Repository navigation
Codex fork 0.158.0
Codex fork 0.158.0
This is a main-based fork release following fork 0.156.0, not a byte-for-byte rebuild of OpenAI's official 0.158.0 tag. The pinned OpenAI main snapshot and the official release tag follow different branch histories; commit-SHA differences alone do not imply missing features.
Comparison baseline and provenance
- Previous fork runtime and SDK: 0.156.0, source
e8bab0d47f. - Previous pinned fork base:
cdd1d9e1bb. - New pinned fork base:
1a840f77ea, merging OpenAImain@44fe510ce3. - Final reviewed semantic source:
25aac31a3a. Version-stamped release source:5066ec32ee; only the workspace/local package versions change beyond the reviewed semantic source. - Matching tags:
rust-v0.158.0andsdk/go/v0.158.0. Earlier tags are not moved. The mistakenly dispatched 0.157.0 build was cancelled and is not a source of these packages. The corrected manual five-target run 36428114388, attempt 1, checked out exact source5066ec32eea7bd37952337e2f35359db7445a629on every runner and completed successfully on 2026-09-28 at 16:34:43 UTC.
The upstream section below covers all 303 non-merge commits between the pinned bases, not every change in OpenAI's official 0.158.0 release branch. The full exact commit list follows at the end.
Upstream changes since the previous pinned base
Agents, Guardian, and context
- Agent operations and V2 child lookup/loading now route through
AgentControl, including host-provided controllers. Pending inter-agent messages and ephemeral message boards survive their relevant lifecycle transitions; agent status and spawn observability improved. - Guardian now retains ordered assistant and authorization context, deduplicates retained instructions, keeps its own history across parent compaction, binds reviews to the target environment, and applies computer-use review to the Browser connector. Circuit-breaker interruptions gained opt-in structured errors.
- Compaction preserves model/access-program pairs and user text, resumes context from the latest boundary, and applies the unchanged-model shortcut to more session sources. New user input can preempt a response; Code Mode can yield early for observations or opt in to user-input yielding.
- Code Mode and tool metadata retention are more selective under outgoing budgets, including late truncated results and provider endpoint overrides. MCP/Code Mode input schema budgets are configurable.
App-server, networking, and security
thread/items/listaccepts an item anchor as well as its existing opaque string cursor;mcpServerStatus/listcan discover one server and reuse a thread's MCP connection. Realtime V3 adds opt-in backend reasoning status, and the executor connection API accepts an optional bearer token.- Application network policy now reaches HTTP/WebSocket, app-server, remote control, AWS auth/telemetry, and embedded startup paths. Executor authentication, reconnect credential boundaries, request bounds, and socket/process handling received targeted hardening.
- Server
Retry-Afterdeadlines are preserved. Flex capacity failures have a distinct terminal error; plugin extension metadata was deliberately removed from discovery and summaries. Pro Max plan support and updated plan labels were added. - Windows sandbox startup/provisioning and child-process launches, macOS Seatbelt trust, Linux descriptor cleanup, and daemon socket masking received platform fixes. Diagnostic uploads and logs include more useful context while sensitive WebSocket headers and tool payloads are kept out of info logs.
Terminal UI and developer experience
- Transcript selection/copying, links, scrolling, warnings, status, session switching, reconnect notices, startup drafts, and the fullscreen composer received focused fixes. Mermaid flowcharts support more native syntax, labels, shapes, and relationships; Markdown tables and math rendering/copying are more faithful.
- Voice sessions remain active across thread navigation, and RTP timing is aligned. Prompt suggestions, working/completion tips, turn durations, and the welcome screen were refined.
- Build/test infrastructure adds prebuilt V8 use, Rust debug-profile alignment, more reliable executable fixtures, and several race/fixture fixes. These are upstream source changes, not claims that optional voice resources or external service integrations are bundled in the archives below.
Fork reconciliation and Go SDK
All 134 fork commits were replayed: 90 patch-identical, 44 adapted, 0 dropped or unmatched. The adaptations preserve fork behavior on upstream AgentControl, host-owned child controllers, current Code Mode/context/history owners, independent Guardian history, and absolute Retry-After deadlines. The canonical collaboration tool remains wait_agent.
