OpenAI Codex 0.142 Adds Rollout Token Budgets, Multi-Agent Delegation, Indexed Web Search, and a Reorganized Plugin Marketplace

OpenAI Codex 0.142 Adds Rollout Token Budgets, Multi-Agent Delegation, Indexed Web Search, and a Reorganized Plugin Marketplace

lschvn

The OpenAI Codex 0.142 line shipped across the week of June 22 to June 26, 2026: 0.142.0 on June 22, 0.142.1 and 0.142.2 on June 25, and 0.142.3 on June 26. It lands one week after the 0.141 Noise relay and cross-OS exec release, and the contrast between the two is telling. Where 0.141 was a security release whose headline was a default end-to-end encrypted transport between the orchestrator and remote executors, 0.142 is a governance and orchestration release. The question it answers is no longer "is the wire safe" but "can a team run many of these agents and stay in control of cost, delegation, and tool sprawl."

Rollout token budgets: cost control at the agent layer

The most consequential feature is the rollout token budget, implemented across #28746, #28494, #28707, and #29423. A rollout is a single agent run. The budget is a configurable token cap that Codex tracks across threads as the rollout progresses. At a configurable threshold it emits a remaining-budget reminder, and when the budget is exhausted it aborts the turn cleanly instead of letting a runaway task keep sampling. Turn-scoped context contributions (#28911) keep the accounting honest, and the budget window context is wrapped separately (#29494) so the model sees what it has left.

This is the first built-in cost-control primitive at the Codex agent layer. Before 0.142, bounding what a background task or a delegation could spend meant an external wrapper script or a hard model-side limit. Now a team can hand an agent a budget and trust that a stuck loop will stop on its own. For anyone running Codex against paid models, or running it unattended inside a CI or a multi-agent orchestration setup, that is the difference between an agent that is safe to leave running and one that is not.

Two orchestration capabilities build on the new budget. Multi-agent delegation lets a parent thread hand work to a subagent, and 0.142 makes the behavior configurable at both the thread and the turn level as disabled, explicit-request-only, or proactive (#28685, #28792, #29324). The "explicit-request-only" mode is the conservative default a team wants when it is evaluating delegation: the parent has to ask before spawning a child. Proactive mode hands the parent the discretion to delegate when it judges a subagent fits the work. A companion fix (#28375) surfaces terminal subagent errors back to the parent, so a child that crashed no longer reports as an empty successful completion and leaves the parent continuing on a false premise.

The indexed web-search mode (#28489) gives the agent live web access without unrestricted browsing. Codex can run live searches, but direct page retrieval is restricted to server-approved URLs. The model gathers fresh information from the web but cannot follow arbitrary links off the approved set. Codex distinguishes cached and live web-access terminology in the config (#29095), and the mode sits alongside the existing filesystem and network approval scopes rather than replacing them. The shape is deliberate: trust the search aggregation, do not trust the open browser.

MCP tool search by default, and a reorganized plugin marketplace

MCP tooling gets two changes that matter for installs that have grown past a handful of servers. MCP tools now use tool search by default when the provider supports it (#29486). Instead of dumping every tool from every connected MCP server into the model's context up front, Codex searches for the relevant tool on demand. That preserves compatibility with older models and providers that do not advertise tool search, but for a mature install it is the difference between a tool list that fits the context window and one that does not.

The /plugins view reorganizes remote plugins into three sections, OpenAI Curated, Workspace, and Shared with me (#26703), and a turn can now recommend and install a relevant plugin mid-run (#28399, #28400, #27704, #28403). 0.142.2 adds dedicated dark-mode logos for plugins (#29488) and curated featured-plugin rankings from remote catalogs (#29485). The reorganization turns the plugin browser from a flat list into something a team can actually navigate, and the in-turn recommendation closes the loop between "I have a plugin for that" and the agent actually using it.

System proxy, time awareness, and remote-executor reliability

The remaining work is the kind that does not make a headline but decides whether Codex runs on a given network. 0.142.1 adds opt-in Windows system proxy support for authentication, including PAC, WPAD, static proxies, and bypass rules (#26708). 0.142.2 adds the macOS equivalent: authentication clients honor system proxy, PAC, and WPAD when respect_system_proxy is enabled (#26709). 0.142.0 also honors startup custom CA bundles with the managed MITM path (#29014), which matters on corporate networks that intercept TLS. Together these cover the proxy-and-CA matrix that previously broke the auth handshake for a large class of enterprise users.

Codex also gains real time awareness: scheduled UTC time reminders and the ability to query the current time directly, including through client-provided app-server clocks (#28822, #28824, #28835, #29011). The /usage command can now show and redeem earned usage-limit reset credits (#28154, #28793).

