Ground truth check before writing this off as too big: Handler is already a
ThreadingHTTPServer (SSE-capable for free), and thinking_payload() -- one sic
hertz SSH round-trip per call -- was already being hit independently by every
open tab every 1.5-3s via client-side setInterval. A naive "just hold the
connection open and loop the existing fetch" SSE implementation would have
made that worse: N open tabs = N concurrent SSH-polling loops running forever.
So: one background daemon thread polls thinking_payload() on the same 1.5s
cadence the client already used, into a version-counted shared cache
(_THINK_CACHE + _THINK_LOCK). GET /api/thinking (snapshot, unchanged
contract) and the new GET /api/thinking/stream (SSE) both read from it --
zero additional sic calls per client, and strictly FEWER total SSH
round-trips than todays per-tab polling, not just parity. Cache-update logic
pulled out as _update_thinking_cache() specifically so it is unit-testable
without a thread/sleep in the way (5 new tests, tests/test_bullseye_thinking.py).
Client (page.py): EventSource replaces the hand-rolled setTimeout poll loop
-- simpler, not just live, since EventSource reconnects on drop natively.
Falls back to the old poll loop when window.EventSource is absent.
69/69 bullseye tests green (64 existing + 5 new). Live-smoke-tested standalone
(snapshot + SSE endpoints both verified with curl) before wiring into the
real service.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Markus Fritsche
aff56bb582