Skip to content

CHANGELOG.md

Latest commit

 

History

History
548 lines (407 loc) · 24 KB

CHANGELOG.md

File metadata and controls

548 lines (407 loc) · 24 KB
Raw

Changelog

All notable Codex VL changes are tracked here.

Codex VL tracks OpenAI Codex upstream, but this changelog only covers fork-specific work.

0.153.2-vl.2 - Upstream rust-v0.153.2

Codex VL fixes

  • Fleet TUIs now require a verified cell identity before attaching to a shared app-server; unverified sessions use an embedded fallback with a visible diagnostic, while explicit shared endpoints are rejected.
  • Catalog-supplied persistent_instructions and Guardian node_repl_policy have an 8 KiB byte cap at every ingress and consumer: oversized values are rejected, never truncated, and Guardian fails closed with a clear diagnostic.
  • Release infrastructure keeps V8 guarded, runs SDK checks on hosted runners, refreshes zlib snapshot URLs, and preserves the README ASCII policy.
  • A dedicated CI gate covers the identity and model-message cap paths.

0.153.2-vl.1 - Upstream rust-v0.153.2

Upstream

  • Rebases Codex VL on OpenAI Codex rust-v0.153.2, including the GPT-6-Astra model catalog and the latest public model-tier updates.
  • Hooks activity is visible in the TUI, with Vim undo/redo and app-server reconnect improvements for longer-running sessions.
  • App-server adds asynchronous user-input requests and richer thread metadata, while [tui].disable_paste_burst and experimental context management expand configuration control.

Codex VL fixes

  • Loop and Vivling slash commands keep the TUI alive when a user precondition fails, including owner or delegation requests with the brain switched off.
  • manage_loops retains access to its jobs after resume, and loop ticks accept fenced or prefixed JSON replies from any provider.
  • The expression planner no longer floods the log database, while terminal turn reconciliation removes phantom “Working” state and stale queued input.
  • Esc interrupts a running task even when a long tool hides the status row, and /loop usage now lists delegate, strategy, delegation, and owner commands.
  • Termux: the npm launcher now fixes its own shebang on install, so codex-vl starts without termux-exec tricks.

0.150.2 - Restore embedded TUI loop routing

Fixed

  • Restored the pre-merge routing for the manage_loops dynamic tool in the embedded TUI app server: manage_loops calls reach the loop controller again instead of being rejected with "TUI dynamic tools require an active external task". The upstream gate still applies to every other dynamic tool in embedded sessions.
  • Added two regression tests guarding the embedded TUI routing boundary.

0.150.1 - Upstream rust-v0.150.1

Update onto the upstream OpenAI Codex rust-v0.150.1 stable release, published on the latest channel. The complete Codex VL workflow layer remains present across goal state, loop jobs, Vivling, the VL interface, remote control, app-server integration, and the fork-owned package and update channels.

Fixed

  • The manage_loops built-in dynamic tools are now injected only for the TUI client. Every thread/start used to receive them, including app-server clients with no loop controller on their side of the connection.
  • Update banners render the fork's own channels again. The history cell snapshots had pinned the upstream package name and installer, so the redirect the updater already performed was never reflected in the copy.
  • Two fork test sites imported ImageDetail from its pre-0.150 module path, and the fork retry harness in the installed-apps suite lost its attempt counter to an upstream-shaped merge; both compile and run again.

0.147.0 - Upstream rust-v0.147.0

Update onto the upstream OpenAI Codex rust-v0.147.0 stable release, published on the latest channel. Upstream tagged it a day after 0.147.0-alpha.13; the two are siblings off the same parent and differ only in the workspace version. The full Codex VL workflow layer remains present across goal state, loop jobs, Vivling, the VL interface, remote control, app-server integration, and the fork-owned package and update channels.

Fixed

  • An AGENTS.md created during a session — which is what /init does — is now discovered. Discovery cached its result under the environment selection, which does not change when a file appears, so the session kept reporting that none existed.
  • The model catalog parses again. Upstream added its own legacy base_instructions key whose serializer also flattens ModelInfo, so with the fork field serialized too the catalog carried the key twice.
  • Models whose catalog entry ships no instruction template get the fork's fallback instructions again. The merge had replaced that branch with upstream's, which returns an empty string; a behavioural test now fails if it is emptied again.
  • musl packaging keeps working: the Linux workflows extract exactly two V8 checksums per target from the versioned manifest, and upstream ships that file for 150.4.0 with Windows entries only.

