202610020008

Upstream sync

No upstream changes in this release.

Full Changelog: https://github.com/yuguorui/codex-plus/commits/rust-v202610020008

202610011928

Upstream sync

• No upstream changes in this release.

Full Changelog: https://github.com/yuguorui/codex-plus/commits/rust-v202610011928

202610010642

Upstream sync

Changes synced from openai/codex this release, grouped by topic:

• SQLite and storage resilience: SQLite corruption is now classified with typed error codes and detected during startup, with recovery backups preserved. Diagnostic logs are pruned periodically by age and database size, and graceful-shutdown guard traces are filtered from SQLite logs.
• Async and blocking I/O performance: session index I/O, rollout listing scans, thread history projection, and compressed rollout snippet searches now run on dedicated blocking or cancellable workers. Token-budget truncation avoids full-string scans, reverse JSONL scans use memrchr, and resuming a thread reuses unchanged history snapshots and returns authoritative replay history.
• Authentication and account security: the TUI shows account security setup reminders, API-key cyber access programs are decoupled from model discovery, and explicit cyber access programs work with OpenAI API keys. Explicit remote session commands use server authentication, bootstrap discovery survives authentication changes, MCP authorization servers are validated before ID-JAG exchange, and enterprise-managed token exchanges use the rmcp SDK.
• Credentials and telemetry: credential storage outcomes are tracked with sensitive errors redacted, and attributed MCP OAuth credential storage telemetry is emitted.
• TUI: selected file paths can be copied as plain text, the command center gains a fork shortcut, voice settings gain local audio device selection, recovered question answers escape command drafts, permissions are server-authoritative, and randomized session header greetings were removed.
• Voice: a managed feature gate for in-app voice was added.
• Multi-agent and model history: an opt-in model catalog is available to multi-agent context, experimental thread prediction protocol types were added, and explicit user goal edits persist in model history.
• Windows sandboxing: elevated Windows sandboxes preserve PowerShell relative paths, managed requirements can disable the Windows MXC sandbox, and Windows UNC and opaque-URI path inference handles forward and mixed slashes.
• Networking: Responses routing fields are serialized before large inputs, and server retry advice is honored across Responses retries and fallback.
• exec-server: file reads are cancellable between chunks, sessions recover after environment info timeouts, and executor tool paths are restored after login shell startup.
• Diagnostics and tooling: skill invocation events are exported through OpenTelemetry, multiline ANSI warnings omit payloads, protocol debug output truncates input text, hook matchers compile during discovery, and host skill discovery is skipped for Guardian reviews.
• Other: remote message board notifications reach active turns, repository-local Codex guidance, skills, and environment config were removed, analytics is enabled by default for daemon-launched app servers, and an experimental flag controls bundled tools in login shells.

Full Changelog: https://github.com/yuguorui/codex-plus/commits/rust-v202610010642

202609301047

Upstream sync

No upstream changes in this release.

Full Changelog: https://github.com/yuguorui/codex-plus/commits/rust-v202609301047

202609300650

Upstream sync

• Model catalogs: Added GPT‑6.1 Sol to the Bedrock catalogs and made it the default catalog model, updated the Bedrock catalog tests to expect multi‑agent V2, and enabled multi‑agent V2 plus Ultra reasoning on Amazon Bedrock.
• Sign-in and credentials: Clarified credential storage wording across the authentication UI and docs, added /mcp login <name> to the TUI, and enabled enterprise MCP sign‑in with account‑scoped grant cleanup.
• Sandbox and shell execution: Allowed approved filesystem escalation while still preserving denied reads, retried the Windows sandbox runner logon once on error 1056, ran piped legacy Windows sandbox processes without a console, and restricted capability roots to the captured turn environments.
• Exec server and sessions: Cleaned up canceled exec‑server RPC requests immediately, modeled exec‑server session attachment state as an enum, and fixed a turn teardown race in the Guardian agent message test.
• Guardian and agents: Notified parent agents when Guardian stops a subagent, and recorded Guardian context mode in review and classification telemetry.
• Thread naming performance: Batched metadata reads when resolving thread names and scanned the session index backwards for batch thread name lookups.
• Pasting and buffers: Continued Markdown blockquotes when pasting multiline text, and compacted the inline hidden tag buffer once per chunk.
• Shell metadata: Carried shell invocation metadata and reported executor PATH directories.
• Configuration: Simplified configuration fingerprint canonicalization.
• Connection resilience: Kept remote control reconnect backoff capped during sustained failures.
• Test isolation: Isolated realtime conversation tests from Responses prewarm connections.

