Sources: github.com , raw.githubusercontent.com
-
Google Antigravity deprecated standalone Workflows (`.agents/workflows/*.md` and `~/.gemini/config/workflows/*.md`) in favor of standard Agent Skills (`.agents/skills/{name}/SKILL.md`), with official workflow retirement scheduled for November 1, 2026 and an upstream `/migrate-workflows` conversion command provided.
Already handled addressed 2026-10-01. Prompts are no longer deployed as workflow markdown files for Antigravity (`AntigravityPromptsStrategy.isSupported()` returns `false`). Instead, `AntigravityAdapter` now leverages `installPromptsAsSkills` to convert prompts into instruction-only Agent Skills under `.agents/skills/{name}/` (or `prompt-{name}` when conflicting with an existing skill name), matching the pattern used by Codex and Zed.
Sources: code.claude.com , www.npmjs.com
-
Claude Code 2.1.277 added native support for reading `AGENTS.md` project instruction files. When an `AGENTS.md` file is present in a repository or directory tree and no `CLAUDE.md` is present, Claude Code automatically reads `AGENTS.md` as its source of project instructions. If both files exist, `CLAUDE.md` takes precedence unless configured otherwise.
Already handled addressed 2026-09-24. In `packages/schema/src/editor-support.ts`, the `agents-md` feature for `claude-code` was updated from `unsupported` to `native`. `ClaudeCodeAdapter.detect()` was expanded to detect projects containing `CLAUDE.md`. In addition, legacy fallback `CLAUDE.md` files whose sole content is an import or link to `AGENTS.md` (e.g. `@AGENTS.md`) are detected with an informational notice that `CLAUDE.md` is now redundant and can be removed.
Sources: x.ai , x.ai
-
Grok Build CLI stores MCP configuration in TOML format (`.grok/config.toml` for project scope, `~/.grok/config.toml` for user scope) under `[mcp_servers.<name>]`.
Needs review supported via GrokMcpStrategy with both stdio and HTTP support.
-
Grok Build CLI reads rules from `.grok/rules/*.md` and user rules from `~/.grok/rules/*.md`, alongside native discovery of `AGENTS.md`.
Needs review supported via GrokRulesStrategy.
-
Grok Build CLI exposes custom slash commands in its interactive TUI through Agent Skills.
Needs review supported via prompt-to-skill conversion into `.grok/skills/prompt-{name}/`.
-
Grok Build CLI supports lifecycle hooks in `.grok/hooks.json` and `~/.grok/hooks.json`.
Needs review supported via GrokHooksStrategy.
Sources: raw.githubusercontent.com , raw.githubusercontent.com , github.com , www.npmjs.com
-
Gemini CLI 0.58.0 isolated Docker and container runtime sockets and execution binaries within macOS Seatbelt sandbox profiles to prevent unauthorized container escape.
No repo change no change needed. This is an internal runtime sandbox hardening measure within the Gemini CLI executor.
-
Gemini CLI 0.58.0 enforced consistent symlink evaluation in ignore path handling to prevent bypasses around project ignore files.
No repo change no change needed. aix respects project boundaries and standard ignore mechanisms.
-
Gemini CLI 0.58.0 declared top-level safety checkers in write policy configuration.
No repo change no change needed. aix does not configure write safety checkers directly; it generates standard instructions and MCP configuration.
-
Gemini CLI 0.58.0 optimized conversational history rollback mechanics and retry nudge flows, and resolved stale cancellation errors in the Agent-to-Agent server.
No repo change no change needed. Runtime session resilience improvements.
Sources: code.claude.com , www.npmjs.com
-
Claude Code 2.1.251 added `PreModelSwitch` and `PostModelSwitch` lifecycle hooks. `PreModelSwitch` fires before switching to a new model and can abort or veto the switch via exit code or structured output. `PostModelSwitch` fires immediately after a model switch completes, receiving the previous and current model identifiers.
Already handled addressed 2026-09-07. Generic `pre_model_switch` and `post_model_switch` events were added to `hookEventSchema` in `packages/schema/src/hooks.ts` and mapped to `PreModelSwitch` and `PostModelSwitch` in `packages/core/src/editors/strategies/claude-code/hooks.ts`. Format and round-trip import support is verified in `packages/core/src/__tests__/editors/hooks-strategies.test.ts`.
-
Claude Code 2.1.248 added an `experimental.cacheTtl` frontmatter option to custom agent markdown files (`"5m"` or `"1h"`) to tune prompt caching lifetimes for subagent turns.
No repo change no change needed. aix writes standard agent configuration and does not strip or validate agent frontmatter fields beyond required properties. Custom agent frontmatter passes through.
-
Claude Code 2.1.259 introduced `managedMcpServers` in managed configuration, allowing organizations to deploy required, unmodifiable MCP servers with managed lifecycle.
No repo change no change needed. aix writes project and user MCP configurations to `.mcp.json` and `~/.claude.json`. Managed settings in global configuration are preserved intact by aix's non-destructive JSON merge strategies.
-
Claude Code 2.1.248 added a `--restricted` CLI execution flag that strips project-level capabilities, disallowing execution of untrusted project hooks, project commands, and inline project MCP definitions.
No repo change no change needed. This is a runtime invocation flag for Claude Code. aix writes compliant configuration to project files that will be respected when running under normal permissions or safely bypassed in restricted mode.
-
Claude Code 2.1.257 stopped reading `defaultMode: "bypassPermissions"` from project-level or local settings (`.claude/settings.json`, `.claude/settings.local.json`), restricting global permission bypass to user configuration.
No repo change no change needed. aix does not configure or emit `defaultMode` or `bypassPermissions` in project settings.
-
Claude Code 2.1.251 disallowed setting `CLAUDE_CONFIG_DIR` or `TMPDIR` in project-level `env` settings to prevent directory traversal and arbitrary code redirection.
No repo change no change needed. aix does not emit these environment variables in generated project settings.
-
Claude Code 2.1.251 fixed an issue where hook `stdout` was buffered indefinitely during long-running streaming operations rather than delivered incrementally.
No repo change no change needed. Runtime fix within Claude Code's hook execution engine.
-
Claude Code 2.1.244 resolved PowerShell path quoting inconsistencies on Windows when invoking hook scripts with spaces in directory paths.
No repo change no change needed. Runtime fix within Claude Code's process launcher.
Sources: learn.chatgpt.com , github.com , www.npmjs.com
-
Codex 0.150.0 added an `Interrupt` hook event that executes commands or MCP handlers when an active top-level turn is interrupted by the user or client.
Already handled addressed 2026-09-07. Generic `interrupt` event was added to `hookEventSchema` in `packages/schema/src/hooks.ts` and mapped to `Interrupt` in `packages/core/src/editors/strategies/codex/hooks.ts`. Format and round-trip import support is verified in `packages/core/src/__tests__/editors/hooks-strategies.test.ts`.
-
Codex 0.150.0 stopped loading project-level `AGENTS.md` instructions in untrusted workspaces, and ensured that managed deny-read rules remain enforced after permission state changes.
Needs review supported. aix generates project-level `AGENTS.md` rules through `packages/core/src/editors/strategies/codex/rules.ts`. This change reinforces the need for user trust confirmation before project instructions take effect.
-
Codex 0.151.0 introduced a configurable grace period for discovering tools from optional MCP servers, and allowed extensions to inspect or replace MCP tool results before they reach the model.
No repo change no change needed. aix configures MCP servers via standard Codex `config.toml` entries. Dynamic tool interception and discovery grace periods are runtime extension capabilities.
-
Codex 0.152.0 updated MCP server name validation to permit `:`, `@`, `/`, and `.` characters, supporting package-style names across CLI commands and authentication scopes.
Needs review supported. aix MCP server schemas permit package-scoped names and string identifiers in `ai.json`.
-
Codex 0.152.0 added an `output_token_limit` setting for individual MCP tools to truncate excessive tool outputs consistently across session resumes.
Needs review not applicable. aix does not currently define per-tool output token limits in generic MCP server configurations.
-
Codex 0.153.0 added `features.context_management.experimental_mode` and `tui.disable_paste_burst` configuration toggles, as well as `request_user_input_async` in app-server thread metadata.
No repo change no change needed. These are specific runtime options. aix's configuration management preserves unmanaged keys in `config.toml`.
Sources: raw.githubusercontent.com , docs.github.com , www.npmjs.com
-
Copilot 1.0.83 added support for custom agents to list multiple models in frontmatter `model` (as an ordered fallback array) paired with `model-policy: required`.
Needs review partial. aix currently models the agent `model` field as a single string in `packages/schema/src/agent.ts`. Custom agent passthrough preserves arbitrary frontmatter, but generic fallback model arrays are not yet first-class in schema.
-
Copilot 1.0.81 injects OpenTelemetry trace context into hook executions via `traceparent` and `tracestate` process environment variables and JSON inputs.
No repo change no change needed. This is runtime contextual telemetry supplied by the Copilot CLI engine to hook handler processes.
-
Copilot 1.0.81 resolved agent instruction discovery order when project and user custom agents share identical identifiers.
No repo change no change needed. aix installs custom agents into the standard `.github/copilot/agents/` directory per documented conventions.
