Bernardo Magri
207b6f3c64
All checks were successful
Check / eval (push) Successful in 3m57s
Bernardo: "opening claude-code is too disruptive". SUPER+CTRL+A spent a terminal, an npm fetch and a REPL on "what's the syntax for…" — the prompt was the right surface, the destination was not. The query is now URL-encoded into the provider's web chat and handed to xdg-open. settings.ask.provider (claude|chatgpt|gemini|grok, default claude) with a Preferences row, read from state at RUN TIME so switching needs no rebuild, and any unknown/absent value falls back rather than opening nothing. He asked for chatgpt + grok alongside claude/gemini, and for the CLI to go: claude-code is deleted and pkgs.nodejs leaves with it (carried only for that npx). Keybind is "Ask AI" now the destination is the user's. The failure mode is the design point: prefill is undocumented on all four vendors, each can drop `?q=` unilaterally, and it fails SILENTLY — you land in an empty chat having already typed the question. So the query is also copied to the clipboard (a vendor breaking change costs one paste, not the question), and the URLs sit in one attrset so the repair is one line. Verified against the generated script: bash -n clean; every bad state value (unset/null/garbage) falls back to claude; a hostile query encodes correctly (& → %26, ? → %3F, # → %23, $(whoami) neutralised, café → caf%C3%A9). NOT verified on purpose: whether each vendor fills the box or auto-submits — that means posting into Bernardo's own accounts. Sweep: ROADMAP's shipped-modules line described the npx REPL; nix-ld's comment cited claude-code as its example; both fixed. Menu row + keybind renamed. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Agent instructions + loop state
Everything an AI agent needs to work on Nomarchy, vendor-neutral and git-tracked. Protocol: LOOP.md. Entry point for every harness: repo-root AGENTS.md.
Instructions (how to work)
| File | Who writes | Role |
|---|---|---|
| LOOP.md | Human | One-iteration protocol (orient → pick → work → verify → commit → record) + the V0–V3 ladder |
| VERIFICATION.md | Human (agents propose) | Enforcement: preflight, honesty rules, visual protocol, hardware-blocked checks, reporting |
| DELEGATION.md | Human (agents propose) | Capability tiers, scout/runner roles, token economy, parallel fan-out |
| GOALS.md | Human (agents propose) | Pillars, quality bars, non-goals |
| CONVENTIONS.md | Human (agents propose) | How to write code/menu/state while shipping |
| THEME-DESIGN.md | Human (agents propose) | Theme/visual design instructions |
State (what's happening)
| File | Who writes | Role |
|---|---|---|
| BACKLOG.md | Both | Prioritized queue — only executable work list |
| JOURNAL.md | Agents | Append-only iteration log (read last 3–5 entries; older → JOURNAL-ARCHIVE.md) |
| MEMORY.md | Agents | Curated durable gotchas |
| HARDWARE-QUEUE.md | Agents append, human checks | On-hardware V3 tests only Bernardo can run |
Product / design docs (not a queue)
| File | Role |
|---|---|
| ../docs/VISION.md | v1.0 product themes — agents slice into BACKLOG PROPOSED |
| ../docs/ROADMAP.md | Design history + shipped log |
| ../docs/README.md | Full docs map |
Harness adapters (vendor-specific, thin)
Shared content never lives in an adapter — adapters only register/route into the files above, in whatever format their harness requires.
| Path | Harness | Role |
|---|---|---|
| ../AGENTS.md | any | Entry point (CLAUDE.md is a symlink to it) |
| ../.claude/settings.json | Claude Code | Tool permissions |
| ../.claude/agents/ | Claude Code | nomarchy-scout / nomarchy-runner role defs (contracts in DELEGATION.md) |
Do not put backlog items, vision text, or policy under an adapter directory — it is not shared with other agent runners.
Rules of thumb
- Execute from BACKLOG only (NOW → NEXT; never PROPOSED without human triage).
- Orient with GOALS + CONVENTIONS + MEMORY + last journal + BACKLOG; when the task is product-shaped, also read the relevant VISION §.
- Record lasting design in ROADMAP ✓ when something ships that future humans should know; delete the BACKLOG line.
- v1 branch is human-only — never advance from an agent session.