Repository navigation
Codex fork 0.160.0
Codex fork 0.160.0
This is a main-based fork release following fork 0.159.0, not a byte-for-byte rebuild of OpenAI's official 0.160.0 release. The fork remains based on pinned OpenAI main; the official release follows a different branch history.
Comparison and source
- Previous fork runtime and SDK source:
dfbd20689f, published as 0.159.0. - Previous pinned base:
85714c61d1. New pinned base:2cf9b0875a, merging OpenAImain@b172810921. - Official OpenAI 0.160.0 was published 2026-10-01 at 20:19:13 UTC and points to
a956835d02. This fork does not claim to be identical to or rebuilt from that release branch. - Final reviewed semantic source:
0688bd1c60. The independent review off65d445480found two issues; the same reviewer closed both after targeted repair, with no remaining blocking findings. - Version-stamped 0.160.0 release source:
54fc12f7aa, shared by the annotatedrust-v0.160.0andsdk/go/v0.160.0tags. This changes only the workspace and 161 local package versions; all 1,312 external Cargo records and other fields are unchanged. Exact-lease push and the single pinned five-target build succeeded.
The upstream section below covers all 229 non-merge commits between the pinned bases, not every change on OpenAI's official 0.160.0 release branch. Exact commit links follow at the end.
Upstream changes since fork 0.159.0
Agents, context, and history
- Abort callbacks now complete before terminal turn events. Remote message-board notifications can reach active turns, queued agent mail survives session eviction, and upstream fresh V2 subagents gain opt-in dynamic-tool inheritance; this fork retains its unconditional fresh-child inheritance. The subagent picker and delegated-task previews track thread/archive and task-input state more accurately.
- Resume/replay gained authoritative history, reusable unchanged-history snapshots, ordered world-state response items, persisted world-state snapshots, and persisted additional tool definitions. Explicit user goal edits can be recorded with explicit provenance; live tool-call metadata survives request windows.
- Guardian V2 adds retained conversation support in async classification, bounded Decisions comparison transport, sender-review context, omission deduplication, and decisions agreement/latency telemetry. Host skill discovery is skipped for reviews and user restrictions survive handoff context.
MCP, app-server, protocol, and security
- Stable
thread/attachmentOwner/listfinds owning threads by exactattachmentTypeandidentityKey, with archived filtering and cursor pagination. Experimental thread-prediction request/notification types were added, but this upstream snapshot does not implement prediction runtime behavior. - MCP follows legacy tool pagination, validates enterprise authorization servers before ID-JAG exchange, uses rmcp for managed token exchange, preserves Windows environment variables, keeps resource helpers and shared types available in Code Mode, and emits attributed OAuth credential-storage telemetry.
- The app-server protocol now accepts unknown
CodexErrorInfostrings or objects, adds an advisory experimental Bedrock GovCloud requirements check, and carries optional explicit goal-mutation provenance.ModelProviderCapabilitiesReadResponse.namespaceToolsis removed and tool namespaces are no longer provider-gated. Separately, custom providers gain capability overrides. - Bedrock adds GovCloud support, Ultrafast service tiers for Astra models, Sol catalog defaults, multi-agent V2/Ultra reasoning, and API-key Daybreak/cyber program paths. Model visibility remains server-, provider-, catalog-, and account-controlled; this release does not globally guarantee Sol visibility or access.
- Exec/server paths gained writable streaming, bounded opens and request cleanup, cancellable reads, session recovery, permission catalogs, retry/jitter behavior, diagnostics, and stronger path/sandbox handling. Responses retries honor failed-event
Retry-Afteradvice while preserving the fork's capacity-retry floor.
Terminal UI, daemon, platform, and build
- Transcript selection/copying, Find, pager and configured keybindings, Markdown links/tables/blockquotes, literal HTML copy, owned-transcript keyboard selection, and modal-period selection behavior improved. Command Center task renaming, retained backdrops, selection adjacency, and fork shortcuts were refined.
- Daybreak gained persistent TUI selection, status/title state, continuation handling, API-key access, and
codex execsupport. Voice adds audio-device/microphone selection, local-device settings, and realtime transcript-tail persistence; managed in-app voice remains feature-gated. - Daemon update/startup diagnostics, release identity, cwd recovery, Windows-mounted WSL handling, updater stderr, remote-control socket setup, and diagnostic pruning improved. Windows sandbox/process/path and macOS/Linux sandbox handling received targeted fixes.
