All Articles
Swapping Tools Mid-Conversation Without Losing Your Prompt Cache
Tools can now be added and removed between turns, but the tool block sits at the very front of the cache prefix. I measured 12 mutation patterns with fingerprints and built a stable-core plus volatile-tail registry with a guard that refuses unsafe changes.
The 400-Character Preview That Was Still Holding On to a Megabyte
Truncating tool output in your own MCP server does not free the original. Node's sliced strings keep the parent alive. Here is the measured 13-character boundary, which flattening tricks actually work, and a heap-snapshot audit script that counts the retained parents for you.
The Line That Disappears at 100K: Measuring What Tool-Output Spill Actually Keeps
When agent tool output passes 100,000 characters, the full text spills to a file and the model sees only a head-truncated preview. Here are measured survival rates from a real repository, and the output envelope I built to push decision-relevant lines to the front.
Rebuilding a Remote MCP Server That Never Needed Mcp-Session-Id
The MCP spec release candidate drops the session header. Here is how I audited my own remote server for session coupling and moved it to signed cursors, with measurements.
Running Four Sites From One Managed Agent Definition
A design for collapsing four near-identical Managed Agents into one base definition, using agent version pinning and session-local overrides — with a validated factory and the traps I hit along the way.
When Memory Store Listings Returned Half the Rows: Migrating to agent-memory-2026-07-22
agent-memory-2026-07-22 changed memories.list: fixed ordering, stricter depth, segment-based path_prefix matching. How an audit quietly halved, an access layer that survives it, finding every call site with ast, and measured depth=1 traversal cost.
When Prompt Caching Was On but the Bill Didn't Move — Field Notes on Instrumenting cache_read to Close the Hit-Rate Gap
You added cache_control but your Claude API bill barely changed. Before guessing at fixes, record cache_read and cache_creation on every request and split the hit-rate gap into three causes: non-deterministic prefixes, TTL expiry, and sub-threshold blocks.
A Two-Stage Pre-Publish Gate for User-Facing AI Text in Consumer Apps
Design a two-stage pre-publish gate for short AI-generated text you ship to end users: a deterministic rule layer plus a Claude classifier, with fail-closed handling, generation-time vetting, and a cost model. Full implementation code included.
Precision Lives in the Second Stage: Reranking a Personal Knowledge Search with Claude
Embedding search alone leaves 'semantically close but not the answer' passages at the top. This is a two-stage design that gathers candidates broadly, then lets Claude reorder them by answerability, with structured scoring, abstention, and a cost estimate.
Coalescing Concurrent Claude API Calls: Single-Flight Against Duplicate Inference and Cache Stampede
A design for collapsing identical prompts that fire at the same instant into a single upstream Claude call, using single-flight (request coalescing). In-process and distributed implementations, jittered retries, and negative caching, with measured results.
Full-Size or Downscaled? A Per-Image Resolution Rule for Opus 4.7's High-Resolution Vision
Opus 4.7 finally read the fine texture in my wallpapers, so I sent everything at full size. My weekly image tokens jumped 2.4x. Here is the preflight that decides resolution and model per image, with the measured savings.
When Extended Thinking Flattened My Accuracy but Doubled the Bill — Field Notes on Measuring the Marginal Utility of Thinking Tokens
You pinned budget_tokens high 'to be safe,' accuracy barely moved, and the bill kept climbing. These field notes show how to ledger real thinking-token usage by p50/p95/hit-rate and measure how much accuracy each extra 1,000 thinking tokens actually buys.