Changed

  • The state runtime keeps the fork's VL migration reconciliation around upstream's refactored DB open path.
  • manage_loops still resolves as a dynamic tool call, after upstream's new rejection path for requests belonging to abandoned side threads.
  • MCP startup keeps its bounded tools-list retry, adapted to upstream's new catalog item limit.

Removed

  • The Termux TLS root patch carried here for parity: upstream removed the client it wrapped and reqwest left the crate's dependencies, so it no longer compiled.
  • The fork's by-name session lookup test, superseded by upstream's named_session_lookup and its own tests.

0.146.1 - Upstream rust-v0.146.1

Update on the upstream OpenAI Codex rust-v0.146.1 patch release (backported safer cyber-model auto-review defaults). The merge touched no fork-owned path. Published on the next channel; latest later moved to 0.147.0. (Amended: this entry originally described a gated candidate.)

The full Codex VL workflow layer remains present across goal state, loop jobs, Vivling, the VL interface, remote control, app-server integration, and the fork-owned package and update channels.

Fixed

  • The Windows installer resolves releases from the fork only. It previously preferred upstream's distribution channel with an upstream GitHub fallback, so running it installed upstream Codex under the fork's name.
  • The message-history batch reader falls back when advisory file locking is unsupported, matching every other lock call site. Without it, batch history reads failed outright on Android/Termux storage.

0.146.0 - Upstream rust-v0.146.0 candidate

Candidate update on the final OpenAI Codex rust-v0.146.0 tag. The full Codex VL workflow layer remains present across goal state, loop jobs, Vivling, the VL interface, remote control, app-server integration, and the fork-owned package and update channels.

The merge retains the no-upstream-installer contract, MCP startup retry and provider tool-namespace support, while adapting the state runtime to upstream's resolved SQLite configuration. Legacy Vivling loop migrations remain reconciled before upstream state migrations run. Publication remains subject to the feature-register audit and sanitized artifact verification.

0.145.0 - Upstream rust-v0.145.0 stable

Stable release based on the final OpenAI Codex rust-v0.145.0 tag. The full Codex VL workflow layer remains available across goals, loop jobs, Vivling, the VL interface, remote control, app-server integration, fork-owned update surfaces, and the Linux, macOS, and Android package lanes.

This release also carries forward the explicit MCP environment-name allowlists introduced in 0.144.8. Linux x64, Linux arm64, and Android arm64 ship native packages; macOS arm64 remains a source-build package with its fail-closed prerequisite checks.

0.144.8 - Explicit MCP environment allowlists

Stable patch on the OpenAI Codex rust-v0.144.6 base. The mcp add command now accepts repeatable --env-var NAME options for local stdio servers whose identity or routing data already exists in the launching process environment.

Only variable names are validated and persisted. Values are resolved from the live process environment when the MCP server starts, remain behind the existing cleared-environment boundary, and are never copied into config.toml. Duplicate names collapse to their first occurrence.

0.144.7 - Darwin npm 12 install diagnostics

Stable patch on the OpenAI Codex rust-v0.144.6 base. On Apple Silicon, both launchers now distinguish a missing platform package, an installed platform package whose native binary is absent, and a native binary that cannot be executed. Each state reports an actionable, path-aware error.

The macOS recovery guidance now uses the complete npm 12 command with an explicit install target, a script allowlist limited to @mmmbuto/codex-vl, and foreground lifecycle output. README, postinstall diagnostics, fork identity pins, and CI fixtures enforce the same fail-closed contract.

This release also carries upstream's refreshed bundled model metadata and GPT-5.6 prompt/context update from rust-v0.144.6.

0.144.6 - Verified ripgrep packaging hotfix

Stable fork hotfix on the final OpenAI Codex rust-v0.144.5 base. Linux x64 and Linux arm64 packages now install the target-specific ripgrep 15.1.0 artifact from the canonical DotSlash manifest instead of copying the host binary or generating a GNU grep compatibility stub.

Every artifact is verified against its pinned size and SHA-256 digest before extraction. Missing targets, malformed manifests, download failures, and integrity mismatches fail the package build closed.

0.144.5 - Upstream rust-v0.144.5 next candidate

Candidate release on npm next, based on the final OpenAI Codex rust-v0.144.5 patch release. Upstream expands dangerous-command detection to cover additional forced recursive deletion forms before execution.

