标签：#CLI

A self-improving memory loop sounds like a clear win until you watch it rewrite something correct with something outdated. The agent remembered a fact. You verified it. A later consolidation pass ran against a stale context window, decided the memory was imprecise, and replaced it with a weaker version. The original was better. You lost ground. This is the failure mode that belief-aware memory was built to prevent. Not "agents write wrong things" — that's a model quality problem. The specific failure is: the improve loop, running unsupervised, overwrites correct content it should have left alone. A loop that can degrade its own best work is worse than no loop at all. akm 0.8.0 ships captureMode and beliefState as first-class frontmatter fields on memory assets. Together they tell the consolidation pass what each memory is, what the agent believes about it, and whether it is eligible to be rewritten. The Two Capture Modes Every memory asset now carries a captureMode field. It has two values. hot means the memory was written or explicitly confirmed by a human. The improve loop treats hot memories as read-only. No consolidation plan, no merge proposal, no rewrite. If every memory in a chunk is captureMode: hot , the consolidation pass skips the LLM call for that chunk entirely — the chunk is counted as judgedNoAction before a single token is spent. This is the all-hot chunk early-exit. background means the memory was generated by an agent — promoted during a prior consolidation run, written by an inference pass, produced by akm remember without explicit human review. Background memories are eligible for improvement. The consolidation pass can propose merges, rewrites, deletions, or upgrades. When no captureMode is set, the memory is treated as eligible for consolidation. Memories that existed before 0.8.0 are treated this way on first encounter. --- captureMode : hot beliefState : asserted description : Primary LM Studio endpoint moved to Shredder (192.168.0.99:1234) -

This is part eleven in a series about managing the growing pile of skills, scripts, and context that AI coding agents depend on. Part nine covered workflow assets and resumable procedures. Part ten introduced the improve pipeline that continuously curates your stash. Earlier parts addressed teams, distributed stashes, and community knowledge. Most automation with AI agents is reactive. You open a session, give the agent a task, wait for the result, close the session. The agent's clock runs when you run it. Task assets flip that model. A task is a YAML file in your stash that defines a workflow — what to run, when to run it, what environment it needs, and how long it's allowed to take. Once registered, the task runs on schedule without your involvement. The OS scheduler calls akm tasks run <id> , which executes the task and writes the result to state.db . You find out what happened when you check akm health or look at the log. This is the piece of akm 0.8.0 that makes continuous operation possible. The improve loop runs twice an hour because a task asset says it does. The hourly Discord health report fires because a task asset says it does. Neither requires an open terminal. The Task Asset Format Task assets live at <stash>/tasks/<id>.yml . The filename is the task ID. A minimal task looks like this: schedule : 0 * * * * command : akm improve --auto-accept 90 enabled : true That's enough to install a cron entry and run akm improve at the top of every hour. The full schema adds metadata and per-task timeout control: schedule : " 7,37 * * * *" command : akm improve --auto-accept 90 --timeout-ms 1620000 enabled : true timeoutMs : 1800000 name : akm-improve description : Run the improve pass at :07 and :37 — reflect, distill, consolidate, lint, and eval. when_to_use : Twice per hour; leaves ~23 minutes of idle headroom between completions. tags : - improve - maintenance The fields that matter most: Field Required Purpose schedule yes Standard cron expression. Maps to cro

This is part ten in a series about managing the growing pile of skills, scripts, and context that AI coding agents depend on. Part nine covered workflow assets, vault assets, and the writable git stash. Part eight tackled multi-wiki support for structured research. Earlier parts addressed teams, distributed stashes, feedback scoring, and community knowledge. This one is about entropy. You ship a feature. Your agent writes several memories during the session — partial findings, a workaround, a note about the build step that kept failing. Those memories are accurate when written. Three sprints later, the workaround is no longer needed, two of the memories say slightly different things about the same subsystem, and the note about the build step refers to a CI config that was replaced. None of this is catastrophic. But it accumulates. After six months, a significant fraction of your stash is stale, redundant, or quietly wrong. You could audit it manually. In practice, you won't — the stash is too large, the relevance of any given memory is hard to assess without the context where it was created, and the judgment calls (merge these two? promote this? delete that?) are exactly the kind of work that's tedious for a human and tractable for an LLM. akm improve is the answer to that problem. It is a multi-phase pipeline that reads your stash, evaluates asset quality, consolidates scattered memories, extracts structured facts, and maps entity relationships — on a schedule, without manual intervention, producing proposals you can review before anything changes. The Five Phases akm improve is not a single LLM call. It is a sequenced pipeline where each phase produces inputs for the next. Reflect evaluates asset quality. For each asset in scope, the reflect pass reviews the content against usage signals — search hits, retrieval counts, feedback — and produces a quality assessment. Low-quality assets are flagged as candidates for improvement. Since 0.8.0, reflect can run as a dire

Would you take a payment to ramp down your electricity use? Would it change anything if you were doing so to help power a local data center? Google just signed a new deal to help pay for a virtual power plant (VPP) in the largest power grid in the US. The agreement is with Voltus,…

Fusion startup Xcimer fired up the world's largest privately owned laser.