Two independent final-review findings were corrected and closed on the reviewed semantic source. R1 restores raw rollout output persistence and applies originating byte/token/line policy during replay, including old serialized rollouts, rather than bypassing caps after a processed-output reconstruction. R2 makes V2 wait inspect the selected AgentControl, so a host-owned running child is visible even without local runtime registration. No new rollout wire marker or migration was added. Fork-specific overloaded-server retry floors remain separate from upstream terminal Flex failures.
The generated Go SDK and runtime protocol are synchronized. Existing opaque string pagination cursors remain JSON strings; a new item anchor is an object. In source, ThreadItemsListParams.Cursor changes from Optional[string] to Optional[ThreadItemsListCursor], with typed string/anchor constructors. This is an intentional raw-Go source API change even though the old string wire shape remains compatible; callers assigning protocol.Some("cursor") must wrap the string with protocol.NewThreadItemsListCursorString("cursor").
RealtimeStartOptions.BackendReasoningStatus opts in with true; omitted/default false is not sent. Generated types also include MCP serverName, environment authBearerToken, FlexUnavailable/TooManyDenials, and Pro Max. Upstream-removed plugin extension bindings and their obsolete test were removed instead of restoring the server feature; surviving PluginSummary defaults remain. The Go module path stays github.com/openai/codex/sdk/go. Strict runtime/SDK digest matching and explicit compatibility overrides remain in force; use matching runtime and SDK versions together. Existing Gateway OAuth admission, cancellation, opt-in, and connection ownership behavior is retained.
Existing fork capabilities retained
These are cumulative fork differences, not all new in 0.158.0:
- Independent
mcp_max_linesfor direct MCP and whole emitted Code Mode results, including exec/wait/notify, while ordinary tools keep the general cap; raw/encrypted/media outputs retain originating byte/token policies through model switches, persistence, and replay. - Custom providers, typed/fallback context-window and auto-compaction overrides, configurable plaintext messages across providers, and inherited dynamic tools.
- Cumulative per-turn spawn budgets, active/idle follow-up semantics, quiet waits, parent completion notifications, bounded compaction history, headroom/no-progress handling, and at least three capacity retries.
- Typed Go SDK resource clients, strict digest handshake, Gateway OAuth lifecycle, realtime closure, filtered completion backlog, image paths, and JSON Schema forwarding.
Downloadable packages
The release contains exactly five bin-only system archives — codex-package-aarch64-apple-darwin.tar.gz, codex-package-aarch64-unknown-linux-gnu.tar.gz, codex-package-x86_64-apple-darwin.tar.gz, codex-package-x86_64-unknown-linux-gnu.tar.gz, and codex-package-x86_64-pc-windows-msvc.zip — each with bin/ and unsuffixed executable names:
- Linux GNU/glibc x86_64 and ARM64:
codex,codex-app-server,codex-code-mode-host,codex-responses-api-proxy,bwrap. - macOS x86_64 and ARM64:
codex,codex-app-server,codex-code-mode-host,codex-responses-api-proxy. - Windows x86_64:
codex.exe,codex-app-server.exe,codex-code-mode-host.exe,codex-responses-api-proxy.exe,codex-command-runner.exe,codex-windows-sandbox-setup.exe.
Linux uses GNU/glibc, not musl. Optional native voice/audio resources and provisioned Windows sandbox-service packaging are outside this bin-only layout. All five artifact sizes and SHA256 digests matched GitHub metadata. Outer ZIP CRCs, package integrity, exact contents, ELF/Mach-O/PE architectures, and Unix 0755 permissions were verified.
SHA256 of the five release archives:
29ac8c2e3b164b69baf471275e2f4bd4f086f2d26dcdf59a2791c325d6797dd3 codex-package-aarch64-apple-darwin.tar.gz
79c56dcee6ad36836b19956943152f61e5c9d95e7908d852904cb0b02ac97966 codex-package-aarch64-unknown-linux-gnu.tar.gz
477fc854e49fd03cf3667fb5b273ae2c3c4f13e7b3ce962e8e9486038b11a458 codex-package-x86_64-apple-darwin.tar.gz
3377e33b80026af109a0dbdd33208b38a51c7c8d01bd91e6db4db522c259b321 codex-package-x86_64-pc-windows-msvc.zip
5ef31921b896deec30b6504c408e598012a80385185c2b310ab4939c1e813f00 codex-package-x86_64-unknown-linux-gnu.tar.gz
Upgrade note: replacing binaries on disk does not update or restart an already-running app-server/daemon. An older process may still cause configuration or strict protocol-handshake errors. Run matching components, restart the daemon yourself if appropriate, or use --no-daemon to bypass it. This release does not automatically alter the installed CLI or running processes.