Loop-job refreshes now reuse the process-owned state database handle instead of reopening and migrating SQLite from the event consumer path. A passive refresh failure is logged and leaves the TUI running rather than terminating the app-server session.

0.144.4 - Upstream rust-v0.144.4 next candidate

Candidate release on npm next, based on the final OpenAI Codex rust-v0.144.4 patch release. Upstream reports no user-facing changes in this patch; the complete Codex VL feature layer and Android/Termux carry-forward remain intact.

Fork compatibility hardening allows model catalog entries to omit base_instructions. Explicit configuration overrides remain authoritative, while missing or blank catalog instructions safely fall back to the embedded base instructions instead of producing an instructionless session.

0.144.3 - Upstream rust-v0.144.3 next candidate

Candidate release on npm next, based on OpenAI Codex rust-v0.144.3. The complete upstream delta is combined with the complete Codex VL feature layer, including goals, loop jobs, Vivling runtime and delegation, VL UI, remote control, app-server integration, fork identity, and Android/Termux carry-forward. npm latest remains on 0.144.1 and stable on 0.143.0.

Upstream adds the advanced reasoning picker and persisted per-thread reasoning effort, and rolls Guardian review prompting back to the validated layout.

0.140.0 - Upstream rust-v0.140.0 final

Stable release on the npm latest tag, based on the OpenAI Codex rust-v0.140.0 release line. Upstream-tracking release: no fork-specific feature changes beyond the standing Codex VL workflow layer and the Termux/Android carry-forward. Validated on device (Linux, macOS, Termux — AI-guided surface reports, PASS). The stable dist-tag stays on 0.135.0.

0.133.0 - Upstream rust-v0.133.0 final

Stable release on the npm latest tag.

Based on the OpenAI Codex rust-v0.133.0 release line.

Vivling improvements

  • Vivlings now actually speak. The CRT footer phrase and every /vl reply flow through the same LLM channel the Vivling chooses for chat, gated by an atomic daily budget (stage defaults: Baby 50, Juvenile 100, Adult 200).
  • The greeting at session start now comes out in the configured language instead of a generic English placeholder.
  • CRT phrases and proactive replies no longer cut mid-word.
  • Ctrl+J opens a bordered, scrollable chat panel that wraps long replies cleanly and adapts to narrow terminals (Termux portrait to wide desktop).
  • New /vivling crt-brain controls: show exposes mode, budget, calls today and remaining head-room; budget unlimited lifts the cap for unmetered wrappers; reset-budget zeroes daily counters without waiting for UTC rollover.
  • Bond mechanics rate tuning makes companion progression more noticeable during real work sessions.

Platform coverage

  • New Linux arm64 prebuilt target: @mmmbuto/codex-vl-linux-arm64 (aarch64-unknown-linux-musl) is now built in CI alongside Linux x64 and Android arm64. The launcher (bin/codex.js, bin/codex-exec.js) auto-resolves the right native package on Raspberry Pi (64-bit OS), AWS Graviton, Apple Silicon Linux VMs, and other ARM64 Linux hosts.
  • The macOS arm64 post-install build (the only target that compiles on the user's machine) now runs a preflight that checks for Xcode Command Line Tools, the Rust toolchain, rustup, and the aarch64-apple-darwin target before invoking cargo build. Each missing item is reported with the exact command to install it, so a fresh Mac no longer fails partway through compilation with cryptic linker or toolchain errors.

Upstream merge

  • Goals are now enabled by default, backed by a dedicated goals.db store, with create_goal, update_goal, and get_goal exposed as model tools.
  • codex remote-control runs as a foreground command, waits for readiness, reports machine status, and keeps explicit daemon-style start / stop commands.
  • Permission profiles gained list APIs, inheritance, managed requirements.toml support, runtime refresh behavior, and stronger Windows sandbox integration.
  • Plugin discovery is easier to inspect, with marketplace-aware list output, installed versions, visible marketplace roots, and remote collection support.
  • Extensions can observe more lifecycle events, including subagent start / stop, tool execution, turn metadata, and async approval / turn processing.

Full upstream release notes: https://github.com/openai/codex/releases/tag/rust-v0.133.0

Changed

  • npm package metadata moved from the pre-release lane to 0.133.0.
  • Workspace and lockfile package versions aligned with the upstream 0.133.0 release.
  • Public install, source-build, and native artifact staging surfaces remain fork-owned for DioNanos/codex-vl and @mmmbuto/codex-vl.