If you've used other browser-based code editors, you've probably noticed that the terminal feels off. You can run a script. You can print to stdout. But the moment you try to install a package and then actually use it in the next command, something breaks. The environment doesn't carry over. It feels like every command starts from scratch in a vacuum. That was the problem I wanted to solve when building the terminal for Cloudpen. Not just a place to run isolated snippets, but a proper environment where you can install dependencies, run build tools, and have everything you did in one command still be there for the next one. What most cloud terminals get wrong The core issue is that running code in the browser is hard to do without cheating somewhere. A lot of tools use sandboxed environments that look like a terminal but don't behave like one. They're good enough for demos. They fall apart in real work. The thing developers actually need is simple: if I install something, it should be there when I run the next command. That's it. That's the whole requirement. Surprisingly few cloud tools actually deliver it. How Cloudpen handles it Without going into the full technical detail, the short version is this: every command runs in a completely isolated environment, but all commands within your session share the same filesystem. So when you run npm install, those files are written somewhere. When you run your next command, that somewhere is exactly where it looks. Package installs work. Build tools work. Multi-step workflows work. And because each command runs in a clean, isolated environment, there's no bleed between users or sessions. The current terminal is optimized for commands that run to completion, while live application previews are handled through Cloudpen's deployment system. On the free plan, you can run any file in your project and see the output directly in the terminal. The live coding environment where you type commands yourself is on the Pro plan. Both use

The National Center for Atmospheric Research won't be losing its supercomputer.

Another fusion startup has raised another a massive round to make this type of power a reality.

Switching from fossil fuels to electricity for heating is no longer covered.

Hello, I'm Maneshwar. I'm building git-lrc, a Micro AI code reviewer that runs on every commit. It is...

Pacific Fusion's sub-scale prototype delivered enormous amounts of power in a flash, setting the company up for its demonstration power plant.

Google has released a new CLI for Google Workspace, offering a unified interface for various services like Drive, Gmail, and Calendar. Built in Rust, the tool dynamically adjusts to API changes and features over 100 bundled skills. It requires Node.js and a Google Cloud project for setup. Initial community feedback is mixed, highlighting both its dynamic capabilities and setup challenges. By Daniel Curtis

Mike Schroepfer's Gigascale Capital has raised a large fund to back founders building climate-friendly solutions for the world's energy and material shortages.

⚠️ この記事はアフィリエイト広告（プロモーション）を含みます。リンク先で発生した収益の一部が運営者に支払われますが、読者の購入価格には一切影響ありません。 By the end of this article you'll have two runnable Python scripts: a CLIP-based pre-filter that re-checks SDXL Turbo output before it ever hits Pinterest, and a prompt sanitizer that strips artist names + trademarked characters so you don't eat a DMCA. I ran this pipeline for 41 days, generated 6,180 images, and went from a 9.7% Pinterest rejection rate down to 2.6%. Here's exactly what broke and what fixed it. Why SDXL Turbo (1-step, ~0.3s on a 4090) beats SD 1.5 for Pinterest volume First, the conclusion: if you're mass-producing pins, SDXL Turbo's single-step guidance_scale=0.0 generation is the only thing that makes the unit economics work. On my RTX 4090 I clock 0.31s per 512x512 image with Turbo vs 4.8s for a 30-step SDXL base run. That's 15x. Over 6,180 images that's the difference between 32 minutes and 8.2 hours of GPU time. But Turbo has a nasty side effect nobody warns you about: because it's distilled and runs at low resolution by default, its built-in StableDiffusionXLPipeline safety checker (when enabled) throws far more false positives on perfectly benign images — beaches, lingerie-free fashion flatlays, even close-up food. In my first 600-image batch, 58 images came back as black squares from the NSFW checker. 51 of them were photos of latte art and knitted sweaters . So I ripped out the default checker and built my own two-stage gate. Stage 1: Replacing the diffusers safety_checker with a tunable CLIP gate in Python The default safety_checker in diffusers is a binary black box — you get a black image and zero signal about why . For a production loop you need a confidence score so you can set your own threshold. I use OpenCLIP's ViT-B-32 to score each output against a small set of NSFW concept prompts, then compare to a safe-concept baseline. This code actually runs (tested on diffusers==0.27.2 , open_clip_torch==2.24.0 ): import torch import open_clip from PIL import Ima

As temperatures rise, some creatures pick fights while others struggle to learn.

