspended before that context was written could lose their attribution entirely.

## What changed

- Persist regular-turn attribution in `TurnStartedEvent` and compaction resume metadata, including the trigger, parent and root turn IDs, and initiating agent path.
- Restore matching attribution during recovery, keep unknown triggers absent, and retain the model-context root fallback for older records.
- Reconstruct attribution from surviving turns across rollback and late completion or abort events, and clear inherited attribution when spawning a forked agent.

## Testing

Add coverage for repeated cold recovery, suspension before context persistence, legacy rollouts, fresh attribution on subsequent turns, compaction checkpoints, rollback, and late terminal events.

GitOrigin-RevId: 85b5763e8fbd48a1ea2c07cfda8196e63c2cd33a
2026-10-06 15:50:27 +00:00
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
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 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
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
jif 5c07856e25 Separate local agent runtime state from controller handles (#47301)
## What changed

Extract `LocalAgentRuntime` from `LocalAgentControl` to hold the shared registry, residency state, execution limiter, rollout budget, service tier, instruction provider, and thread-manager bindings. Keep session identity on the controller handle.

Store the runtime in `SessionServices` and use it to bind controllers for child startup, resume, delegation, and guardian reviews. Route ownership checks and residency handling through the shared runtime, preserving registry identity across handles. Update existing test fixtures to keep the controller and runtime paired.

GitOrigin-RevId: 846094657cbdf736da3149f0a1c96b4e01423dfe
2026-09-22 14:48:28 +00:00
jif c026e7a622 Expose a backend-independent agent control contract (#46547)
## What changed

Add a public, object-safe `AgentControl` trait and request/response types for coordinating an agent tree independently of where its threads run. Define contracts for lifecycle operations, message delivery, status observation, execution admission, usage accounting, shared configuration, and guardian evidence.

Move shared agent types out of the local controller and export them from `codex-core`. Let `AgentExecutionGuard` own a backend-provided reservation while preserving the local limiter's release-on-drop behavior. Include `root_thread_id` in `GuardianRootSnapshot` to identify the source of authorization evidence.

## Testing

Add tests for reservation release when a guard is dropped or a turn is cancelled, and extend the guardian authorization test to assert the root thread identity. Check trait object safety at compile time.

GitOrigin-RevId: 9385c9fbfd189f08429539fd89061ef8c9efe0d9
2026-09-19 01:03:55 +00:00
jif 7fb599de8c Rename AgentControl to LocalAgentControl (#46513)
GitOrigin-RevId: a8e8a5c818a1751c5d2667f5af87e481482fc7db
2026-09-18 23:39:30 +00:00
Eric Traut f2b5b81f39 Continue interrupted work after managed daemon restarts (#45820)
## Why

Daemon recovery reloads threads but leaves interrupted work unfinished. Resume eligible work automatically from the saved conversation, even without a connected client.

## What changed

- Attempt one new continuation turn immediately after restoration, including in Plan mode. Mark the old turn interrupted and supply recovery context without creating a user message or granting new authorization.
- Require an idle thread, matching permissions, and the same single local environment configured by the thread. Skip completed, aborted, or superseded work and snapshots without environment identity.
- Add `continue_turn_if_idle` with an atomic previous-turn check so newer tasks or standalone settings changes invalidate pending continuation.
- Preserve the output schema, service tier, and root turn ID, and emit a “Resuming interrupted work” warning when continuation starts.

## Testing

Add daemon restart coverage for continuation without a client, Plan mode, permission and environment mismatches, and legacy snapshots. Add core coverage for preserved continuation metadata, absence of user-message events, rejection of superseded continuations, and exclusion of remote execution from recovery snapshots.

GitOrigin-RevId: 2b290b75d9399fc58258bcda2e85c74e9b5b3b09
2026-09-16 00:54:58 +00:00
felixxia-oai c51cb968e4 Preserve Guardian evidence during checkpoint migration (#45789)
## Why

A checkpoint produced by a previous model may be incompatible with the selected Guardian reviewer. Legacy review must retain user restrictions and verified answers across compaction and restart until a compatible checkpoint is available.

## What changed

- Expose retained user evidence independently of whether Guardian reviews the legacy transcript or parent context.
- Resolve the selected reviewer's compaction compatibility hash during resume and remote compaction. Preserve the legacy transcript when checkpoint compatibility is unknown or mismatched.
- Activate parent-context review immediately after compatible compaction and invalidate pending reviews bound to the previous evidence policy. Ordinary compaction preserves pending reviews.
- Keep strict compatibility checks for previously migrated checkpoints whose complete legacy transcript is no longer available.

## Testing

Add coverage for mismatched checkpoint hashes, retained evidence during legacy review, and pending-review cancellation on migration. Add a request-history scenario covering a model switch, incompatible automatic compaction, restart, and compatible manual compaction with preserved user restrictions and verified answers.

GitOrigin-RevId: b52bfd7c21ce7ebed482174c8d7c327e50b1ada8
2026-09-15 20:41:39 +00:00
Krish Chainani 446b771049 Add attachment upload and resolution APIs and pass stores into sessions (#45556)
## What changed

Replace `AttachmentStore::persist` with `upload` and `resolve`. Uploads return inline bytes or a file ID; resolution returns optional file metadata and a download URL only when a minimum URL lifetime is requested. Add image dimensions, digest, size, MIME type, and categorized errors to the API.

Make `InlineAttachmentStore` return the original bytes and report `NotFound` when resolving file IDs. Redact attachment bytes and file URLs in debug output.

Pass the configured image store from the thread manager into sessions and inherit it in delegated sessions. Allow `TestCodexBuilder` to accept a custom image store.

## Testing

Update unit coverage for preserving PNG and JPEG bytes, redacting bytes and URLs in debug output, and rejecting file resolution in the inline store.

GitOrigin-RevId: f313e0048925998d84394c7e5a278e0d39d86a45
2026-09-15 00:29:13 +00:00
jif 18d7ace221 Move Guardian reviewer lifecycle into the extension (#45537)
## Why

Guardian reviews must stop when their parent shuts down or their history is reset. Shutdown must also finish reviewer cleanup before closing the parent's persistent history, including when a review is waiting to retry after a rate limit.

## What changed

- Let the Guardian extension own reviewer startup, prewarming, and shutdown through `ThreadManager`, tracking and joining outstanding work during teardown.
- Cancel reviews on history reset or parent shutdown and reject decisions returned after cancellation, including cached extension decisions.
- Keep background prewarm previews from overwriting the active turn's model metadata.

## Testing

Extend regression coverage to verify reviewer cleanup after parent shutdown, prompt shutdown during a 60-second Guardian retry with the network request denied, and preservation of active model review requirements during prewarming.

GitOrigin-RevId: 05efc369b9a1642a10365eac8a09da77f3f6e28c
2026-09-14 22:58:51 +00:00
Owen Lin 3052bbcf8c Remove the deprecated thread/rollback API (#44915)
## What changed

Remove `thread/rollback`, its request and response types, generated bindings, and the core `Op::ThreadRollback` operation. Requests now follow the generic unknown-method rejection path. Document `thread/revert` as the alternative for paginated threads.

Keep historical `ThreadRolledBack` markers and legacy error deserialization so existing rollouts remain compatible with replay and migration.

## Testing

Adapt retained-context and Guardian history tests to append legacy rollback markers and resume threads, preserving coverage of surviving instructions, answers, and review history.

GitOrigin-RevId: b3da1becdf86b1869275aacb0ffc2817cee5af2e
2026-09-11 21:32:04 +00:00