Plugin

reflect

Session-local metacognition harness for long-running Claude Code work. When suggestions get reverted 3+ times within 10 tool calls, reflect invokes Opus 4.7 to reason about why the rejections clustered. The model reads back the last 20 tool calls, the rolled-back diff, and active rules, then writes structured guidance — pattern, signal, adjustment, confidence — to .reflect/session-guidance.md. The next turn auto-loads it via path-scoped rule injection and adjusts. Key features: • 3-tier weighted signal taxonomy: git revert/restore (Tier 1, +100), user-file deletes (Tier 2, +70), utterance negation regex (Tier 3, +50). Cumulative threshold 240 + 5-turn cooldown — fires only when reverts genuinely cluster. • Single-shot Opus 4.7 with adaptive thinking (effort: high) and 3-layer prompt cache (L1 1h, L2 5m, L3 ephemeral). Measured: cold ~$0.05, warm ~$0.01, 95.9% L1 cache hit. • Two cross-platform hooks: PostToolUse (Tier 1+2) and UserPromptSubmit (Tier 3). Pure-bash arithmetic — no jq/bc dependency on Windows git-bash. • Session-local by design: no persistence, no cross-session sync, no telemetry. The guidance file evaporates at session end. • Failure modes characterized: cold-start refusal (< 5 turns), false-trigger flagging on intent shifts, abstention on regulatory/opaque domains. • Optional localhost Next.js companion Viewer (binds 127.0.0.1, never deploys, PII redactor on every read) for live state + reflection history + cross-session drift clustering. • MIT-licensed. SLSA v1 provenance via GitHub Actions OIDC trusted publisher (Sigstore-signed).

• Published source
• Repository

Analysed

11 October 2026

Last observed

12 October 2026

What the code does

• Runs hooks automatically Hooks that execute without being called explicitly.

Having a capability is not wrongdoing; many tools need these to do their job. It tells you what to check before you install.

This is public metadata, not a safety certification or complete assessment. Static analysis cannot establish what a tool does at runtime.