cchu@nycu:~/blog$ ls -lt ~/writing
Writing
Notes, build logs, research fragments, and the occasional sharp-edged opinion.
2026.09.23A Stable Prefix at 53 Percent Cache HitReasonix's local HTTP tests separate message-prefix stability from cache ratios, session totals, and rewrite diagnostics.[agents][prompt-caching][go][testing]5 min2026.09.23A Zero-Length Tail That Keeps the Whole PromptA small compression decoder exposes how context slicing and compressor state become part of generation behavior.[language-models][compression][python][testing]5 min2026.09.21A Private MCP Port, Three Network BoundariesLocal tests trace a tool call through an outbound MCP tunnel and check where control-plane keys, connector tokens, and local headers can travel.[agents][mcp][networking][security][go]4 min2026.09.21Three Choice Calls to Write Two LettersAn offline Jevchat replay turns choice scores into a two-letter reply, then counts the questions hidden behind a larger alphabet.[language-models][decoding][python][testing]4 min2026.09.19Empty Detections and Offline DetectorsA local Fugleramme outage probe checks the difference between a quiet bird detector and a broken connection, including a Windows startup failure.[edge-ml][fault-tolerance][python][testing]6 min2026.09.19Tiled Inference Without Full-Image MasksA Supervision probe follows segmentation masks through tiling, RLE storage, suppression, and merging. The memory savings depend on the next operation.[computer-vision][segmentation][memory][python][testing]6 min2026.09.16Deterministic Gates Before AI Code ReviewA local OpenCodeReview inspection shows where ordinary programs should constrain an LLM review loop, and where the evidence still stops.[ai-agents][code-review][llm][go][testing]6 min2026.09.13Claims Need Replayable EvidenceA local Reverify run shows why an agent's durable memory needs receipts, verdicts, and known limits instead of a prose summary.[ai-agents][verification][context-engineering][python]5 min2026.09.12Litelm Keeps the Tool Call, Drops the FallbackA loopback HTTP test of Litelm's provider adapter: tool-call IDs survive a round trip, while a fallback list disappears before a simulated 429.[llm][python][tool-calling][api-design]5 min2026.09.12A One-Line Read, 4,745 Tokens SavedA local RTK reproduction and OmniRoute fidelity checks expose the gap between shorter tool output, preserved meaning, and cheaper agent tasks.[ai-agents][context-engineering][testing][token-accounting]7 min2026.09.06350 Lines Is a Routing DecisionSix offline probes of Spotify's Shunt hooks separate delegation policy from model quality, token accounting and enforceable read limits.[agents][model-routing][claude-code][shunt]4 min2026.09.06Forecast Step 129A TimesFM property test crosses the first output patch and checks whether negating the input also reverses the forecast's quantile channels.[time-series][timesfm][pytorch][testing]4 min2026.08.30Dataset Version 1 Still Contains the Deleted RowA four-row Lance experiment traces append, delete and pinned reads to show what a reproducible training-data reference must include.[machine-learning][datasets][lance][reproducibility]4 min2026.08.30The Arrow Had No ProtocolRemoving one label from an Archify deployment diagram produces a precise failure. Replacing it with an invented protocol reveals the validator's actual boundary.[architecture][agents][validation][diagrams]4 min2026.08.28Microduck's Safety Layer Stops at the ServoTwelve passing Rust tests and a MuJoCo model expose the difference between rejecting a bad tensor and respecting a robot's joint geometry.[robotics][reinforcement-learning][rust][mujoco]4 min2026.08.28Two Agents, Two Indexes, One Git ConfigA scratch-repository experiment with Worktrunk shows which parts of parallel agent work are isolated, which remain shared, and why dirty removal refuses to proceed.[agents][git][developer-tools][worktrunk]4 min2026.08.19File Handles for Agent MemoryA field note on ai-memory, OpenViking, Munder Difflin, Turbovec, and the storage contract forming around agent context.[ai-agents][agent-memory][context-engineering][retrieval]11 min2026.08.18The 151-Tool Trap in Security AgentsA field note on Strix, HexStrike AI, Anthropic's Defending Code harness, and why security agents need scope, sandboxes, egress control, and artifacts before they get shell access.[ai-agents][ai-security][security-testing][agent-infrastructure]11 min2026.08.15Spec Files, Hook Receipts, and Agent PluginsA field note on Spec Kit, Cursor plugins, Graft, Mole, Opencomplai, and the contract layer forming around agent workflows.