Sources: cursor.com , docs.cursor.com
-
Start from scratch (2026-08-27): Introduced enhanced initial prompt options and empty-state workflow suggestions in the Composer agent interface.
No repo change no change needed. This is an internal UI enhancement for Composer and does not alter configuration formats.
-
Self-hosted machines (2026-09-02): Enabled running Cursor Cloud Agents on private infrastructure and self-hosted worker pools.
No repo change no change needed. Infrastructure configuration for cloud agents is managed via the Cursor web dashboard and does not affect local workspace files.
-
Cursor hook definitions remain standardized at `.cursor/hooks.json` with `version: 1` envelope formatting.
Needs review supported. `packages/core/src/editors/strategies/cursor/hooks.ts` emits compliant `version: 1` hook configuration files.
Sources: github.com , opencode.ai , www.npmjs.com
-
OpenCode 1.18.29 updated Codex OAuth model filtering to recognize integer GPT versions such as `gpt-6`.
No repo change no change needed. Model resolution bugfix in OpenCode's provider adapter.
-
OpenCode 1.18.28 sent the session identifier as GitHub Copilot's interaction header to improve cross-turn request correlation.
No repo change no change needed. Runtime networking enhancement.
-
OpenCode 1.18.27 defaulted provider header timeouts and streamed chunk timeouts to five minutes (with `false` supported to disable), and added Anthropic `thinking.blockBinding` configuration support for Claude 5.1+ models.
No repo change no change needed. These are model provider execution settings and do not alter OpenCode workspace rules or MCP files.
-
OpenCode 1.18.26 fixed tool call timing during live metadata updates and prevented `apply_patch` from emitting empty move paths in permission metadata.
No repo change no change needed. Runtime execution fix.
-
OpenCode 1.18.24 enabled V1 runtimes to read supported V2 configuration fields so newer configuration files operate across mixed environments.
Needs review supported. aix generates standard OpenCode workspace configuration files compatible with current stable and forward runtimes.
-
OpenCode 1.18.23 added native dashed slug conversion for Anthropic models routed through Cloudflare AI Gateway.
No repo change no change needed. Provider gateway routing fix.
Sources: docs.devin.ai , docs.devin.ai
-
Windsurf configuration formats continue to use the established Cascade paths: `.windsurf/rules/`, `.windsurf/workflows/`, and `~/.codeium/windsurf/mcp_config.json`.
Needs review supported. `packages/core/src/editors/adapters/windsurf.ts` correctly targets these paths.
-
As documented in `docs/editor-research/windsurf/3.8.20.md`, Devin Local subagents remain isolated within `.devin/agents/` and are not currently targeted by the Windsurf adapter.
No repo change no change needed. Architectural namespace boundary maintained.
Sources: zed.dev , github.com , zed.dev
-
Zed 1.18.1 fixed an issue where integrated terminal inputs were dropped on Windows for certain keyboard layouts, and prevented `goToDefinition` from navigating out-of-worktree.
No repo change no change needed. Core editor stability bugfixes.
-
Zed 1.18.0 refreshed Markdown preview styling and added persistent File Finder history across workspace sessions.
No repo change no change needed. UI and preview enhancements.
-
Zed 1.18.0 fixed the inline assistant becoming unresponsive when an LLM model emitted reasoning blocks prior to initiating tool calls.
No repo change no change needed. Assistant runtime execution fix.
-
Zed 1.17.2 introduced tabular data preview support and enabled low reasoning effort configurations for DeepSeek V4 Flash and Pro models.
No repo change no change needed. Editor capability improvements.
-
Zed's assistant configuration, context server schema (`context_servers` in `settings.json`), and rules loading (`.zed/rules.md`) remain stable and unchanged.
Needs review supported. `packages/core/src/editors/adapters/zed.ts` targets `.zed/settings.json` context servers and rules files.
Sources: github.com , github.com , github.com , www.npmjs.com
-
0.52.0: "fix(core): simplify plan mode write policy to support relative paths".
No repo change no change needed. This is the agent's own write-permission policy in plan mode, not a config schema aix generates.
-
0.53.0: "fix(a2a-server): enforce workspace trust and task isolation to prevent RCE".
No repo change no change needed. Server-side hardening of the agent-to-agent server. It does affect when project config is trusted, the same way the equivalent Claude Code and Cursor trust changes do, but does not change what aix writes.
-
0.53.0 added an eval coverage report command, and 0.54.0 added session ID rotation on model fallbacks.
No repo change no change needed. Neither is a config surface.
Sources: code.claude.com , www.npmjs.com
-
Claude Code 2.1.219 added a `DirectoryAdded` hook that fires after `/add-dir` or the SDK `register_repo_root` control request registers a new working directory mid-session.
Already handled addressed 2026-08-24. A `directory_added` event was added to `hookEventSchema` in `packages/schema/src/hooks.ts` and mapped to `DirectoryAdded` in `packages/core/src/editors/strategies/claude-code/hooks.ts`. It is a new event rather than a reuse of `cwd_changed`, which maps to `CwdChanged` and fires when the working directory changes rather than when one is added; mapping both to one name would make the hook fire at the wrong time. Every other editor reports it through `getUnsupportedEvents`, so targeting them warns instead of dropping it. Covered by three tests in `packages/core/src/__tests__/editors/hooks-strategies.test.ts` and an end-to-end install test in `packages/cli/src/__tests__/commands.test.ts`.
-
Claude Code 2.1.239 fixed agents, skills, and commands whose `.md` file starts with a UTF-8 byte-order mark being silently ignored.
No repo change no change needed, but worth a regression test. aix writes markdown through the runtime adapter's file writes and does not prepend a BOM. A rule, prompt, agent, or skill copied from a source file that already carries a BOM would have been affected before this fix; it now loads either way.
-
Claude Code 2.1.239 required that a folder's trust dialog be accepted before a project `.mcp.json` `headersHelper`, or an inline MCP server in a project or `--add-dir` agent file, will run. The same release stopped a `headersHelper` from a project `.mcp.json`, plugin, or agent file inheriting credential env vars, and moved user, managed, and claude.ai-scope helpers to run from the Claude config directory.
No repo change no change needed. aix writes static `headers` into MCP config, not `headersHelper`; `grep -rn "headersHelper" packages/core/src` returns nothing. The header handling aix does own was changed in commits 05993a7 and cc68cbc, which keep MCP auth headers across sync and import.
-
Claude Code 2.1.239 fixed `claudeMdExcludes` not excluding a symlinked `.claude/rules` file.
No repo change no change needed. aix writes real files into `.claude/rules` (`packages/core/src/editors/strategies/claude-code/rules.ts`), not symlinks, so the exclusion path this fixed does not apply to aix output. aix does symlink native skills, which this fix does not cover.
-
Claude Code 2.1.232 added `additionalMarketplaces` and `allowedMarketplaces` as aliases for the `extraKnownMarketplaces` and `strictKnownMarketplaces` settings keys, and 2.1.223 added owner wildcard entries (`"owner/*"`) to the strict and blocked marketplace managed settings.
No repo change no change needed. aix does not generate marketplace settings. These are additional unknown keys that aix's `.claude/settings.json` merge must preserve, which the merge tests in `packages/core/src/__tests__/merge.test.ts` cover.
-
Claude Code 2.1.232 fixed nested git repositories inheriting trust from a parent directory; each repository now requires its own trust confirmation.
No repo change no config change needed. This affects when aix-written project config takes effect for a user, not what aix writes. A monorepo user who installs into a nested repo now has to confirm trust there before the config loads.
-
Claude Code 2.1.222 fixed `PreToolUse` auto-allow hooks bypassing tool restrictions in background agent tasks, and 2.1.239 fixed OpenTelemetry trace fragmentation for tool executions deferred by a `PreToolUse` hook.
No repo change no change needed. Both are runtime behaviors of hooks aix already writes. The `PreToolUse` output shape from `packages/core/src/editors/strategies/claude-code/hooks.ts` is unaffected.
-
Claude Code 2.1.222 stopped repo-local settings (`.claude/settings.json` or `.claude/settings.local.json`) from turning Remote Control auto-start on; they can still turn it off.
No repo change no change needed. aix does not write Remote Control settings. Noted because it is a narrowing of what project-scope `.claude/settings.json` can do, which is the file aix writes project hooks into.
-
Claude Code 2.1.221 changed plugins to accept `"."` as a `skills` path.
No repo change no change needed. aix installs skills as directories under `.claude/skills/`, not through plugin manifests.
-
Claude Code 2.1.216 added a `sandbox.filesystem.disabled` setting, 2.1.219 added `sandbox.network.strictAllowlist`, 2.1.224 added `crossSessionInbound`, `dialogExpiry`, and sandbox credential-masking options, 2.1.235 added `spellcheck`, and 2.1.238 added `keybindingFlavor`.
No repo change no change needed. None of these are keys aix emits. As with every settings addition, aix's section-scoped writes to `.claude/settings.json` must leave them alone, which the merge tests cover.
