[rmstdope]

Bead: cb-70mv — R15's suite waited on the 30-second work-pane refresh.

What changed

tests/fleet-view-restart.sh only:

• The view now runs on the shared tests/lib/pty-drive (as fleet-keys.sh and fleet-role-triggers.sh do), which gives it a keys fifo; the suite's private pty.py is gone.
• A refreshing <predicate> wrapper presses g (refresh everything) at most every two seconds, then asks the predicate. Steps 2 (start on cb-b2 after the closing pass) and 3 (held_by_guard after the give-back) wait through it, bounded at 60s, instead of waiting on app::WORK_REFRESH_INTERVAL. Pressed repeatedly because a g landing on an in-flight work read is dropped, and a read begun before the pass ended answers the old board.
• The failure evidence also prints the view's stderr.

Validation

• Before: per-step timing locally, step 2 = 26s and step 3 = 10s (both on the timer); 71s on CI run 42803911818.
• After: time bash tests/fleet-view-restart.sh — 16.5s, 17.5s, 19.6s locally, all three passes; 28 consecutive green runs of the suite while chasing a flake (below).
• bash tests/gate: every suite and the Rust and UI legs green except tests/fleet-keys.sh, which timed out ("the stub agent to start", a 20s bound) in both runs. It is untouched here, passed twice alone, and the same failure hit two other producers' worktrees (cb-15cw, cb-yf8h) on this machine in the same hour; CI decides.

Deviation from the plan

• The planned RED (a 25s bound on the old suite) stayed green because the timer's phase lined up; per-step timing is the evidence instead. Recorded in the bead's design.
• Seen once in early runs: the view stayed read-only ("could not take charge") for the whole 90s. main.rs resolves the lease endpoint once at startup (read_supervisor_endpoint(...).ok()), so one slow fleet-supervisor under load leaves a view unable to supervise for life. Not reproduced in 28 later runs and not in this diff's scope; noted for the navigator.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WCDKexJtfAU67v3sjW244j

[rmstdope]

Review (cold read, reviewed_head: 2081a294f069471a5241ab19196db92f91474fca) — reviewed before merge under the producer's review practice, by an agent given the diff and the bead, and not the producer's reasoning. Reviewer key default: claude / opus / medium.

No findings.

Checked: the plan is met (only the suite changed, pty-drive + fd-7 keys fifo, steps 2–3 through refreshing, 60s bounds, header names WORK_REFRESH_INTERVAL only as no longer waited on); g is honoured (reviewer's run 15.5s; step 2 took 3s, step 3 4s, against 26s/10s on the timer); teardown leaves nothing behind after pty-drive's SIGKILL (the pkill on the marker covers it); fifo opened after the fork, never blocks; predicates still key on timestamps after the second end, so a stale read cannot satisfy step 3. Non-code note: the plan's Known traps still said wait_refreshing; corrected in the bead.

[Copilot]

Note

Copilot was unable to complete this review at the selected tier due to model policy or plan restrictions.

🟢 Approval recommended

The change is a contained refactor to use the shared pty-driving helper while adding bounded refresh/polling and enhanced stderr-based diagnostics, with no review evidence indicating unresolved problems.

No open issues 🎉

What changed in this PR

Speeds up and stabilizes the R15 system suite by actively triggering fleet-view refreshes instead of waiting on the work pane’s fixed 30s refresh interval.

Changes:

• Switch tests/fleet-view-restart.sh from a private pty helper to shared tests/lib/pty-drive (keys FIFO + screen capture).
• Add a refreshing wrapper that throttles g (refresh-all) presses while polling predicates within bounded waits.
• Improve failure diagnostics by including the view’s stderr tail in the suite’s evidence output.

File	Description
tests/​fleet-view-restart.sh	Migrates the test to pty-drive, adds a throttled g-driven refresh/polling helper to avoid 30s timer waits, and expands failure evidence to include fleet-view stderr.

_{Review effort: Balanced}