feat(antigravity-cli): add supervised background task skill

[scarletkc]

Add a self-contained Antigravity CLI skill so supervising agents can delegate bounded work, inspect progress, and continue the same conversation without waiting in a foreground terminal. Native AGY can report SUCCESS and exit 0 after a permission denial or print timeout; the client collects both output channels and classifies these outcomes before reporting completion.

Changes

• Add start, status, wait, reply, cancel, and close, with UTF-8 prompt files/stdin and ASCII-safe JSON output. Each turn uses official streaming JSON and resumes its exact conversation ID in a fresh AGY process.
• Bind both cwd and --add-dir; inherit authentication, models, and permissions unless explicitly overridden. Document effort selection, authorized edit modes, artifact targets, result verification, and visible failures.
• Preserve partial responses, tool errors, permission denials, and native results. Enforce startup, turn, exit, and idle deadlines, including cancellation while a large prompt is blocked on stdin. Stop task-owned process trees using Windows Job Objects or Unix process groups.
• Extract Grok's existing task store, worker launcher, prompt handling, and Windows process ownership into one maintained runtime, with checked bundled copies so both skills install independently. Grok's command surface and task database schema are unchanged; Antigravity uses a separate state directory.
• Update the generated README catalog and add offline process/protocol regressions plus isolated ZIP execution coverage.

Validation

• uv run --python 3.12 -m unittest discover -s tests -v: 77 tests passed, including all 20 existing Grok tests.
• uv run --python 3.12 scripts/sync_agent_runtime.py --check
• uv run --python 3.12 scripts/update_readme_skills.py --check
• gh skill publish --dry-run
• Skill Creator validation, Ruff checks, and git diff --check passed.
• uv run --python 3.12 scripts/package_skill.py antigravity-cli: ZIP contents and standalone imports verified.
• Real Windows AGY 1.2.1 calls with a per-task claude-sonnet-4-6 override verified file creation/readback, exact conversation recovery, blocked permissions, and cancellation. The final version wrote readiness.txt, read it back, recalled its marker in a subsequent process, and closed cleanly.
• An independent agent using only the skill, reference, and help completed Chinese conversation recall, observed and recovered from a blocked operation, and used the revised edit guidance to write and independently verify a UTF-8 workspace file. No global configuration or permission bypass was required.

Selected JSON fields from the final live CLI test:

{"status":"completed","closed":false,"turn":1,"text":"VERIFIED\n","error":null,"denied_actions":[]}
{"file_matches":true}
{"status":"completed","closed":false,"turn":2,"text":"AGY_READY\n","error":null,"denied_actions":[]}
{"status":"completed","closed":true,"turn":2,"text":"AGY_READY\n"}

Boundaries

This is supervised delegation, not an OS sandbox. AGY does not expose programmatic headless approvals; denied tools remain explicit failures requiring an appropriate scoped permission or workflow change. Cancellation stops the task's process tree and does not undo prior actions. There is no automatic prompt retry, authentication switch, or provider fallback. Each reply pays CLI startup cost so trailing diagnostics and process exit can be checked reliably.