Why We Migrated from Cursor to Claude Code for Complex Refactors — And Where Cursor Still Wins

TL;DR: Cursor Composer became unreliable for sessions >90 min — lost tunnels, DB connections, file context. Claude Code (Opus 4.5) maintains terminal state, processes, and 200k context across 4+ hour sessions. We use Claude Code for refactors/migrations; Cursor only for <30 min TypeScript edits. Full comparison →

The Context

Three-dev team, production workload: Stripe webhook migrations, auth system rewrites, PostgreSQL schema changes, Go API refactors. Standard session: 2–4 hours. Cursor was daily driver Jan 2025–May 2026. Failure modes compounded: Composer context reset on forced model updates (3×/week), RAM spikes to 12GB crashing Electron, agent forgetting running stripe listen tunnel and local Supabase PID. June 2026: Team tested Claude Code (Opus 4.5) side-by-side on identical 3-hour tasks.

What We Tested

Task	Cursor Composer	Claude Code (Opus 4.5)	Winner
3-hr Stripe webhook migration (14 files, live tunnel + Supabase)	❌ Lost tunnel 3×, 45 min recovery	✅ Kept tunnel, DB, context entire session	Claude Code
2-hr Auth refactor (JWT→OPA, 22 files, integration tests)	❌ Context reset on model update popup	✅ Completed in one sitting	Claude Code
30-min TypeScript type-fix sprint (8 files, heavy LSP reliance)	✅ LSP catches errors mid-edit	⚠️ No inline LSP diagnostics; runs tsc after	Cursor
45-min Greenfield API endpoint (new route, tests, docs)	✅ Fast Composer multi-file	✅ Equivalent; better test generation	Tie
Debugging flaky CI (reading logs, patching, re-run)	❌ Forgets log context after 60 min	✅ Maintains full log + code context	Claude Code

The Pivot Point

June 5, 2026: Dev running 3-hour PostgreSQL migration (ALTER TABLE + data backfill + index rebuild). Cursor Composer: forgot the running psql session after 75 min, lost the transaction state, rolled back partially applied migration — 2 hours recovery. Same task in Claude Code: agent kept the psql PID, transaction state, and full schema context. Completed in 2.5 hours, zero recovery. Team rule locked: Claude Code for any task touching running infrastructure (tunnels, DBs, long processes); Cursor only for pure code edits <30 min.

What We Use Now

Claude Code (Opus 4.5 via Anthropic API) for refactors/migrations/debugging:

• claude-code CLI with --dangerously-skip-permissions for trusted repos (internal only)
• Project instructions in CLAUDE.md: “Never assume server restarted. Preserve tunnel PIDs. Ask before killing processes.”
• Typical session: 3–4 hours, 150k–200k tokens, ~$8–12/session on Opus 4.5

Cursor (Pro, $20/mo) for TypeScript-heavy quick edits:

• Keep for: fixing type errors, adding props, quick component tweaks
• Hard limit: 30 minutes max session. Timer enforced by convention.

Zed + Continue.dev (local models) as backup for long sessions when Opus 4.5 rate-limited.

When You’d Choose Differently

• Cursor Composer if: 90% of work is TypeScript/React component edits <30 min, LSP diagnostics critical, team refuses terminal-first workflow.
• Windsurf ($15/mo) if: want Cascade’s persistent context in familiar VS Code UI, don’t need Opus 4.5 reasoning depth.
• Aider (terminal, BYOK) if: live in tmux, want git-native workflow, accept less polished UX.

Tool Crucible Rating

Dimension	Cursor Composer	Claude Code (Opus 4.5)
Overall	3	4
Long-session Reliability	2	5
TypeScript LSP Integration	5	2
Terminal/Process Awareness	1	5
Context Window (effective)	~50k (loses context)	200k (maintains)
Cost per Heavy Session	$20/mo flat	~$10 (Opus 4.5 API)
Setup Friction	Low (IDE)	Medium (CLI + API key)

---

This is part of our AI Coding Tool Evaluation series. See full comparison: Claude Code vs Cursor 2026: Opus 4.5, Composer Reliability, and the Terminal-First Alternative