Why We Stopped Trusting Cursor Composer for Multi-Hour Work — The Context-Loss Pattern We Documented Across 12 Sessions

TL;DR: Cursor Composer loses terminal state (running servers, tunnels, DB connections) and file context after 60–90 minutes. Across 12 documented sessions (May–Jun 2026), average context-loss incidents: 2.3 per 3-hour session; average recovery: 18 min each. We restrict Composer to <30 min tasks only. Full incident log →

The Context

Two-dev team. Cursor Pro was our daily driver Jan–May 2026. Composer felt magical for multi-file edits — until sessions passed 90 minutes. We started logging every context-loss incident May 15 to quantify the pattern.

What We Tested

Session	Task	Duration	Context Losses	Recovery Time	What Was Lost
1	Stripe webhook migration	3.2 hrs	3	45 min	stripe listen tunnel, Supabase local, 14-file context
2	Next.js 13→14 app router	4.1 hrs	2	30 min	Dev server PID, middleware context
3	Supabase RLS refactor	2.8 hrs	3	40 min	supabase start instance, migration file context
4	Auth0→Clerk migration	3.5 hrs	2	25 min	Clerk dev server, env var context
5	Prisma schema overhaul	2.5 hrs	1	15 min	DB connection, schema diff context
6	Tailwind v3→v4 migration	2.0 hrs	2	20 min	Dev server, config file context
7	API versioning (v1→v2)	3.0 hrs	3	35 min	Test server, route context
8	React Query v4→v5	1.8 hrs	1	10 min	Dev server only
9	WebSocket refactor	2.2 hrs	2	22 min	WS server, connection context
10	i18n restructure	1.5 hrs	1	8 min	Config context only
11	Error boundary overhaul	2.0 hrs	2	18 min	Dev server, component tree context
12	Bundle analyzer audit	1.2 hrs	0	0 min	—
Average (sessions >90 min)	—	2.6 hrs	2.0	24 min	Server/tunnel + file context

The Pivot Point

Session 4 (Auth0→Clerk): 2 context losses, 25 min recovery. Same dev, next week, similar scope (Clerk→custom auth) in Codex Persistent Agent: 0 losses, 0 recovery. The pattern was undeniable: Composer’s context window management evicts terminal state to make room for file edits. We added “Composer max 30 min” to team conventions May 28.

What We Use Now

Cursor Composer — Restricted to <30 min, TypeScript-surgical tasks only:

• Team rule (enforced via PR template): “Composer used? Duration <30 min? Y/N”
• Allowed: Fix type errors, extract component, rename symbol, add props, single-file edits
• Forbidden: Multi-file refactors >5 files, tasks needing running server/tunnel/DB, sessions >30 min

Migration Path for Long Tasks:

• “This needs Composer” → Check estimated duration
• If >30 min: cx (Codex Persistent Agent) or cc (Claude Code terminal-autonomous)
• If TS-heavy AND long: cx for architecture + cursor for final type cleanup (<15 min)

When You’d Choose Differently

• Solo dev, small codebase: Composer context loss less painful (simpler server setup). May tolerate up to 2 hrs.
• Windsurf Cascade: Early reports suggest better context retention than Composer; eval planned Q3 2026.
• Cursor 1.0+ (rumored): If Cursor fixes context persistence, we’ll re-evaluate. Currently no roadmap commitment.

Tool Crucible Rating

Dimension	Rating (1–5)	Notes
Overall	2.5 (for long sessions) / 4.5 (for <30 min TS)	Mode-dependent; not a general-purpose tool
Ease of Use	4	Excellent UX for allowed use cases
Value	2 (long) / 5 (short)	$20/mo justified only if you stay in lane
Support	3	Cursor team responsive; context persistence not prioritized

---

This is part of our AI Coding Tool Evaluation series. See full incident log: Cursor Composer Context Loss: 12-Session Incident Log & Migration Path