-
Claude Code 2.1.233 stopped exposing todo/task-tracking tools (`TaskCreate`, `TaskGet`, `TaskUpdate`, `TaskList`, `TodoWrite`) on Opus 4.8, Sonnet 5, Fable 5, Mythos 5, and newer models unless `CLAUDE_CODE_ENABLE_TODO_TOOLS=1` is set.
Already handled addressed 2026-08-24 as a documentation change. aix still writes `TaskCreated` and `TaskCompleted`, which is correct: the events exist and the user may have the tools enabled. A note in the Claude Code profile in `packages/schema/src/editor-support.ts` now records that the hooks only fire when the todo and task tools are available, and names the env var and the model set.
-
Claude Code 2.1.233 fixed skill and command argument substitution so argument values are not re-expanded.
No repo change no change needed. aix copies prompt and skill file contents unchanged; the substitution happens at invocation time inside Claude Code.
Sources: github.com , learn.chatgpt.com , www.npmjs.com
-
Codex 0.148.0 added asynchronous command hooks: "Hooks can now run commands asynchronously and invoke MCP tools." The hooks documentation describes an `async` boolean on a command handler, up to 8 concurrent async hooks per session, output delivered at the next safe point in the conversation, no ability to block the triggering operation, and `SessionEnd` hooks always running synchronously regardless of the flag.
Already handled addressed 2026-08-24. `async` is now in `CODEX_COMMAND_FIELDS` and on `CodexCommandHookEntry`, `buildCodexHook` emits it, and `parseCodexHook` reads it back, all in `packages/core/src/editors/strategies/codex/hooks.ts`. `async_rewake` is deliberately still unsupported: it is a Claude Code behavior with no Codex equivalent, so it keeps being reported rather than silently written. Covered by "writes the async flag Codex 0.148.0 added" and "round trips async and commandWindows through an import" in `packages/core/src/__tests__/editors/hooks-strategies.test.ts`.
-
The Codex hooks documentation lists a `commandWindows` field, the Windows counterpart to `command`.
Already handled addressed 2026-08-24. `buildCodexHook` now derives a POSIX command and a Windows command separately, mirroring `buildWindsurfEntry`, and honors the `shell` selector: `shell: 'bash'` suppresses `commandWindows`, `shell: 'powershell'` promotes `command` into it. `parseCodexHook` maps `commandWindows` back to `powershell` and now accepts an entry that has only a Windows command, which the old guard rejected. Covered by three tests in `packages/core/src/__tests__/editors/hooks-strategies.test.ts`.
-
The Codex hooks documentation lists eleven events: `PreToolUse`, `PermissionRequest`, `PostToolUse`, `PreCompact`, `PostCompact`, `UserPromptSubmit`, `SubagentStop`, `Stop`, `SessionStart`, `SubagentStart`, and `SessionEnd`.
Already handled addressed 2026-08-24. All five missing mappings were added to `EVENT_MAP`, which now covers every native event Codex documents. The source URL in the comment above the map was corrected at the same time: `https://developers.openai.com/codex/hooks` now 308-redirects to `https://learn.chatgpt.com/docs/hooks`. Covered by "maps every event Codex documents" in `packages/core/src/__tests__/editors/hooks-strategies.test.ts`, which asserts the full native event set rather than a sample.
-
The Codex hooks documentation lists an `additionalContextLimit` field, a token threshold defaulting to about 2500 for how much hook output is fed back to the model.
No repo change no change needed, decision confirmed 2026-08-24. aix has no generic equivalent and no other supported editor exposes the same control, so adding a Codex-only field to the shared schema is not yet justified. Revisit if a second editor gains one.
-
The Codex hooks documentation confirms hooks observe MCP tools through the normal tool-call path, matched by names such as `mcp__filesystem__read_file`.
Already handled already supported. `TOOL_MATCHER_MAP` maps `pre_mcp_tool` and `post_mcp_tool` to `PreToolUse` and `PostToolUse` with an `mcp__.*` matcher, which is the documented shape.
-
The Codex hooks documentation confirms the four discovery locations: `~/.codex/hooks.json`, `~/.codex/config.toml` inline `[hooks]` tables, `<repo>/.codex/hooks.json`, and `<repo>/.codex/config.toml`.
Already handled verified, no change needed. `CodexHooksStrategy.getConfigPath` returns `hooks.json` under the adapter's `.codex` config directory and `getGlobalConfigPath` returns `.codex/hooks.json`, which are the two JSON locations. aix does not write the inline TOML form, which is correct: writing hooks into `config.toml` would collide with the MCP server config aix already writes there.
-
Codex 0.147.0 added opt-in support for the MCP 2026-07-28 protocol revision, including paginated discovery, multi-round requests, and non-blocking server startup, and 0.147.0 upgraded the MCP SDK to 3.0.0.
No repo change no change needed. This is client-side protocol negotiation. aix writes server definitions (command, args, env, headers, transport), which the revision does not change.
-
Codex 0.145.0 fixed "correctly quoted hook commands" on Windows.
No repo change no change needed, and it strengthens the `commandWindows` follow-up above. Codex is treating the Windows hook command path as its own execution surface.
-
Codex 0.145.0 through 0.147.0 expanded `/import` to migrate Cursor and Claude Code settings, MCP servers, plugins, sessions, commands, and project-scoped memories, and 0.147.0 added importing Cursor-managed skills.
No repo change no change needed. This is Codex's own one-way migration command, which overlaps with `aix sync` but does not change any config format aix reads or writes.
-
Codex 0.147.0 removed the deprecated `codex exec --full-auto` flag in favor of `--sandbox workspace-write`.
No repo change no change needed. aix does not invoke `codex exec`.
-
Codex 0.146.0 and 0.147.0 added Agent Plugins manifests, workspace plugin publishing, and plugin catalogs, including a Claude Code plugin marketplace.
No repo change no change needed today. aix writes loose files under `.codex`, not plugin manifests. Worth watching: if plugins become the way Codex distributes skills and hooks, aix would need a plugin-shaped output target.
Sources: raw.githubusercontent.com , www.npmjs.com
-
Copilot CLI 1.0.76 added "support for Open Plugin Spec v1 plugin manifests and mcp.json configuration".
No repo change no change needed for the config aix writes, but the file name is worth confirming on the next pass. aix writes Copilot MCP servers to `.mcp.json` (`CopilotMcpStrategy.getConfigPath` in `packages/core/src/editors/strategies/copilot/mcp.ts`). The changelog entry is about `mcp.json` inside a plugin manifest, which is a separate distribution surface, not a rename of the project file. If a later release makes plugin-scoped `mcp.json` the primary location, aix would need a second output target.
-
Copilot CLI 1.0.73 changed relative links in custom agent instructions to resolve from the agent file's own location rather than the working directory.
No repo change no change needed, and it makes aix's output more correct rather than less. aix writes agent files into Copilot's agents directory; a relative link an author wrote next to the agent source now resolves the way they meant it to. Anything aix generated that relied on working-directory resolution would break, but aix does not rewrite link targets.
-
Copilot CLI 1.0.72 fixed an `agentStop` hook that always blocks looping indefinitely.
No repo change no change needed. It confirms the native event name aix already emits: `EVENT_MAP` in `packages/core/src/editors/strategies/copilot/hooks.ts` maps the generic `agent_stop` to `agentStop`. This was a runtime loop-guard fix, not a config format change.
-
Copilot CLI 1.0.72 gave custom agents that request a shell tool by alias the matching read, list, and stop shell tools, and made plugin custom agents honor deferred-tool-loading frontmatter.
No repo change no change needed. aix passes agent frontmatter through; it does not compute tool sets. The tool-alias expansion happens inside Copilot after it reads the file aix wrote.
-
Copilot CLI 1.0.79 changed prompt pinning to be off by default, with a `pinnedPrompts` setting to turn it back on, and 1.0.76 moved the pinned prompt one row up.
No repo change no change needed. `pinnedPrompts` is a UI preference aix does not write, and this does not affect the prompt files aix installs.
-
Copilot CLI 1.0.79 allowed an enterprise-managed sandbox policy to enforce a proxy URL while leaving credentials user-controlled, and showed where sandbox settings are stored in `settings.json` from the `/sandbox` dialog.
No repo change no change needed. aix writes no Copilot sandbox or proxy settings. Recorded because it confirms Copilot keeps those in `settings.json`, a file aix must continue to leave alone except for the sections it manages.
-
Copilot CLI 1.0.74 changed subagent timelines to identify whether a prompt came from the main agent or another subagent.
No repo change no change needed. This is transcript presentation.
Sources: cursor.com , cursor.com , cursor.com
-
None of these touch a local config surface aix writes.
No repo change no change needed. Model routing, pricing plans, the Slack and iPad clients, and cloud agent startup are all outside the file formats aix generates. Google Workspace Plugins and Origin code hosting add connectors and a code host, not a change to `.cursor/rules`, `.cursor/commands`, `.cursor/agents`, `.cursor/skills`, `.cursor/mcp.json`, or `.cursor/hooks.json`.
-
The hooks documentation still describes project hooks at `<project-root>/.cursor/hooks.json` and user-level hooks, spawned as processes that communicate over stdio with JSON in both directions.
Already handled verified, no change needed. That is what `packages/core/src/editors/strategies/cursor/hooks.ts` writes, including the `version: 1` wrapper and camelCase event names.
