68.9% of Claude Code Commits Never Ran, Plus Lazy MCP Tools and Qwen3.8 Recipes

Today's findings

1. #1 Stop Dumping Every MCP Tool Schema Every Turntechnique

   Progressive tool discovery - expose a small search tool plus deferred names, then load full schemas only when needed - avoids context bloat and wrong-tool picks.

   MCP tool loading

   From full-catalog dump to progressive discovery

   Dump it all

   • Full catalog injected
   • Reloaded every turn
   • Thousands of tokens upfront
   • Colliding tool names
   • Wrong-tool picks

   → Thousands of tokens kept out of context before the user types

   Discover on demand

   • One small search tool
   • Deferred names visible
   • Schemas loaded only when needed
   • Self-contained responses

   The model was fine — the schema dump wasn't.

   Why it matters: MCP servers can add thousands of tokens before the user types; colliding tool names and hidden steps cause agents to fail even when the model is fine.

   How to apply: Audit your harness: replace full catalog injection with a search/list tool, keep deferred tool names visible, and design tool responses to be self-contained and unambiguous.

   mcpagentstoolingcontext

   Read more: Dumping every MCP tool schema every turn is a harness bug · The worst agent mistakes I've seen were my tool's fault, not the model's

2. #2 Verify Claude Code Commits Actually Ran the Codetip

   A measured 68.9% of Claude Code commits touched files never run after the last edit; add a run/test gate before commit and check scheduled tasks actually fired.

   AGENT COMMIT AUDIT

   Most Claude Code commits were never run

   68.9%

   of measured Claude Code commits touched files that were never executed after the last edit

   Gate commits on actually running the changed code — verification must be external to the agent's claim.

   Why it matters: Agent-generated commits can look plausible while shipping untested changes; scheduled tasks can silently miss, so verification must be external to the agent's claim.

   How to apply: Add a pre-commit hook or CI step that runs the changed file/tests, log scheduled-task outcomes, and require explicit evidence before merging agent work.

   claude-codeagentstestingci

   Read more: I measured how often Claude Code commits a file it never ran after the last edit. On my laptop: 68.9% of commits ! · I let Claude Code run my two YouTube channels overnight. The failures taught me more than the output. · 9,000 commits in two months with Claude Code: 5 things we learned building an agent that's allowed to touch real WordPress and Joomla sites

3. #3 Qwen3.8 27B Quant and KV Cache Cheat Sheettechnique

   Across 189 benchmarks, q5_1/q5_1 KV quant was safest; pick the highest model quant at q4_0/q4_0 before spending VRAM on KV precision.

   Qwen3.8 27B · 16GB VRAM · GGUF

   Spend VRAM on the model quant before the KV cache

   4 checks · 189 benchmarks

   Max out model quant first

   KV cache at q4_0 / q4_0

   Upgrade KV to q5_1 / q5_1 if headroom

   q8_0 / q4_0 KV without benchmarking

   pass warn fail

   q5_1/q5_1 measured safest overall — q8_0/q4_0 the outlier to avoid unless benchmarked (KLD/perplexity sweep).

   Why it matters: Local 16GB VRAM users must trade model quant vs KV quant; this gives a concrete default that minimizes KLD/perplexity.

   How to apply: For Qwen3.8 27B on 16GB, start with highest model quant at q4_0/q4_0, then try q5_1/q5_1 if room; avoid q8_0/q4_0 unless benchmarked.

   quantizationlocal-llmggufkv-cache

   Read more: I tested Qwen3.8 27B KLD for 21 Models at 9 different KV quants that fit on 16GB VRAM (189 benchmarks). · I tested Qwen3.8 27B KLD for 21 Models at 9 different KV quants that fit on 16GB VRAM (189 benchmarks).

4. #4 Qwen3.8 Flash-Next Serving Recipes for Single DGX Spark and Stratatool

   A vLLM + MTP setup hits 76.8 tok/s single-stream and 265.9 tok/s at 8 streams on one DGX Spark, while a three-machine Strata field review compares it to 27B for daily tasks.

   Qwen3.8 Flash-Next · Local Serving

   Real-time single-stream speed on one DGX Spark

   76.8

   tok/s single stream · vLLM + MTP

   265.9 tok/s at 8 streams

   3-node

   Strata field review vs 27B for daily tasks

   GB10

   ultrafast repo ships the exact serving flags

   Clone the GB10 ultrafast repo for the vLLM + MTP flags; read the Strata review before choosing Flash-Next over 27B.

   Why it matters: Shows what local serving can look like on modest single-node hardware and when Flash-Next beats a larger 27B model.

   How to apply: Clone the GB10 ultrafast repo for vLLM/MTP flags, and read the Strata field review before choosing Flash-Next vs 27B for your workload.

   local-llmvllmperformanceqwen

   Read more: Fastest Qwen3.8 Flash-Next on a single DGX Spark (GB10): 76.8 tok/s single stream, 265.9 tok/s at 8 streams (vLLM + MTP) · I ran Qwen3.8-Flash-Next on Strata across three machines. Here's what it's actually like for daily tasks vs 27b

5. #5 Strata Dual-RTX 3090 Regression Traced to Auto GPU Placementtip

   A dual-3090 Strata regression came from automatic GPU placement; the fix is tested, PR'd, and benchmarked.

   Strata · multi-GPU tip

   Auto GPU placement vs manual pinning

   vs

   Auto placement

   Manual pinning

   Layer → GPU mapping

   Auto-guessed

   Pinned explicitly

   Dual-3090 throughput

   Silent regression

   Restored (PR'd)

   Setup effort

   Zero config

   Manual edits

   Auto placement wins the row Manual pinning wins the row

   If Strata slows across multiple GPUs, check auto placement first — then pin assignments.