Full Changelog: https://github.com/yuguorui/codex-plus/commits/rust-v202609300650

202609292249

Upstream sync

This release syncs the changes below from upstream openai/codex, grouped by topic.

• Configuration and policy: preserve thread overrides and cloud policy validity during config reloads; restrict enterprise MCP auth and fail closed on config refresh; honor app-server provider defaults in the TUI; treat explicit provider model catalogs as authoritative; simplify cloud task base URL normalization.
• Reasoning, retries, and errors: preserve server reasoning summary and verbosity settings in the TUI while hiding those settings in /status for server connections; add recovery guidance to content-filter retries and move them into the shared Responses retry handler; expose original error details to turn lifecycle contributors.
• Multi-agent sessions and history: support remote agent message boards in multi-agent sessions; add history pagination to the agent command center; trace turn phases and correlate accepted input with turns.
• TUI experience: support projectless TUI sessions with workspace defaults; fix model provider lookup for TUI history; add X11 primary selection and middle-click paste; omit blockquote markers when copying quoted selections; resume unsent input after reconnecting; tidy key hint separators.
• Sandbox and Windows: preserve Windows sandbox runner launch errors; resolve Windows sandbox PowerShell fallbacks on the exec server; suppress Windows console windows for background subprocesses; allow Guardian cached approvals with incomplete root context.
• Plugins, skills, and storage: reuse the HTTP connection pool for remote plugin requests; cache parsed plugin manifests across plugin workflows; deduplicate cloud and executor skill listings before budgeting; preserve SQLite vacuum modes and surface pool initialization errors.
• Telemetry and lifecycle: attribute analytics requests to each thread's product SKU; notify lifecycle extensions of compaction usage limits.

Full Changelog: https://github.com/yuguorui/codex-plus/commits/rust-v202609292249

202609290759

Upstream sync

This release syncs a large batch of upstream changes from openai/codex, grouped by topic below.

TUI experience

• Follow-up directive labels now render in the TUI and in copied responses, and TUI subscription labels are updated and centralized.
• Added several status-line and footer improvements: a short, neutral interruption notice, short turn durations in completion footers, the Plan mode cycling hint in the fullscreen status line, and updated Pro plan display names.
• Transcript scrolling with the mouse wheel now works while a modal is open, and a hand pointer is shown over transcript links in Ghostty and Kitty.
• Selections inside inline code copy as plain text, ordered Markdown list markers follow the terminal palette, and SGR mouse reporting works again in Windows terminal capture.
• Realtime voice catalog failures are now surfaced in the TUI, and startup promotions were simplified to platform-specific desktop app tips.
• Sign-in success copy for ChatGPT was clarified.

Mermaid and Markdown

• Native Mermaid flowchart syntax support was expanded, and punctuation and semicolons are now preserved in Mermaid labels.

Sandbox, Windows, and macOS

• Configuration values are kept out of Windows sandbox policy events, Windows sandbox ACL repair now handles long runtime paths, and the Windows MXC sandbox uses a compatible PowerShell fallback.
• The Windows sandbox provisioning service is now given a brief wait on startup, macOS sandbox policy uses the numeric ioctl value, and remote Git discovery is skipped for Guardian diff paths.

Guardian reviews

• Guardian handoff and messaging snapshots were updated for separate and native agent messages, handoff-aware root context was added, and encrypted agent messages are preserved during reviews.
• Added opt-in conversation history retrieval for Guardian reviews, plus opt-in structured errors for circuit-breaker interruptions.