Sources: github.com , opencode.ai , www.npmjs.com
-
1.18.6 fixed "legacy MCP state refreshing when opening a V1 workspace" and 1.18.9 "Restored compatibility with legacy MCP SDK clients".
No repo change no change needed. Both are client-side connection handling for servers already defined in config. The server definition shape aix writes is untouched.
-
1.18.12 "Skipped legacy config reads against v2 servers to avoid spurious config loading issues".
No repo change no change needed, but this is the entry to watch. It is the first sign that OpenCode reads config differently between its V1 and V2 servers. Nothing in the release notes says the file format or location differs, and `https://opencode.ai/docs/config/` still documents the single config file aix targets. Re-check on the next pass whether V2 introduces a second config location.
-
1.18.20 "Surface failed subagent tool calls with a resumable `task_id`" and "Answer permission requests triggered by subagents during `opencode run`".
No repo change no change needed. Runtime subagent behavior, not the agent file format aix generates.
Sources: docs.devin.ai , docs.devin.ai , docs.devin.ai
-
Devin Desktop 3.6.21 (2026-07-29): "Custom subagents can now be defined as flat `agents/<name>.md` files, and per-agent directories also accept `AGENTS.md`, `agent.md`, and `agents.md`."
No repo change no change needed to the Windsurf adapter, and the reason matters more than the answer. This entry sits under the changelog's **Devin Local** section, not Cascade. Devin Local is the agent that replaced Cascade as Devin Desktop's default local agent at the June 2026 rebrand, and the Devin Local documentation states it "shares the same agent harness as Devin CLI" and "uses the same skills format and discovery mechanism." Its subagents therefore live where the Devin CLI puts them: `.devin/agents/` and `.agents/agents/` at project scope, `~/.config/devin/agents/` (`%APPDATA%\devin\agents\` on Windows) at user scope. Accepted layouts are flat `agents/<name>.md` or a directory whose file resolves in the order `AGENT.md`, `AGENTS.md`, `agent.md`, `agents.md`. Frontmatter is `name`, `description`, `model`, `allowed-tools`, and `max-nesting`.
-
Devin Local reads the config aix already writes. Devin imports rules, skills, and MCP servers from other tools by default: "If you don't explicitly configure `read_config_from`, all imports are enabled by default." For Windsurf the documented import covers rules from `.windsurf/rules/*.md` and `.windsurf/global_rules.md`, skills from `.windsurf/skills/` and `~/.codeium/<channel>/skills/`, and MCP servers from `~/.codeium/<channel>/mcp_config.json`. Stable reads `~/.codeium/windsurf/`, which is the directory aix writes.
No repo change no change needed. This settles the namespace question. aix's Cascade paths are a supported input to Devin Local, not a stale namespace it ignores, so there is no reason to migrate the adapter to `.devin/` or to write both namespaces. Doing either would duplicate config that Devin already imports.
-
One documented exclusion: "Windsurf workflows are explicitly excluded from import."
Needs follow-up follow-up needed, and it is a real gap. aix installs Windsurf prompts as workflows, to `.windsurf/workflows/` and `~/.codeium/windsurf/global_workflows/` (`packages/core/src/editors/strategies/windsurf/prompts.ts`). Rules, skills, and MCP servers all reach Devin Local through the import; prompts are the one section that does not. Recorded in the Windsurf notes in `packages/schema/src/editor-support.ts` so a user who moved to Devin Local knows why their aix-managed workflows stopped appearing.
-
Devin Local subagents remain outside anything aix writes. There is no Windsurf import path for subagents, so they can only come from `.devin/agents/`, `.agents/agents/`, or `~/.config/devin/agents/`.
No repo change no change needed, and this is why Windsurf agents support is not simply a missing strategy. Supporting it would mean writing outside the namespace the `windsurf` adapter owns, which is a larger decision than adding a strategy. aix already treats `devin` as an alias that normalizes to `windsurf` (`packages/schema/src/normalize.ts`), and that alias plan deliberately kept output paths unchanged; this finding is consistent with that choice.
-
Devin Desktop 3.8.20 (2026-08-21): "`mcp_config.json` no longer shows spurious 'Property is not allowed' warnings."
No repo change no change needed, and it likely removes a user-visible complaint about aix output. aix writes Windsurf MCP servers to `~/.codeium/windsurf/mcp_config.json` (`packages/core/src/editors/strategies/windsurf/mcp.ts`). Any aix-written key the editor's schema did not recognize would have produced exactly this warning, on a file that worked. Nothing about the file aix writes has to change.
-
Devin Desktop 3.8.20: "Cascade, Devin Local, and every other ACP agent are now unavailable while a workspace is open in Restricted Mode, and hooks no longer load or run there." The hooks documentation repeats this: "Hooks do not load or run while a workspace is open in Restricted Mode."
No repo change no config change needed. This is the Windsurf counterpart to Claude Code's safe mode: aix writes the hooks correctly and the editor declines to run them. Added 2026-08-24 to the Windsurf notes in `packages/schema/src/editor-support.ts`, so a user whose hooks never fire has somewhere to look.
-
Devin Desktop 3.7.16 and 3.8.20: "Permission rules from different levels (enterprise, mode, user, project, subagent) now compose predictably — an explicit `deny:` always wins."
No repo change no change needed. aix generates no Windsurf permission rules. Recorded because it establishes a five-level precedence order that a future portable permissions feature would have to match.
-
Devin Desktop 3.6.21 fixed "Permission rules written with relative globs (e.g. `Read(**/*.pem)`) being pinned to the directory the CLI was launched from."
No repo change no change needed, same reason.
-
Devin Desktop 3.5.17 (2026-07-17): "The `devin.*` settings for gitignore access, completion mode, and auto-continue are now honored."
No repo change no change needed. aix writes none of the `devin.*` settings. Worth noting only as evidence that the rebrand is reaching the settings namespace, which is the direction to watch for a future path change away from `.codeium/windsurf/`.
Sources: github.com , zed.dev , zed.dev
-
1.12.0 fixed "MCP servers configured in `.zed/settings.json` not being discovered when multiple project folders were open in a workspace".
No repo change no change needed, and it confirms the path aix targets. This was a discovery bug for a correctly-written file. A user with a multi-root workspace who reported aix's Zed MCP output "not loading" before 1.12.0 was hitting this.
-
1.13.1 fixed "Agent MCP servers running from the wrong directory when started before a worktree was available", and 1.15.0 fixed "MCP servers missing from new project threads after a worktree opens".
No repo change no change needed. Both are startup-ordering fixes for servers aix already writes correctly. Relevant only in that an MCP server aix configures with a relative command path would have resolved against the wrong directory before 1.13.1.
-
1.14.2 added sandboxing for the agent's terminal commands and web fetches.
No repo change no change needed today. Worth watching: Zed has no aix-managed hooks support, but if it gains lifecycle hooks that run commands, the sandbox will govern them and aix would need to say so in the support matrix.
-
1.14.2 added the `agent_ui_font_family` and `agent_buffer_font_family` settings.
No repo change no change needed. Both are appearance keys in `.zed/settings.json` that aix's section-scoped merge must preserve, which it does.
-
1.12.0 improved skill deletion with a confirmation prompt and moved deleted skills to the system trash instead of deleting them outright.
No repo change no change needed. This is Zed's own UI for removing a skill. It does mean a skill a user deletes in Zed goes to the trash while `aix remove skill` deletes it, which is a difference in behavior, not a defect in either.
-
1.13.1 improved discoverability of the `create_worktree` hook, and 1.14.2 added a "Skip Hooks" toggle for commits.
No repo change no change needed, and neither is an agent lifecycle hook. Zed's `create_worktree` sets up a git worktree, and "Skip Hooks" skips git `pre-commit` and `commit-msg`. Neither is the hooks surface aix manages, and Zed remains without aix-managed hooks support.
-
1.16.1 removed OpenCode Free models from Zed's built-in OpenCode provider.
No repo change no change needed. Model availability, not config.
Sources: github.com , github.com , www.npmjs.com , releasebot.io
-
Gemini CLI 0.49.0 migrated the `coreTools` setting to `tools.core` in `settings.json`.
Already handled verified 2026-07-19, no change needed. aix does not emit `coreTools`, and the existing test "preserves user-authored keys in .gemini/settings.json across installs" in `packages/core/src/__tests__/editors/adapters.test.ts` covers unknown-key preservation for both old and new key shapes.
-
Gemini CLI 0.49.0 added path-traversal prevention during skill install, and 0.51.0 fixed a "symbolic link directory escape in memory import processor".
Needs follow-up needs follow-up for verification only. aix installs Gemini skills as symlinks from `.gemini/skills/{name}/` (and `.agents/skills/`) into `.aix/skills/`. Verify on a current Gemini CLI that symlinked skill directories and GEMINI.md imports still load under the stricter symlink handling.
-
Gemini CLI 0.47.0 shipped Antigravity transition banners and migration commands for the Antigravity CLI (externally sourced from release summaries).
