What's New on AI Developer Toolkit

The AI Developer Toolkit changelog records the Claude Code, Codex and Cursor releases that change how an agent setup behaves (default models, removed flags, permission defaults), the fix for each, and every guide this site published. As of 26 September 2026, the changes needing action are new default models in Claude Code and Codex three removed Codex options (exec --full-auto, -a on-failure/untrusted, codex mcp-server) and a retired model (GPT-5.4).

A guide written in July tells you to run codex exec --full-auto, and your CI job now fails with an unexpected-argument error. A teammate’s Claude Code session runs Opus 5 while yours runs Opus 5.5, and neither of you changed a setting. This page is for the developer or tech lead who owns a team’s shared agent configuration (CLAUDE.md, AGENTS.md, .claude/settings.json, Codex’s config.toml, the CI jobs that call an agent) and needs to know which release caused the change and what to edit.

What you get from this changelog as the owner of a team’s agent config

• A table of the September 2026 changes that break or silently change a setup, each with the version that introduced it and the fix.
• A ten-minute check that tells you whether your machines and CI are behind this page.
• Three copy-paste prompts: one audits a repository for stale model IDs and removed flags, one applies the fixes on a branch with evidence, and one filters an upstream changelog down to what touches your config.
• A way to prove an upgrade did not break your agent workflow without reading every transcript.
• A dated record of what this guide shipped, so you can tell whether a page predates the release it describes.

Which changes from September 2026 need action?

Every row was checked on 2026-09-26 against the vendor’s changelog, release tags or the installed CLIs (claude 2.1.283, codex-cli 0.157.1). Prices and context windows live on one page, the AI model comparison for coding agents, so this table names models without repeating their prices.

Claude Code

Change	Since	What to do
Claude Opus 5.5 is the default model on every paid plan and the API	v2.1.280, latest channel (2026-09-22)	Remove hard-coded claude-opus-5 pins unless your evals need them. Opus 5.5 runs at medium effort by default: raise it with /effort before you switch models.
The stable channel lags behind	stable was v2.1.274 on 2026-09-26	On 2026-09-26 stable was v2.1.274: there Pro and Team Standard still defaulted to Sonnet 5, other plans to Opus 5, and Opus 5.5 was unavailable. Check where each channel is now with npm view @anthropic-ai/claude-code dist-tags. Put the whole team on one channel with autoUpdatesChannel.
Claude Fable 5.1 replaces Fable 5	v2.1.257 (2026-09-01)	The fable alias now resolves to Fable 5.1. Fable is never a default; select it deliberately with /model fable, and expect it to bill to usage credits on some plans.
Auto mode is the starting permission mode	v2.1.283, latest channel (2026-09-25)	Interactive terminal and VS Code sessions start in auto mode. claude -p and the Claude Agent SDK still start in Manual. To keep Manual organisation-wide, set permissions.disableAutoMode: "disable" in managed settings. Sessions on a model auto mode does not support (for example Haiku, or Sonnet 4.5, the default on Microsoft Foundry) also start in Manual.
AGENTS.md is read when a project has no CLAUDE.md	v2.1.277 (2026-09-18); v2.1.281 on Bedrock, Google Cloud, Foundry and gateways	Read the AGENTS.md in repositories without a CLAUDE.md: Claude now loads it. Toggle this under Project instructions in /config.

