Only a merged or closed pull request is landed; unread and idle are visited once on change, awaiting again on age (#1707)

[suleimansh]

Implements items 1 and 3 of #1707 as decided in #1707 (comment).

Item 1 — content.js. A row showing only a pull request is landed only for #N · Merged / #N · Closed; #N · Open is idle. A status word on the row still wins. Any other pull-request wording stays unknown with the label carried, as before for labels the bridge does not know.

Item 3 — driver-plan.js. PARKED + one REVISIT_MS becomes two sets: ON_CHANGE = {awaiting, unread, idle} (visited when never seen or when the list's word changed) and ON_AGE = {awaiting} (also visited again after five minutes unchanged). A queued answer still forces a visit whatever the status. running, landed, missing are never visited on their own.

Note on idle: the decision says an open-PR session "is visited on change, never on age". On main the planner never visits idle at all, so honouring that puts idle — the list's own "Idle" rows included — in the on-change set. Cost: one visit per session when it turns idle, and on a fresh worker (empty seen) idle sessions join the first cycles' backlog, four visits a beat. If you'd rather keep idle unvisited, drop it from ON_CHANGE and item 1 becomes a report-only change (idle instead of landed on the dashboard).

check.mjs — three cases, each red on main before the fix and each verified to go red again when its rule is reverted:

• an open pull request alone reads as idle, a merged or closed one as landed
• awaiting/unread/idle are visited when never seen or on a status change; running/landed/missing are left alone
• after five minutes unchanged only awaiting is visited again; unread and idle are not

Specs: content.SPEC.md, driver-plan.SPEC.md (glossary "parked" → "stopped", now including idle), check.SPEC.md, the extension SPEC.md, FEATURES-SPEC.md:170.

Nothing in dist/ changes — the daemon needs no rebuild; the extension needs a reload.

🤖 automated · Fable 5, effort high

[suleimansh]

Dogfood on the live Driver, 2026-08-26 00:28–01:15Z (daemon pid 74851 on d3ce73a6, extension reloaded twice). Evidence = onBridgeStatus polled every 3 s (lastContact route + the page's hello with its session id) and the live list's aria-labels read from a second claude.ai tab.

Driver keeps cycling on the new script — /_bridge/statuses every 30 s, page reload every ~90 s, questions 0, version 0.12.0.

Item 1 — one bug found, fixed in 29684d3a. To get a real not-landed row I reopened draft PR #1706 (session session_01TU8F5…, inside the 12-hour window). claude.ai's row read #1706 · Draft — a state the first commit did not cover (only Open → idle), so the session came back unknown and was never visited. Now Merged/Closed → landed and every other pull-request state → idle, exactly the decision's wording; check.mjs carries a Draft case that is red on the Open-only rule. After the reload the worker's first cycle visited it once as idle (01:02:58), together with the list's Idle session (01:03:01). Closed #1706 again → row #1706 · Closed → no visit across the next 7 cycles.

Item 3 — no age revisit. Neither idle session was visited a second time across 16 cycles (7½ min, the revisit is 5). The awaiting half — revisited on age — was not exercised: no session is awaiting right now; the check.mjs case covers it.

Side observation, not this PR's: claude.ai refreshes a row's pull-request state lazily — the row said Closed for six minutes after the reopen until the session page was opened. A landed session is never visited, so a pull request reopened later flips on the list only when claude.ai refreshes it on its own.

🤖 automated · Fable 5, effort high