Needs follow-up needs follow-up as a watch item. If Gemini CLI rebrands or merges into Antigravity the config directories (`.gemini/`, `~/.gemini/`) and context file (`GEMINI.md`) could move, similar to the Windsurf → Devin Desktop rebrand. No path change is documented yet, so no aix change now.
-
Gemini CLI 0.50.0 added tool registry discovery for automated tool detection.
No repo change no config shape change identified. MCP config in `.gemini/settings.json` is unchanged.
-
Gemini CLI 0.49.0 standardized tool output formatting, fixed zero-quota fail-fast behavior, and pinned dependencies; 0.51.0 added caretaker service work and escape sequence handling fixes.
No repo change no implementation change needed. These are runtime behaviors.
-
No changes to GEMINI.md handling, TOML command files, agent frontmatter, or hook event names were found for 0.46.0–0.51.0.
No repo change no change needed to Gemini rules, prompts, agents, or hooks output. The 0.46.0–0.48.0 release notes are thinly documented upstream (the changelog index only carries announcements for select versions); re-verify if a bug report points at one of those versions.
Sources: code.claude.com , www.npmjs.com
-
Claude Code 2.1.214 changed hook `if:` glob semantics: a single-segment `dir/**` condition now matches only `<cwd>/dir`, and `**/dir/**` is required for any-depth matching.
Already handled addressed 2026-07-19. aix passes the `if` field through unchanged (`packages/core/src/editors/strategies/claude-code/hooks.ts`); the new matching semantics are now documented in the Claude Code hooks notes in `packages/schema/src/editor-support.ts`, which feeds the support matrix docs.
-
Claude Code 2.1.214 made `SessionStart` hooks report `source: "fork"` when a session begins as a fork, and 2.1.198 added `Notification` hook payloads `agent_needs_input` / `agent_completed` for background agents.
No repo change no config schema change needed. aix writes hook definitions, not runtime payload schemas, and `SessionStart` / `Notification` are already valid event names in the supported list.
-
Claude Code 2.1.178 added nested `.claude/skills/` discovery (with `<dir>:<name>` disambiguation) and closest-wins resolution for nested `.claude/` agents, workflows, and output styles; 2.1.211 and 2.1.195 fixed nested `.claude/rules/*.md` loading.
No repo change no implementation change needed. aix writes top-level `.claude/` directories, which remain valid; nested discovery is additive editor behavior.
-
Claude Code 2.1.178 added `Tool(param:value)` permission-rule syntax and 2.1.200 renamed the "default" permission mode to "manual" (`"defaultMode": "manual"`).
No repo change no implementation change needed today. aix does not generate Claude Code permission rules or `defaultMode`. Re-check if aix adds a portable permissions feature.
-
Claude Code 2.1.196 stopped `claude mcp list` / `get` from spawning `.mcp.json` servers that were self-approved via a committed `.claude/settings.json`.
No repo change no format change needed. aix `.mcp.json` output is unchanged; users may see extra approval steps at runtime.
-
Claude Code 2.1.169 added `--safe-mode` / `CLAUDE_CODE_SAFE_MODE` (disables all customizations) and `disableBundledSkills`.
No repo change no config change needed, but safe mode disables aix-written rules, prompts, skills, agents, and hooks at runtime. Worth a docs note if users report aix output "not loading".
-
Claude Code 2.1.207 stopped reading plugin option values from project-level `.claude/settings.json` and rejected `${user_config.*}` in shell-form plugin hook commands.
No repo change no implementation change needed. aix does not generate Claude Code plugin manifests or plugin options; aix-managed hooks in `.claude/settings.json` are not plugin hooks.
-
Claude Code 2.1.172 / 2.1.198 changed subagent execution: subagents run in the background by default, can nest up to 5 levels, and are capped per session (`CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION`).
No repo change no config schema change needed. aix agent frontmatter (`description`, `mode`, `model`, `tools`, `permissions`, `mcp`) is unchanged by these runtime behaviors.
-
Claude Code 2.1.166–2.1.215 added many new settings keys (`axScreenReader`, `vimInsertModeRemaps`, `sandbox.credentials`, `autoMode.classifyAllShell`, `language`, `attribution.sessionUrl`, `enforceAvailableModels`, and others).
No repo change no implementation change needed, but aix's section-managed writes to `.claude/settings.json` must keep preserving unknown keys. Existing merge behavior should be covered by tests before any hooks change.
Sources: learn.chatgpt.com , github.com , www.npmjs.com , releasebot.io
-
The official Codex changelog moved: `https://developers.openai.com/codex/changelog` now returns a 308 redirect to `https://learn.chatgpt.com/docs/changelog`.
Already handled addressed 2026-07-19. The `skills/editor-release-research/SKILL.md` source hints now list `https://learn.chatgpt.com/docs/changelog`. No generated editor config is affected.
-
Codex 0.144.0 added a `writes` app-approval mode (declared read-only actions run without prompting; writes still prompt).
No repo change no implementation change needed. aix does not write Codex approval or permission config. Re-check if aix adds a portable permissions feature.
-
Codex 0.144.0 let MCP tools request interactive authentication without an experimental opt-in, and 0.142.2/0.143.0 turned on MCP tool search by default.
No repo change no config shape change needed. `~/.codex/config.toml` MCP entries that aix writes are unchanged; these are runtime behaviors.
-
Codex 0.140.0 made disabled MCP servers persist across runtime overlays, resolved MCP server registrations through a catalog system, and added encrypted local storage for CLI and MCP OAuth credentials.
No repo change no implementation change needed. aix writes server definitions, not enable/disable state or credentials.
-
Codex 0.140.0 decoupled the skills extension from core and improved parallel skill metadata reads (0.142.0); 0.142.0 also improved plugin loading across "multiple skill paths".
No repo change no path change identified. aix-managed skills at `.agents/skills/{name}/` and `~/.codex/skills/{name}/` remain the documented surfaces. Re-verify skill paths in the Codex docs if a future release note names new locations.
-
Codex 0.142.0 extended remote environments to support executor-native `AGENTS.md` discovery, and 0.140.0 made AGENTS.md loadable "from all bound environments".
No repo change no implementation change needed. aix writes root and directory-level AGENTS.md files, which remain the native rule surface.
-
Codex 0.143.0 enabled remote plugins by default with npm marketplace sources.
No repo change no implementation change needed. aix does not manage Codex plugins.
-
Detailed official release notes for 0.136.0–0.141.0 are no longer exposed on the changelog page or the visible GitHub releases pages. Externally sourced summaries (releasebot.io, havoptic.com) describe 0.136 (archived sessions), 0.137 (TUI and enterprise credit limits), 0.138 (Codex Desktop handoff), and 0.139 (code-mode web search) with no config-surface changes.
No repo change no implementation change identified for 0.136–0.141. These facts are externally sourced; re-verify against official docs if any related bug reports arrive.
-
No hook-related changes were found for 0.136.0–0.144.6.
No repo change no change to `.codex/hooks.json` output. The supported event list (`PermissionRequest`, `PostToolUse`, `PreToolUse`, `SessionStart`, `Stop`, `UserPromptSubmit`) remains current.
Sources: raw.githubusercontent.com , www.npmjs.com
-
Copilot CLI 1.0.61 auto-loads MCP servers from the `.github/mcp.json` workspace config file.
Already handled verified 2026-07-19, no change needed. The Copilot CLI config-dir reference documents both `.mcp.json` and `.github/mcp.json` as project-level MCP files with equal standing ("Project-level MCP configurations (in `.mcp.json` or `.github/mcp.json`) take precedence over user-level definitions"). aix keeps writing `.mcp.json` (shared with Claude Code) and imports either path; the support-matrix note was updated to say so.
-
Copilot CLI 1.0.63 rejects custom-agent names that would create hidden files.
Already handled verified 2026-07-19, guard already present. `sanitizeFileName` in `packages/core/src/editors/adapters/base.ts` converts non-alphanumerics to dashes and strips leading dashes, so a leading-dot filename cannot be emitted; a regression test was added in `packages/core/src/__tests__/editors/adapters.test.ts`.
-
Copilot CLI 1.0.66 lets skills with the same name from different plugins coexist, and 1.0.69 keeps embedded `/skill` tokens in skill prompt context.
No repo change no implementation change needed. aix-managed skills under `.github/skills/{name}/` and `~/.config/github-copilot/skills/{name}/` are unaffected by plugin-skill namespacing.
-
Copilot CLI 1.0.67–1.0.69 preserve custom-agent tool filters in nested subagents and show warnings when host custom agents fail to load.
No repo change no config schema change needed. This is runtime enforcement of the existing `tools` frontmatter aix already writes.
-
Copilot CLI 1.0.70 lets custom agents that request a shell tool by alias also receive the matching read, list, and stop shell tools.
No repo change no implementation change needed. aix passes `tools` lists through; alias expansion happens in the editor.
-
Copilot CLI 1.0.70 persists GitHub MCP toolset/tool config via `settings.json`, and 1.0.61 added a `/settings` dialog plus a `tabs` setting.
No repo change no implementation change needed. These are editor-managed settings files, not surfaces aix generates.
-
Copilot CLI 1.0.62 lets plugins ship extensions via the plugin marketplace.
No repo change no implementation change needed. aix does not manage Copilot plugins.