- Rollout attachments are bundled more compactly, paginated command output is bounded, reverse JSONL scans use
memrchr, blocking history/index work is moved off async workers, and SQLite corruption/recovery handling is typed. CI/build updates include toolchain, Windows bindings, age, signing-page, and test-infrastructure changes; they are not claims about bundled optional resources.
Fork reconciliation and Go SDK
All 146 fork commits were replayed: 105 patch-identical, 41 adapted, 0 dropped or unmatched. The new base remains an ancestor. Cumulative fork capabilities are retained, including independent direct-MCP/Code-Mode mcp_max_lines and originating output caps through replay; custom providers and context/compaction overrides; plaintext provider messages and inherited dynamic tools; cumulative spawn budgets, active/idle follow-ups, quiet waits, parent completion, bounded compaction history and capacity retry floors.
Final review of f65d445480 found two blockers. R1 was a duplicate abort_all_tasks lifecycle callback after upstream moved callback completion ahead of the terminal event; the duplicate outer emit was removed and the existing red regression now passes four targeted cases. R2 lost unconditional inherited dynamic tools for fresh V1/default V2 children; fresh-child inheritance was restored and the three targeted cases now pass. Guardian tool isolation and fresh-history behavior remain unchanged. The combined Guardian/abort/inheritance selection passed 14 tests. A fresh debug CLI/helper rebuild and isolated required-real-runtime strict SDK handshake passed, followed by scoped fix/format checks and the same reviewer's targeted closure of both findings.
The Rust stable/experimental schemas, serde manifest, precomputed exports, and generated Go protocol are synchronized from one frozen protocol snapshot. The Go generator now renders the open CodexErrorInfo anyOf as its existing typed union: known strings and known object variants keep typed fields, while unknown strings or objects round-trip through RawJSON instead of degrading the whole type to json.RawMessage.
Raw Go protocol changes include:
- Breaking raw-field removal:
ModelProviderCapabilitiesReadResponse.NamespaceToolsis gone because the upstream wire response no longer containsnamespaceTools. - Stable raw
ThreadAttachmentOwnerListbinding and types. - Experimental raw
ThreadPredictionRequest,ThreadPredictionUpdatedNotification, and result types. These are protocol bindings only; no prediction runtime workflow is implemented by this snapshot. - Experimental raw
BedrockCheckGovCloudRequirementsbinding and response types. ThreadGoalSetParamsandThreadGoalClearParamsgain optional nullableorigin: "user" | "automatic". Generic wrappers do not infer or substituteuser; callers must set it explicitly when they genuinely represent user provenance.
No new high-level SDK wrappers are claimed for these methods; generated raw-client coverage is the intentional scope. Existing MCP OAuth loginId correlation and legacy name/thread fallback, typed ThreadItemsListCursor, opt-in BackendReasoningStatus, strict digest handshake, resource clients, Gateway OAuth lifecycle, realtime behavior, and module path github.com/openai/codex/sdk/go remain in place.
Use the runtime and Go SDK from the same matching 0.160.0 release. Strict digest validation is not implicitly bypassed. Update the entire binary package, including codex-code-mode-host; do not mix a new CLI with an old daemon/app-server.
Downloadable packages
The pinned five-target release build succeeded on all targets. All five runner checkouts and codex-cli 0.160.0 version assertions passed. The release contains exactly five bin-only system archives:
codex-package-aarch64-apple-darwin.tar.gzandcodex-package-x86_64-apple-darwin.tar.gz:bin/codex,codex-app-server,codex-code-mode-host,codex-responses-api-proxy.codex-package-aarch64-unknown-linux-gnu.tar.gzandcodex-package-x86_64-unknown-linux-gnu.tar.gz: the same four binaries plusbin/bwrap; Linux is GNU/glibc, not musl.codex-package-x86_64-pc-windows-msvc.zip:.exeversions of the four binaries pluscodex-command-runner.exeandcodex-windows-sandbox-setup.exe.
Downloaded workflow artifacts matched their GitHub sizes and SHA256 digests, and their outer ZIP CRC checks passed. Inner archive integrity, exact binary layout, ELF/Mach-O/PE architecture, and Unix 0755 permissions passed. The downloaded Linux x86_64 CLI reported codex-cli 0.160.0; TestRealAppServerInitializeStrictDigest -count=1 passed in 0.231 s with the real downloaded runtime required, isolated/mock homes, and no compatibility override. No real OAuth login or paid provider call was used. Optional native voice/audio resources and provisioned Windows sandbox-service packaging remain outside this bin-only layout.