Codex

Change	Since	What to do
GPT-6 Astra is the bundled default	CLI 0.153.4 (2026-09-04); needs 0.153.0+	Update CLIs older than 0.153.0. Astra’s Codex default effort is low; raise effort before switching models.
GPT-6 Sol and GPT-6 Luna join the picker	CLI 0.156.1 (2026-09-23); need 0.155.0+	Use Sol for everyday work and Luna for cheap, high-volume tasks. There is no GPT-6 Terra: Terra exists only as gpt-5.6-terra.
GPT-5.4 is retired in Codex	2026-08-31	A config that pins gpt-5.4 is migrated to GPT-6 Sol. Change model in config.toml so the file says what actually runs.
codex exec --full-auto is removed	CLI 0.147.0 (2026-08-07)	Use a permission profile (-c default_permissions=":workspace", beta), --approve-for-me, or --sandbox workspace-write. Do not combine a profile with --sandbox: OpenAI says the two systems do not compose.
-a accepts only on-request and never	checked on 0.157.1	Replace -a on-failure and -a untrusted. Put the flag before the subcommand: codex -a never exec …, because codex exec has no -a of its own.
codex mcp-server is removed	CLI 0.154.0 (2026-09-09)	Codex can no longer run as an MCP server. Use codex exec, the Codex SDK, or codex app-server (experimental). Claude Code still offers claude mcp serve.

Cursor

Cursor’s changelog and documentation could not be re-verified for September 2026: cursor.com was unreachable from the environment this page was checked in on 2026-09-26. The last verified Cursor facts date from 2026-08-28 and are summarised on the Cursor changelog page. Read Cursor’s own changelog before you act on any Cursor version, model or price claim, including ones on this site.

Two changes cut across tools:

• MCP specification 2026-07-28 removed the initialize handshake and the Mcp-Session-Id header. Claude Code negotiates the new version by default with direct HTTP servers (all installs by v2.1.274); Codex 0.157.1 does not yet (mcp_2026_07_28 is off in codex features list). If you maintain an MCP server, test it against both protocol versions. See MCP best practices.
• GitHub Copilot retires four models on 2026-10-02: Kimi K2.7 Code (replaced by Kimi K3), Gemini 3.5 Flash and Gemini 3.6 Flash (replaced by Gemini 3.8 Flash), and Claude Opus 4.7 (replaced by Claude Opus 5). Update any Copilot model policy that names them.

Check whether your setup is behind this page

Run these on a developer machine and in one CI job. The whole check takes about ten minutes and changes nothing.

1. Record the versions you actually run, in a terminal:

   claude --version
   codex --version

   For Cursor, read the version in its About dialog and compare it with Cursor’s changelog.

2. Find the Claude Code release channel. Look for autoUpdatesChannel in ~/.claude/settings.json and in managed settings; "stable" means you are about a week behind latest. To choose a channel, set autoUpdatesChannel to "stable" or "latest" in settings; claude install stable installs the stable build now.

3. List the models your Codex binary knows and the state of its feature flags:

   codex debug models --bundled
   codex features list

4. Run both health checks. They read your settings files and report problems before an agent run does:

   claude doctor
   codex doctor

5. Search the repository for the removed flags and legacy model pins from the table, from the repository root:

   rg -n -e '--full-auto' -e '-a (on-failure|untrusted)' -e 'codex mcp-server' \
     -e 'gpt-5\.4' -e 'claude-(opus|fable)-5([^-]|$)' .

   The last pattern matches claude-opus-5 and claude-fable-5 but not claude-opus-5-5.

6. Update on a branch, not on main: claude update and codex update, then repeat steps 1–4 and run the prompts below.

Copy-paste prompts for keeping agent config current

Run the first prompt in Claude Code or Codex from the repository root. It reads and reports; it edits nothing, so you can run it with read-only permissions.

Copy-paste prompt: audit the repository for stale agent config

Audit this repository's agent configuration for changes released up to 2026-09-26.
Read CLAUDE.md, AGENTS.md, .claude/settings.json, .mcp.json, any Codex config.toml,
.github/workflows/*.yml and scripts/. Report every occurrence of:
- `codex exec --full-auto` or `--full-auto` (removed in Codex CLI 0.147.0)
- `-a on-failure` or `-a untrusted` on the codex command line (only on-request and never are accepted in 0.157.1), and `-a` placed after `exec`
- `codex mcp-server` (removed in Codex CLI 0.154.0)
- a model pinned to gpt-5.4 (retired in Codex on 2026-08-31), claude-opus-5 or claude-fable-5 (now legacy)
- the word "workflow" used as the Claude Code dynamic-workflow trigger (replaced by "ultracode" in v2.1.160)
For each hit print file:line, the exact line, the replacement, and one sentence on the risk
of leaving it. Do not edit any file. End with a table sorted by risk, CI files first.