-
No hook event or hook schema changes were found in 1.0.61–1.0.71.
No repo change no change to `.github/hooks/hooks.json` output or the supported event list.
Sources: cursor.com
-
Cursor 3.9 added a Customize page that consolidates plugins, skills, MCPs, subagents, rules, commands, and hooks, with user, team, or workspace-level customization.
No repo change no generated config change identified. The changelog describes a management UI over the existing surfaces aix writes (`.cursor/rules`, `.cursor/commands`, `.cursor/agents`, `.cursor/skills`, `.cursor/mcp.json`, `.cursor/hooks.json`). Re-check Cursor docs if team/workspace scopes gain file-backed local config aix could target.
-
Cursor 3.10 added team MCP server configuration distributed across cloud agents, IDE, and CLI through team marketplaces.
No repo change no implementation change needed. Team MCP distribution is cloud-managed, not a local config file. aix project and user `mcp.json` output is unchanged.
-
Cursor 3.11 added side chats (`/side`, `/btw`) and conversation search; 3.9 added the iOS app and marketplace leaderboard.
No repo change no implementation change needed. These are editor UI features.
-
No changes to rules, commands, agents, hooks events, skills paths, or AGENTS.md compatibility were found in the 3.8–3.11 changelog entries.
No repo change no change needed. The Cursor hook event list and config paths in `packages/schema/src/editor-support.ts` remain current as documented.
Sources: github.com , opencode.ai , www.npmjs.com
-
OpenCode 1.18.2 stopped subagents from launching nested subagents by default and added a configurable `subagent_depth` limit.
No repo change no implementation change needed today. aix does not write OpenCode agent-depth settings. Re-check if aix's portable agent config grows a nesting or depth field that should map to `subagent_depth`.
-
The OpenCode skills docs now list a native `skill` tool plus pattern-based skill permissions in `opencode.json` (`allow` / `deny` / `ask`, per-agent overrides, and `skill: false` to disable).
Needs follow-up needs follow-up only if aix adds portable skill permissions. Skill discovery paths are unchanged and match aix output: `.opencode/skills/`, `.claude/skills/`, and `.agents/skills/` at project scope, with `~/.config/opencode/skills/`, `~/.claude/skills/`, and `~/.agents/skills/` at user scope.
-
OpenCode 1.17.10 introduced namespaced plugin hook APIs and a V2 plugin API.
No repo change no implementation change needed. These are JavaScript plugin surfaces, not the lifecycle hook config files aix models; OpenCode hooks remain `unsupported` in the aix matrix.
-
OpenCode 1.17.12–1.17.14 added MCP resource template listing/read tools, a code-mode MCP adapter, scoped MCP auth status per server URL, and preserved metadata in paginated tool catalogs; 1.17.7 sends the current workspace as the MCP client root.
No repo change no config shape change needed. `opencode.json` `mcp` entries that aix writes are unchanged; these are runtime MCP behaviors.
-
OpenCode 1.18.0 completed the Desktop v2 migration; 1.17.16–1.17.19 are composer and model-picker UI changes.
No repo change no implementation change needed.
-
No changes to AGENTS.md handling, command files, or agent frontmatter were found in 1.16.1–1.18.3.
No repo change no change needed to OpenCode rules, prompts, or agents output.
Sources: docs.devin.ai , docs.devin.ai
-
Windsurf's rebrand to Devin Desktop is now fully reflected in its web presence: `https://windsurf.com/changelog` returns a 308 redirect to `https://docs.devin.ai/desktop/changelog`.
Needs follow-up partially addressed 2026-07-19. The `skills/editor-release-research/SKILL.md` source hints now point at `docs.devin.ai`. Display naming stays "Windsurf" with the `windsurf` editor ID and `.windsurf` paths, which the Devin docs still use for all config surfaces; revisit only if the editor changes its on-disk directories.
-
The Devin Desktop hooks docs list the user-level hooks file at `~/.codeium/windsurf/hooks.json` (Devin Desktop IDE) and `~/.codeium/hooks.json` (JetBrains plugin), with workspace hooks at `.windsurf/hooks.json`.
Already handled fixed 2026-07-19. `WindsurfHooksStrategy.getGlobalConfigPath()` and the support matrix now use the documented `~/.codeium/windsurf/hooks.json` path (the site docs already said this), with a test locking the path in `packages/core/src/__tests__/editors/hooks-strategies.test.ts`.
-
The hooks docs confirm the same twelve events aix already maps (`pre_read_code` … `post_setup_worktree`), hook fields `command`, `powershell`, `show_output`, and `working_directory`, and exit-code-2 blocking for pre-hooks only.
Already handled verified 2026-07-19, no change needed. Event names and blocking semantics match aix output, and `WindsurfHooksStrategy` already passes `show_output` and `working_directory` through on both install and import.
-
Devin Desktop 3.4.22 made skill `permissions:` frontmatter apply to auto-approvals.
Needs follow-up needs follow-up. aix-managed skills are symlinked from `.windsurf/skills/{name}/` into `.aix/skills/`; if aix skill definitions carry permission metadata, decide whether to emit Windsurf-compatible `permissions:` frontmatter. No change if aix skills stay permissionless.
-
Devin Desktop 3.3.18 made subagents configurable with a default model, and 3.2.16 let subagents call MCP tools directly.
Needs follow-up needs follow-up for investigation. aix currently marks Windsurf agents as `unsupported` because no safe file-backed agent config was documented. Check whether Devin Desktop now documents a subagent config file aix could write; if so, Windsurf agents support could be upgraded.
-
Devin Desktop 3.4.22 added `sandbox.excluded` allow/ask/deny config, 3.3.18 added an `attribution` option to the Devin Local config file, and 3.1.7 added `.devinignore` support alongside `.windsurfignore` and `.codeiumignore`.
No repo change no implementation change needed. aix does not manage sandbox settings, attribution config, or ignore files for Windsurf.
-
Devin Desktop 3.2.16 added an opt-in enterprise plugin system (preview).
No repo change no implementation change needed. aix does not manage Windsurf plugins.
Sources: zed.dev , zed.dev , zed.dev , releasebot.io , byteiota.com
-
Zed 1.6.3 added a manual trigger for the Rules-to-Skills migration, and 1.7.2 fixed old threads that mention Rules being inaccessible after migration. Zed's docs now describe always-on guidance as "Instructions" (project `.rules` / AGENTS.md files and the global AGENTS.md) while reusable on-demand rules became Skills.
No repo change mostly no change, verification needed. aix already writes project `.rules` and user `~/.config/zed/AGENTS.md`, and both remain documented instruction surfaces per search results (`zed.dev/docs/ai/rules`, `zed.dev/docs/ai/skills` were unreachable through the research proxy, so the docs facts here are externally sourced). Verify the docs pages directly before renaming any aix-facing terminology from "rules" to "instructions".
-
Zed 1.6.3 added support for symlinked global skill directories and improved project skills for remote workspaces.
No repo change no change needed; this strengthens aix's approach. aix installs Zed skills as symlinks into `.agents/skills/{name}/` and `~/.agents/skills/{name}/`, which is now explicitly supported at global scope.
-
Zed 1.7.2 moved agent skills management into the settings UI and now allows loading skills whose descriptions exceed the 1024-byte limit (warning instead of failure).
Already handled addressed 2026-07-19. `validateSkill` in `packages/core/src/skills/validate.ts` now warns when a skill description exceeds 1024 bytes (the documented Zed and OpenCode limit), with tests in `packages/core/src/__tests__/skills/validate.test.ts`.
-
Zed 1.10.0 moved key agent-related settings (LLM providers, external agents, MCP servers) into the settings editor.
Needs follow-up needs follow-up for verification only. Externally sourced summaries describe this as a settings-editor consolidation; confirm the `context_servers` key that aix writes into `.zed/settings.json` and `~/.config/zed/settings.json` is still the on-disk format and was not renamed.
-
Zed 1.8.2 added `agent.terminal_init_command` and Anthropic-compatible provider support; 1.9.0 added a remote MCP server quick-add entry in the Agent Panel.
No repo change no implementation change needed. These are editor settings and UI affordances; aix MCP output is unchanged.
-
Zed 1.11.3 added terminal-thread search and turn-end slash commands (`/helpful`, `/not-helpful`).
No repo change no implementation change needed. Zed still has no file-based prompt format, so aix's prompt-to-skill conversion remains correct.
-
No hook support was found in any release through 1.11.3.
No repo change no change; Zed hooks remain `unsupported` in the aix matrix.
Sources: github.com , github.com , github.com
-
Gemini CLI 0.45.0 includes a "context simplification" change in its release notes.
No repo change no generated config change identified from the published release note. Re-check Gemini docs if this becomes a documented settings or memory file surface.
-
Gemini CLI 0.45.0 includes a fix that prevents `mcp list` blacklist bypass.
No repo change no config shape change needed. This affects runtime MCP behavior after aix writes `.gemini/settings.json`.
-
Gemini CLI 0.45.0 includes routing and session history fixes.
No repo change no implementation change needed for aix output.
-
The GitHub releases page also shows 0.46.0 preview builds after 0.45.0.
