Upgrading from v0.73.0

Stop sync --watch. Before a postinstall can sync, align your requirement and release-pinned schema with v0.74.0. requires: "0.74.0" now pins the generator exactly; >=0.74.0 permits newer releases. A v0.73 binary cannot parse exact pins or ranges, so install the new release when it reports AAI-004. Choose your existing channel:

Standalone macOS/Linux: agnostic-ai upgrade --version v0.74.0.

Project npm dependency: npm install --save-dev --save-exact agnostic-ai@0.74.0.

Use the local runner for project dependencies and check agnostic-ai --version. Remove Claude's retired outputs.claude.settings.taskOutputMaxChars if configured. Regenerate, review and stage tracked changes, then check what you staged:

agnostic-ai sync
agnostic-ai sync --check --against index

In CI, use --against HEAD if installation or initial sync may repair the working tree before checking. Fetch two commits to identify leftovers from deleted specs. Installation, CI.

What agnostic-ai changed

Pin the generator and check the actual Git state

requires accepts exact releases and ranges, preventing a different generator from rewriting committed outputs:

requires: "0.74.0"
# Or a bounded range: ">=0.74.0 <0.75.0"

Update a versioned schema comment to the v0.74.0 tag too. AAI-005 names the matching install command, including the project's package manager for a local dependency. Requirements.

New Git-state checks render the selected specs and compare the outputs tracked in that state:

agnostic-ai sync --check --against index
agnostic-ai sync --check --against HEAD

The first catches unstaged regeneration; the second catches a stale commit even after CI repairs its working copy. Ignored outputs are skipped. Proven generated leftovers also fail, including headerless files identified through the previous commit.

Preserve ownership in fresh checkouts

sync.output-manifest: true writes .agnostic-ai/outputs.lock with generated paths and checksums. Commit it with outputs so fresh checkouts can detect deleted-spec leftovers without .sync-state; edited files remain for you to resolve. sync --keep-edits also protects uncommitted tracked-output edits without a local sync record. Plain sync still regenerates them. Sync configuration.

Reuse environments in the Codex app

Portable environments now produce Codex worktree setup, cleanup and action buttons:

name: local
target: codex
setup: pnpm install --frozen-lockfile
cleanup: docker compose down
dev-commands:
  - name: Docs
    command: [pnpm, dev]
    cwd: apps/docs

The action keeps its working directory; setup-windows supplies Windows setup. import codex reads the environment back. Codex has no action equivalent for port, auto-port, env or URL, and does not consume the older install/terminals fields. Codex environments.

Keep shared review text and imports intact

Reviews can inline a project file with a standalone @path line, and explain --inputs includes those dependencies for hook triggers. Codex import preserves hand-written instruction sections; Cursor environment import retains JSON comments. Codex lint also checks the root-to-scope instruction chain against 32 KiB: changing the lint threshold does not change the tool's loading limit. Complete release notes.

What changed in the targets

  • Kiro (September 28): experimental Opus 5.5 is available on paid plans in us-east-1 and eu-central-1, at 2.0x credits. Refresh the client. Existing agent model passthrough supports selection; access depends on Kiro's plan and region. Announcement.
  • Amp (September 28): medium switches to Opus 5.5, while ChatGPT Only retains GPT-5.6 Sol. Choose the model through Amp modes; portable Settings do not configure its Dial. Announcement.