Tutorial: Proposal to the fairyfox system — adopt an "agent tooling / execution" standard for all projects

Proposal to the fairyfox system — adopt an "agent tooling / execution" standard for all projects

date: 2026-06-30 procedure: propose-standard (node → hub suggestion) node: random-ai-prompt outcome: proposal-only (hub NOT modified)

Proposal to the fairyfox system — adopt an "agent tooling / execution" standard for all projects

A node-originated suggestion for the hub to adopt, written here for the owner to carry upstream. Per the anti-recursion / stay-inside-this-repo rules, the hub repo was not touched. Acting on this at the hub is the owner's separate, manual step.

One-line ask

Make the machine's tooling rule a shared, cross-project standard: agents use PowerShell (the Windows-MCP PowerShell tool) + the file tools, never the Cowork mcp__workspace__bash sandbox — and they EXECUTE git/build/release work directly rather than handing the user a script.

Why (the trigger)

This came up repeatedly while working on random-ai-prompt. Two recurring failure modes, both costly:

  1. Using the Cowork bash sandbox at all. On this Windows machine the sandbox is actively broken, not just redundant: it serves stale/truncated file views (read a 42-line package.json as 36 lines), it can't reliably touch .git (unable to unlink .git/objects/... Operation not permitted), and it mangles line endings — once staging a 12-line changelog edit as 3,256 phantom CRLF-flip lines. Every slip risks a bad edit or a corrupted commit. (It happened again this session — bash used for "quick" repo reads — which is what prompted the owner to re-emphasize the rule.)
  2. Handing off instead of executing. The agent has PowerShell + full git control on the machine at all times, but kept writing "here's a PowerShell script for you to run" for verify/commit/release. The owner's expectation is the opposite: when they say ship / clean up and commit / verify, the agent should run the gate, stage, commit, branch, merge, and release itself — pausing only for the confirmations the workflow actually requires (e.g. the go-ahead to release to main, which "ship" already grants).

Neither is specific to this project; any fairyfox node operated in this environment hits both. Hence a shared standard rather than per-repo CLAUDE.md notes that each have to rediscover.

Proposed standard (sketch — for the hub to refine)

Suggested name: hub/standards/agent-tooling.md (or a section in an existing operating-guidance standard). Rough shape:

  1. Never use the Cowork bash sandbox (mcp__workspace__bash) on this environment — zero exceptions, including "quick" reads. Use the file tools (Read/Edit/Write/Glob/Grep) for files and the Windows-MCP PowerShell tool for everything else (npm, git, node, builds, doxygen).
  2. Execute, don't hand off. With PowerShell + git available continuously, the agent performs verify/commit/branch/merge/release directly. A handed-off "run this script yourself" is the wrong default; reserve user steps for genuine approvals.
  3. Known PowerShell gotcha to encode: in-place edits via Get-Content -Raw | Set-Content (and positional Set-Content) can silently no-op on this machine; use [System.IO.File]::ReadAllText/WriteAllText for scripted bulk replaces (or the Read/Edit tools).
  4. Respect the repo's core.autocrlf=true — PowerShell stages real diffs cleanly; the bash sandbox does not.

Reference implementation (in this repo)

  • Memory: no-bash-use-powershell (the rule + the "execute, don't hand off" point + the CRLF/.NET gotchas).
  • Both local CLAUDE.md files already direct PowerShell + file tools over the sandbox; this proposal lifts that to the hub so every node inherits it.

Guardrails honored

On-request only; no hub pull/push; reference clone untouched; no edit to any other repo. This is a report for the owner to take to the hub manually.