Validation and limitations
- Rebase accounting: 134/134 fork commits replayed (90 identical, 44 adapted), pinned new base is an ancestor, and independent final review plus targeted R1/R2 closure found no remaining blockers.
- Rust checks were focused, not full suites: app-server-protocol 350 passed/1 skipped; adjacent selections 927 passed; code-mode-runtime 88 passed; R1 old-wire/history checks 150 passed plus four metadata/encrypted cases; R2 host-owned wait plus existing cases 20 passed. These selections overlap and are not summed as unique coverage.
- Stable and experimental schema/manifest/generator checks passed. The full Go race-enabled suite passed: 598 tests in 14 packages. A freshly rebuilt debug CLI passed an isolated strict real-runtime SDK handshake in 0.410s without compatibility override or paid model call.
- All five runners passed the packaged
codex-cli 0.158.0version assertion. The downloaded Linux x86_64 release CLI passed an isolated version smoke and a fresh required-real-runtime strict SDK handshake (-count=1, 0.283s), without compatibility override. No installed binary or running daemon was changed. - Full Rust core, workspace, and app-server suites were not run. Focused checks are not full-suite coverage. The cancelled 0.157.0 run is excluded; only artifacts from corrected run 36428114388 are included.
Source and full upstream changelog
- Previous fork release 0.156.0.
- Pinned base comparison.
- OpenAI official 0.158.0 release (a different release-branch history).
All 303 upstream non-merge commits since the previous pinned base
44fe510ce3Expand native Mermaid flowchart syntax support (openai#48895)1cc7e23612Show a short, neutral TUI interruption notice (openai#48830)e6f4af1d92Wait briefly for the Windows sandbox provisioning service to start (openai#48829)81d5405882Allow archiving threads before their first turn (openai#48828)6af89155d0Show a hand pointer over transcript links in Ghostty and Kitty (openai#48827)84cc4b3fb5Keep voice RTP timestamps aligned to 20 ms packets (openai#48824)456212ca21Use explicit histogram buckets for tool and skill context metrics (openai#48819)659b35f131Preserve punctuation and semicolons in Mermaid labels (openai#48814)3f4668da20Add history-aware prewarming for idle threads (openai#48812)99f7758a57Show short turn durations in TUI completion footers (openai#48807)d9487a2930Allow transcript wheel scrolling while a modal is open (openai#48805)abc8f0c9a1Use the terminal palette for ordered Markdown list markers (openai#48800)4c8cf3964dFix SGR mouse reporting for Windows terminal capture (openai#48799)d7748e1185Add opt-in structured errors for Guardian circuit-breaker interruptions (openai#48796)ea64727556Add single-server MCP status discovery with thread connection reuse (openai#48783)21eb35513dPreserve independent Guardian history across parent compaction (openai#48779)89bf86d0bdRemove thecurrentbadge from TUI task rows (openai#48776)596f8c5c3cMatch pinned transcript headers to the original prompt style (openai#48775)fdbce2080cFix Unix socket connections through long symlink paths (openai#48772)32f5784851Preserve MCP app resource URIs without defaulting display mode (openai#48764)e75b6b1e0cShow hidden output line counts in compact terminal activity (openai#48761)71b38795d8Match TUI status shimmer timing to desktop headers (openai#48757)819cdb726dRender/statuswithout borders and wrap long values (openai#48754)18344a972dCentralize executable fixture creation to avoid Linux ETXTBSY races (openai#48727)88235f881dRetain confirmed Code Mode messages for Guardian reviews (openai#48725)274d41a398Prevent Linux ETXTBSY races in MCP stdio tests (openai#48724)41f9084b30Remove WebSocket headers and tool payloads from info logs (openai#48686)67a709665aFix the session-start helper call in the command center test (openai#48646)d8ec479c34Set the provisioned macOS CLI bundle name to ChatGPT (openai#48643)8f195c93d7Preserve blank TUI sessions when switching tasks (openai#48628)449d42ced9Stop showing previous-session summaries when switching TUI sessions (openai#48626)98072cf5f6Preserve empty Markdown list markers in the TUI (openai#48623)334b6e7321Remove follow-up prompt suggestions from the TUI (openai#48621)814de47b69Centralize persistent mode enablement checks (openai#48611)9db8162d65Remove the