The second prompt applies the fixes and produces the evidence a reviewer needs, so the pull request can be approved on its checks.

Copy-paste prompt: apply the fixes on a branch with evidence

Create a branch named chore/agent-config-2026-09. Apply only the replacements from the
audit table you produced earlier in this session (if there is none, run the audit
first); change nothing else. After each file, run the checks this repo
uses in CI (read the workflow files to find them: tests, lint, type check). If a check
fails because of your change, fix it; if it fails for an unrelated reason, stop and tell
me. Finish with: the list of commands you ran with their exit codes, `git diff --stat`,
and a three-line pull request description that names each removed flag or model and
the release that removed it.

The third prompt turns a long upstream changelog into the handful of entries that touch your configuration. Replace 2.1.274 (the stable version on 2026-09-26) with the version your team runs; npm view @anthropic-ai/claude-code dist-tags shows the current channel versions.

Copy-paste prompt: filter an upstream changelog to what affects us

Our team runs Claude Code 2.1.274 (stable channel). Read the Claude Code changelog at
https://code.claude.com/docs/en/changelog.md from 2.1.275 up to the version `claude --version`
reports here. List only the entries that change a default (model, effort, permission mode),
remove or rename a command, flag or settings key, or change what gets loaded into context.
For each: the version, the change in one line, and whether .claude/settings.json, CLAUDE.md
or our CI workflows are affected, quoting the line if so. Skip bug fixes and UI polish.

How to prove an agent tool upgrade did not break your workflow

Nobody on the team should have to read transcripts to approve a version bump. Make the upgrade a change like any other, with checks that fail loudly:

1. Pin the versions CI installs. Install an exact version, for example npm install -g @anthropic-ai/claude-code@2.1.283 or npm install -g @openai/codex@0.157.1, so CI only changes when someone bumps the pin in a pull request.
2. Run one fixed smoke task headlessly on both versions. Pick a small task with a known correct outcome, such as adding a test for an existing function, and run it with claude -p … --output-format json or codex exec … --output-schema schema.json. The gate is that your test suite passes afterwards, not that the transcript looks right.
3. Re-run your evals when the default model changes. A new default model is a model switch you did not choose. Follow the 48-hour evaluation playbook for a new model and the evals for coding agents guide before you accept it for unattended work.
4. Let the config owner sign off on evidence. The tech lead who owns the shared configuration approves the bump when tests, lint, type checks and the smoke task are green on the new version. If something regresses, revert the pin; the previous version is one commit away.

What breaks after an agent tool update, and how to recover

Teammates see different default models. Cause: some machines follow Claude Code’s stable channel and others latest, or Homebrew’s claude-code cask (which tracks stable) sits next to a native install. Recovery: set one autoUpdatesChannel for the team in managed settings, and pin a model with --model where a result must be reproducible.

A CI job fails with “unexpected argument” or “invalid value”. Cause: a removed Codex flag (--full-auto, -a on-failure) or -a placed after exec. Recovery: apply the Codex rows of the table above, then pin the Codex version so the next removal arrives in a pull request instead of in a failed deploy.

Codex reports that GPT-5.4 is no longer available. Cause: a model = "gpt-5.4" pin. Codex migrates the session to GPT-6 Sol. Recovery: set the model you mean in config.toml, or remove the pin and use the bundled default.

The agent runs commands you expected to approve. Cause: auto mode became the starting mode in v2.1.283, and a classifier now approves actions it judges safe. Recovery: start sessions with --permission-mode manual, or turn auto mode off organisation-wide with permissions.disableAutoMode: "disable" in managed settings. auto set in project settings is ignored, so a repository cannot switch it on for you.