Archive sizes and SHA256 (these describe the downloadable packages, not the enclosing workflow artifact ZIPs):
codex-package-aarch64-apple-darwin.tar.gz: 199,280,891 bytes;817cc9d837e4df6c3f0252f04114f1d52589a6bef21b217e43632521033d9df8.codex-package-aarch64-unknown-linux-gnu.tar.gz: 206,294,385 bytes;5076f696b6a737f54d41d1ce13ea073100929cc4a5b515607e8547afce7db5b6.codex-package-x86_64-apple-darwin.tar.gz: 212,538,780 bytes;00310d3de032dc15a58f568dbfe898d38f56c5d68bfba1888b954e06f6a1b606.codex-package-x86_64-pc-windows-msvc.zip: 230,645,735 bytes;e83b22ff04132898d9e53c50f9a945fbffbde043620abe6a539ea34fe6899ce3.codex-package-x86_64-unknown-linux-gnu.tar.gz: 219,507,103 bytes;6c23fb5abd32609f8a82d574674d85f5d6da75ae089937eae3ce5c7c10da540f.
Upgrade note: replacing binaries on disk does not update or restart an already-running app-server/daemon. An older process can still cause configuration or strict protocol-handshake errors. Restart the daemon yourself if appropriate, or use --no-daemon; this release never changes a running process or installed CLI on its own.
Validation and limitations
- Rebase accounting and preservation checks passed: 146/146 fork commits retained, 105 patch-identical, 41 adapted, 0 dropped, with the pinned base as ancestor and protected user files unchanged.
- Adjacent selected Rust tests: 2,515 total, initially 2,510 passed, 5 failed, and 9 skipped. Both capacity failures were repaired and their existing checks passed; all three rmcp remote failures were closed after a fresh CLI/exec-server helper and passed. One intrinsic flaky retry was not treated as a source failure. The complete adjacent suite was not rerun as one final pass.
just test -p codex-app-server-protocol: 356 passed, 1 skipped. Stable/experimental schema and precomputed generation, Rust Go-manifest check, and Go generator checks in stable, experimental, and combined modes passed.- Go SDK:
go test -race ./...passed with 602 tests across 14 packages, without inherited runtime-path/required-real-runtime environment variables. - Core/Code Mode targeted behavior selection: all 226 selected behaviors are green after eight targeted repairs. Targeted app-server checks passed 20 cases. R1/R2 closure added the 14-test combined check and a fresh isolated strict SDK handshake (0.276 s, required runtime, no compatibility override). Scoped fix/clippy/format checks passed; these selections overlap and are not summed as unique coverage.
- Version-only stamp, exact-lease push, all five release build jobs, downloaded artifact/archive checks, and the fresh packaged-runtime strict digest handshake passed. Runtime and SDK tags point to the same exact source; no existing release tag was moved.
- Full Rust core, workspace, app-server, platform-wide, Docker/Wine, and official-release test suites were not run. Focused checks are not full-suite or all-platform coverage. No result from the previous 0.159.0 release is counted as verification of this candidate.
Source and full upstream changelog
- Previous fork release 0.159.0.
- Pinned base comparison.
- OpenAI official 0.160.0 release (a different release-branch history).