The reliability fixes target the remote-executor path that 0.141 made the default. Exec-server processes and stdio MCP sessions now survive transient disconnects, with signed-URL refresh on reconnect and retry-safe stdin writes (#28512, #28374, #28546, #28895). The Linux TUI no longer breaks after a Ctrl+Z suspend and fg resume (#28342). Remote environments now preserve executor-native paths, shells, AGENTS.md discovery, and sandbox behavior across operating systems, which is the payoff from the PathUri layer 0.141 introduced.

Security hardening and what to watch

0.142.2 hardens a few spots worth noting. PowerShell commands containing executable AST regions the safety classifier cannot inspect now require approval (#24092). Remote HTTP(S) image inputs return clear, model-visible validation errors instead of failing silently, while inline data URLs and local images keep working (#29417, #29419). Expired Amazon Bedrock credentials now produce actionable recovery guidance instead of a generic authorization error (#28992). The release also refreshes the bundled OpenSSL to 3.6.3 (#29487) and esbuild to 0.28.1, the latter being the same version that shipped the Deno RCE and Windows path-traversal fixes we covered earlier this month.

Read alongside the 0.141 Noise relay release and the broader move of runtimes into Anthropic's AI coding infrastructure, the 0.142 cycle is the moment Codex stops being a single-agent coding tool and starts being an agent platform you can govern: budgets to bound cost, delegation modes to control who spawns whom, a web-search mode that trades reach for safety, and tool search that scales to a real MCP estate. The full changelog between 0.141 and 0.142.0 is on the compare view, and the 0.143.0 alpha line is already on its twenty-sixth pre-release at time of writing, which is the signal that the cadence is not slowing down.

Frequently Asked Questions

Deno 2.9 Ships 1.98x Faster Cold Start, 2.2-3.1x Less RSS Under Load, Default-On npm Minimum Release Age, No-Downgrade Trust Policy, and Built-In Snapshot Testing

Deno 2.9 (Bartek Iwańczuk, published 2026-06-25 on deno.com/blog/v2.9) is the largest Deno release of the cycle. Cold start drops from 34.2 ms to 17.3 ms (1.98x), peak RSS on the Deno.serve realworld workload drops 2.2x (142 MB → 64 MB) and 3.1x on 1 MiB bodies (197 MB → 63 MB), and Deno.serve throughput climbs 1.27x realworld (56.8k → 72.4k req/s), 1.11x plaintext, and 1.18x on 1 MiB bodies. Supply chain hardening: npm minimum-release-age is enabled by default with a 24h window (PR #35458), and a new opt-in no-downgrade trust policy (PR #34927) refuses to resolve any version whose trust evidence (staged publish, trusted publishing, provenance attestation) is weaker than the strongest evidence on any earlier-published version of the same package. Test runner parity: built-in t.assertSnapshot() (#35139), Deno.test.each (#34938), --shard for CI fan-out (#35057), retry and repeats (#35053), change-aware --changed and --related (#35199), and coverage thresholds (#35056). Lockfile interop: deno install seeds deno.lock from package-lock.json, pnpm-lock.yaml, yarn.lock, or bun.lock (#34296, #35330, #35346, #35350, #35394), pnpm-workspace.yaml auto-migrates to deno.json / package.json (#34993), and git merge conflict markers in deno.lock auto-resolve (#34726). Plus: deno desktop graduates from experimental (the June 16 PR #33441), deno link / deno unlink / deno list / deno watch subcommands, stable --unsafe-proto (#34738), Web Locks API (#31166), Happy Eyeballs v2 (RFC 8305) (#31726), navigator.userAgentData (#34743), the WebCrypto Modern Algorithms proposal (ML-KEM, ML-DSA, SLH-DSA, ChaCha20-Poly1305, SHA-3 family, KMAC, Argon2) (#34447, #34448, #34914, #35223), Node 26.3.0 compat (#34746, #34747), Node-API v10 (#35270), and CSS module imports under --unstable-raw-imports (#35093). 165+ PRs land in this cycle.

Cline 4.0 Migrates the VS Code Extension Onto the Shared Cline SDK, Adds ClinePass, a Customize Marketplace, Plugins, and Queued Prompts

Cline shipped v4.0.0 on June 26, 2026 (github.com/cline/cline), a major version that migrates the VS Code extension from its legacy standalone task implementation onto the shared Cline SDK, the same TypeScript engine that runs the Cline CLI, Kanban, and JetBrains plugin. The release adds ClinePass (built-in onboarding, provider selection, subscription handoff, and entitlement handling), a Customize marketplace for Skills, MCP servers, and the new Cline Plugins, queued prompts, edit-and-regenerate, a provider and model configuration rework around providers.json and a shared model catalog (Fireworks GLM 5.2, Kimi K2.6 Fast, Kimi K2.7 Code, Qwen 3.7 Plus, MiniMax M3, SAP AI Core, LiteLLM, Codex OAuth), command auto-approval disabled by default, and an extension build and package workflow moved to Bun. Subagents are temporarily disabled in the extension while the SDK-backed experience stabilizes.

