phuryn

Free repo (GitHub): phuryn/product-engineering If it helps, star the repo so the next person finds it.

pm-skills: This AI agent can create new skills while it works pm skills turns successful steps into reusable capabilities, builds a personal knowledge base, and applies what it learns to harder developer tasks. Instead of starting from zero on every prompt, it can improve the workflow as the job continues. #github #opensource phuryn/pm-skills

As promised, Staygreen is now open source: phuryn/staygreen You can host it yourself or get free uptime monitoring from San Francisco and Toronto, 50 monitors, checked every 3 min at staygreen.io It automatically detects common API health check endpoints + JSON fields inside HTML to suggest the monitors.

Allowance table update for GPT & competitors and Terra opinion My commentary about the model at the bottom of the post. For some time now I have been measuring GPT model's allowances and several others. Because a new GPT model was released, it is time for an update. Unfortunately, the trend continues. With each new model release, the new model gets less quota in API-equivalent compared to the previous ones. Also, unfortunately, the claim that we get 2x the usage out of Terra compared to Sol is a straight-up lie. Measured on a fresh context window on a Plus account for accurate measurements, like always. Context limited to 272k to make sure I don't go past the double price. I didn't check Luna or Astra yet, but Sol remains as it was. Generally the allowance for models is constant after their release from my measurements; only the new ones have less. |Name|Per 5 hours|Per week|Per month (30 days)|API cost÷Sub price ($20) ratio| |:-|:-|:-|:-|:-| |GPT-5.6 Sol|$14.4|$86.4|$370|18.5| |GPT 6 Luna|$6.5|$39|$167|8.35| |GPT-6 Sol|$11|$66|$283|14.15| |GPT-6.1 Terra|$5.33|$32|$137|6.85| |GPT-6.1 Sol|$7.8|$46.8|$200|10| |GPT-6 Astra|$10|$60|$257|12.85| For Pro, simply multiply by 5/10/25. As a bonus, here is a table of the "real cost" per million tokens and total amount of tokens per real dollar spent, also compared to DeepSeek v4.1 Flash (off-peak pricing), GLM 5.3 (off-peak), and Muse Spark. If you use OpenCode, Ollama Cloud, or CommandCode subscriptions, the price for DeepSeek goes down. At peak hours, the price for DeepSeek and GLM is 2x. Ollama has its own infrastructure with ZDR and uses different peak hours. For z.ai plans, you can subscribe quarterly or yearly for better ratings. |Model|Input|Cached Input|Output|Tokens per $1| |:-|:-|:-|:-|:-| |GPT-5.6 Sol ÷ 18.5|$0.216216|$0.021622|$1.081081|31.46M| |GPT-6 Luna ÷ 8.35|$0.011976|$0.001198|$0.059880|567.90M| |GPT-6 Sol ÷ 14.15|$0.141343|$0.014134|$0.706714|48.13M| |GPT-6.1 