agent status --json

[graywolf336]

atomic agent status had no machine-readable form. Tools that need to decide something — "does this agent have hooks installed, so should I install them before recording a run?" — could only scrape the ✓/○ lines, which are prose.

$ atomic agent status --json
{"agents":[{"name":"claude-code","display_name":"Claude Code","detected":true,"hooks_installed":true},
           {"name":"codex","display_name":"Codex","detected":true,"hooks_installed":false}, ...],
 "sessions":[...], "totals":{"sessions":203,"turns":2,"files_touched":0}}

Shape

• Every registry agent, sorted by name. Absent ones are listed too — a caller choosing whether to install must tell "Atomic does not know this agent" from "known, and not set up", and an omitted entry cannot say which.
• Sessions carry what --verbose prints; --verbose is ignored in JSON.
• sessions_error appears only when the session store cannot be read. The human output prints that and continues; dropping it would make a broken store look like an empty one.

Follows agent lifecycle's existing --json: same flag, serde_json::to_string, snake_case.

Why: Sherpa gates on this. Without the integration installed, an agent's turns are never recorded, so runs complete and the ledger stays empty — a silent gap. Detecting it needs a stable signal.

Checks: cargo fmt --all --check, cargo clippy -- -D warnings, cargo check, and 9 tests in agent::status (4 new) all pass. Verified against a real repo — output matches the text form exactly.

One note: the new tests do not assert that a bare fixture detects nothing. Several agents are configured in $HOME rather than the repository, so that would depend on the machine running the test; they assert the flags agree with the registry instead.