This is a submission for the GitHub Finish-Up-A-Thon Challenge What I Built pq — jq for Parquet. A 50 MB Rust single binary that wraps DuckDB's query engine in a jq-style expression DSL, optimized for terminal one-liners and unix pipes. $ pq sales.parquet 'group_by .country | sum .revenue | top 3 by sum_revenue' ┌─────────┬─────────────┐ │ country ┆ sum_revenue │ ╞═════════╪═════════════╡ │ US ┆ 19065.00 │ │ FR ┆ 999.99 │ │ DE ┆ 312.00 │ └─────────┴─────────────┘ Where it started. I work in adtech. I look at parquet files dozens of times a day — campaign deliveries, partner exports, audience snapshots. Every existing option was painful: Tool Pain pyarrow / pandas 5-second cold start, 200 MB virtualenv parquet-tools JVM, slow, no query support pqrs Inspector only — can't filter or project duckdb CLI Great engine, but SELECT email FROM 'file.parquet' WHERE country='US' is too verbose to type 50 times a day Spark Are you serious pq is the tool I actually want — single binary, no JVM, no Python, jq-style syntax for piping into the rest of the unix toolbox. It's been my default cat for parquet since v0.5. Demo Repo : github.com/thehwang/parq Latest release : v0.14.0 (this submission) Install : brew install thehwang/parq/pq Tutorial : doc/tutorial.md — 30-minute hands-on walkthrough A taste of what shipped in v0.14: # Streaming JSON output (was the only buffered format until v0.14) $ pq big.parquet '.id, .country' -o json | head -c 200 # returns instantly even on a 40 GB file # Schema-drift gate for CI $ pq diff baseline.parquet candidate.parquet # Schema diff - a: ` baseline.parquet ` - b: ` candidate.parquet ` ## Added (1) | column | type | nullable | |-----------|---------|----------| | ` country ` | VARCHAR | yes | $ echo $? 1 # exits non-zero on drift, slots into CI without scripting And the new TUI Explain panel — press capital E for EXPLAIN ANALYZE , get row-group pruning per scan (this is exactly the panel you see on the cover image at the top of this post): Expla

AI coding tools have moved way past autocomplete. Today's CLI agents read your entire codebase, plan changes across files, run tests, and even open pull requests - all from the terminal. Picking the right one matters, and in 2026 there are several solid options worth knowing. Why CLI Over IDE? IDE plugins work within a single editor and optimize for in-file completions. CLI agents operate at the shell level - they run commands, manage files across your whole repo, handle Git, and work in remote servers or CI pipelines. They don't lock you into one editor either. You keep your existing setup and layer the agent on top. Claude Code (Anthropic) Claude Code is Anthropic's official terminal agent and the top-ranked CLI tool in 2026. It handles complex, multi-file tasks better than most - analyzing architecture, coordinating edits across files, reviewing PRs, and running multi-step refactors. Supports custom slash commands and sub-agents for team workflows. Pay-per-token pricing with no free tier. Codex CLI (OpenAI) OpenAI's open-source terminal agent. The standout feature is sandboxed execution - code runs in isolation before touching your filesystem, reducing risk of irreversible changes. Fast to start, minimal footprint, and supports one-shot mode for CI pipelines. Best for OpenAI-stack teams that want a safety net around agentic execution. OpenCode A fully open-source agent supporting 75+ model providers - Anthropic, OpenAI, Google, Mistral, and local models via Ollama. Switch providers mid-session. Uses a dual-agent system: a Plan agent for structured reasoning and a Build agent for implementation. LSP integration brings real code intelligence into the terminal. Free with local models. Aider Aider has the largest installed base of any open-source CLI agent - over 4.1 million installs. Its Git-native design is the key differentiator: every change gets auto-committed with a descriptive message. If something breaks, git revert gets you back instantly. Supports any model