Terra ÷ 6.85|$0.145985|$0.007299|$0.729927|69.54M| |GPT-6.1 Sol ÷ 10|$0.200000|$0.010000|$1.000000|50.76M| |GPT-6 Astra ÷ 12.85|$0.778210|$0.077821|$3.891051|8.74M| |Claude Opus 5.5 ÷ 42*|$0.095238|$0.004762|$0.476190|106.60M| |DeepSeek V4.1 Flash (API)|$0.150000|$0.003000|$0.600000|103.52M| |DeepSeek V4.1 Flash (Ollama) ÷ 3|$0.050000|$0.001000|$0.200000|310.56M| |DeepSeek V4.1 Flash (CommandCode/OpenCode) ÷ 8|$0.018750|$0.000375|$0.075000|828.16M| |Muse Spark 1.3 Contributor (OpenCode) ÷ 6|$0.016667|$0.000333|$0.033333|1,103.33M| |GLM-5.3 (Ollama/OpenCode $40 Plan) ÷ 3|$0.466667|$0.086667|$1.466667|9.70M| |GLM-5.3 (Z.ai Lite) ÷ 7.668|$0.182577|$0.033907|$0.573813|24.80M| |GLM-5.3 (Z.ai Pro) ÷ 10.352|$0.135240|$0.025116|$0.425039|33.48M| |GLM-5.3 (Z.ai Max) ÷ 11.502|$0.121718|$0.022605|$0.382542|37.20M| |GLM-5.3 Flash (Z.ai Lite) ÷ 2.657|$0.056455|$0.011291|$0.188182|75.16M| |GLM-5.3 Flash (Z.ai Pro) ÷ 3.586|$0.041829|$0.008366|$0.139431|101.44M| |GLM-5.3 Flash (Z.ai Max) ÷ 3.985|$0.037641|$0.007528|$0.125471|112.73M| |GLM-5.3 Flash (OpenCode) ÷ 6|$0.025000|$0.005000|$0.083333|169.73M| |GLM-5.3 Flash (CommandCode) ÷ 8|$0.018750|$0.003750|$0.062500|226.31M| Tokens per $1 assumes 97% cached input, 2.5% uncached input, and 0.5% output. * - Claude charges for cache writes. The pricing is averaged. I used these numbers. Their findings for Codex match mine, so I trust them. My commentary about the new model. It is the first and only model that simply doesn't want to work. During a simple mechanical refactor, it stopped 3 times, telling me it is only finished partially. At the end it changed a part of the code, ran tests, saw several errors, reverted the change, and reported that it is only partially done. I ran it on XHigh reasoning. I doubt Max would change much, but it is not out of the question. This is not a model for agentic coding, that's for sure. It is quick, but very unreliable. I double-checked 6.1 Sol allowance; it is still the same as it was, so I recommend staying with 6.1 Sol or Luna. I have no use for 6.1 Terra. I see no reason in using it for coding when it gives us almost the same usage as 6.1 Sol. If you don't want your job done, use the new Terra, I guess. It may have a potential as an orchestrator, but that requires further testing. Unfortunately, OpenAI continues its tactic to nerf the usage with each new model. By my own measurements, they don't nerf the usage of each model after they are released; they are nerfed from the release (compared to previous ones). You can see the trend with 5.6 Sol -> 6 Sol - 6.1 Sol -> 6.1 Terra. We are at the third of the 5.6 allowance now, which was also lowered compared to 5.5.