bundledplugin-creatorskill (openai#48604)985cf47a4eAllow provisioned executors more time to come online (openai#48575)0f9a731adePreserve deferred tool namespace names before descriptions (openai#48574)b8d5e3f12eAllow exec-server to proxy permitted private IPs upstream (openai#48568)228ae3da8dAllow macOS TLS trust evaluation in network-enabled Seatbelt profiles (openai#48565)e8fdbf1f7cUse a consistent borderless session header in the TUI (openai#48562)6a39914e37Keep working tips stable during transcript interaction (openai#48560)7b7d934408Fix TUI math rendering for zero and big wedge expressions (openai#48551)75a714843bPreserve Markdown tables and whitespace when copying TUI responses (openai#48549)16c21e3f36Preserve table cell source metadata through TUI rendering (openai#48548)d6f25a54baFade blossom replays back to the idle state (openai#48547)7ed14a27dbMake onboarding login links easier to copy (openai#48544)06f97622f8Add context to Windows sandbox runtime registration errors (openai#48531)7f6c0f9387Refresh the TUI welcome screen for new sessions (openai#48513)12de0e395dPreserve WebSocket continuations when steering a turn (openai#48508)0fbf0bedc2Fix ChatGPT browser sign-in for local app servers (openai#48502)a6bd19261cFall back to embedded mode under restrictive Windows launchers (openai#48491)6e1ab4d294Fix Mermaid shape, relationship, and state description parsing (openai#48489)1a89aec960Prevent console windows for piped Windows child processes (openai#48483)b334d5b3f2Default to copying transcript selections in more terminals (openai#48469)e72da2b538Stabilize skill catalogs across executor availability changes (openai#48353)ca41ed337fShow turn tips while working and after completion in the TUI (openai#48352)09a3d2108fDisplay reconnect commands on a separate line (openai#48350)c9e2520707Preserve tool metadata for OpenAI provider endpoint overrides (openai#48344)25270df261Keep TUI reconnect attempts running until the shared deadline (openai#48318)dfdb40cd0bPrevent Windows daemon launches from retaining launcher stdio (openai#48272)4b9e0cc77fSuppress console windows for local Windows MCP servers (openai#48238)10fd75bba7Extract Responses failure parsing into a dedicated module (openai#48229)5f3180c793Preserve model and access program pairs during compaction (openai#48224)63eb71c9daPreserve late result metadata for truncated code-mode calls (openai#48222)58670eeac4Isolate executable fixture copies in CLI tests on Linux (openai#48213)d5355e95efKeep Codex visible during external editor handoff (openai#48211)f5ffa46959Preserve queued output for observers during code-mode termination (openai#48207)4193f1acfbAdd a keep-and-next action to the warnings viewer (openai#48206)7703f8b354Dismiss viewed TUI warnings when closing the viewer (openai#48205)dd9a7101d9Extract Responses header conversion into a shared module (openai#48200)7596ab19a6Keep archived threads with empty previews visible in listings (openai#48199)8f5387d104Honor execution environment proxy requirements (openai#48198)888948e522Optimizeblake3in Bazel fastbuilds (openai#48197)782826663dBound agent message board SSE frames before parsing (openai#48190)8f0b05910eFix zsh alias quoting in sourced shell snapshots (openai#48187)1d804e91b7Protect.awsdirectories under sandbox writable roots (openai#48176)44a9bfa145Preserve usage limit windows in turn and compaction analytics (openai#48174)f6a4bd81e3Generate unique exec-server process IDs for every request (openai#48168)a0b85c7a66Fix Guardian retained context spacing and empty assistant handling (openai#48158)8b78f47966Allow Windows daemon launches with residual job membership (openai#48157)645b683a9ePreserve filesystem denials when preparing approved commands (openai#48155)de9e78e3e7Add item anchors tothread/items/listpagination (openai#48151)c7e80f873fPreserve executor MCP credential boundaries across reconnects (openai#48143)f92655d07fPreempt model responses when new user input arrives (openai#48141)17f37f56beSuppress console windows when spawning the code-mode host on Windows (openai#48138)1bf73324caAdd opt-in code-mode yielding on new user input (openai#48135)d9275a63edAllow Left to open the command center from read-only conversations (openai#48132)0b430baeb4Use request notifications in the cloud config loader lifetime test (openai#48130)55543d8772Add