Audit: 5 dashboard tabs broken on real-store data (yield/messages/stats/optimize/agents)

[0bserver07]

Audit run on the real-store SutroYaro project (95 sessions, 374 Task tool calls, 7M tokens). All wave-1 / v0.9.0 surfaces tested individually with urllib.request.

Real failures

Tab	Symptom	Impact
yield	/api/yield times out at 15s	tab unloadable
agent-teams	Returns 10 teams when project has 0 sub-agent activity (uses ?project= filter but data model is wrong)	shows random sessions from other projects
messages	Returns 37 MB body — every message in 95 sessions, no pagination	dashboard crawls / runs out of memory
stats	Returns 3.2 MB	every chart re-fetches this; cumulative latency
optimize	13.2s response time	tab visibly hangs

Working

overview (40ms), cost-by-provider, compare, commands, search, qa, tags, bookmarks, playback-project.

Root causes (where known)

• yield: needs profiling; likely git log per cwd over a project with many sessions
• agent-teams: _list_team_sessions_scan synthesizes "teams" from any session in any project that has sidechain anywhere. SutroYaro has 0 sidechain messages but 374 Task tool calls. The scan logic is solving the wrong problem.
• messages: route returns full message list without limit cap — 26,158 rows × ~1.4KB each = 37MB
• stats: payload includes the full daily_stats / hourly_pattern + per-tool counters. Probably fine in absolute terms but loaded on every tab switch
• optimize: scans messages.raw_json across 95 sessions × 16K msgs each

Plan

Each fix dispatched as a parallel agent. Tracked as a wave — single v0.9.1 patch when all merge.

Closes part of the broader "is the dashboard actually correct" question raised after the meta-agent + wave-1 ship.