Preserved fork features

  • All Codex VL workflow commands and fork-owned modules are preserved across the upstream stable merge: /vivling, /vl, /loop, /remote-control, Vivling Memory V2 (V10 schema, bond mechanics, brain expression channel, daily LLM budget, language-aware boot, Ctrl+J chat panel, CRT footer).
  • The fork-safe managed-update channel and the disabled standalone auto-updater are kept (the daemon never fetches the upstream installer script).
  • The fork feedback channel points to DioNanos/codex-vl/issues and the announcement tip source is DioNanos/codex-vl/main/announcement_tip.toml.

Per aspera ad astra.

0.132.0 - Upstream rust-v0.132.0 final

Stable release on the npm latest tag.

Based on the OpenAI Codex rust-v0.132.0 release line.

Changed

  • npm package metadata moved from the pre-release lane to 0.132.0.
  • Public install, source-build, and native artifact staging surfaces remain fork-owned for DioNanos/codex-vl and @mmmbuto/codex-vl.

Preserved fork features

  • Existing Codex VL workflow commands and fork package identity are preserved across the upstream stable merge.

Upstream merge

  • Python SDK authentication now supports API key login, ChatGPT browser and device-code flows, account inspection, and logout APIs.
  • Python turn APIs are easier to use for text-only workflows and return richer turn results for handle-based runs.
  • codex exec resume accepts --output-schema, so resumed automations can keep session context while still enforcing structured JSON output.
  • TUI startup probes are batched before the first interactive frame.
  • Remote executor registration can use standard Codex auth.
  • App-server turns preserve requested image fidelity, including original local image detail, across user inputs and image-producing tools.

0.131.0 - Upstream rust-v0.131.0 final

Based on the OpenAI Codex rust-v0.131.0 release line.

Added

  • Public development journal pointer in the /vivling README section, linking to the Codex VL dev journal at dev.mmmbuto.com/vivling.
  • Upstream additions inherited from rust-v0.131.0, including data-driven service-tier controls, blended token usage display, permissions/approval mode surface, effective workspace roots, responsive Markdown tables, unified @ mention picker across files, plugins and skills, plugin marketplace CLI flows, daemon-managed codex remote-control with runtime enable/disable and status reads, and the new codex doctor diagnostic command.

Changed

  • npm latest line moved from 0.130.0 to 0.131.0 with matching platform packages for Linux x64, Termux Android arm64, and a macOS arm64 source-build package that compiles locally with Cargo during npm postinstall.
  • README release-channels section rewritten around 0.131.0 latest. The npm next tag is now described as reserved for the next upstream alpha lane after a stable release rather than tracking the current pre-release line.

Preserved fork features

  • /loop, /goal, /vivling, /vl and the Vivling runtime, lifecycle, and CRT layer are preserved across the merge.
  • Fork-owned update, doctor and install surfaces stay on @mmmbuto/codex-vl with the fork repository as the package source. No upstream installer URL is reintroduced into fork-owned scripts.

0.131.0-alpha.23 - Fork-safe remote-control bootstrap

Pre-release on the npm next tag. latest stayed on 0.130.0.

Changed

  • npm-installed and bun-installed Codex VL now keep autoUpdateEnabled=false for the app-server daemon and short-circuit the managed updater path so the current fork binary stays in control of remote-control sessions.
  • Standalone, Brew and other install contexts keep their previous managed path, so this is a targeted change for npm/bun installs only.

0.131.0-alpha.22 - Fork identity hardening (F-bis)

Pre-release on the npm next tag. latest stayed on 0.130.0.

Changed

  • Doctor, updater, npm registry hints and release links all point to @mmmbuto/codex-vl and the fork repository, with regression tests pinned against silent upstream reintroduction.
  • install_native_deps derives the owner/repository from the workflow URL passed by fork pipelines instead of a hardcoded upstream value, while the historical default URL is preserved as a documented placeholder.

0.131.0-alpha.20 - Upstream sync rust-v0.131.0-alpha.20

First Codex VL alpha aligned with the upstream rust-v0.131.0 alpha line.

Changed

  • Merge target is the explicit upstream tag rust-v0.131.0-alpha.20, not the upstream branch head, to keep the conflict surface bounded.
  • TUI session resume is adopted from upstream and the previous inline helpers are gated to test builds where appropriate.
  • Release profile uses thin LTO for slightly smaller binaries.