Needs review no preview release note was used for implementation planning. Re-check once a stable 0.46.x release is published.
Sources: code.claude.com , code.claude.com , code.claude.com
-
Claude Code 2.1.163 added managed `requiredMinimumVersion` and `requiredMaximumVersion` settings.
Needs follow-up needs follow-up only if aix starts writing Claude Code managed settings. Current aix output covers rules, prompts, MCP, skills, agents, and hooks.
-
Claude Code 2.1.163 lets `Stop` and `SubagentStop` hooks return `hookSpecificOutput.additionalContext`.
No repo change no config schema change needed. aix writes hook definitions, not runtime payload or response schemas.
-
Claude Code 2.1.160 added more prompts before editing shell startup files and build tool config files in `acceptEdits` mode.
No repo change no generated config change needed. This is Claude Code runtime safety behavior after aix writes editor files.
-
Claude Code 2.1.154 introduced dynamic workflows and `/workflows`.
No repo change no current implementation change. aix does not model Claude workflow files as a portable feature yet.
-
Claude Code 2.1.154 added plugin `defaultEnabled: false`.
No repo change no current implementation change. aix does not generate Claude Code plugin manifests.
-
Claude Code 2.1.154 changed unapproved `.mcp.json` servers to show as pending approval instead of being auto-approved when output is piped.
No repo change no format change needed. Existing MCP output remains the same, but users may see an approval step inside Claude Code.
-
Claude Code 2.1.154 sends `CLAUDE_CODE_SESSION_ID` and `CLAUDECODE=1` to stdio MCP server subprocesses.
No repo change no config change needed. This affects runtime server environments, not the MCP config files aix writes.
Sources: github.com , github.com , developers.openai.com
-
Codex 0.135.0 added named permission profiles to `/permissions`.
Needs follow-up needs follow-up only if aix adds a portable permissions feature. Current Codex support writes AGENTS.md, MCP, skills, prompt-as-skill shims, and hooks.
-
Codex 0.135.0 added richer `codex doctor` diagnostics for environment, Git, terminal, app-server, and thread inventory state.
No repo change no implementation change needed. This improves diagnostics for users after aix writes config, but it does not change config shape.
-
Codex 0.135.0 made `/status` show remote connection details and server version.
No repo change no config change needed.
-
Codex 0.135.0 added non-interactive `install.sh` and `install.ps1` behavior through `CODEX_NON_INTERACTIVE=1`.
No repo change no current implementation change. aix does not install Codex itself.
-
The current GitHub release page also shows 0.136.0 alpha builds after 0.135.0.
Needs review no alpha release note was used for implementation planning. Re-check once a stable 0.136.x release is published.
Sources: raw.githubusercontent.com , docs.github.com , docs.github.com
-
Copilot CLI 1.0.60 added `builtInAgents.rubberDuckAutoInvoke`.
No repo change no current implementation change. aix does not manage Copilot runtime settings for built-in agents.
-
Copilot CLI 1.0.60 added LSP server config support for `bash`, `powershell`, and `cwd`.
No repo change no current implementation change. aix does not generate Copilot LSP server config.
-
Copilot CLI 1.0.60 changed Windows executable discovery so bare command names no longer search the working directory.
No repo change no format change needed. Generated hook and MCP commands should already avoid relying on current-directory executable discovery.
-
Copilot CLI 1.0.60 added `-r` as shorthand for `--resume`.
No repo change no implementation change needed.
-
Copilot CLI 1.0.54 added `deferred-tool-loading` frontmatter for custom agents.
Needs follow-up needs follow-up if aix should expose this field in portable agent config. Current Copilot agent support passes editor-specific fields through `editor.copilot`.
-
Copilot CLI 1.0.54 migrates legacy nested MCP OAuth keys to `oauthClientId` and `auth.redirectPort`, and MCP OAuth re-authentication honors `redirectPort`.
Needs review inspect before implementation if aix starts emitting Copilot OAuth MCP fields. Current shared MCP normalization does not appear to generate those legacy nested keys.
-
Copilot CLI 1.0.53 fixed `/skills` picker persistence with `--config-dir`.
No repo change no generated config change needed.
-
Copilot CLI 1.0.51 added `preMcpToolCall` hooks and lets `postToolUse` hooks inject `additionalContext`.
Needs follow-up needs follow-up if Copilot hook docs expose stable config for `preMcpToolCall`. Current aix Copilot hook support maps the events already covered by `packages/core/src/editors/strategies/copilot/hooks.ts`.
Sources: cursor.com , docs.cursor.com
-
Cursor 3.7 added Design Mode improvements in the Cursor browser, including multi-select elements and voice input while an agent is running.
No repo change no generated config change needed. This is editor UI behavior, not a rules, commands, MCP, skills, agents, or hooks format change.
-
Cursor 3.7 added Cursor SDK support for custom metadata stores, custom local tools, local-tool auto-review, and nested subagents.
No repo change no current implementation change. These SDK surfaces are separate from the local editor config files aix writes.
-
Cursor 3.7 added a context usage report and canvas Design Mode changes.
No repo change no generated config change needed.
-
Cursor 3.7 SDK releases changed run metadata and local shell packaging behavior.
No repo change no current implementation change. Re-check only if aix starts generating Cursor SDK scripts or SDK config.
Sources: github.com , opencode.ai , opencode.ai
-
OpenCode 1.16.0 added skill discovery and file-based agent loading.
Needs follow-up needs follow-up. aix already writes OpenCode skills under `.opencode/skills/{name}/` and agents under `.opencode/agents/*.md`, but this release should be checked against the current OpenCode docs to confirm whether new discovery paths or frontmatter fields are now preferred.
-
OpenCode 1.16.0 added managed workspace cloning that keeps dirty and untracked files.
No repo change no generated config change needed.
-
OpenCode 1.16.0 added moving sessions between workspaces and directories.
No repo change no generated config change needed.
-
OpenCode 1.16.0 fixed Windows path normalization in migrated storage.
No repo change no direct config change needed, but keep Windows fixture coverage for OpenCode config discovery.
-
OpenCode 1.16.0 fixed the GitHub extension so it refuses to commit without an existing git author identity.
No repo change no generated config change needed.
Sources: docs.devin.ai , windsurf.com
-
Windsurf is now Devin Desktop as of the 3.0.12 release note.
Needs follow-up needs follow-up for naming and detection only if the rebrand changes config directories, binary names, or docs URLs. Current aix editor ID remains `windsurf`.
-
Devin Desktop 3.0.21 added a command to rerun migration from Windsurf.
No repo change no generated config change needed.
-
Devin Desktop 2.3.15 increased remote server startup timeout.
No repo change no generated config change needed.
-
Devin Desktop 2.3.15 updated the bundled Devin Local agent. The note says Devin Local is aware of open editor files, adds MCP permission choices for all tools on a server, repairs hooks so they can block user prompts, improves plan mode in the OS sandbox, and persists "Always Allow" permission grants.
Needs follow-up needs follow-up for hook behavior only if the Devin Desktop hook docs changed event names, payloads, or config shape. Existing aix Windsurf hook output should be checked against the current Devin Desktop docs before changing code.
Sources: zed.dev , zed.dev
-
Zed 1.5.3 fixed commit message generation so it respects instructions from global `AGENTS.md`.
No repo change no generated config change needed. This confirms that the Zed 1.4.2 global rule path `~/.config/zed/AGENTS.md` remains relevant.
-
Zed 1.5.3 fixed deprecated Rules being automatically included in new agent requests.
No repo change no generated config change needed. aix already moved Zed skills to native `.agents/skills/{name}/` support in the prior research pass.
-
Zed 1.5.3 fixed global skill edits appearing without a file path in change summaries.
No repo change no generated config change needed. This affects runtime display after aix writes global skills.
-
Zed 1.5.3 fixed UTF-8 parsing issues when loading skill frontmatter with multi-codepoint graphemes.
No repo change no generated config change needed. aix writes skill frontmatter as plain YAML; keep tests focused on ASCII unless the repo intentionally supports non-ASCII skill metadata.
-
Zed 1.4.4 and 1.4.3 only show Copilot-chat model dropdown fixes in the stable release page.
No repo change no generated config change needed.
Sources: zed.dev , github.com , github.com
-
Zed 1.4.2 introduced native Agent Skills, replacing the Rules Library.
Needs review switch from `PointerSkillsStrategy` to `NativeSkillsStrategy` with `editorSkillsDir: '.agents/skills'`. aix copies skills to `.aix/skills/` and symlinks into `.agents/skills/<name>`. Symlinks are explicitly supported.
-
Zed 1.4.2 removed the Rules Library UI entirely.
Needs review pointer rules are no longer the right approach for skills. The `PointerSkillsStrategy` usage in the Zed adapter is removed.
-
Zed 1.4.2 added `~/.config/zed/AGENTS.md` as the user-scope global rules file.
Needs review `ZedRulesStrategy.getGlobalRulesPath()` now returns `.config/zed/AGENTS.md`. User-scope rules install to `~/.config/zed/AGENTS.md` instead of being dropped with a limitation warning.
Sources: github.com
-
Gemini CLI 0.42.0 introduced an Auto Memory Inbox with a canonical-patch contract for skill management.