How individual runs scored: - max: 34, 42, 46 - xhigh: 11, 9, 14 - high: 11, 8, 10 - medium: 8, 6, 8 - low: 7, 5, 6 Source data you can give to your agent (just point it to that repo), stars appreciated: phuryn/bug-hunt-bench

One gateway for Grok Build, Claude Code, and GPT The agent CLIs are useful, and they are a poor thing to install on every laptop. Each machine needs its own login, its own copy of grok or claude, and a key that should never have left the server. A team that wants one subscription, a token budget, and a single outbound proxy ends up repeating that setup per person. Remote LLM Gateway (AlanForester/remote-llm-gateway) keeps the project on the machine where the person is editing, and runs the agent on a server. The VS Code and Cursor plugin and the remote-llm CLI are thin clients. They speak ACP over a WebSocket. They do not call Grok, Claude, or OpenAI themselves. The gateway starts the agent, holds the keys, and writes the resulting edits back into the local tree. This project is not affiliated with or endorsed by xAI, Anthropic, OpenAI, or Meta. The names Grok, Grok Build, Claude Code, Codex, and Muse Code are used only to describe compatibility. What stays local Open a repository in the editor, or cd into it and run remote-llm. At the start of a session the client copies the workspace to a directory that belongs to that session. After the reply, changed files come back. The path on your laptop is only a key for that directory. The gateway does not clone your remote and does not open a checkout on the server at the path you sent. Commands the agent runs happen inside the session directory. Gitignored files stay behind, and so do dependency trees such as node_modules and vendor. A single file larger than 1 MB is skipped. The rest is uploaded in chunks, so a workspace that does not fit in one HTTP request still gets a session. There is a cap on the session copy: 20,000 files and 256 MB. Past that cap the overflow is skipped and the chat says so, instead of refusing the prompt. --no-sync on the CLI leaves the project entirely local. The plugin has no local grok or claude process to fall back on. If the gateway is down, the chat is down. One subscription, several people Every user connects to the same gateway. A session id is the user plus the project, on wss://<gateway>/ws/{sessionId}. The token travels in the Authorization header, not in the URL. An administrator creates the user and sees the access token once. From the admin UI they can pin a model, set a token limit, choose when the counter resets (a day, a week, a month, or a custom interval), and set the outbound proxy. Without a pinned model, the person picks one in the client. With a pin, the plugin and the CLI show that model and do not offer a different one. The limit counts tokens for the current period. When it is spent, new requests stop until the period ends and the counter resets. Older usage stays in the admin view. Provider traffic leaves the server through the proxy configured in the admin UI or in .env. Changing the address changes the exit path. The service does not need a rebuild. Sign-in for Claude and the other provider CLIs happens on the gateway machine. A login from a laptop does not authorize the agent, because the agent is not running there. Three backends, one protocol | Agent | On the server | |---|---| | Grok Build | grok agent serve, ACP over WebSocket | | Claude Code | claude agent serve | | GPT | OpenAI Responses API and tool calling inside the gateway | Grok Build and Claude Code bring their own tools and read MCP configuration from the gateway host. GPT has no separate Codex process. The gateway implements the tools: read, write, and show files in the session directory, and run a command there. Extra tools can be added from stdio MCP servers listed in MCP_CONFIG. The command environment is a short allow-list. Gateway secrets such as ADMIN_TOKEN, OPENAI_API_KEY, and REDIS_URL are not passed through. The same three entry points sit in front of that: - The plugin is a chat in the right sidebar of VS Code or Cursor. Set remoteLlm.gatewayUrl, store the token in Secret Storage, and open the chat. The command palette titles are currently Russian: «Remote LLM: задать токен», «Remote LLM: открыть чат», and «Remote LLM: показать назначенную модель». - remote-llm is the terminal client. Point it at the same gateway with --token or GATEWAY_TOKEN. The short name llm is intentionally unused. - The admin UI is users, tokens, models, limits, and usage. It listens on port 8081 when you run it locally. api accepts calls with user tokens. gateway accepts the agent sessions. The chat UI is based on GUI for Grok Build & Muse Code. That project has its own license. This repository is MIT. Try it Release v0.1.2 (github.com/AlanForester/remote-llm-gateway/releas…) has remote-llm binaries, a .vsix, and container images. The plugin is not in the Marketplace yet. Install the .vsix from that release, or run the services from source: cp .env.example .env # set ADMIN_TOKEN in .env set -a && source .env && set +a go run ./gateway/cmd/gateway go run ./admin/cmd/admin go run ./api/cmd/api Create a user at http://127.0.0.1:8081, then: go run ./cmd/remote-llm --gateway http://127.0.0.1:8080 --token <token> GATEWAY_ECHO=1 answers with an echo and does not need grok or claude installed. Unset it when you want a real agent. For a single host, deploy/docker-compose.yml brings up Redis, the three processes, nginx, Prometheus, and Grafana. Nginx is the public door: the admin UI on /, the gateway on /v1/ and /ws/, the API on /api/. The processes themselves listen on localhost. A local go run without Redis stores users, limits, and the proxy in DATA_FILE. WebSocket sessions are not written to Redis. A restart drops a live conversation. What this does not do The gateway is a shared place to run agents you already pay for. It is not a new model, and it does not move the editor to the server. Provider login belongs on the gateway host. GPT goes through the Responses API in-process, so there is no Codex CLI in the loop. The plugin is distributed as a .vsix in the GitHub release until it is published to the Marketplace. The code, the plugin, and the deploy files are at AlanForester/remote-llm-gateway. If the project is useful, star it: github.com/AlanForester/remote-llm-gateway/starga….

One more test of the work-humanizer skill. In a business context it's critical that AI doesn't: - Invent facts - Makes commitments - Changes the meaning Beats every other method. Free, open source: phuryn/work-humanizer t.co

phuryn/work-humanizer Install: npx skills add phuryn/work-humanizer Then: "Use work-humanizer on this: <paste text>" Every scoreboard output, verdict and fact check is in the repo. t.co