reflect

reflect is session-local metacognition for Claude Code: the deliberate opposite of cross-session memory. Long /loop sessions drift; I watched it happen in my own multi-hour runs, hour two, the model repeating the same misjudgment. When suggestions get reverted 3+ times within 10 tool calls (weighted threshold, not single fail-retry), Opus 4.7 reads back recent tool history, active rules, and the rolled-back diff, infers why the user kept pushing back, not what failed, and injects structured guidance into the next turn: pattern, signal, adjustment, confidence. I'd previously shipped a static-rules version (stetkeep on npm). reflect is the dynamic recalibration layer static rules can't reach. Single-shot. Prompt-cached (3-layer: 1h + 5m + ephemeral, measured 95.9% warm cache hit). No persistence, no telemetry, by design. MIT-licensed, published as @chanjoongx/reflect. Failure modes characterized: cold-start refusal, false-trigger flagging, abstention on ambiguous prompts. Built solo over 6 days: npm package, Claude Code plugin, localhost viewer, dogfooded on the build itself.