Sessions, threads, and app server

• App-server running turns are now tracked incrementally, threads can be archived before their first turn, and idle threads support history-aware prewarming.
• Thread timestamp updates avoid full metadata rewrites, and message-board notifications no longer reopen final answers.
• Pending environments are preserved when spawning subagents, and skill analytics no longer collect unused Git metadata.

Performance, storage, and observability

• Unused SQLite log database pages are reclaimed in the background, and SQLite stalls from connection setup and stderr span logging were addressed.
• Explicit histogram buckets were added for tool and skill context metrics.

Platform and build

• h2 was updated from 0.4.16 to 0.4.19 in the Cargo and Bazel lockfiles, and Cargo package versions are now propagated to Bazel Rust targets.
• Single-server MCP status discovery now reuses thread connections.

Full Changelog: https://github.com/yuguorui/codex-plus/commits/rust-v202609290759

202609280530

Upstream sync

Changes synced from upstream openai/codex in this release, grouped by topic:

• Guardian reviews: Guardian history is now preserved independently across parent compaction, and confirmed Code Mode messages are retained for Guardian reviews.
• TUI rendering: Removed the current badge from task rows, matched pinned transcript headers to the original prompt style, aligned TUI status shimmer timing with desktop headers, showed hidden output line counts in compact terminal activity, and rendered /status without borders while wrapping long values.
• TUI sessions: Blank sessions are preserved when switching tasks, previous-session summaries are no longer shown on session switches, and empty Markdown list markers are kept intact.
• TUI prompts: Removed follow-up prompt suggestions.
• Connections: Fixed Unix socket connections through long symlink paths.
• MCP: MCP app resource URIs are preserved without defaulting the display mode.
• Logging: Removed WebSocket headers and tool payloads from info logs.
• Persistent mode: Centralized persistent mode enablement checks.
• Skills: Removed the bundled plugin-creator skill.
• Packaging: Set the provisioned macOS CLI bundle name to ChatGPT.

Full Changelog: https://github.com/yuguorui/codex-plus/commits/rust-v202609280530

202609271126

Upstream sync

• Networking and sandboxing: exec-server can now proxy permitted private IPs upstream, provisioned executors get more time to come online, and macOS TLS trust evaluation works in network-enabled Seatbelt profiles.
• Tooling: deferred tool namespace names are now preserved before descriptions.
• TUI session UX: a consistent borderless session header, working tips that stay stable while the transcript is interacted with, and faded blossom replays that return to the idle state.
• TUI rendering and copy: math rendering is fixed for zero and big wedge expressions, Markdown tables and whitespace are preserved when copying responses, and table cell source metadata is kept through rendering.
• Onboarding: login links are easier to copy.

Full Changelog: https://github.com/yuguorui/codex-plus/commits/rust-v202609271126

202609270532

Codex++ changes

• TUI: added a typing-reactive Bongo Cat pet that reacts to your keystrokes.
• Added support for dynamic workflows.
• CLI commands and the release version are now branded as Codex++.
• Custom model providers: added support for Chat Completions and Anthropic Messages wire formats.
• Custom model providers can now send model-specific extra body fields with requests.
• Custom model providers now retry 429 rate-limit responses.

Upstream sync

• Windows: more reliable sandbox and daemon behavior, with clearer sandbox runtime registration errors, embedded-mode fallback under restrictive launchers, and no stray console windows for piped child processes or local MCP servers.
• TUI: refreshed welcome screen for new sessions, turn tips while working and after completion, reconnect commands on their own line, reconnect attempts that continue until the shared deadline, and transcript selections that default to copy in more terminals.
• Session handling: WebSocket continuations are preserved when steering a turn, and ChatGPT browser sign-in works again for local app servers.
• Skills: skill catalogs now stay stable as executor availability changes.
• Rendering: fixed Mermaid parsing for shapes, relationships, and state descriptions.
• Providers and responses: tool metadata is preserved for OpenAI provider endpoint overrides, and Responses failure parsing moved into a dedicated module.
• State integrity: model and access program pairs are preserved during compaction, and late result metadata is preserved for truncated code-mode calls.

Full Changelog: https://github.com/yuguorui/codex-plus/commits/rust-v202609270532