early yielding for code-mode observations (openai#48123)66c4e5a20dKeep startup drafts visible during command center session handoff (openai#48121)68e0c9f5d8Prevent worker completion races in the guardian authorization test (openai#48119)b8a1fe4afcAdd configurable right-click paste to the fullscreen TUI (openai#48118)bf832f4678Allow reasoning shortcuts to reach Max (openai#48116)bd3d4d1436Preserve user text parts during local compaction (openai#48115)85928697a4Deduplicate retained instructions in async Guardian context (openai#48110)d5cbfe1455Deduplicate retained instructions against Guardian transcripts (openai#48109)e75b36efdeShow multiline command previews in/ps(openai#48101)6d083e3bc4Add an HTTP client for remote agent message boards (openai#48100)b63a296775Honor shell environment policy in legacy snapshots (openai#48099)b35a7afbe8Preserve recent authorization context for Guardian reviews (openai#48098)c98e263fb5Replay exec-server shell snapshots through unnamed files (openai#48078)466d5f583eAdd Serde support to agent message board request types (openai#48077)60713126eeAvoid stdin approval for runtime-only permission grants (openai#48073)ca60d6fa72Skip message-board notification previews when there are no recipients (openai#48072)69fa3881cbHandle early command yields in the Guardian network approval test (openai#48069)86be5320b0Deduplicate retained instructions across Guardian reviews (openai#48060)4b1c0c30daRemove plugin extension metadata from discovery and summaries (openai#48035)d7b07d4551Test same-cell permission grants and strict review in code mode (openai#48017)d0a64e91aeValidate tool suggestion install URLs before showing the app link (openai#48015)8ae55c863dRespect configured authentication in the thread manager sample (openai#48004)aa380897f6Add startup-only PID namespace inheritance to exec-server (openai#47989)0f0efab8d3Reuse MCP handlers across equivalent bindings (openai#47988)9ef08dcf2bAdd multi-agent spawn latency and failure metrics (openai#47984)75e0e0aad9Prepare MCP calls directly from advertised tool identities (openai#47981)1f17a0a04bPrevent stale voice answers from reappearing during speech recovery (openai#47975)a92ccbde53Preserve Git directory protections across writable roots (openai#47974)b725da3b6dAdd Pro Max plan support and update Pro display names (openai#47971)97da50c1aaExpose current environment selections for a running turn (openai#47970)a708fc6839Handle Btrfs device mismatches when masking daemon sockets (openai#47968)dafb133c5bSurface Flex capacity failures as a distinct terminal error (openai#47967)8f103417efPreserve the client-agent header for Amazon Bedrock Runtime (openai#47964)8dd0a08160Request transparent huge pages for Cargo and eligible Bazel rustc jobs (openai#47962)dbb875d23cBound tool-call observations to the outgoing Responses message budget (openai#47957)51d66fdd3bSupport file references in image edit requests (openai#47956)40e13e6281Move fullscreen startup tips into the transcript (openai#47954)9442a175baPrune expired in-memory message board registry entries (openai#47952)4a358b1ae9Use prebuilt V8 archives for Bazel on macOS and GNU Linux (openai#47951)cf792c3a25Expand root authorization context to 16 messages (openai#47947)dda227891dAdd an in-memory agent message board for ephemeral sessions (openai#47946)5d80031172Parameterize the thread initialization analytics test by originator (openai#47945)549455f3ecRemove unused Windows world-writable audit code (openai#47943)e8098eb406Separate selected plugin identities from MCP contributions (openai#47939)fbd0c0d96cAdd direct replies for thread settings updates (openai#47937)339e981ba7Make MCP and Code Mode input schema budgets configurable (openai#47936)b62fcd3c4bAllow cached catalogs to satisfy MCP startup readiness (openai#47935)af09e926d6Apply the unchanged-model compaction shortcut to all session sources (openai#47934)694d8d45bdRemove GPT-5.4 from bundled catalogs and preserve migration prompts (openai#47932)e154f7b807Generate TUI prompt suggestions after successful turns (openai#47929)4b97832cfbUse127.0.0.1for local login redirects (openai#47927)0037b39cc1Retry file blob uploads on HTTP 502 and 504 (openai#47926)1866e51677Make project trust lookup paths explicit and defer root resolution (openai#47924)731d4a5ca8Allow full-access