perf: tapes deck eliminates N+1 queries, extend cache coverage, add created_at index

[bdougie]

Summary

• Fix N+1 in SessionDetail / SessionAnalytics — both now check the session candidate cache by ID before falling back to loadAncestry(), eliminating per-session parent-chain DB round-trips
• O(1) session candidate index — added a map[string]*sessionCandidate index to the cache struct, replacing O(n) linear scans in SessionDetail and SessionAnalytics with direct map lookups on the hot path
• Overview() and AnalyticsOverview() use cache — switched from allowCache=false to true; rapid TUI/dashboard refreshes reuse the in-memory candidate set instead of triggering a full DB scan
• Deduplicate extractToolCalls in analytics loop — called once per node instead of twice, result reused for both counting and error attribution
• Bump sessionCacheTTL 10s → 30s — aligns with TUI refresh cadence and reduces cold-load frequency
• Add created_at index to Node schema — enables future time-range predicate pushdown; applied automatically via client.Schema.Create on startup
• Exponential backoff for facet polling — replaces fixed setInterval(fn, 3000) with a recursive setTimeout that grows 1.5× per tick up to 30s, reducing background network pressure during slow backfills

Benchmark results (Apple M4 Max, 6×5s runs, benchstat)

Benchmark	main	branch	Δ
Overview warm	240µs	13µs	−94%
SessionDetail warm	255µs	15µs	−93%
SessionAnalytics warm	234µs	10µs	−95%
AnalyticsOverview warm	305µs	89µs	−70%
Overview warm (50 sessions)	1074µs	65µs	−93%
SessionDetail warm (50 sessions)	161µs	7µs	−95%

Overall geomean: −70.74%

Cold-path calls (cache miss) incur a small constant overhead (+9–11%) from writing to the cache on each full load — acceptable given warm-path savings.

Test plan

• go test ./pkg/deck/... — 121 specs pass, no regressions
• go build ./... — clean build

---

Continue Tasks: ✅ 1 no changes — View all

[oppegard]

This looks awesome :D

My primary concern is lack of test coverage. I prompted codex with:

Based on the changes this branch made to query.go , can you suggest some high value tests?

It came up with this perf/deck-query-cache-and-index...oppegard:tapes:perf/deck-query-cache-and-index

(It's wild that it can just generate this)

Before this new test suite, query.go had 27% test coverage. These tests push it to 78% coverage, and they run in ~250 ms (I don't love that). There are still some conditionals not being exercised, and functions like groupSessionDetail() and appendGroupedText() have nearly zero coverage.

That's the extent of what I've reviewed.

[bdougie]

Tests pushed — 166 specs now (up from 126), covering the new cache index, buildSessionMessages, buildGroupedMessages, buildSessionAnalytics, matchesFilters, and the text grouping helpers. All pass in ~30ms.

[continue]

No docs update needed. This PR is a pure performance optimization that doesn't change any user-facing behavior:

• Internal caching improvements: N+1 query elimination and session cache index are implementation details
• Cache TTL change (10s → 30s): This is an internal constant, not user-configurable
• Database index: Applied automatically on startup via schema migration
• Frontend polling optimization: Exponential backoff is transparent to users

All API responses and CLI behavior remain unchanged. Users will experience faster dashboard/TUI performance without any action required.

PR #132 was merged: perf: tapes deck eliminates N+1 queries, extend cache coverage, add created_at index
PR URL: #132
Merged by: bdougie