v7.2.3 — the status line stops overstating a viewed agent's context

When you view a subagent, the status line estimates that agent's context window. It used ~200K for any agent on a different model from the main conversation, and that is no longer true for current models. A Haiku 5.5 agent read 15% (31K/~200K) when it was at about 3%.

The estimate now comes from Claude Code's local model catalog cache and keeps its ~. sonnet-4-0, sonnet-4-5, sonnet-4-6, opus-4-6 and CLAUDE_CODE_DISABLE_1M_CONTEXT stay at 200K, and so does anything the catalog cannot answer. The status line reads one more local file; the plugin README says so.

This release also records a full sync with Claude Code 2.1.316. There is no new hook event and no new status line field. viewed_agent still does not exist, so the view sensor stays.

Verified

• Live on Claude Code 2.1.316, Windows: a Haiku 5.5 agent read 3% (31K/~1M), a Sonnet 4.6 agent 12% (25K/~200K), and the main view kept its exact window.
• The view sensor followed 12 of 12 transitions.
• --resume loads the installed plugin version, with a mock model API. A session already open keeps its version until /reload-plugins.
• CI: 884 tests on three operating systems and three Python versions; the validator on 2.1.316.

Not verified: macOS or Linux live, other accounts' context windows, sonnet-4-0 and sonnet-4-5 (read from the binary, not run).

Not added: a variant for the new plugin_notification type, which needs a sound in both themes.