[ai-agents][agent-infrastructure][developer-tools][ai-governance]13 min2026.08.15VRAM Budgets, Browser Handles, and Model RoutersA field note on ShoeHorn, Needle, Switchyard, Chrome DevTools MCP, OpenCode Senses, RAGless, Lumabri, RTK, and the adapter layer forming around local AI work.[local-ai][ai-agents][ml-systems][agent-infrastructure]14 min2026.08.13Flight Recorders for Agent RunsA field note on Decant, Prism-Eval, aakit, async-bulkhead-llm, LongHorizon-Harness, AgentSight, ActPlane, Debroid, and why agent systems need run records that survive the transcript.[ai-agents][observability][agent-infrastructure][ai-security]16 min2026.08.13Native Artifacts After Transcript-First AgentsA field note on OJCP, Needle, Diagram Design, Orca, PPT Master, Embabel, LTX-2, Bough, Trunchbull, and why useful agent systems now need outputs that other tools can inspect.[ai-agents][ml-systems][developer-tools][agent-infrastructure]16 min2026.08.12Private Reasoning, Public ProofA field note on stolen reasoning traces, model routing, local model desktops, provenance graphs, agent budgets, document extraction benchmarks, and why AI systems need private state boundaries plus public proof.[ai-agents][ai-security][ml-systems][local-ai]16 min2026.08.11Control Planes for Local AgentsA field note on Muse Glimmer, Docker Sandboxes, CtxRay, agent skills, Harvey LAB, Qwen multimodal plugins, and why local agent work now needs hard boundaries around context, tools, data, and evaluation.[ai-agents][local-ai][agent-infrastructure][ml-systems]15 min2026.08.10Runtime Receipts After Hero ModeA field note on WeatherNext, Claude Code auto mode, Tura, code-graph-rag, Agent Governance Toolkit, and why useful agent systems now need evidence trails around each action.[ai-agents][agent-runtime][ml-systems][ai-infrastructure]13 min2026.07.30Work Surfaces Before Autonomous ActionA field note on voice agents, robot control, cheaper model routes, agent governance, MCP tools, executable skills, offline knowledge, and the interfaces agents need before they touch real work.[ai-agents][agent-ui][voice-ai][model-serving]14 min2026.07.29Contracts at the Agent Runtime BoundaryA field note on Codex Security, MCP stateless transport, Kimi K3 serving, FlashKDA, sandboxed execution, browser agents, formal verification, voice models, and the contracts agents need before they touch real systems.[ai-agents][model-serving][agent-security][mcp]10 min2026.07.28Chain of Custody for AI WorkA field note on data provenance, agent trust metadata, Kimi K3 serving, security harnesses, design specs, video tools, token ledgers, and risky credential gateways.[ai-agents][model-serving][agent-security][data-provenance]12 min2026.07.27Receipts for Agent HandoffsA field note on AI handoffs, token fraud, proof, context, cost traces, governance toolkits, retrieval layers, skill routers, voice agents, and worktree surfaces.[ai-agents][agent-governance][observability][context-engineering]19 min2026.07.25Tracing the Agent Supply ChainA field note on FLUX 3, Claude Cookbook, Hetzner inference, open weights, token compression, and agent projects that expose model-work control points.[ai-agents][inference][agent-governance][model-routing]14 min2026.07.24The Work Ledger Behind Agent SessionsA field note on Tao's ChatGPT math thread, open-weight policy, AI capex, tokenization, credential gateways, semantic IDE tools, spec workflows, and agent workspaces.[ai-agents][agent-infrastructure][agent-governance][local-ai]17 min2026.07.23Sandbox Incidents Expose the Agent Control PlaneA field note on the OpenAI-Hugging Face evaluation incident, model routing, MCP usability, AI backlash, and evidence-gated agent systems.[ai-agents][agent-security][model-routing][control-plane]14 min2026.07.22Agent I/O Became the Product SurfaceA field note on Gemini's agent-shaped model tiers, coding-agent judgment, local web and code context, model routing, structured outputs, and domain agents that leave replayable work.[ai-agents][context-engineering][agent-runtime][web-agents]13 min2026.07.16The Kernel Around Local AgentsA field note on small local models, Claude memory leakage, DSL-shaped reliability, and the agent infrastructure that outlasts demo hype.[ai-agents][runtime][agent-security][local-ai]10 min2026.07.14Managed Agents Split Brain from HandsA practical note on why Claude Managed Agents split the brain from the hands, how BYOI sandboxes change the trust boundary, and why E2B, CubeSandbox, Kubernetes Agent Sandbox, Cloudflare, Modal, Vercel, Daytona, and NAP all feel like different answers to the same data-plane question.