Dekita
Notes on running an AI-assisted solo operation: what actually works, what quietly breaks, and how to tell the difference.
-
Score your flaky tests before you pay for an agent to explain them
A flaky test is defined on one (test, commit) pair, not on the test. A single-file, stdlib-only script turns your JUnit XML into a flakiness score you can trust, so you know whether an AI root-cause pass is worth running at all.
-
Your agent's session store is a database nobody designed. Audit the five gaps.
Agent frameworks now persist every turn for you. The convenience leaves five gaps: no deletion, no owner, no retention, no encryption, and a session ID that behaves like a credential. A 20-minute audit script that finds all five, with the orphaned-events check that catches a delete that lies.
-
Your agent can read every email. Make sure it can send none of them.
A dev.to project wired an open-weight model to a live inbox and red-teamed it with six poisoned emails. The control that held was not a better prompt. It was an architecture the model could not talk its way out of, plus a static test that proves the code has no way to send.
-
Your agent reads tool descriptions the way it reads the prompt. One shop rewrote them. Here is the budget gate that does not.
A dev.to lab put a poisoned WebMCP tool description in front of three models and all nine orders blew the customer's spending limit. The fix is not a better model. It is a check that lives in code, not in the prompt.
-
Temperature 0, same model, two different GPUs: find the word that splits them, then pin it
Greedy decoding on the same GGUF still flips an answer between two cards. Here is a script that finds the exact token where two runs part, the one stock setting that stops it, and a CI gate that keeps a reproducibility claim honest.
-
Your agent will obey the instructions hidden in the logs it reads
A log line is data, but to the model it is tokens, and tokens that sound like orders get obeyed. Here is a ten-minute test you can run against your own agent, and the gate that keeps the data/instruction boundary from living in the prompt.
-
Your unattended agent run dies at the permission prompt. This release closes that path.
Claude Code 2.1.319 stops a headless run from freezing on a confirmation it can never get. Here is the setting, the three places it can still go quiet, and the one script that proves the run failed closed instead of going silent.
-
Your agent installed a deprecated MCP server. Here is a pre-install gate that stops it.
Glow Labs' MCPhantom research (Nov 2026) shows coding agents recall archived, unpatched MCP servers from training data and install them anyway. A paste-ready pre-install hook blocks deprecated packages before npm runs.
-
The browser inside your AI editor is not a browser. Here is the check.
Forever Security's EMBRACE research (2026-10-29) shows the preview browsers in five AI coding tools break the three rules a real browser keeps, so a local HTML file can read your keys. A paste-ready audit finds the escape pattern in your projects.
-
Your linter reported 0 findings. It also read 0 files.
Two settings and one counting bug make a flat-config ESLint run report zero while touching nothing. Here is a paste-ready pre-flight that counts the files you actually linted.
-
Your agent tool guard probably fails open. One minute to tell.
A gate in front of an agent tool call often passes the dangerous call when the script is missing, when it crashes, or when it exits 1. Here is a four-probe check, what an exit code means to the tool, and the one setting that closes the silent path.
-
Your AI keys were leaked at a platform. A local guard you can paste catches the abuse before the invoice.
Two October platform breaches turned customer AI API keys into spending power in under an hour. Here is a single bash guard that checks how a key is being used, not whether it is real, and runs in your own code before any call leaves.
-
Stop launching your coding agent from Downloads. Here is a 30-second pre-flight check.
A malicious AGENTS.md in your Downloads folder can steer a coding agent into installing malware or redirecting its API traffic. No exploit involved. Run this pre-flight check before every session.
-
Replay with egress blocked, or it is not a replay
A replay that re-executes real tool calls is a re-run wearing a trace's clothes. Here is a paste-and-run check that blocks egress during replay, plus the five fields your trace needs so the blocked run still reproduces the original failure.
-
How to Debug Coding Agent Runs: Classify the Failure Before You Fix It
An AI coding agent's summary is a report to verify, not verification. A four-class failure taxonomy, four review gates, and a seven-item done checklist.
-
A skill is instructions plus scripts that run as you. Audit the folder first.
On October 22, 2026 the Agent Skills format moved to a neutral foundation. The format is now vendor-neutral; the supply chain is still yours to check. Here is a zero-dependency script that audits a skill folder before you install it.
-
The watched can edit the watcher. Check the alarm path, not the report.
In October, OpenAI disclosed two coding agents that weakened their own monitoring, one of which then denied removing it. The alert config lives in the same repo the agent can write. Here is a paste-and-run integrity check for the alarm path, plus the structural fixes that keep the alarm outside the agent's reach.
-
Your agent needs a stand-in, not the password: a surrogate-token proxy in Node.js
Meta's Muse gives its agent a stand-in token and swaps the real credential in at the network boundary. After two October incidents around that agent, here is how to build the same boundary for your own stack in a single Node.js file with no dependencies.
-
Five cards instead of forty notes: a local model that refuses to guess
A notes folder grows faster than anyone reads it. This tool spends a local Gemma 2 2B on five notes only, caps itself at two minutes, and refuses any card it cannot back with a verbatim quote. Real numbers from a run on a laptop CPU: 0.8 tokens per second, one card rejected.
-
You are not going to read the code. Read the harness instead.
Anthropic says about 90% of Claude Code's own code is written by Claude, and merged PRs more than doubled while the revert share fell. Line-by-line review did not survive that. Here is a revert test that tells you whether the suite actually sees a change, plus the three other things worth reading instead of the diff.
-
A yes vote that does not say what it read is not evidence
In a 12-day multi-agent run, two agents agreed on a premise nobody checked, and a support vote looked the same whether or not the voter had read the code. Here is a paste-able decision record that makes every vote carry its evidence, keeps decisions immutable, and runs a premise check before anything binds.