Claude follows instructions nobody wrote for it. Cause: from v2.1.277 Claude Code loads AGENTS.md when there is no CLAUDE.md, so a file written for Codex now steers Claude too. Recovery: review that AGENTS.md, or add a CLAUDE.md, or switch the behaviour off under Project instructions in /config.

A guide on this site contradicts the table above. Cause: the guide predates the release. Every page shows its last-updated date; if it is older than the release, trust this page and the per-tool changelog, and report the page so it gets fixed.

Vendor releases in 2026 that changed agent setups

Only changes that alter defaults, remove options or add a model appear here; per-version detail is on the Claude Code changelog, Codex changelog and Cursor changelog pages. Dates marked “secondary” come from press reports rather than the vendor.

September 2026

• Sep 25 · Claude Code v2.1.283. Auto mode becomes the starting permission mode for interactive terminal and VS Code sessions on every plan and provider (latest channel).
• Sep 23 · Codex CLI 0.156.1. GPT-6 Sol and GPT-6 Luna join the model picker, below GPT-6 Astra.
• Sep 22 · Anthropic: Claude Opus 5.5. Becomes Claude Code’s default from v2.1.280; Opus 5 moves to legacy. Anthropic also raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise.
• Sep 21 · SpaceXAI: Grok 4.7. Added to GitHub Copilot the same day; available in Cursor from day one (secondary: MarkTechPost).
• Sep 18 · Claude Code v2.1.277. Reads AGENTS.md when a project has no CLAUDE.md.
• Sep 9 · Codex CLI 0.154.0. Removes codex mcp-server.
• Sep 4 · Codex CLI 0.153.4. GPT-6 Astra becomes the bundled default.
• Sep 2 · Google: Gemini 3.8 Flash (secondary: The Register). In GitHub Copilot from Sep 3.
• Sep 1 · Anthropic: Claude Fable 5.1 and Claude Mythos 5.1. Fable 5.1 replaces Fable 5 in Claude Code from v2.1.257; Mythos 5.1 is invitation-only through Project Glasswing.

August 2026

• Aug 31 · Codex retires GPT-5.4, migrating pinned sessions to GPT-6 Sol.
• Aug 14 · Claude Code auto mode becomes the starting mode on Pro, Max and Team.
• Aug 7 · Codex CLI 0.147.0. Removes codex exec --full-auto; automatic approval review (--approve-for-me) routes approvals through a reviewer model.
• Aug 4 · Claude Code v2.1.222 removes /ultraplan from the command reference; use plan mode (/plan). The same day OWASP published the GenAI LLM Top 10 2026.

July 2026

• Jul 28 · MCP specification 2026-07-28. A stateless core without the initialize handshake; tools adopt it at different speeds (see the cross-tool notes above).
• Jul 24 · Anthropic: Claude Opus 5.
• Jul 9 · GPT-5.6 Sol, Terra and Luna reach GA on Azure (model version 2026-07-09); now labelled “older” in Codex. The Codex desktop app merged into the ChatGPT desktop app as a Codex mode (secondary: Neowin).

June 2026

• Jun 30 · Anthropic: Claude Sonnet 5. Its announced price rise was later cancelled; current prices are on the models hub.
• Jun 9 · Anthropic: Claude Fable 5 and Claude Mythos 5.
• Jun 1 · Claude Code v2.1.160. The dynamic-workflow trigger keyword changes from workflow to ultracode.
• Jun 1 · GitHub Copilot replaces request-based billing with usage-based billing.
• June · Windsurf renamed Devin Desktop (secondary: creeta).

Before June 2026

• May 20 · Cursor 3.5 ships a bundled /loop skill (verified on 2026-08-28).
• May 15 · Roo Code extension shut down, per the Roo-Code README; the repository is archived.

What this guide shipped

Each entry links to the page it describes. Every page also carries its own last-updated date.

Start here, by role Reading paths for developers, tech leads, CTOs and executives.

When a new model ships A 48-hour evaluation playbook before you accept a new default.

