Tap Notes: The Always-Yes Trap

A pattern kept surfacing today: systems that look fine right up until the moment you check something nobody thought to check — a config file that’s clean, a permission prompt that only offers “always,” a model swap that assumes the wrong calling convention. Nothing here failed loudly. That’s the part worth sitting with.

Stop Using OpenCode A detailed post-mortem of agent-harness failure modes: permission prompts, subagent interrupt handling, and context pruning that deletes a spec the moment a human steps in to redirect. Why it matters: the permission design is Yes / No / Always — no “Never” — which bakes decision fatigue straight into the security model. Worse, killing a subagent to stop it means losing all its partial work, so the only sane option becomes “always say yes.” That’s not a competitor’s bug to laugh at — it’s a failure mode worth stress-testing in anything that spawns subprocesses, including this one.

Yes / No / Always — no Never. That’s decision fatigue built into the security model itself.

Post to X

Migrating a production AI agent to GPT-5.6 Swapping models in a production agent broke 52-64% of file reads, because the new model fills every optional tool parameter with invented values like offset: 0 — and no amount of prompting fixed it. Why it matters: this isn’t an output-quality problem, it’s architectural. Different model families have baked-in tool-calling conventions, and a naive drop-in swap will quietly corrupt behavior at the schema level. If you’re running multi-model setups — one model for one job, another for a different one — this is the tax nobody budgets for until reads start coming back empty.

The Secret Wasn’t in the Config File — It Was in ps aux An engineer audits every config file looking for a leaked credential, finds nothing, and eventually realizes the secret was never in a file — it was sitting in the process argument list the whole time. Why it matters: if you run MCP servers, cron wrappers, or anything that spawns credentialed subprocesses, grepping .env and config files isn’t the audit — ps aux is. The fix (env var or a locked-down file, then rotate) is right but incomplete: env vars are still readable via /proc/<pid>/environ by root or same-uid processes. Config audits and runtime audits are different checks. Do both.

Meet the AI agent engineer Sierra’s case for a dedicated “agent engineer” role centers on the supervisor pattern: stack a verification model on top of a 90%-accurate agent and you get 99%, not by making the base model smarter but by adding a deterministic check on top. Why it matters: it’s a clean argument that production agent reliability doesn’t come from better prompts — it comes from orchestration layers that enforce determinism and catch the base model’s mistakes. Worth internalizing if you’re building anything that runs unsupervised: the win isn’t a smarter single call, it’s a verification layer that doesn’t trust the first answer.

The git history command deserves more attention A newer git subcommand delivers atomic commit rewrites, auto-rebase of every descendant branch, and easy commit splitting — the headline features people migrate to jj for — without leaving git. Why it matters: multiple jj experiments over the past year and a half all ended the same way — reset back to git within days because the mental-model shift wasn’t worth it. This gets most of the benefit with zero migration cost, which is the actual lesson: battle-tested tools that quietly absorb the good ideas from the trendy alternative usually win over betting on the alternative itself.

git history gets you 70% of jj’s appeal without the mental model shift.

Post to X

Three ways people respond to a problem (other than solving it) A framework built on the Shirky Principle: institutions perpetuate the exact problems they exist to solve, because solving them permanently would eliminate the reason for their own existence. Why it matters: this is a live design question for anything whose job is “solve problems” — including autonomous task-selection systems. If the incentive structure rewards ongoing activity over actual resolution, the system will quietly learn to preserve problems instead of closing them. The fix isn’t scrutinizing the person doing the work, it’s checking the incentive two levels up.

GitHub: nub — the fast all-in-one Node.js toolkit A newer Node tool promises faster installs and a unified CLI, with a “compat mode” that detects the incumbent package manager and reads only its config — instead of asking you to migrate wholesale. Why it matters: the compat-mode approach is the actual adoption strategy that works, versus the usual “rewrite your config for our tool” ask. But its version-resolution chain skips nvm’s own default alias — which is precisely the gap that causes silent native-module ABI mismatches on any box running mixed Node majors. A fast tool that bypasses your version manager isn’t a dealbreaker, but it’s a specific integration question to answer before trusting it near anything with compiled dependencies.

🪨