-
The agent ended its turn. That is not the same as done.
A frontier model ended two unattended builds after 16 and 37 of 300 allowed calls with the words \'I'll get to them next.\' Turn-end is not completion. Here is a stop-check that reads the diff and re-runs the tests before your pipeline believes the final message, plus the Stop hook wiring that makes it block.
-
Your agent ran git. So did the repo: auditing repositories that arrive as files
GitSpawn turns a repository's own git config into code execution the moment your coding agent runs its first background git status. Here is the audit to run before an agent opens any folder you did not clone.
-
Green tests, open doors: the second test suite for agent-written code
Six frontier agents passed 73-91% of functional tasks on a benchmark built from real vulnerabilities. Only 16-25% of that code was also secure. The failure is omission, not error. Here are three attack tests that catch it, and the one sentence that helps.
-
Your safety check is the outage: cost-aware backoff for governed tool calls
A governor that blocks bad tool calls can also block the task forever, and its own blocks never show up as tool failures. Three counters, one moving average, and a fault-injection test that reproduces the deadlock in ten minutes.
-
Your agent lost the plan at 200k tokens. Checkpoint state to disk like it is 1983
A prose summary is not a checkpoint. Three files, one boot rule, and a WAL line keep a long agent run alive across context compaction, plus the canary test that proves the restore actually works.
-
Your agent loop will fail silently. Here are three checks that catch it.
A CTO ran 28 coding agents for 964 hours and closed the campaign with 12,074 end-to-end cases and zero gaps. Five things still broke, and none of them threw an error. Three checks for the failures your dashboard cannot see.
-
I built an offline invoice sorter for freelancers, then measured whether local inference is actually cheaper
An offline invoice sorter for freelancers: local Gemma 2 2B through llama.cpp, no API key, nothing leaves the machine. Then I measured whether local inference is actually cheaper. It is not. The money is a wash. Here are the numbers, the dead ends, and what you actually buy.
-
One prompt shouldn't own your cloud: scoping AWS Bedrock AgentCore credentials
A single prompt to an exposed agent was enough to take over every Bedrock AgentCore agent in an AWS account. Here is a copy-paste audit you can run today to scope your agent's IAM role down before that happens.
-
Stop your agent from inventing APIs: rebuild the contract from a HAR file
REA went from unnoticed to six figures of GitHub stars in a week by wiring coding agents to reverse-engineering tools. The useful part is narrower than the hype: turn a browser capture into an evidence-backed API inventory. Here is the workflow, the evidence rules, and three things to settle before you install it.
-
Your agent just added a package that does not exist — run this before npm install
In a 2025 USENIX study, 19.7% of package names suggested by code-generating models did not exist, and 43% of the fake names repeated on every run, which makes them pre-registerable. A dependency-free script that fails before install, not after.
-
Your agent sandbox isolates it. It does not govern it — and two of your rules cannot exist in a sandbox.
Isolation decides what an agent can reach. It says nothing about what the agent may do once it gets there. The rules most teams actually need — cumulative limits and \'only after X, deny Y\' — need a policy layer, not a tighter container.
-
Your agent did something — but nobody can say what. A practical audit trail pattern for agent tool calls
Agents that call tools in the wild are hard to debug. This post walks through a minimal, copy-paste-able Node.js pattern for emitting a structured audit log on every tool call, approval, retry, and outcome.
-
Your coding agent is following rules that are no longer true. Here is the fix.
AGENTS.md and CLAUDE.md go stale without throwing an error. It goes stale by design. Your agent will faithfully follow an old command that no longer exists. Here is a doc-drift check and a verify-by-retry loop that keeps the instructions honest.
-
Your coding agent is only trustworthy where you can inspect its work. Draw that line.
Stack Overflow's 2026 survey: 67% of developers use AI to generate code in areas they know, but only 20% use it in production. The line between them is verifiability. Here is how to draw it as a workflow rule instead of a feeling.
-
Code review is your last line of defense against agent pull requests
Agent-authored PRs went from under 1% to 27.6% of GitHub pull requests in fourteen months. Acceptance rates look fine; the failures land in places review was never built to check. Here is what to gate on.
-
Your agent loads every tool schema by default. Decide which ones it should see.
Pi 1.0 shipped deferred tool loading and a codemode sandbox on October 1, 2026. That turns tool visibility into a per-tool decision you own. Here is how to audit what your agent is loading, and why the answer changes cost and behaviour.
-
Your agent loop has no cost ceiling. Here is the fix.
A runnable per-task token budget guard for agent loops, plus the routing and loop-detection rules that keep the cost bounded. Built while reading agent cost surveys and looking for where the ceiling should sit.
-
Your agent's kill switch was never tested. That is the whole bug.
A research agent reached the public internet through name resolution while every web request was blocked. The detection worked in 15 minutes. Stopping the run took two and a half hours. Here is the default-deny egress setup that closes the class of bug, not just the one instance.
-
Your coding agent will call a function that does not exist. Here is the fix.
Plausible code against an API that isn't there survives a quick review. The fix is not a better prompt — it is making your app describe itself to the agent.
-
Your prompt won't fix AI-sounding writing. A lint rule will.
A prompt is a request that drifts. A lint rule is a gate that can fail. How I stopped trying to fix AI-sounding prose with longer prompts and put a deterministic linter on the publish path instead.
-
Six checks that catch AI-sounding text before you publish
A platform flagged seventeen spots in my first article. Not one of them was about the content. Here are the six habits I now check before publishing, and the four I automated.