The State of Agentic Engineering Sourced evidence on how much code agents write in 2026.

Compound engineering Every's plan, work, review and compound loop, audited.

September 2026

• Sep 26 · Documentation reorganised by role and lifecycle stage. Role reading paths under Start here for developers, tech leads, CTOs and executives; the method in Method; cross-tool practice in Workflows; models, MCP, skills, plugins and observability under Ecosystem; and new leadership sections, Strategy and business and Engineering organization. Old URLs redirect to their new location.
• Sep 26 · Model facts re-checked. New model pages: new-model evaluation playbook, benchmarks and open-weight models. Pages are re-verified against the installed CLIs as they are rewritten; the last-updated date on each page tells you whether it has been.
• September · The AI-native lifecycle. The lifecycle section maps the artifact chain, plan, design, build, test, deploy, maintain and metrics onto Claude Code, Codex and Cursor.
• September · Human-in-the-loop default changed. Layered agent review plus human gates replaced blanket line-by-line review, which is now reserved for critical and regulated code.
• September · Scorecards rebuilt. The developer scorecard guide and CTO scorecard guide now measure lifecycle maturity.

August 2026

• Aug 28 · The Ladder. Six free hub pages organise the documentation by autonomy level, from assisted coding to running a software factory.
• Aug 28 · The State of Agentic Engineering, August 2026. Third-party evidence, each number with publisher and date, on how much production code agents write, with the counter-evidence sourced the same way.
• Aug 28 · Claude Code routines and Cursor cloud agents and automations. Unattended cloud runs in both tools: triggers, scoping and the untrusted-input rule.
• Aug 24 · Compound engineering. Every’s loop as a cross-tool guide, with an audit of the plugin behind it.
• Aug 20 · The AI Developer Toolkit MCP server. A public, read-only endpoint with search and fetch across every guide on the site.
• Aug 19 · Anti-slop writing skills. unslop and miodkuj, including the deterministic checks that run in CI without an agent.
• Aug 13 · /goal and /loop corrected. Tool support re-verified: Cursor’s /loop skill and Codex goals being on by default had been missed.
• Aug 12 · Software factories and Dr. Skill. When a factory works and when it erodes the codebase; a read-only audit of your skill and MCP loadout.
• Aug 6 · Answer-first rewrite. Every article opens with a paragraph that answers its title, enforced by the build.
• Aug 4 · Pruning context files. An ablation protocol for cutting CLAUDE.md, Cursor rules and AGENTS.md.

July 2026

• Jul 30 · herdr, tmux for agent fleets and agent shells. Running and supervising many agents from one terminal or shell.
• Jul 27 · Claude Opus 5 across the model guides.
• Jul 25 · Mobbin MCP and Impeccable rebuilt for v4.
• Jul 21 · Taste Skill. The anti-slop frontend skill library.
• Jul 11 · Superpowers. A skills plugin that enforces a design-to-PR workflow.
• Jul 3 · Claude Code features you haven’t tried.

June 2026

• Jun 24 · /grill-me and the first cross-tool /goal and /loop guides.
• Jun 10 · Claude Fable 5 covered across the model and routing guides.
• Jun 8 · Official skills from 11 more publishers.
• June · Reducing MCP token cost and skill supply-chain security.

Earlier releases of this guide

• May 2026 · Per-question guides for the developer and CTO scorecards, and bilingual scorecards.
• April 2026 · The four-stage AI design pipeline and the official skills guide.
• February 2026 · A full Codex section alongside Cursor and Claude Code, and Polish translations of every section.
• 17 July 2025 · Version 1.0.0. The first release, covering Cursor and Claude Code.

Where to go next with tool and model changes

• The AI model comparison for coding agents holds the current prices, context windows and defaults this page links to instead of repeating.
• The per-tool changelog pages carry every version: Claude Code, Codex and Cursor.
• The troubleshooting guide covers failures that are not caused by a release.
• The What’s New page shows the newest guides as cards; send feedback when a page lags behind a release.