agnostic-ai v0.82.0: guided upgrades and hooks that match each tool

Agnostic-ai released v0.82.0, which makes previously ignored Codex sandbox settings in `outputs.codex.config.sandbox` take effect and defaults interactive upgrade offers to Yes. The release adds optional `rtk` and `caveman` builtins for command output reduction and shorter responses, plus a `before-tool` hook that maps to Kiro CLI V3 and lets a script exit 2 to block a shell tool call. Users upgrading from v0.81.0 must stop any running `agnostic-ai watch` process, install via `npm install -g agnostic-ai@0.82.0`, and run `agnostic-ai sync --check` before restarting watch.

Upgrading from v0.81.0 Stop any running agnostic-ai watch process before changing the installation or project config. Codex sandbox settings now take effect. If your config sets outputs.codex.config.sandbox , review the value before the first sync. Use read-only , workspace-write , or danger-full-access . An invalid value stops sync. A valid value was previously ignored and now controls the generated session setting. See Codex configuration https://agnostic-ai.org/docs/targets/codex/ . Interactive upgrade offers default to Yes. When a terminal command offers a release, pressing Enter accepts the update and its listed project changes. Type n to continue your original command. Review the offer before accepting. Install the release through your existing installer. For a global npm installation: npm install -g agnostic-ai@0.82.0 If your project uses an exact version pin, run: agnostic-ai upgrade --requires agnostic-ai sync --check upgrade --requires sets requires and the schema link to the installed release, then syncs. It replaces a deliberate minimum or range too, including an existing local override. To preserve a minimum or range, keep that value, update the schema URL to v0.82.0 , review agnostic-ai migrate --dry-run , and run agnostic-ai sync followed by agnostic-ai sync --check . See guided upgrades https://agnostic-ai.org/docs/cli-reference/maintain/ upgrade . Restart watch after the check passes. RTK and Caveman are optional additions; upgrading does not turn either on. What agnostic-ai changed Upgrade from the command you already use Interactive commands offer available releases with their release notes and project guidance. Accepting the offer updates the installation, applies the migrations your project needs, and checks its generated files. Project version and schema changes appear in the prompt. See upgrade behavior and installer support https://agnostic-ai.org/docs/cli-reference/maintain/ upgrade . Choose command output reduction or shorter responses Add rtk or caveman to your existing builtins list, keeping the names already there. For a project with no other built-ins: builtins: rtk RTK adds a Claude Code hook that rewrites supported commands to reduce their output. Install RTK separately. Caveman adds a response skill to tools that support skills; use builtins: caveman to choose it instead, or add both names. Neither is installed or enabled automatically. Keep one RTK hook when you already have a global or project installation. Claude Code applies permissions to the rewritten command, so an approval rule for git status does not cover rtk git status . The published tests found no additional cost benefit from adding Caveman to RTK; choose its response style separately. See setup, approval rules, and measured results https://agnostic-ai.org/docs/rtk-and-caveman/ . Check Kiro tool calls before they run agnostic-ai now maps a shared before-tool check to Kiro CLI V3. Create your own executable .agnostic-ai/scripts/check-shell.sh , which reads the tool call from stdin, exits 0 to allow it, or exits 2 with a reason on stderr to block it. Then add: name: check-shell on: before-tool match: shell command: .agnostic-ai/scripts/check-shell.sh Sync writes PreToolUse with an anchored list of Kiro's documented shell IDs. The other built-in kinds also select known IDs, so a read guard does not accidentally run on an MCP tool whose name contains read . edit selects file write tools and excludes code . mcp:<server selects that server, including one named mcp , builtin , or powers . Only before-tool maps portably to Kiro in this release. Other portable events have different blocking behavior, and the portable decision: stdout wrapper is not supported there. Use Kiro's own event names when you need those behaviors. hook run now follows CLI V3 prompt and Stop exits, and leaves ambiguous MCP server/tool IDs untested with an explanation. See portable hooks https://agnostic-ai.org/docs/spec-format/hooks/ portable-events and Kiro https://agnostic-ai.org/docs/targets/kiro/ . Share hooks and servers across workspaces Home specs now reach Kiro hooks and Antigravity and Warp MCP servers: agnostic-ai sync --global --only kiro agnostic-ai sync --global --only antigravity agnostic-ai sync --global --only warp Kiro hooks land in ~/.kiro/hooks/ , Antigravity servers in ~/.gemini/config/mcp config.json , and Warp servers in ~/.warp/.mcp.json . agnostic-ai also uses KIRO HOME as an output-root override for its Kiro files; the vendor docs do not confirm that Kiro loads hooks from that override. Global import supports these surfaces too. Sync preserves unrelated user entries and reports conflicts. Kiro import skips hook files one spec cannot reproduce, including mixed-trigger files. Disabled Warp servers are omitted because Warp starts its global servers automatically. See global output https://agnostic-ai.org/docs/target-behavior/ global-output . Make failed guards and preserved settings visible Claude Code hook specs accept failClosed: true to block a supported synchronous action when its handler fails or times out. Sync names events and background handlers where Claude ignores that setting, rather than implying they protect the action. Imported handler settings and onFailure values survive sync without duplicate groups. See hook fields https://agnostic-ai.org/docs/spec-format/hooks/ . Trae preserves an explicitly empty agent tool list and imports shared .agents/skills/ with native Trae skills taking priority. Codex import preserves manual skill sources. Repo memory gives OpenCode and Devin CLI clearer save instructions, and warns when a hand-written Codex config lacks access to the personal memory folder. Commands also do less repeated work on 500-spec projects. Sync and import preserve more hand-written content, graph --target selects one tool, and doctor lists installed tools in name order. Full list: v0.82.0 release notes https://github.com/Chemaclass/agnostic-ai/releases/tag/v0.82.0 . What changed in the targets Claude Code's failure setting has event limits. Its current reference, checked on 2026-10-10, requires Claude Code 2.1.295 or later for onFailure: "block" . The setting does not affect Stop, SubagentStop, TaskCompleted, TeammateIdle, or background command hooks. This release expresses the setting and reports those limits. Use a synchronous tool or prompt gate when failure must prevent an action. Claude Code hooks reference https://code.claude.com/docs/en/hooks block-the-action-when-a-hook-fails . Kiro CLI V3 treats prompt and Stop hooks differently. Documentation checked on 2026-10-10 says prompt hooks cannot reject a prompt, and Stop exit 1 requests another agent turn. CLI V3 also loads global hooks alongside workspace hooks. The shipped Kiro changes above follow these contracts. Its agent references now agree on category tags; knowledge and invoke sub agent use direct IDs, while keyboardShortcut belongs to CLI 2.x. Those are documentation discoveries, not claimed launch dates. Hook types https://kiro.dev/docs/hooks/types , CLI V3 hook migration https://kiro.dev/docs/cli/v3/hooks-migration.md , hook locations https://kiro.dev/docs/hooks , agent configuration https://kiro.dev/docs/custom-agents/configuration-reference , built-in tools https://kiro.dev/docs/tools . Antigravity and Warp document global MCP files. Their references were checked on 2026-10-10. Antigravity uses serverUrl for remote connections, which global sync now writes from portable url . Warp starts native global file-based servers by default; project servers require approval. The new global mappings use each tool's native file and preserve that distinction. Antigravity MCP https://antigravity.google/docs/mcp?tab=ide , Warp file-based MCP https://docs.warp.dev/agents/capabilities/mcp .