Development environment

How this repository is wired into the machine and the agent harness around it: which MCP servers are enabled and why the rest are not, which ECC skills and agents are installed, what .ignore excludes, how brand assets are deployed, and where the dependency caches live.

None of this is about the chain. It is about the cost of working on it — a tool nobody calls is a tax on every session, and a cache in the wrong place fills the wrong volume.

Moved out of CLAUDE.md on 2026-09-20, verbatim.

MCP Servers

Project scope, .mcp.json, enabled in .claude/settings.local.json. Tool search defers full schemas, but every server still puts its tool names and its instructions block in the prompt on every turn. So a server nobody calls, or a tool nobody calls, is a permanent tax. The set below was cut to what 18 sessions of transcripts show being used (2026-09-11), and each server has one job so that two of them are never competing for the same question:

Question	Server and tool
Where is X defined? What is in this file?	serena find_symbol / get_symbols_overview (rust-analyzer: exact)
Who uses X? Rename X everywhere.	serena find_referencing_symbols / rename_symbol
What calls what, across crates? What is the shape of this subsystem?	codebase-memory-mcp trace_path / get_architecture / search_graph
A crate’s API, before depending on it or calling a part of it this repo does not use yet	context7 resolve-library-id then query-docs (the docs-lookup agent runs on it)
One known URL (an RFC, a FIPS spec, docs.rs)	fetch

• serena: its file-read, shell, memory and text-insertion tools are excluded in .serena/project.yml, because the built-ins already do those jobs.
• codebase-memory-mcp (0.10.8): .mcp.json overrides the user-scope entry with --tool-profile=scout, which cuts it from 15 tools to 7 and from 24.9K to 13.8K schema characters.
  • Re-indexing is not in the scout profile. Re-index with codebase-memory-mcp cli index_repository '{"repo_path":"D:/Maya2C","mode":"full"}'.
  • claude mcp list warns that the server is defined in two scopes. That is the override working, not a fault.
  • Version 0.9.0 silently skipped files.

Disabled deliberately:

• Project servers, via disabledMcpjsonServers:
  • filesystem duplicates Read/Write/Edit/Glob.
  • git duplicates Bash git, which is allowlisted in .claude/settings.json.
  • memory duplicates codebase-memory-mcp and the file memory under ~/.claude/projects/.
  • sequential-thinking duplicates built-in extended thinking.
  • headroom was never called. Its headroom_compress takes the text as an argument, so that text is already in context before anything is compressed. It cannot save tokens from inside the conversation.
• User servers and connectors, via disabledMcpServers:
  • rustrover was never called, and adds about 40 tool names (SQL, database, run-configuration tools) that serena and the shell already cover. It is also only live while the IDE runs.
  • The claude.ai connectors Shopify and Viewmax add about 80 tool names between them and have nothing to do with a blockchain.

All of these are reversible: remove the name from the list. The pre-trim configs are in ~/.claude/backups/mcp-trim-20260911/.

Research tooling is shell-side, not MCP, so it costs nothing until it’s used:

Tool	Use it for
agent-reach skill (~/.claude/skills/agent-reach/)	Multi-source research. It routes to the tools below. The CLI is pinned to upstream Panniantong/Agent-Reach@da5044d; do not run check-update
mcporter call 'exa.web_search_exa(query: "...", numResults: 5)'	Semantic web search: papers, advisories, standards. Exa is configured in ~/.mcporter/mcporter.json
yt-dlp --write-auto-sub --skip-download	Conference-talk transcripts. ~/.config/yt-dlp/config sets --js-runtimes node
gh	Issues, PRs, releases, gh search code. Needs gh auth login once

Social channels (X, Reddit, …) are deliberately unconfigured: they need the user’s browser cookies. Agent-Reach’s own MCP server exposes only get_status, so it is not registered.

Graphify is a skill, not an MCP server — invoke with /graphify.

Harness Surface (ECC + plugins)

~/.claude runs ECC 2.2.1 as a manual install, not the ecc@ecc plugin — never run /plugin install ecc@ecc on top of it, that duplicates every skill, command and hook. The same tax that applies to MCP tool names applies here: every skill, agent and command puts its name and description in the prompt on every turn. The full developer profile measured ~18.1K tokens/turn, so the surface is curated to this stack and the rest is archived, not deleted, in ~/.claude/backups/curated-20260912/:

Surface	Installed	Archived	Cost
Skills	43	85	~3.7K
Agents	27	41	~1.8K
Commands	45	49	~1.2K
Plugins (4)	—	—	~1.3K

Kept language packs are the ones this repo actually contains: rust-*, cpp-* (cuda-miner), python-* (scripts/), typescript-reviewer (sdk-js, docs/site). Everything Django/Laravel/Vue/Flutter/Kotlin/Swift/SEO/marketing is archived. Restore one by moving the file back — do not re-run ecc install --profile developer to get it, that restores all 240 files and the full 24-entry hook graph.