[ai-agents][sandbox][infrastructure][security]15 min2026.07.12Trace-Based Evaluation for AI AgentsA practical survey of mainstream agent evaluation and observability tools, from trace viewers and event streams to datasets, scorers, and production feedback loops.[ai-agents][evaluation][observability][llmops]13 min2026.06.30Context Platforms for Enterprise GenAIA shorter version of my research note on why enterprise GenAI needs governed context, source authority, evaluation, and workflow boundaries before it can move beyond demos.[enterprise-ai][rag][data-platform][plm]6 min2026.06.30Hidden Markers Break Developer Tool TrustThe Claude Code request-marking controversy shows why AI developer tools need visible telemetry, explicit metadata policy, and inspectable client behavior.[developer-tools][ai][privacy]6 min2026.06.29Fast Multimodal Inference on CerebrasA technical read on why wafer-scale speed plus open multimodal models changes the product envelope for visual agents, document AI, and real-time copilots.[multimodal-ai][inference][cerebras][gemma]10 min2026.06.28Abuse Budgets for Open RepositoriesMass malware repositories and anonymous zero-day drops show why open development platforms need risk lanes, provenance, and rate limits without destroying low-friction collaboration.[security][open-source][github]6 min2026.06.26AI as an Instrument for Slow KnowledgeThe Herculaneum scroll breakthrough shows a valuable AI pattern: models as inspectable instruments that extend observation instead of replacing scholarly judgment.[ai][research][history]6 min2026.06.24Age Checks as Identity InfrastructureAge verification proposals should be judged as identity systems: by data minimization, unlinkability, abuse resistance, and the secondary uses they enable.[privacy][policy][identity]7 min2026.06.22Appliance Computing Returns to the DesktopSteam Machine and Deno Desktop point at the same design pressure from opposite sides: users want general-purpose power without constant system administration.[computing][desktop][runtime]6 min2026.06.20Version Control After AI Code VolumeLore is a good excuse to ask what version control should provide when AI agents increase branch count, diff size, and review load.[version-control][developer-tools][ai]7 min2026.06.18The Local Model ThresholdLocal models are becoming daily developer tools. The technical story is about latency, privacy, model routing, hardware budgets, and workflow design.[local-models][ai][developer-tools]6 min2026.06.16Job Offers as Supply-Chain Attack PathsThe LinkedIn backdoor story is a reminder that recruiting, take-home assignments, GitHub repos, and local dev machines form an attack path.[security][supply-chain][developer-tools]6 min2026.06.14Spend Limits Belong in Agent SandboxesThe DN42 cloud-bill incident shows how cost, network, time, and resource limits act as first-class permissions inside agent runtimes.[ai-agents][cloud-cost][security]7 min2026.06.12Human Effort as an Attention FilterAI has made polished text cheap. Technical teams now need stronger attention filters: reproducible evidence, reviewed diffs, labeled AI output, and clear ownership.[ai][communication][product]7 min2026.06.10HTML-First Distribution EngineeringA technical read on why the HTML-first success story is less about framework taste and more about reach, resilience, conversion, and operational simplicity.[frontend][web][product]7 min2026.06.08LLMs Change the Training LoopA technical career essay on how AI coding tools change the training path for software engineers, and what teams should preserve when automation eats the easy tickets.[software-engineering][ai-tools][career]7 min2026.06.06Helpful Defaults Spend User TrustA technical product read on AI features in email clients, why inbox automation feels invasive, and how teams can design defaults that preserve user control.[product-design][email][trust]8 min2026.06.04Product Jobs for Small ModelsGemma 4 12B and the discussion around model weights point to a practical design question: which bounded jobs should small multimodal models own inside real products?[ai][local-models][multimodal]8 min2026.06.02Threat Modeling Account RecoveryA technical read on why account recovery, support escalation, and exception handling deserve the same threat modeling as login, MFA, and session security.[security][product][identity]9 min