   Why it matters: Multi-GPU local inference can silently regress when auto-placement misassigns layers; knowing the failure mode saves hours.

   How to apply: If Strata performance drops on multi-GPU, check auto GPU placement first, pin layer/GPU assignments manually, and follow the submitted PR benchmarks.

   local-llmmulti-gpustrataperformance

   Read more: I traced a dual-RTX 3090 Strata regression to auto GPU placement — fix tested, PR submitted, full benchmarks inside

6. #6 JetBrains Mellum 2.1 12B A2.5B MoE Runs Well on Small GPUstool

   A 12B MoE with 2.5B active parameters hits ~60 tok/s on M2 Pro 16GB and 140 tok/s on 9070 XT, with 130K context and Q8 on 16GB.

   Why it matters: A local coding model that fits 8-16GB cards and runs fast enough for quick task execution.

   How to apply: Try Mellum 2.1 as a fast local coder/executor; pair it with a larger planner model for detailed task decomposition.

   local-llmcodingmoeperformance

   Read more: JetBrains Mellum 2.1 12B A2.5B MoE is a very good model.

7. #7 Jais 3 Open Weights Span 0.9B to 375B, Arabic-Focused and Agenticrepo

   IFM released Jais 3 open-weight models from 0.9B to 375B, tuned for Arabic and agentic use.

   Why it matters: Adds a strong open-weight option for multilingual/agentic workloads, especially Arabic-language products.

   How to apply: Pull the Hugging Face collection, test the smaller sizes locally, and evaluate the larger variants for agentic Arabic tasks.

   open-weightsagentsmultilinguallocal-llm

   Read more: Jais 3 is out · IFM/Jais-3

8. #8 Agent Praxis Tests Whether Agents Actually Fix Systemsrepo

   Six adversarial environments check traps, metric gaming, and fake authority so agents can't pass by claiming success.

   Agent evaluation

   Did the agent fix the system — or just claim it did?

   Typical benchmarks

   vs

   Agent Praxis

   A fix counts only if the system state proves it.

   Run local LLM agents through Praxis before trusting them with tool use.

   Why it matters: Most agent benchmarks reward plausible claims; this suite verifies real system state and exposes reward hacking.

   How to apply: Run local LLM agents through Agent Praxis before trusting them with tool use, and use the failure cases to harden evaluators.

   agentsevaluationsafetylocal-llm

   Read more: I built Agent Praxis: 6 adversarial environments to test how LLM agents handle traps, metric gaming, and fake authority · I built Agent Praxis: 6 adversarial environments to test how LLM agents handle traps, metric gaming, and fake authority

9. #9 Enforce Agent Limits Outside the Prompttechnique

   Move destructive capabilities out of tools, use scoped credentials, and mediate all tool access through a gateway rather than relying on approvals or system prompts.

   Why it matters: Prompt-level rules and rubber-stamped approvals fail under injection or fatigue; enforcement must live in credentials and gateways.

   How to apply: Remove delete/payment/deploy from agent tools where possible, scope API keys per task, and require a gateway for any sub-agent tool call.

   agentssecuritypermissionsmcp

   Read more: I started rubber-stamping agent approvals, so I changed the tool permissions · Where do you enforce limits on what your agent can do? Prompt, app code, or outside the agent entirely? · Can runtime security actually protect AI agents if they can bypass the gateway?

10. #10 RAG Quality Is Won Before Embeddingstechnique

    The hard part is document structure, table/chart handling, and conflict resolution - not just chunking and embedding.

    Why it matters: Enterprise RAG fails when PDFs, tables, and contradictory chunks are flattened into text; retrieval can't fix bad ingestion.

    How to apply: Build a document AST, preserve structural hierarchy, add freshness/version metadata, and resolve contradictory chunks before generation.

    ragingestionretrievaldocuments

    Read more: The hardest part of building a RAG knowledge base happens before embeddings · Can a RAG system tell when two individually correct chunks contradict each other? · Do we really need wrappers like LangChain/LlamaIndex?

11. #11 Aidlc: Adaptive Workflow Steering Rules for AI Coding Agentsrepo

    AWS Labs open-sourced adaptive workflow steering rules for AI coding agents.

    Open-source release · AI coding

    Adaptive steering rules for AI coding agents

    Aidlc

    tool Open source

    From AWS Labs

    GitHub repoWorks with Claude Code & other local agents

    runClone the repo, adapt the rules to your stack, test with your coding agent

    A concrete rule set that keeps agents on track — without hardcoding every step

    Why it matters: Gives a concrete rule set to keep coding agents on track without hardcoding every step.

    How to apply: Clone the repo, adapt the steering rules to your stack, and test them with Claude Code or other local coding agents.

    agentscodingworkflowsopen-source

    Read more: Aidlc: Adaptive workflow steering rules for AI coding agents

12. #12 Apple Silicon Matrix-Unit Formats: M6 Does FP8 at ~2x FP16, M5 Does Nottechnique

    Per-chip GPU matrix-unit benchmarks show M6 runs fp8 at about 2x fp16 while M5 does not, with help wanted for more variants.

    Why it matters: Choosing quantization and inference engines on Apple Silicon depends on chip-specific matrix-unit support, not just unified memory.

    How to apply: Check the published per-chip fp8/fp4/int8 numbers before picking a Mac for local inference, and contribute benchmarks for your variant.

    apple-siliconquantizationlocal-llmperformance

    Read more: Apple Silicon GPU performance on matrix-unit formats per chip (fp8/fp4/int8): M6 does fp8 at ~2× fp16, M5 does not. Help wanted to run performance test on more Apple Silicon variants.