Windows setup to provision through registered Core (openai#47922)c19dcd975dAllow directory moves under global Seatbelt basename denies (openai#47920)c8c1ee5da8Transport large Windows sandbox launch payloads through the environment (openai#47919)3e27195f2dParameterize the turn-start originator header test (openai#47918)edd0df9084Reuse verified V8 checksum manifests from the artifact cache (openai#47915)766d2377a8Add an opt-in flag to defer mailbox preemption (openai#47913)a8a4f14087Fix attestation routing during thread startup (openai#47912)6743672ad7Add TUI composer support for prompt suggestions (openai#47911)5c8fc15cc9Aliastui.whimsytotui.effects.starfield(openai#47908)f0f38b68f7Support nested canonical paths in config key aliases (openai#47904)0b78ddb035Move config key alias normalization ahead of merging (openai#47903)975d30c043Avoid repeated table clones during config merging (openai#47902)2347a61d0dAdd bounded credential-storage telemetry helpers (openai#47901)12cb14f7b7Default local threads to paginated history (openai#47900)62197a65e5Add diagnostic reasons to MCP attribution errors (openai#47899)2457571e4cPreserve local-binding inheritance in environment network policies (openai#47898)a33fb9751cPreserve Markdown formatting when copying transcript selections (openai#47896)f6fd7f17edClean up temporary Codex homes after TUI tests (openai#47894)83b56bc5adSupport client secrets for pre-registered MCP OAuth clients (openai#47891)b3ea9fa632Include TUI client logs in diagnostic uploads (openai#47889)3d6428510fWarn users when SQLite diagnostic log writes fail (openai#47887)a285066737Preserve diagnostic logs when SQLite logging fails (openai#47886)90b67c96caUse Tokio's clock for TUI paste timing (openai#47881)6f51c65958Fix macOS system-alias matching in patch permission checks (openai#47879)4021746a15Measure deferred tool namespace fragments before and after truncation (openai#47873)42266240fdFix PID reservation test race and update guardian heartbeat snapshot (openai#47871)30758e5c77Add opt-in OTLP logging for Guardian assessments (openai#47870)2deb542491Render remote permission paths using executor context (openai#47867)3e9d1d2937Avoid nested read locking when rendering browser sign-in (openai#47861)58838dfa5bValidate loaded plugins outside the cache lock (openai#47858)df06bb054cInitialize media estimates outside the global cache lock (openai#47856)421082e79dAvoid blocking async proxy resolution on the system proxy cache (openai#47852)e7f119819aPreserve human overrides across repeated heartbeat instructions (openai#47851)12fd929f72Keep the TUI responsive during clipboard copies (openai#47847)282cd7b019Stabilize Rosetta test timing and retry delay telemetry (openai#47832)61e23bc6a1Bind Guardian async scores to target environment permissions (openai#47830)b15718acfcWait for login completion in recommended plugin tests (openai#47828)4083a68f88Allow four concurrent threads in the multi-agent resume test (openai#47824)b19cebecc0Add integration coverage for host agent controllers (openai#47820)49862b62beRetry Guardian reviews when authorization changes (openai#47819)16b20547f6Stabilize thread resume and memory dual-write tests (openai#47817)869b5527cdFix a lost wakeup in the unified exec termination test (openai#47814)8ac495b799Fix sleep interruption test event handling and fixture lifetime (openai#47813)1cd253c987Preserve explicit user goal updates in Guardian authorization (openai#47811)fdbe23da45Allow hosts to provide agent controllers through ThreadManager (openai#47808)c098f97e53Enable terminal input approval by default (openai#47799)51d4562070Support close-on-exec attachments without changing PTY semantics (openai#47797)29f056c26cHonor catalog schemas for asynchronous user input (openai#47773)53446f90a5Preserve more tool metadata within outgoing message budgets (openai#47758)b0a9dcb843Refactor tool telemetry product SKU matching to use an allowlist (openai#47757)a83ba61249Centralize typed app-server response decoding (openai#47755)e0ef5a1a0fReduce generic code duplication in RPC and Markdown rendering (openai#47751)a16381c445Align Cargo and Bazel Rust debug information defaults (openai#47748)a98a07759aSkip startup prewarm preparation when the WebSocket is ready (openai#47745)1d87af5faaHonor network policy in history notes and image generation