Working notes from consolidating a Proxmox cluster down to one node and fixing the storage layer so containers can be snapshotted and migrated cleanly going forward.
The only value masked for this public write-up is the NVMe drive serial (
<YOUR-NVME-BY-ID>). Substitute your own/dev/disk/by-id/...name there.
Reproducible setup notes for sharing the tank ZFS pool over SMB to Windows clients,
alongside an existing Linux/cachefs access path, on a Proxmox (Debian 13) host.
Host: aorus — Proxmox VE, Debian 13, ZFS pool tank mounted at /tank.
Goal: Add a password-protected SMB share of /tank/shared for Windows, without
breaking the existing container access, and surface the daily ZFS snapshots as Windows
"Previous Versions".
The narrow-scope Obsidian plugin recommendations have been merged into the broader problem-statement-and-solutions booklet.
See the merged document at https://gist.github.com/YoraiLevi/893a2e5f0725457685e4cd0664ed9cc6
The Obsidian plugin recommendations now live inside the problem-brief's three-frictions structure: each plugin recommendation slots under the friction it addresses, with motivated reasoning, devil's-advocate counter-arguments, and evidence-base citations inline.
| title | Engineer↔Agent Workflow Infrastructure for a Dyslexic, Voice-First Operator — Problem Brief and Recommended Solutions |
|---|---|
| created | 2026-05-27 |
| updated | 2026-05-28 |
| status | SPEC — not actively implemented. A booklet describing the workflow gap and a stack of recommended solutions with evidence and counter-arguments. |
| audience | solo engineers working with executing AI agents who face the same friction, and AI agents asked to research or build the system |
Reflections from a session building a company-foundation operator template. Captured 2026-05-27 before a break.
What stands out across this session is a constant tension between making the operator's reading experience cheap and making our own authoring experience honest. Every choice we made favors honest authorship — research before prose, descriptors before content, decisions captured at the point of impact — even when a shortcut would have produced apparent progress faster. We have been willing to be slower today to be cheaper tomorrow.
The recurring move is "build the artifact that prevents the failure mode." When the engineer would have to grep five files → one landing page. When a recommendation would have aged badly within months → research before prose. When a quirk burns time → file the pitfall. When new content scatters across a directory → restructure. Each artifact exists because a specific failure mode was noticed and the artifa
How /goal actually works under the hood, and how to get the same effect yourself with a one-line file edit — no plugin, no restart, no patched binary.
Claude Code has a private, in-memory sessionHooks registry that only the built-in /goal command and Skills can write to. You can't access it from user-land.
But the runtime re-reads your .claude/settings.local.json file on every single hook event, so dropping a hooks block into that file is effectively the same thing: the hook becomes active on the next tool call, and deleting it from the file removes it on the call after that. No restart needed. Both directions were verified empirically.
The user invoked /agent-pain-points with scope = internet-research-agent,
electing the inverted variant where the IRA introspects on its own behavioral
definition rather than the calling agent introspecting from outside. Plan-mode
constraint required the subagent to return everything in-message (no file
writes). The forthcoming filename (...-sharded-hopcroft.md) signals a
downstream sharded/partitioned redesign of the IRA; pain points surfaced here
Two related quality-of-life upgrades to ~/.config/zellij:
Ctrl+Shift+A is currently a session-only picker. It opens a list of zellij sessions (via dispatch.py:_list_sessions) and switches to whichever the user picks, landing on whatever tab was last focused in that session. The user wants it upgraded to a flat list of every tab in every live session, with the picker jumping directly to the chosen (session, tab) pair. Without this, you cannot navigate by intent — only by session, then by tab via a second action.Tab #N). For the unified picker to be useful, tab names must reflect the currently running foreground program (e.g. vim, python, cargo) and fall back to the working directory when the shell is idle. This must work for both shells the user runs: WSL bash and PowerShell 7+ (pwsh).These are designed together because they're
What this is. A design proposal for an orchestrator that runs an army of live-TUI agentic CLIs (Claude Code, Codex, Gemini, etc.), controllable simultaneously from a web browser and from a local/SSH terminal — the same PTY visible to both.
Audience. The engineer who's going to build this. Assumes they've already read the companion catalog Gist (link below) and know the ~50 mechanisms surveyed there.
Companion Gist. This document references the catalog Gist by chapter throughout. The catalog answers what mechanisms exist; this document answers which mechanisms to compose and how.
Companion gist URL. https://gist.github.com/YoraiLevi/ba3d3b72374153909c0feeeb1b046c32 >
Audience. Engineers designing a system that programmatically controls one or more interactive AI coding CLI processes (Claude Code, Codex, Kiro, opencode, Copilot CLI, Gemini CLI, Amp, Cline, etc.).
Format. Capability-keyed. Chapter 1 surveys every capability such a system may need. Chapters 2–16 take one capability each and present every observed mechanism in detail. Chapter 17 composes the recommended Windows-native stack end-to-end.
Verification. Every load-bearing claim cites
file:linein cloned source or a URL. Claims marked[unverified]could not be sourced to primary evidence.Style. See
research/gist-style-guide.md.