All 229 upstream non-merge commits since the previous pinned base
b172810921Let the transport create the Windows remote-control socket directory (openai#50700)3a69ec3ef8Preserve local Markdown link labels in the TUI (openai#50695)58ae3ba611Keep third-party tools deferred in strict Code Mode Only (openai#50687)b741e480e2Allow transcript selection and copying while bottom modals are open (openai#50564)58ca099b03Keep Code Mode tool discovery guidance stable across catalog changes (openai#50562)1a169eda11Distinguish daemon release identity from executable contents (openai#50559)19e554bb70Avoid reading the current directory when resolving absolute paths (openai#50558)6b43e6fe1fSkip daemon auto-start for Windows-mounted WSL homes (openai#50555)55b6f282a8Keep MCP resource helpers available in code mode (openai#50546)6326163b9aSend incremental tool catalog updates in Responses Lite (openai#50540)e1b5b56a4fKeep shared MCP types stable in Code Mode exec descriptions (openai#50536)7d5f55bdadPersist realtime transcript tails before closure without inference (openai#50531)b65ab465ceReject unknown TUI keys in strict config validation (openai#50525)86a54b051cAdd scenario coverage for remote/compactcontext preservation (openai#50516)af5d95f255Require GovCloud guidance acknowledgment after Bedrock setup (openai#50510)c542fb93efRecord Windows sandbox service stop diagnostics (openai#50507)47379efd52Keep Command Center selection adjacent after task removal (openai#50505)d42aecc56bCenter TUI confirmations over their retained backdrop (openai#50504)9ce35d337aUse Enter to accept transcript Find results and Escape to cancel (openai#50503)55922b7984Include installer stderr in daemon update failures (openai#50499)8f7a0f7a87Skip managed config loading for registered Windows sandbox refreshes (openai#50480)6c15cc4aafUse the app-server default output cap for TUI workspace commands (openai#50477)604061ce51Enable Ultrafast service tiers for Amazon Bedrock Astra models (openai#50472)be48ae396eAccount for JSON overhead when truncating MCP tool results (openai#50470)f5a2272c6cCopy transcript selections as literal text while preserving rich HTML (openai#50467)ef8cfe5e96Retry registry authentication outages and jitter executor reconnects (openai#50465)a0bfaac70cAdd theincremental_toolsfeature flag (openai#50464)09bced5ad9Populate thread previews from delegated task inputs (openai#50462)c6049ebd96Add capability overrides for custom model providers (openai#50459)820f85cf59Truncate oversized MCP results in paginated thread history (openai#50458)3c3a990da0Measure rollout persistence size reductions (openai#50454)fd75aa116cRemove the provider capability gate for tool namespaces (openai#50447)9b0a676d5dBundle rollout attachments into a gzip tar archive (openai#50446)d4eed6dca5Assert that only direct tool calls emit timing events (openai#50445)a4bfd07d51Stabilize paused-time code-mode service tests (openai#50443)3629508849Preserve native USD amounts in thread usage responses (openai#50442)0df76892b1Support ordered response items in world-state context updates (openai#50441)12a30d4e6dAdd a CLI command to uninstall the legacy Windows sandbox (openai#50437)602d2add6ePersist additional tool definitions in rollout history (openai#50435)c536ffcb18Add keyboard copy selection to the owned transcript (openai#50434)66561d0301Allow API-key accounts to use Daybreak in the TUI (openai#50433)b6903c0669Preserve terminal hyperlinks in agents overview previews (openai#50431)bee28e8a06Cap persisted command output in paginated history at 64 KiB (openai#50427)44dd77b71eHonor Retry-After headers in failed Responses events (openai#50418)dff5270b29Clarify Git worktree choices for new and forked conversations (openai#50416)c5d242fa79Consolidate command execution output intoaggregated_output(openai#50402)d61c7a824fHonor pager bindings for transcript page keys (openai#50396)f88a6efe43Honor configured keybindings before transcript navigation (openai#50389)2739e82858Allow opting into 16 KiB ARM64 code pages for macOS signing (openai#50384)1ab8c6ef28Fix thread unloading after disconnect during MCP startup (openai#50380)50b4c5e58cUse printable ASCII terminal titles under GNU Screen (openai#50375)c73775f19eRemove initial messages from session configuration events (openai#50360)91168365a5Render ANSI styles in TUI hook system messages (openai#50359)f4e18a95bbSkip unrelated subtrees during config alias normalization (openai#50354)9d2b60303eBack off automatic remote control reconnects with jitter (openai#50348)84d5437b6eKeep the subagent picker consistent with thread archive state (openai#50345)4cedd0caacAdd