Pi does not follow the XDG Base Directory specification on Linux. Its user configuration lives in ~/.pi/agent, and the only supported way to relocate it is the PI_CODING_AGENT_DIR environment variable. Set it to the agent directory — ~/.config/pi/agent, never ~/.config/pi — or authentication and model settings silently read as empty. This guide covers what is actually inside that folder, why the widely cited PI_CONFIG_DIR variable does nothing, the three layouts that fit real workflows, and the verification commands that prove the move worked. ...

The browser-use macOS Harness is an MIT-licensed Python package that hands an LLM six raw primitives — see, key, type, click, ax, and script — inside one persistent process, so the model writes its own Mac automation instead of picking from a recipe library. It is v0.1.2, alpha, and it has not received a commit since launch day. That last sentence is the review. This is a genuinely clever, genuinely thin control surface published on 2026-08-17 and then left alone, and the most useful thing you can do before installing it is understand exactly which parts are load-bearing and which parts are a demo. ...

Best-of-N coding agents run the same task N times in isolated workspaces, then use an LLM-as-a-verifier to score every full trajectory and select the winning patch. In Oh My Pi, the omp-best-of plugin does this with /best-of --n 5 --apply "Fix the failing authentication test", defaulting to 3 candidates and selection-only mode. That is the short answer. The rest of this guide is the part that decides whether the technique saves you money or burns it: what selection can and cannot fix, why the maintainers’ own benchmark numbers are much smaller than the headline claims circulating about them, and how to choose N without paying for a fourth opinion nobody needed. ...

agent-codemode (janwilmake/agent-codemode) is a MIT-licensed Node 18+ CLI and TypeScript library that lets a script your coding agent writes call the MCP servers that agent has already authenticated. It reads the OAuth tokens and stdio configs Claude Code already created, speaks Streamable HTTP or stdio JSON-RPC directly, and returns an exact answer instead of routing every intermediate result through the model. What Is agent-codemode and Which Number Does It Lead With? The project is small, and it is deliberately pitched as small. Its own README says: “Be clear about what is and isn’t new here.” What it does is give a coding agent a way to skip the tool-call loop. Instead of the model calling 40 MCP tools in sequence, feeding each result back into context, the agent writes one script that calls those tools as ordinary functions and prints one result. ...

Codex Command Center is a free, open-source, Windows-first desktop app that organizes Codex work rather than executing it. It registers your repositories, tracks engineering tasks across four states, indexes local Codex session metadata without modifying it, surfaces Git diffs, and shows your Skills and MCP configuration — all in local JSON with no telemetry, no accounts, and no cloud sync. That is the entire thesis of the product, and it is a narrower claim than the name suggests. The repository’s README is explicit that Codex does the reasoning, editing, commands, and tests, while Codex Command Center “keeps the surrounding engineering workflow understandable and resumable.” It says outright that it is intentionally smaller than an IDE, a terminal emulator, a chat client, and an advanced Git GUI. In a category where every launch promises to become your new operating system for agents, a tool that publishes its own non-goals deserves a fair reading before a verdict. ...

BugbountyRules is a Claude Code skill that keeps 42 always-active behavioral rules in context — scope, coverage, evidence, persistence — instead of payloads. It exists because an agent already knows what SQL injection is; it fails by shipping false positives, inflating severity, and quitting early. The rules govern method, not knowledge. That inversion is the whole point of the project. Nearly every “AI security skill” on the market is a knowledge pack: a longer list of things to look for. BugbountyRules (github.com/sanjarbiy/bugbountyrules) takes the opposite position, and its README states the thesis bluntly — the agent’s problem is not that it does not know what an IDOR is, it is that it will not reload prior state after a context compaction, will not distinguish a confirmed finding from a plausible one, and will declare a surface clean the moment it runs out of ideas. Those are behaviours. Behaviours are what the 42 rules target. ...

Sharing memory across every user of a public AI is three different architectures wearing one name: a private per-user layer, an attributed shared layer, and a de-attributed wisdom layer. Raw pooling without those boundaries produces 57-71% cross-user contamination from benign interactions alone, and 80-99% memory-poisoning success under attack. That is the short answer, and it is deliberately unflattering to the idea. The longer answer is more useful: cross-user sharing genuinely works where local experience is scarce, the failure mode nobody budgets for is not forgetting but confidently remembering somebody else’s local convention, and the security model changes so completely that standard prompt-injection controls do not catch it. This guide walks through what the 2026 research actually measured, what it costs, and the design rules that hold up when one memory pool serves an entire user base. ...

For most teams evaluating dsh agent observability, the TencentCloud AgentObs SDK for DeepSeek Harness is worth adopting only if you already run Tencent Cloud Log Service. Six weeks after its first release, the evidence is mixed: solid engineering, thin adoption (14 stars, 519 npm downloads in 30 days), a privacy default that ships content capture ON, and a zero-collector architecture that replaces infrastructure you operate with metered consumption you must now watch. If you are not already a CLS customer, the community OpenTelemetry plugin is the safer default. ...

The Apex Inference Chip (SigmanticAI’s APEX, or tinyNPU) does run a real LLM on an FPGA, and it does so at 0.56 tokens per second on its fastest measured image — roughly 1.78 seconds per token, confirmed on two separate builds. That number is deliberately slow, honestly published, and the entire point of the project is the verification trail behind it. That paragraph is the honest summary, and this review exists because almost every other sentence about this repository can be made misleading with one small edit. Strip the words “measured on silicon” and APEX looks like a hardware failure. Strip “one decoder layer” and it looks like a chip. Strip “projected from an analytic model” and its 7B performance claims look like results. The repository itself refuses to let you do any of those things, which is why it is worth your time even if you never build a bitstream. ...

The fx coding agent is Vercel Labs’ experimental coding-agent harness and CLI, written in Zig and open sourced under Apache-2.0 in August 2026. It ships as a single native binary of roughly 6.2 MiB on Apple Silicon — no Node.js, no Python, no node_modules — and drives four interfaces (shell CLI, one-shot JSON, an ACP server, and a WebAssembly SDK) from one agent core. As of 2026-10-01 it is at 3,237 GitHub stars, v0.0.12, and is still explicitly labeled experimental with breaking changes between releases. ...