extensions (openai#47742)5f2ce5e02cAttribute tool telemetry to the invoking turn's product SKU (openai#47741)7dae8c53d9Avoid recursive TUI event dispatch for model picker selections (openai#47717)e7165c2dbbPreserve tool result metadata more selectively under size limits (openai#47714)f5960fcc22Reduce dependency coupling in shared configuration crates (openai#47713)4891c4e35fUpdate unified exec output buffers atomically (openai#47712)f5f08c54cbRoute resume prewarm through the cached WebSocket session (openai#47709)6989c6548bFix spawn flag typing and isolate project configuration tests (openai#47704)d04b0561d3Preserve account network policy for ChatGPT backend requests (openai#47703)d838c2346dAllow idle threads to prewarm and repair WebSocket connections (openai#47701)4b3b569f39Allow WebSocket test server shutdown while waiting for requests (openai#47698)8215b81920Avoid the shutdown timeout in the lagged-event test (openai#47696)17cd2834e3Repair rejected Windows sandbox credentials during provisioning (openai#47695)80a0c1da88Configure curl retries for DotSlash installation in CI (openai#47693)b4b055cfc8Materialize rollout persistence for pending inter-agent messages (openai#47691)7b66c21f58Remove obsolete Guardian context capture mode branches (openai#47690)920e1a34dcMake Guardian thread context capture unconditional (openai#47689)1d516d8b3dRemove legacy Guardian authorization evidence paths (openai#47688)a6df7385e6Make thread-owned Guardian context always enabled (openai#47686)ab7439231cAdd executor capability discovery V2 infrastructure (openai#47683)6fa3bcaae8Add exec-server RPC timing and process startup tracing (openai#47680)8de2d336b8Add extension hooks for model requests and response streams (openai#47679)9477081ccdSupport quoted labels and ampersands in Mermaid flowcharts (openai#47678)b2aa42e6f1Support model catalog overrides for MCP resource tool specs (openai#47677)c530382cf3Clarify registered Windows sandbox setup errors (openai#47673)5f371ba30aFix no-reparse directory opens on Windows 10 (openai#47672)f2654ea6c9Support model-specific descriptions for agent message board tools (openai#47670)d47b9a8c00Preserve early unified exec output in completion events (openai#47665)8e17909b27Preserve managed network policy in route-aware transports (openai#47663)a79a18cd0eExpose tool dispatch and timing observations to extensions (openai#47662)5bae5b563eRestrict the default Bedrock GovCloud model catalog (openai#47657)7e5054d32fBump the exec-server stable compatibility test to Codex 0.156.1 (openai#47655)b086ad5493Make Linux descriptor cleanup fork-safe (openai#47654)5bfb04b247Attach inherited rollout history to diagnostic reports (openai#47653)e6da6fd7c7Add opt-in OTLP logging for final agent responses (openai#47649)ad26b2520aSupport bearer tokens for app-server executor connections (openai#47648)e9b4c6967eApply Guardian computer-use review to the Browser connector (openai#47647)8c66a5a243Add model-specific prefixes to indirect tool descriptions (openai#47642)9d8de19674Honor Retry-After and preserve server retry deadlines (openai#47641)ab021e9f13Add configurable copy-on-select for transcript selections (openai#47639)c847294c60Classify retryable exec-server preparation errors by type (openai#47638)ce7df36237Overlap startup WebSocket preconnect with tool discovery (openai#47635)dce62a50a2Route message board agent resolution through the selected controller (openai#47633)851d9e95f1Bind Guardian reviews to the action's target environment (openai#47630)e1f3fafad8Route V2 child loading throughAgentControl(openai#47629)fdeaa6c1ecAllow history and notes without experimental context capability (openai#47625)f07b26c9c9Recognizeuser_messagetools in Guardian authorization context (openai#47624)8bbe8f8702Fix read-only metadata mount ordering for nested writable roots (openai#47623)13869ca39eAdd portable project trust lookup APIs (openai#47620)0ed77c274bAdd bounded buffering for global operation metrics (openai#47619)591ffb1aaePrefer Shift-arrow hints for queued messages and questions (openai#47618)88b498ded4Route Linux PTY launches through the process setup helper (openai#47617)15922a50aaExpand Linux spawn-helper lifecycle test coverage (openai#47613)df3ecee6f0Launch Linux pipe processes through a fresh setup helper (openai#47612)60f0ab6cbfUse the