end-to-end coverage for MCP sandbox state enforcement (openai#50339)ca466061d6Record Guardian V2 Decisions agreement and latency metrics (openai#50273)14a477ea89Bound tmux option probes to one second (openai#50219)37f53199a6Use the shared text editor for command-center task renaming (openai#50216)d64360f628Support Ctrl+Insert for copying TUI selections (openai#50215)4dd51f4a5fMake transcript mouse scroll speed configurable (openai#50209)e3c2a83937Release stable Markdown tables into scrollback during streaming (openai#50207)a75987455aReport the configured TUI mode incodex doctor(openai#50200)3f97f2b3bdRestore the account email in/statusafter account updates (openai#50199)a20fe6335fReplace the Figma OAuth exception with Mercado Pago (openai#50189)b997c72a99Adddotsto issue labeler guidance (openai#50183)c39bfa4c8fEnable writable file streaming in exec-server (openai#50177)7135b303d9Upgradeageto 0.12.1 and remove the obsolete advisory exception (openai#50166)d25c114d49Bound in-flight file opens in exec-server (openai#50162)e7ea5f4a86Add managed worktree tools to the TUI (openai#50148)cb6da58876Use the server permission catalog for TUI permission shortcuts (openai#50140)8d44977aa2Add opt-in JSON diagnostics for TCP tunnels (openai#50131)7d3e696c4cPreserve Windows environment variables for remote MCP servers (openai#50129)4f99920b8cExpose the model selected for a running turn's next step (openai#50128)b707714ae4Add a native gRPC client for cloud thread resume and attach (openai#50113)a0a7a63002Centralize TUI loading glyphs and frame scheduling (openai#50112)6ece7bfc21Keep fullscreen prompts bounded and scrollable (openai#50109)e58b493329Consolidate chat composer footer logic infooter_state(openai#50105)59f18e8133Add opt-in Decisions comparison for Guardian V2 (openai#50099)2635431edbAdd attachment owner lookup to the app-server (openai#50094)9d53dd91b5Prevent shared instruction providers from delegating to themselves (openai#50093)960e878df4Preserve queued agent mail across session eviction (openai#50087)26a49b3882Add paginated reverse lookup for thread attachments (openai#50083)e4e33a56b0Enable dynamic tool inheritance for fresh V2 subagents (openai#50082)d91294c39eAdd a bounded Decisions transport for Guardian comparison classification (openai#50066)ae7aad586eFix Linux sandbox startup with multiple denied files (openai#50059)b06b7d2f77Upgrade Windows bindings towindows-sys0.61.2 (openai#50058)d9cf203890Check token estimate filtering directly on the tracing subscriber (openai#50054)57a38c1bebPreserve question context in recovered TUI answer drafts (openai#50052)9561a34531Keep plugin and skill snapshots scoped to each step (openai#50050)5fa5aaf0ffStabilize Windows voice build cache keys across tool reinstalls (openai#50046)1cc9917508Keep transcript Find expansion scoped to the current match (openai#50045)a97fb78085Keep transcript Find results readable after closing the query (openai#50039)cc31e374fcFollow MCP tool pagination in legacy protocol mode (openai#50035)2685e3a4cePreserve user restrictions in Guardian handoff context (openai#50026)9552906b2bProtect the guardian decisions API key from environment forwarding (openai#50019)595534314fUse descriptor-safe helpers for executable test fixtures (openai#50018)b527ce4734Honor server model defaults when starting fresh TUI threads (openai#50013)57ac6f5163Preserve the async Guardian history prefix as retained context changes (openai#49993)ecc78e4cf5Add renewable EMA HTTP authentication and credential versioning (openai#49987)a933dd77dbShare byte buffers across exec-server output chunks (openai#49972)d6c3b448a4Test session index thread-name append and removal (openai#49959)8a400e7812Cache the placeholder regex for MCP hook argument expansion (openai#49956)f70810bcd2Include preceding assistant context in Guardian sender reviews (openai#49951)dd90f160edPrevent stale file search results from being labeled with a new query (openai#49946)6b4daafdb4Add per-turn Cyber access program selection to exec and the SDK (openai#49939)444da310e1Respect approval policies in temporary structured threads (openai#49912)819efd7273Preserve validation errors for invalid TUI keybindings (openai#49910)da2e174a66Scope extension filesystem access to callback permissions (openai#49898)08a7031b6aReturn world-state snapshots and context updates together (openai#49894)1f52d40704Bind permission grants to the originating turn (openai#49880)d2b254fd17Remove personality plumbing from the TUI (openai#49876)7b88d09d61Decouple TUI startup presentation from execution configuration (openai#49875)bdfbab9f4fPoint usage and credit links to