Related articles

More coverage with overlapping topics and tags.

Claude Code Issue #74066: Users Report Cross-Workspace Context Bleed on Sonnet 5, Anthropic Has Not Yet Responded
ai

Claude Code Issue #74066: Users Report Cross-Workspace Context Bleed on Sonnet 5, Anthropic Has Not Yet Responded

An open bug filed against Claude Code on 2026-07-04 by an [Enterprise ZDR](https://docs.anthropic.com/en/docs/build-with-claude/zero-data-retention) user describes a working session on Sonnet 5 that suddenly starts referencing an unrelated Minecraft temple build, then doubles down on the wrong task in its recap. The reporter (GitHub: [@milesrichardson-edb](https://github.com/milesrichardson-edb), issue [anthropics/claude-code#74066](https://github.com/anthropics/claude-code/issues/74066)) is on Enterprise Zero Data Retention, the tier Anthropic specifically advertises as session-isolated. Triage on the reporter's local session JSONL at `~/.claude/projects/<encoded-cwd>/<session-id>.jsonl` finds the leaked text is not in the transcript, ruling out a local context bleed by file overlap. Four other users in the comments (with work histories going back to last year) describe near-identical behavior across Claude Code, Claude Mobile, and Claude deep research. The most plausible architectural fit is shared KV-cache state in inference ([per @yv3nne in the comments](https://github.com/anthropics/claude-code/issues/74066#issuecomment-4880448776)), but no Anthropic engineer has commented on the issue in the 22 hours since it was filed, and the issue reached the top of [Hacker News](https://news.ycombinator.com/item?id=42481789) on 2026-07-04. The tone in the thread is split: half suspecting a real platform cache reuse, half suspecting a [sonnet-5-specific hallucination triggered by a Pygments lexer](https://github.com/anthropics/claude-code/issues/74066#issuecomment-4880334711). Both readings are credible.
Claude Code 2.1.199 Adds Stacked Slash-Skill Invocations, Fixes Subagent Error Reporting, the Background-Agent Daemon Crash Loop on Linux, and the Streaming-Response Discard Bug
ai

Claude Code 2.1.199 Adds Stacked Slash-Skill Invocations, Fixes Subagent Error Reporting, the Background-Agent Daemon Crash Loop on Linux, and the Streaming-Response Discard Bug

Anthropic released Claude Code [v2.1.199](https://github.com/anthropics/claude-code/releases/tag/v2.1.199) on 2026-07-02, one day after the v2.1.198 feature release (Chrome GA, background-agent auto-PR, /dataviz, Gateway on AWS) and two days after v2.1.197 made Sonnet 5 the default model. v2.1.199 is a bug-fix release with 24 entries: one new feature (stacked slash-skill invocations load up to 5 leading skills in one prompt), and a long list of reliability fixes for subagent error reporting, the background-agent daemon on Linux, streaming-response preservation on mid-stream API errors, SSL/TLS error handling behind corporate proxies, plan mode tool-call handling, and a per-env retry policy that no longer caps the user-tunable retry count at 15. Most of the items are in the background-agent / subagent area, which is consistent with the v2.1.197 + v2.1.198 direction of treating delegation as the product surface for Sonnet 5.
Claude Code 2.1.198 Makes Claude in Chrome Generally Available, Lets Background Agents Commit, Push, and Open Draft PRs, Adds `/dataviz` Skill, and Brings Claude Platform on AWS to the Gateway
ai

Claude Code 2.1.198 Makes Claude in Chrome Generally Available, Lets Background Agents Commit, Push, and Open Draft PRs, Adds `/dataviz` Skill, and Brings Claude Platform on AWS to the Gateway

Anthropic released Claude Code [v2.1.198](https://github.com/anthropics/claude-code/releases/tag/v2.1.198) on 2026-07-01T20:45:36Z, the second consecutive day of Claude Code shipping after v2.1.197 (the Sonnet 5 default release) the day before. The headline change is that Claude in Chrome is now generally available inside Claude Code: the browser-side assistant that was a research preview moves to a stable channel and is no longer gated behind a feature flag or a waitlist. The release also lands the first end-to-end automation for background agents in `claude agents`, which now commit, push, and open a draft pull request when they finish code work in a worktree, instead of stopping to ask for permission. Other notable additions: a `/dataviz` skill for chart and dashboard design with a runnable color-palette validator, a Gateway upstream for Claude Platform on AWS (`anthropicAws`) with model-not-found responses advancing the failover chain, a built-in Explore agent that now inherits the main session's model (capped at opus) instead of running on Haiku, and subagents that inherit the parent session's extended thinking configuration. The release also removes the `/agents` wizard and replaces it with the conversational flow, fixes a long list of reliability bugs around background agents, network drops, and platform auth, and improves syntax highlighting by upgrading to highlight.js 11.

Comments

Log in Log in to join the conversation.

No comments yet. Be the first to share your thoughts.