The problem I often switch between Codex, OpenCode, Cline, Claude Desktop, scripts, and terminals. The annoying part is not starting a new tool. The annoying part is explaining the same workspace state again: what changed what is still pending what should not be touched what tests passed what the next agent should read before editing What I built AgentContextBus (acb) is a local-first CLI for handing off workspace context between coding agents. It saves a local handoff packet, then lets the next agent read it through: paste-ready prompts brief prompts a local dashboard JSON output explicit MCP tools First run npx @xiaoshuo1988/acb verify first-run For Chinese output: npx @xiaoshuo1988/acb verify first-run --lang zh-CN A normal handoff From the agent that has context: acb handoff --from codex --summary "Ready for the next agent" --git From the receiving side: acb receive --latest After the receiving agent summarizes the packet: acb ack --latest --by opencode What ACB intentionally does not do no hidden prompt injection no traffic interception no third-party client config mutation no cloud sync no background daemon Why local-first I want the user to be able to inspect the packet store, copy text manually, and decide exactly when context crosses from one agent to another. What I want feedback on Is the handoff packet concept clear? Is verify first-run enough to understand the tool? Is receive --latest the right receiving-side command? Which client path needs the most work? Would you trust this workflow in a real project? Repo: https://github.com/xiaoshuo1988130/acb Feedback discussion: https://github.com/xiaoshuo1988130/acb/discussions/1

fd vs find vs ripgrep: I Created 10,000 Files to Settle This Debate TL;DR: fd is ~2.5x faster than find for filename searches, rg demolishes grep by ~3x for content searches, and find + grep combined lose on every single benchmark I ran. But there's a catch: both fd and rg skip hidden files by default, which can bite you if you're not paying attention. Here are the receipts. Why I Did This Every time someone posts a shell one-liner using find on Reddit, there's always that guy in the comments: "jUsT uSe fD, iT's fAsTeR." Then someone else chimes in with "actually ripgrep can do that too." I got tired of the anecdotes. I wanted numbers. Real ones. On real files. So I fired up WSL, generated 10,900 files across 1,506 directories (~143 MB of mixed content), and ran actual benchmarks with hyperfine . No synthetic microbenchmarks, no "I feel like X is faster" — just cold, hard terminal output. Methodology The Test Bed I created a directory at /tmp/fd-benchmark containing: Category Count Details Plain text files 2,000 file_*.txt — 20 bytes each, contains "test content line N" Binary files 2,000 data_*.bin — 15 bytes each Log files 1,500 match_*.log — contains unique "match_this_test_N" strings Config files 1,000 nested_file_*.cfg Nested dir files 1,000 level1_*/level2/level3/deep_*.txt + level1_*/shallow_*.txt Hidden root files 1,500 .hidden_* + .config_*.yml Hidden dir files 500 .hidden_dir/subdir/deep_hidden_*.txt Git objects 500 .git/objects/obj_* Multi-ext source files 800 src_*.{py,js,ts,rs,go,java,rb,php,cpp,h,css,html,json,xml,yaml,md} (50 each) Large binary files 100 large_*.dat — 1 MB each (random data) Total 10,900 $ du -sh . 143M . $ find . -type f | wc -l 10900 $ find . -type d | wc -l 1506 Tools Tested Tool Version What It Does find (GNU) 4.9.0 The OG. Ships with every Linux distro. fd 10.2.0 Rust-based find alternative. Smarter defaults, colored output. grep (GNU) 3.11 Content search. Also the OG. rg (ripgrep) 15.1.0 Rust-based grep alternative. Respects .gi

Let Your AI Agent Manage skill itself Most developer tools still assume the human is the operator. You read the documentation. You install the CLI. You decide where files should go. You copy commands from a README, paste them into a terminal, check the output, fix the path, and then explain the final state back to your AI coding agent. That made sense when tools were only built for humans. But AI coding agents now run commands, inspect files, follow project conventions, and repair broken local state. If a tool is meant to help agents, the better question is not: How does a human use this CLI? It is: Can the agent operate the CLI by itself? That is one of the quiet but important ideas behind aweskill : it is a CLI-first Skill package manager that AI agents can operate themselves. Website: aweskill.webioinfo.top It is already used as supporting infrastructure for several Webioinfo projects: awescholar — AI-agent-operable scientific literature discovery and curation. Search, annotate, filter, and report on academic papers. aweshelf — Session bookmark manager for Claude Code and Codex. Bookmark, categorize, and restore sessions with aweswitch profiles. Awesome AI Meets Biology — A curated survey of AI applications in biology, bioinformatics, and biomedical research, powered by awescholar. The Old Workflow: You Manage the Agent's Tools When a new AI Agent needs a Skill, the usual workflow looks like this: You find the Skill. You download or copy it. You locate the agent's Skill directory. You place SKILL.md in the right folder. You restart the agent. You hope the next agent uses the same layout. This is manageable once. It becomes messy when you use Codex, Claude Code, Cursor, Gemini CLI, Windsurf, Qwen Code, OpenCode, or any other coding agent side by side. Each one has its own directory layout and conventions. The human becomes the package manager. That is backward. If the agent is already capable of editing your repo, running tests, and diagnosing failures, it should