ChatGPT settings (openai#49874)90d7f2715aUpdate elevated-launch warning snapshot to use⌃ofor copy (openai#49867)8ea2428c38Add Daybreak state to the status line and terminal title (openai#49861)cd4a9cbd25Honor Daybreak settings in TUI continuations and background tasks (openai#49859)e1522188e9Add a persistent/daybreaktoggle to the TUI (openai#49858)f58ed54a9dUse the model catalog to select TUI cyber refusal guidance (openai#49857)c41a72cd3fSupport Daybreak selection incodex exec(openai#49856)bc197b77bcUse embedded mode for elevated Windows TUI sessions (openai#49855)236be1ad9fImprove diagnostics for report attachment failures (openai#49852)e2a7d53c9cLaunch Windows daemon children in a dedicated working directory (openai#49850)b44ca872b1Persist world-state snapshots alongside rendered context (openai#49847)408f48ce0cCapture host-supplied extension data for each turn (openai#49846)08b07eef12Preserve daemon diagnostics and include updater logs in reports (openai#49843)e53e932dc8Allow microphone channel selection for voice conversations (openai#49836)08e2b58b07Clarify service tier default save errors in the TUI (openai#49835)799324821dBox the resume future in the agents overview permissions test (openai#49822)cda82a2c68Recover daemon startup and updater re-exec after cwd deletion (openai#49819)a73898c249Use dedicated parameters for sandboxed file opens (openai#49818)a5eac80150Add an advisory Bedrock GovCloud requirements check (openai#49817)1182834d13Remove browser-open success messages from the TUI (openai#49816)8e44ad94b0Add coordinated shutdown for local agent trees (openai#49814)d4a475addaSupport AWS GovCloud regions for Amazon Bedrock Mantle (openai#49813)0f8df3d214Move shadow skill ranking off the turn preparation path (openai#49812)c51f5bfb82Handle unsupportedfs/writeBlockrequests in exec-server (openai#49811)cca440697fFlush expired paste bursts before handling Enter (openai#49810)2fde968de8Preserve local launch permissions across TUI sessions and reconnects (openai#49809)7d4c7a0767Enable API-key model discovery by default (openai#49807)a5d56d8120Accept unknown Codex error variants in the app-server protocol (openai#49806)e7798c9944Add capability-gated writable file streams to the exec-server client (openai#49805)d2f2c40095Use platform-specific modifier labels in TUI shortcut hints (openai#49804)33aea33b39Update the Rust toolchain action for argument-comment linting (openai#49801)ec0cfa5da8Allow cleanup of replay-only side conversations with missing threads (openai#49800)606b139565Preserve server web-search settings in the TUI (openai#49799)c538fbabe5Share cached exec-server environment info with Arc (openai#49798)5ca55db09cDeduplicate Guardian retained-context omission notices (openai#49796)47a8bd7321Avoid duplicate sync reviews in Guardian classifier continuations (openai#49795)726f1492dbAdd conversation mode to Guardian v2 async classification (openai#49793)5f300d3f74Add retained conversation support to Guardian async sampling (openai#49792)83cf88306eRemoveAGENTS.mdfrom Bazel core test data (openai#49787)f542ba68e5Clarify V2 spawn model override guidance for context catalogs (openai#49786)1f77c0cfa6Persist empty paginated threads when naming them (openai#49785)3bbf8ec3a1Add a requirements feature gate for the browser annotation API (openai#49784)106772e3c6Preserve background thread requests when forking in the TUI (openai#49783)1a913493dcClean up process groups for failed shell snapshot captures (openai#49782)8953de1f1aInclude the environment's MXC backend in MCP sandbox metadata (openai#49781)875bf9209bDefine exec-server protocol types for streamed file writes (openai#49778)60947e2341Add account security setup reminders to the TUI (openai#49715)f151a0f5c2Decouple API-key cyber access programs from model discovery (openai#49714)18131270feRemove repository-local Codex guidance, skills, and environment config (openai#49713)5e96aabd68Avoid full-string scans in token-budget truncation (openai#49712)596ae94fb0Classify SQLite corruption using typed error codes (openai#49710)4f699cd642Move session index I/O off async runtime threads (openai#49708)17e5a4dbcbUpgrade the argument comment lint toolchain and Dylint (openai#49706)5602705c96Prevent npm alpha dist-tags from moving backward (openai#49704)3b16b5a5b0Rename exec-server file handle management identifiers (openai#49702)3620b2caf8Detect SQLite corruption during startup and preserve recovery backups (openai#49701)f2b2e5b2a9Make exec-server file reads cancellable between chunks (openai#49696)5aa92804d2Batch rollout listing scans on cancellable blocking workers (openai#49694)ecc205edacMove