shared process launcher for Unix shell snapshots (openai#47611)7aa6519909Use native POSIX spawning for command hooks (openai#47610)50fc133d05Route pipe processes through the shared child launcher (openai#47605)b8d365b18eExtend child commands with session and descriptor controls (openai#47604)897b7ae285Add a reap-only drop policy for child processes (openai#47603)2210190435Add opt-in WebSocket authentication to exec-server (openai#47601)6824dabe03Guard prerelease channel and canary updates against older versions (openai#47597)c114f232f9Refresh realtime context for each model request (openai#47596)8f1490eabeStream daemon executable hashing off the async runtime (openai#47591)ae132dc50aSerialize numeric custom reasoning effort as JSON numbers (openai#47590)e0a64cf2bcKeep unfinished link destinations out of rich streaming previews (openai#47589)7342991f0cPreserve delivered assistant messages in Guardian retained context (openai#47585)876ce590e1Preserve streamed assistant message order in retained context (openai#47584)29aa1df62eRetain assistant context for Guardian authorization reviews (openai#47582)cb1eea3e98Explain Mermaid rendering fallbacks in the TUI (openai#47572)acd5ed9207Wait for idle before injecting remote compaction test history (openai#47571)6f6af0fce4Record sandbox backends in command execution analytics (openai#47568)d6c4c6aea4Classify rollout read failures by reason and progress (openai#47565)30fc6864ccAdd an option to disable multi-agent v2 direct messaging (openai#47540)ec4d27ae80Test interrupted one-shot command launch failure persistence (openai#47539)0b23f365f8Route agent lookups and V2 context throughAgentControl(openai#47536)a69d757cd8Emit command lifecycle events for unified exec launch failures (openai#47529)27969c0ae9Route session agent operations throughAgentControl(openai#47520)40eac3ce8aAdd explicit background control to image generation (openai#47484)0a2eb4696cParallelize instruction refresh and tool preparation (openai#47458)c44deff7b1Extract WebSocket authentication intocodex-websocket-auth(openai#47447)4981037e99Use secondary text styling for the transcript footer shortcut hint (openai#47441)b7add4df3dStabilize retry timing and WebSocket tests (openai#47437)370cd7cba8Allow environment selection updates during an active turn (openai#47435)8a3c4ea3b5Apply inherited environment settings at turn boundaries (openai#47428)408a77dc1aClear pending async question notifications when turns end (openai#47424)3032b387faPreserve history search during background draft recovery (openai#47423)811fe5fa43Recover unsent question drafts when TUI turns end (openai#47422)2925d06f5aUse native scrollback for Terminal.app over SSH in auto mode (openai#47417)39598ed178Support Shift-click to extend transcript selections (openai#47414)53486e5bdcCache decrypted gateway OAuth secrets per backend (openai#47413)9d4d34d436Apply shared network policy throughout embedded Codex startup (openai#47411)db3cb33ecaHonor network policy in remote control and recover remote execution (openai#47410)8f8ace78cbEnforce application network policy for AWS auth and telemetry (openai#47408)22a3f6d5d8Enforce application network policy across app-server requests (openai#47407)4b664e0ef0Respect tmux mouse settings in fullscreen and overlays (openai#47399)24462234b2Refresh bundled model metadata and instructions (openai#47397)2c3295306fRetry transient OpenAI file blob upload failures (openai#47393)888e02db34Enforce network policy throughout HTTP and WebSocket requests (openai#47389)506a328dabRestore composer hints after clearing selection at the transcript bottom (openai#47384)58d722cb5dShow a voice badge in the agents overview (openai#47382)a837a656ffKeep voice conversations running across TUI thread navigation (openai#47381)72ec34ee9aRoute TUI voice controls through the app (openai#47380)90e653014aAdd opt-in reasoning status for realtime V3 delegations (openai#47377)418199f6adAdd an opt-in preference for the local MXC sandbox (openai#47375)78fdcd2311Overlap agent metadata reads when resuming a V2 root (openai#47369)e7bbc79f48Resume model context from the latest compaction boundary (openai#47365)54487a5b61Bound inbound exec-server requests across client transports (openai#47362)c6ad6f33ebRestrict Windows sandbox default object access to the logon session (openai#47361)