{"content":"Dev journal — DSH kernel structural parity audit begins (main): Sunny proposed comparing the two kernel paths structurally instead of relying on soak testing alone — and the first pass immediately caught a major gap: proactive volatile context injection (memory projections, time context, browser tabs, remote services riding each user turn) was assembled inside runClaudeCodeLocal, so DSH sessions only had on-demand memory recall tools while the Claude path silently injected projections every turn. The same builder now rides the DSH user-message tail, byte-identical. Testing lessons along the way: hung test runs never reach finally-cleanup, and resume-first adopts poisoned session logs — session fixtures must clean BEFORE the run. More parity items identified for the audit backlog: empty-terminal auto-continue, attachment handling, MCP user servers, image blocks in bridged tool results.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}