Plugins (user scope, official marketplace only): rust-analyzer-lsp, claude-security (8 scanning agents — the reason to keep the priciest plugin on a consensus/crypto codebase), skill-creator, claude-md-management, claude-code-setup. Deliberately not installed: github and serena plugins (the gh CLI and the serena MCP server already cover them), semgrep and sonarqube (thin Rust rule coverage next to clippy + cargo audit + cargo deny, and a permanent tool-name cost), pr-review-toolkit and code-review (ECC ships the same agents), superpowers (a third overlapping TDD/verification/review system makes agent selection worse), codspeed (wants a CI account), security-guidance (POSIX shell hooks plus an LLM diff review on every Stop).

Project agents (.claude/agents/, ~0.6K tokens/turn): nine curated from msitarzewski/agency-agents@ad9264e (MIT) out of 295 — blockchain-security-auditor, security-architect, codebase-archaeologist, research-synthesist (read-only tools:), plus rust-refactoring-specialist, webassembly-engineer, minimal-change-engineer, sre, desktop-app-engineer. Each has its upstream path in a frontmatter comment and a trailing Maya2C Operating Context section that overrides the generic web/EVM body. Skipped on purpose: upstream code-reviewer and software-architect (collide with ECC’s), reality-checker (Laravel + Playwright screenshots), solidity-* (no EVM here), and every non-engineering division. Add one by hand in the same shape — do not run upstream install.sh --tool claude-code, which copies all 295 into ~/.claude/agents/.

Nine ECC hooks are wired in ~/.claude/settings.json; 15 were removed and are listed in env.ECC_DISABLED_HOOKS. pre:bash:dispatcher in particular blocks the first Bash call of every session with a GateGuard prompt. Re-running any ECC install or repair restores all of them — re-trim afterwards.

Ignore Rules

.ignore at the repo root is the real exclusion file: the Claude Code binary references .ignore, .rgignore, and .gitignore, and contains no reference to .claudeignore — the .claudeignore that used to live here was inert and has been removed. .gitignore cannot exclude tracked paths, which is why .ignore carries apps/wallet-gui/ui/target and target-contracts. .claude/settings.json adds permissions.deny as a hard backstop.

Branding

logo-assets/ is the source of truth and is never edited in place. Three commands deploy it; the two scripts take --check so drift is a failure rather than a discovery:

python scripts/deploy_brand_assets.py    # favicons + wordmarks into each app
python scripts/make_og_card.py           # the 1200x630 social card
cargo tauri icon logo-assets/print/HighRes-Square-2000_2000x2000.png `
  -o apps/wallet-gui/src-tauri/icons          # the desktop icon set

Two pack defects are corrected on copy, not propagated: site.webmanifest ships an empty name/short_name, and paste-in-head.html hardcodes root-absolute paths that only suit the explorer.

Reference checking differs per surface: apps/dashboard/ and apps/wallet-gui/ui/ fail trunk build on a missing data-trunk dir, then scripts/check_brand_refs.py walks each built dist/; apps/explorer/ has no build step, so crates/node/tests/server_tests.rs asks the running server for every path its rendered HTML names. The explorer resolves --assets against its working directory, logs that directory at startup, and warns loudly when it is absent — a ServeDir over a missing path would 404 every icon and look like a browser problem.

Caches

Dependency caches live on D: (UV_CACHE_DIR=D:\Caches\uv_cache, NPM_CONFIG_CACHE=D:\Caches\npm_cache, set at user scope and pinned again in .mcp.json). Do not create parallel cache dirs.

Nothing this project generates should land on C:, which ran down to 26.7 GB free on 2026-09-22. Moved off it that day:

What	Now	Why it is safe
CARGO_HOME (registry, bin/, advisory-db)	D:\Caches\cargo	moved whole, not re-downloaded; C:\Users\EricW\.cargo is a junction to it, so PATH is unchanged
RUSTUP_HOME (toolchains)	D:\Caches\rustup	same, junction at C:\Users\EricW\.rustup
WSL Ubuntu-24.04 disk (55.6 GB)	E:\WSL\Ubuntu-24.04\ext4.vhdx	wsl --manage Ubuntu-24.04 --move, the supported path
TEMP / TMP (user scope)	D:\Temp	rust-analyzer copies every proc-macro DLL into TEMP per session (proc-macro-srv*, up to 887 MB each, 36 were left behind)

CARGO_HOME and RUSTUP_HOME are also set at user scope. The junctions stay anyway: anything that resolved ~/.cargo before the move still finds it.