Internal

  • Merge-safety refactors landed before the upstream sync:
    • chatwidget slash_dispatch boundary extraction so most VL slash logic lives outside the upstream-heavy dispatch surface.
    • app/loop_controller split from a single large module into a small set of focused submodules behind a stable internal facade.
    • bottom_pane VL boundary extraction so the upstream-heavy view module no longer carries VL-specific logic blocks. These changes are not directly user-visible. They exist to keep upstream merges practical without changing public behavior.

0.130.0 - Upstream rust-v0.130.0

Based on the OpenAI Codex rust-v0.130.0 release line.

Added

  • npm latest Linux x64 and Termux Android arm64 prebuilt packages plus a macOS arm64 source-build package that compiles locally with Cargo during npm postinstall.

Changed

  • /goal lifecycle completion is now clear-on-complete to avoid stale completion state across sessions.
  • MCP startup snapshot behavior and stdio retry hardening kept aligned with the upstream 0.130.0 runtime.
  • SQLite state contention hardening retained for multi-session local use.

Preserved fork features

  • Vivling runtime, identity, persistence, lifecycle and brain profile routing.
  • /loop session-scoped recurring jobs.
  • /goal workflow alongside upstream /goal semantics.

0.128.3 - Local Linux rebuild

  • Rebuilds and reinstalls the local Linux package from the aligned Forge/GitHub base.
  • Keeps the 0.128.2 packaging corrections while refreshing the installed CLI payload.

0.128.2 - Corrected npm packaging

  • Publishes Linux x64 and Termux Android arm64 npm packages with native prebuilts.
  • Keeps macOS npm installs on local Cargo builds instead of shipping unsigned macOS binaries.
  • Supersedes the deprecated 0.128.1 candidate packages.

0.128.1 - macOS npm packaging cleanup

  • Changed the macOS npm package to build native binaries locally with Cargo.
  • Removed unsigned macOS binary payloads from the candidate packaging flow.

0.128.0-vl.0 - Upstream Sync

Based on the OpenAI Codex rust-v0.128.0 release line.

Changed

  • Preserves the upstream 0.128.0 feature set as the base, including /goal workflows and related app-server/TUI APIs.
  • Keeps Codex VL additions as additive layers: Vivling, /loop, fork packaging, and platform build paths.

0.126.0-vl.0 - Upstream Sync and Vivling CRT Foundation

Based on the OpenAI Codex 0.126.0 release line.

Added

  • First modular Vivling CRT renderer foundation for the bottom terminal strip.
  • Baby lifecycle CRT scripts for idle, play, eat, sleep, and work states.
  • Focused 15+ZED Vivling roster foundation with Common, Rare, Legendary, and Mythic tiers.
  • Vivling brain model guide for profile-based model resolution.

Changed

  • Vivling CRT output now prioritizes compact visual state and short speech over dense metrics.
  • Lifecycle activity no longer floods the expanded Vivling chat log.
  • Public README/docs were reduced to a smaller release-facing surface.
  • Internal concept art, roadmap notes, and release lane notes moved under .docs/.

Removed

  • Old generated 90-species Vivipendium and EPUB prototype docs from the public docs tree.

0.124.0 - First Public Release

Based on OpenAI Codex 0.124.0.

Added

  • Side-by-side codex-vl CLI packaging under @mmmbuto/codex-vl.
  • /loop session-scoped loop supervision for recurring checks and long-running work.
  • /vivling persistent companion system with local state, levels, species, cards, and work memory.
  • /vl <message> direct Vivling chat shortcut.
  • Adult Vivling brain dispatch through normal Codex profiles and model providers.
  • Vivling loop-awareness and loop-owner experiments.
  • Initial public README positioning for Codex VL.
  • Initial README hero asset under docs/assets/.

Changed

  • /vl now routes to the Vivling brain when the active Vivling is adult, brain-enabled, and has a brain profile.
  • /vl keeps a local fallback reply path when the brain is not ready.
  • /vivling remains the controlled command surface rather than becoming free-form chat.

Experimental

  • Vivling learning from work summaries and loop events.
  • Vivling brain profiles backed by custom model catalog entries.
  • Linux and Termux/Android packaging flow.
  • GitHub public release pipeline.

Known Gaps

  • Public release workflow still needs hardening.
  • npm platform packaging needs cleanup before broad publish.
  • Merge-safety refactor is still pending for slash commands, app events, migrations, and TUI integration hooks.
  • Vivling genetics, bonding, spawn inheritance, and richer roster UX are still future work.

Upstream Codex

For upstream OpenAI Codex changes, see the official OpenAI Codex release notes.