Needs follow-up needs follow-up only if Gemini documents stable memory files or skill patch files that aix should manage. Current aix Gemini support writes rules, prompts, MCP, and hooks.
-
Gemini CLI 0.40.0 introduced MCP resource tools and a four-tier memory management system.
No repo change no current MCP config change found in this run. If resource tool settings become user-configurable, update the Gemini MCP strategy.
-
Gemini CLI 0.24.0 and later release notes describe Agent Skills support, `/skills reload`, and `/skills install/uninstall`.
Needs follow-up needs follow-up. `packages/schema/src/editor-support.ts` should be checked to confirm whether Gemini skills are intentionally unsupported or now should be native.
-
Gemini CLI hooks are configured in `.gemini/settings.json`.
Already handled already aligned. `packages/core/src/editors/strategies/gemini/hooks.ts` writes hooks into `.gemini/settings.json`.
Sources: code.claude.com , code.claude.com
-
Claude Code 2.1.147 added the `Workflow` tool behind `CLAUDE_CODE_WORKFLOWS=1`.
Needs follow-up needs follow-up only if aix decides to model Claude workflow files or workflow-specific permissions. Current aix support is focused on rules, prompts, MCP, skills, agents, and hooks.
-
Claude Code 2.1.145 added `background_tasks` and `session_crons` fields to `Stop` and `SubagentStop` hook input.
No repo change no schema change needed for hook input payloads because aix stores hook config, not runtime payload types. Docs should mention these fields if the hook guide starts documenting event payload examples.
-
Claude Code 2.1.147 fixed hook `if` conditions such as `PowerShell(git push*)`.
Already handled already exposes the Claude Code `if` action field in `packages/schema/src/hooks.ts` and passes it through in `packages/core/src/editors/strategies/claude-code/hooks.ts`.
-
Claude Code 2.1.147 fixed MCP pagination for servers, resources, templates, and prompts.
No repo change no install-format change needed. This affects runtime Claude Code behavior after aix writes MCP config.
-
Claude Code 2.1.147 renamed `/simplify` to `/code-review`.
No repo change no aix implementation change. aix prompt and skill names are user-defined and do not depend on Claude Code built-in slash command names.
Sources: developers.openai.com , github.com
-
Codex 0.133.0 enables goals by default and stores goal progress across active turns.
No repo change no current code change. aix does not model Codex goal storage yet.
-
Codex 0.133.0 expanded permission profiles with list APIs, inheritance, managed `requirements.toml` support, runtime refresh behavior, and stronger Windows sandbox integration.
Needs follow-up needs follow-up research before implementation. Current aix Codex support focuses on AGENTS.md, MCP, skills, prompts-as-skills, and rules; permission profile files are not modeled in `ai.json`.
-
Codex 0.133.0 made plugin discovery easier to inspect and added visible installed versions and marketplace roots.
No repo change no current code change. aix does not install Codex plugins.
-
Codex 0.133.0 added more extension lifecycle events, including subagent start/stop, tool execution, turn metadata, and async approval/turn processing.
Needs follow-up needs follow-up if Codex exposes these events through a stable config file. `packages/schema/src/editor-support.ts` currently marks Codex hooks as unsupported.
-
Codex 0.133.0 made AGENTS instruction loading more reliable, including local global reads and invalid UTF-8 warnings.
No repo change no format change needed. aix already treats `AGENTS.md` as the Codex instruction surface.
Sources: raw.githubusercontent.com , docs.github.com
-
The current npm package version is 1.0.51. The public Copilot CLI changelog inspected during this run listed detailed notes through 1.0.49.
Needs review use 1.0.51 as the checked package version, but compare future work against 1.0.49 release-note details unless GitHub publishes missing 1.0.50 and 1.0.51 notes.
-
Copilot CLI 1.0.49 says repo hooks in `.github/hooks/` load in prompt mode when the folder is already trusted.
Already handled already aligned. `packages/core/src/editors/strategies/copilot/hooks.ts` writes project hooks under `.github/hooks/hooks.json`.
-
Copilot CLI 1.0.49 accepts MCP server configs with no `args` field and treats it as an empty args list.
No repo change no code change needed unless aix validation rejects missing args for Copilot MCP output. Existing MCP normalization should be checked before changing code.
-
Copilot CLI 1.0.49 reports MCP stdio servers as `stdio` instead of `local`.
Needs follow-up needs follow-up only if generated Copilot MCP config still emits a deprecated local type. Inspect `packages/core/src/editors/strategies/copilot/mcp.ts` before changing.
-
GitHub custom agents support `tools`, `mcp-servers`, `metadata`, `disable-model-invocation`, and `user-invocable`; `argument-hint` and `handoffs` are ignored by the cloud agent.
Needs follow-up needs follow-up if aix expands Copilot agent support beyond the currently documented fields.
Sources: cursor.com , prod.cursor.com
-
Cursor 3.5 added Automations to the Agents Window and supports automations with multiple repositories or no repository.
No repo change no current code change. aix does not model Cursor Automations yet.
-
Cursor 3.4 improved cloud-agent environment configuration as code, including Dockerfile-based setup, build secrets, environment version history, and scoped secrets.
No repo change no current code change. These settings are not part of the local editor config files aix writes today.
-
Cursor 2.4 introduced Agent Skills in the editor and CLI with `SKILL.md` files.
Already handled already aligned. `packages/schema/src/editor-support.ts` lists native Cursor skills, and Cursor skill installs are handled through the editor strategy layer.
-
Cursor 2.4 moved MCP server definitions and tools into JSON files under `.cursor`, loaded only when needed.
Already handled already aligned for project MCP output at `.cursor/mcp.json`.
-
Cursor 2.4 expanded hooks, including stop hooks, `beforeSubmitPrompt`, `PreToolUse`, `PostToolUse`, team and MDM hooks in CLI, and Claude Code hook compatibility.
Already handled already partly aligned. `packages/core/src/editors/strategies/cursor/hooks.ts` maps the supported aix hook events to Cursor native names; MDM/team cloud hook config is not modeled.
Sources: github.com , opencode.ai , opencode.ai
-
The current npm package version is 1.15.7. The GitHub releases page inspected during this run showed 1.15.6 as the latest detailed release note.
Needs review use 1.15.7 as the checked package version, but compare release-note details against 1.15.6 until the detailed 1.15.7 notes are visible.
-
OpenCode 1.15.6 fixed agent and command name resolution from relative config paths.
Needs review already likely aligned. aix writes OpenCode `instructions` and `command` config using project-relative paths; keep this note for regression testing.
-
OpenCode config supports JSON and JSONC, merged global/project/remote config, `instructions`, `command`, `mcp`, plugins, permissions, agents, and managed settings.
Already handled already partly aligned. aix supports `opencode.json` and `opencode.jsonc`, imports config-defined `instructions` and `command` prompts, writes MCP under top-level `mcp`, and treats hooks as unsupported.
-
OpenCode agent docs mark `tools` as deprecated in favor of permissions.
Needs follow-up needs follow-up if aix emits OpenCode agent files with `tools`. Current OpenCode support should prefer permission-oriented fields if agent output expands.
Sources: windsurf.com
-
Windsurf 2.3.9 fixed terminal processing performance, restored conversation sharing, and repaired WSL path resolution for the Devin local agent.
No repo change no config-format change found for aix.
-
Windsurf 2.2.17 included MCP server bug fixes.
No repo change no schema change found in this run. Keep existing Windsurf MCP output unless a specific MCP config key changes.
-
Windsurf 1.9577.24 added support for loading `SKILL.md` files from `.windsurf/skills/`.
Already handled already aligned. `packages/schema/src/editor-support.ts` documents native Windsurf skills and the strategy layer handles `.windsurf/skills/`.
-
Windsurf 1.9552.21 added support for reading skills from `.agents/skills`, added `rules_applied` to the `post_cascade_response` hook, and fixed `post_write_code` hooks for all code editing tool formats.
Needs review `.agents/skills` compatibility is documented. Runtime hook payload fields such as `rules_applied` do not require a schema change unless aix starts documenting event payload examples.
-
Windsurf 1.9566.9 added `POST_CASCADE_RESPONSE_WITH_TRANSCRIPT`, team hook visibility, and MCP OAuth improvements.
Already handled already partly aligned. `packages/core/src/editors/strategies/windsurf/hooks.ts` maps `post_response_with_transcript` to `post_cascade_response_with_transcript`.
Sources: zed.dev
-
Zed 1.3.5 added terminal threads in the sidebar and Agent Panel.
No repo change no config-format change found for aix.
-
Zed 1.3.5 added a `subagent_model` setting for spawned subagents.
Needs follow-up needs follow-up only if aix adds Zed agent/subagent config support. Current Zed support focuses on rules, prompts, MCP, and skill pointer rules.
-
Zed 1.3.5 added user confirmation before tools modify `.agents/skills/` or `~/.agents/skills/`.
No repo change no code change needed. This protects skills installed by aix or other agents after they exist on disk.
-
Zed 1.3.5 added inline image and Mermaid rendering in the agent.
No repo change no config-format change found for aix.
