Claude Code Digest — Jun 28–Jul 01

Claude Code’s biggest shift this week: teams are replacing “let the model figure it out” with hard guardrails, and one pair of Bash hooks cut an Anthropic bill from $312 to $156.

80% token reduction with Argus

Trending Now

🔥 Caliper pass@k turns skills into measurable reliability
Instead of judging a skill by vibes, Caliper runs it k times and compares pass@k against a no-skill baseline. If you ship Claude Code skills, start treating them like tests: keep the ones that beat baseline, delete the rest. 🔥 MCP is becoming the default wiring layer for Claude Code
From pgvector search to CRM access to model discovery, more teams are exposing capabilities as MCP servers instead of hardcoding integrations. The practical move: wrap internal services once, then let Claude Code discover and reuse them across repos. 📈 Deterministic workflows are beating freeform agent reasoning
Argus reports ~80% token reduction by replacing open-ended reasoning with structured execution paths. If a task has repeatable steps, encode them as a workflow or hook instead of asking Claude to rediscover the process every run.

Best Practices

Add two Bash hooks to veto retry spirals and block dangerous deploys
Without this: Claude can burn tokens looping on the same failing tool call or push into risky deploy paths. With this: the hook intercepts bad trajectories before they compound, and one team cut spend from $312 to $156. Write a <50-line root CLAUDE.md with Stack, Router Design, DI, and Prohibitions
Without this: FastAPI repos drift into inconsistent patterns across generated files. With this: Claude gets positive patterns plus explicit no-go zones, which reduces architectural drift before it starts. Pin MCP servers and enforce allowed_tools instead of fail-open discovery
Without this: a resolver can silently pull untrusted or changed servers into the session. With this: version-pinned servers plus least-privilege tool allowlists close the exact gaps LiteLLM’s audit found.

Tools & MCP

Caliper — Runs Claude Code skills k times and scores pass@k against a baseline — tells you if a skill actually improves reliability. FastMCP pgvector server — Wraps pgvector search as a reusable MCP server — lets any Claude Code session query vector search without hardcoded RAG glue. TS7 native Go TypeScript server — Swaps a TypeScript LSP plugin for a native Go server — faster setup, lighter runtime, and a 2-minute install path.

Multi-Agent Patterns

Graph-engine orchestration with human-in-the-loop checkpoints
Google ADK Go 2.0 adds a graph-based workflow engine so agents follow explicit branches instead of improvising. The human-in-the-loop step is the important part: use it for approval gates, not just chat supervision. Default-private publish primitives for agent-generated web output
Thryvate’s model is to publish through MCP tools that make artifacts revocable, expiring, and per-viewer tracked. That solves the common agent problem of “it created a public URL too early and now we can’t retract it.” Discovery-driven tool selection for changing infrastructure
Instead of hardcoding model lists, Claude queries live discovery tools like `nvidia_list_foundation_models`. This keeps multi-agent systems from breaking when providers add, rename, or retire endpoints.

Community Requests

• A native Caliper benchmark mode for MCP tools and hooks, not just skills
• A standard Claude Code hook pack for retry detection, deploy blocking, and formatter auto-runs
• A security scanner for MCP servers that flags fail-open resolvers, unpinned versions, and overbroad tool scopes