v0.10.2 — configurable chat watchdog + 14-day chart accuracy

0.10.2 — 2026-05-08

Two limits hit by busy fleets are now configurable / accurate.

Configurable chat watchdog timeout

The chat-session watchdog (kills the Claude CLI subprocess if no stream events arrive mid-turn) was hardcoded to 5 minutes — too short for agents running long tools (large web fetches, builds, big greps), causing spurious "no response from the CLI" errors.

Now configurable via chatWatchdogMinutes in plugin settings (default 10 min, range 1–60). ChatSession reads the value at every watchdog arm, so changes apply on the next turn without recreating sessions.

ExecutionManager keeps using the per-agent agent.timeout for task / heartbeat hard-cap — that's a different concern (wall-clock cap on an entire run) and intentionally untouched.

Dashboard 14-day chart now accurate on busy fleets

The Run Activity bar chart was being fed by runtime.getRecentRuns(), which is capped at 50 runs. On a fleet with frequent heartbeats / scheduled tasks, those 50 runs only spanned a few days — older bars in the 14-day window appeared empty even when they had real activity.

Fix: new FleetRepository.listRunsSince(date) walks _fleet/runs/YYYY-MM-DD/ folders lexicographically >= sinceDate. Cost is bounded by window size, not total historical runs. New parallel chartRuns cache populated alongside recentRuns via a new refreshRunCaches() helper used by all 7 refresh sites. Bar chart uses the new list; success-rate donut and activity timeline keep the recent-50 list (those are "lately" indicators, not 14-day windows).

103/103 tests green. No schema or default-content changes.

See v0.10.0 for the major feature release this patches.