Monday, May 4, 2026

Daily picks

articles scored

#1 GOLDGuideDev.to Claude

How I stopped Claude Code from hallucinating function names on a 4,000-file repo (with a local MCP server)

In a 4,000-file TypeScript monorepo, Claude Code kept inventing function names that looked plausible but didn't exist — `logResponseTime`, `parseConfigFile`, `validateInput` — then calling them confidently in generated code
Root cause: without a real symbol graph, Claude fills gaps with statistically plausible names from training data; the larger the repo, the worse it gets
Fix: run a local MCP server that gives Claude Code actual symbol lookup and ranked code search, replacing grep+read with a proper index
Result: function name hallucinations dropped to near-zero on the same repo — author includes benchmark numbers (NDCG-style) and explicit cases where it still doesn't help

#2 SILVERGuideDev.to Claude

A non-developer ran Claude Code on a 2015 MacBook Air for 6 months — 857 autonomous sessions, 840 self-merged PRs — and open-sourced the framework that made it survivable
Three problems autonomous Claude Code always hits: sessions start blank (no memory of yesterday), failures cascade silently overnight, and git becomes a graveyard of orphan branches
The fix is five small scripts around the existing `claude` CLI — no new agent framework: launchd scheduler, session-start/end hooks to inject/extract context, scope-guard to block protected files, and a healthcheck that opens GitHub Issues on failure
Open-sourced as `claude-autonomous-kit` on GitHub — each piece is under 200 lines

#3 BRONZEReleaseHN Claude Code

DeepClaude plugs DeepSeek V4 Pro in as the "thinking" layer before Claude Code executes — planning and reasoning on the cheap model, actual code actions left to Claude
Claims 17x cost reduction vs. Claude-only loops; 268 HN upvotes and 110 comments with skeptical benchmarks and real comparisons in the thread