thread history projection into one blocking task (openai#49693)d1c4e3c3e6Keep compressed rollout snippet searches on one blocking worker (openai#49692)7e8878f605Preserve PowerShell relative paths in the elevated Windows sandbox (openai#49690)0a76a2520fExport skill invocation events through OpenTelemetry (openai#49689)49be2c7ab0Deliver remote message board notifications to active turns (openai#49686)6996cde697Add a managed feature gate for in-app voice (openai#49683)fcbed044c8Escape command drafts when recovering question answers (openai#49678)ed0cc1a4abSerialize Responses routing fields before large inputs (openai#49675)67727e7cf1Allow managed requirements to disable the Windows MXC sandbox (openai#49642)7219fd735bUse server authentication for explicit remote session commands (openai#49624)92bc601ad6Reuse unchanged history snapshots when resuming threads (openai#49600)d7b0d4aa66Return authoritative replay history when resuming a thread (openai#49599)de02016798Persist explicit user goal edits in model history (openai#49598)1fc8d54807Make the strict network approval test independent of request order (openai#49595)a5cce8895aSkip host skill discovery for Guardian reviews (openai#49584)0b43721d8dCopy selected file paths as plain text in the TUI (openai#49564)2e5fea64eeAdd an opt-in model catalog to multi-agent context (openai#49560)d42056091aAdd a fork shortcut to the TUI command center (openai#49517)bcd6d9ab6bAdd regression coverage for account switches between analytics batches (openai#49489)90abcfac02Add experimental thread prediction protocol types (openai#49480)15fd656ddbDiscover and validate MCP authorization servers before ID-JAG exchange (openai#49478)9ef9cb1d9fComplete turn abort callbacks before emitting terminal events (openai#49475)c9b3924a62Use the rmcp SDK for enterprise-managed token exchanges (openai#49473)b588812e8cUse server-authoritative permissions in the TUI (openai#49472)76a6e55d5aRestore executor tool paths after login shell startup (openai#49467)8c3612fb63Usememrchrto find newlines in reverse JSONL scans (openai#49444)6ba4bf9e64Honor server retry advice across Responses retries and fallback (openai#49441)fc81a7154bAdd local audio device selection to TUI voice settings (openai#49437)d8f69ea8bcPreserve bootstrap discovery across authentication changes (openai#49432)2cc65cdd4cEnable analytics by default for daemon-launched app servers (openai#49426)8ea2c0e0d4Prune diagnostic logs periodically by age and database size (openai#49425)d1d0e89558Infer Windows UNC paths with forward and mixed slashes (openai#49424)ab84d71f57Omit payloads from multiline ANSI warnings (openai#49416)d940fb21f3Truncate input text in protocol debug output (openai#49415)16a7c0f0ebFilter graceful shutdown guard and trigger traces from SQLite logs (openai#49414)eefe0ce1a8Bind the app-server time provider to a local variable (openai#49411)bd5185820dCompare tool call metadata in the recorder refresh test (openai#49408)67352b5b4aRecover exec-server sessions after environment info timeouts (openai#49407)7c35e1551fSupport explicit cyber access programs with OpenAI API keys (openai#49406)17a9df60e4Add an experimental flag for bundled tools in login shells (openai#49403)87d3e06847Preserve live tool-call metadata across request windows (openai#49401)ceea67163fRemove randomized greetings from TUI session headers (openai#49395)05ea5f757eAdd attributed MCP OAuth credential storage telemetry (openai#49392)9212b3eca8Serialize tests that share Windows sandbox accounts (openai#49389)d6571322d2Fix Windows path inference for opaque URIs with slash prefixes (openai#49388)a44afa527cTrack credential storage outcomes and redact sensitive errors (openai#49384)bd4204efc2Compile hook matchers during discovery (openai#49379)2a34aef794Update Bedrock GPT-6 Sol catalog tests to expect multi-agent V2 (openai#49369)94d642d8b4Clarify credential storage wording across authentication UI and docs (openai#49361)995138d71aCarry shell invocation metadata and report executor PATH directories (openai#49360)1983c48fd1Continue Markdown blockquotes when pasting multiline text (openai#49357)804d6306e8Allow approved filesystem escalation while preserving denied reads (openai#49353)8ffd91e42aEnable multi-agent V2 and Ultra reasoning on Amazon Bedrock (openai#49345)a6e9eaa9bdAdd GPT-6.1 Sol to Bedrock catalogs and make it the default (openai#49339)ed9e5a26a8Clean up canceled exec-server RPC requests immediately (openai#49332)5937592c07Keep remote control reconnect backoff capped during sustained failures (openai#49330)26dd19ef47Retry Windows sandbox runner logon once on error 1056 (openai#49325)