All Articles
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 Committed Symlink That Points Outside the Worktree — Auditing Repos Before You Let AI Spin Up Parallel Trees
Claude Code 2.1.212 fixed a bug where a committed symlink under .claude/worktrees could be followed during worktree creation and write outside the repo. The patch closes the following side. Here is an audit script for the committed side, plus a quarantine workflow.
I Could No Longer Remember What I'd Changed in Auto Mode — Where claude auto-mode reset Fits In
Running nightly automation for weeks, I kept nudging my auto-mode settings one at a time. One morning a small oddity made me realize I could not recall what I had changed. Here is how I rebuilt my configuration around claude auto-mode reset as a known-good baseline, from a solo developer's field notes.
I Stopped the Headless Claude Code Job, but the Build Kept Running — Designing Teardown Around exit 143
In headless mode, a SIGTERM received while Claude Code was mid-Bash used to orphan the command's process tree. The 2.1.212 fix changes that to a clean exit 143. Here is how to redesign your supervisor and cleanup around that contract.
I Believed Plan Mode Only Read — Replacing That Belief With Machinery
Claude Code 2.1.212 fixed a bug where plan mode ran file-modifying Bash commands without the permission prompt or the SDK canUseTool callback. Here is what could happen while that assumption was broken, how to verify your own setup, and how to stop leaning on a mode name for safety.
Branch with /fork, Delegate with /subtask — Two Commands That Traded Jobs
In Claude Code 2.1.212, /fork now clones your conversation into a background session, and in-session subagents moved to /subtask. Here is how I decide between them, and the budgets worth setting before you start branching freely.
The MCP Server I Declared Vanished from the List — Designing Config for a Namespace You Share with the Vendor
When a vendor reserves an MCP server name, your .mcp.json declaration goes quiet instead of failing. Here is the preflight check and prefix convention I use to turn that silent gap into a loud stop before an unattended run.
The String I Approved Wasn't the String I Read — Testing a Relayed Permission Prompt with Deceptive Characters
I pushed bidi overrides and zero-width characters through my own approval relay. NFKC normalization caught 0%. Here is why, and the implementation that catches 100% with zero false positives.
My Nightly Subagents Ran, and I Couldn't Retrace Them — Trimming the Logs That --forward-subagent-text Unleashed
Subagent text and thinking can now flow into stream-json output, and unattended log volume climbs sharply with it. Here are measured numbers, a three-tier retention filter that keeps only what postmortems need, and a correlation ID design that survives missing fields.
It Worked on My Machine, but Nobody Could Trigger It — Four Assumptions I Stripped Out of a Cowork Plugin
I bundled four working automation skills into a plugin and shared it. Only one of them ever fired. Here is how I measured skill trigger rate, and the four assumptions — vocabulary, paths, connections, and naming — I had to strip out before it was portable.
The Morning My Table Ended in "… 2,847 more rows" — Separating Render Caps from Token Cost in Tool Output
Claude Code 2.1.209 caps markdown tables at 200 rows plus a remainder count. Only the rendering is capped — the model still receives every row. Here is how to measure the gap and redesign tool output around aggregates.