| name | agent-tooling |
| description | Pick the right code-intelligence tool for a task in this repo — repowise (MCP), the pyrefly LSP tool, or Basic Memory (MCP). Use when deciding how to answer "who calls / where is / find the symbol", "what's risky to change", "what did we decide and why", "what's the architecture", or "what do my notes say" — instead of blind grep-and-read. NOT for turning the tools ON in an agent (install / register MCP / LSP → the `agent-setup` skill); NOT for reproducing the underlying stack (docs site, contributing/agent-tooling.md). |
| metadata | {"project":"saitenka"} |
agent-tooling
Three code-intelligence tools back this repo. Reach for them before grep-and-read — they are exact and
far cheaper. All optional; the repo works without them.
Which tool
- pyrefly
LSP tool — exact symbol graph. findReferences / incomingCalls / outgoingCalls /
documentSymbol / goToDefinition. The canonical first move before touching a symbol (map callers,
not a maybe-incomplete grep). See AGENTS.md § Refactoring. (Backend: pyrefly lsp; basedpyright stays
the type-gate, not the nav.)
- repowise (MCP) — whole-repo understanding from the built index.
get_answer (how/where/why,
content-grounded), search_codebase (semantic/symbol/path), get_context (a file's map), get_risk
before editing a hotspot, get_why for decision rationale, get_overview once in an unfamiliar area.
Prefer these over re-reading files. Freshness: repowise update after big changes; it never edits
AGENTS.md.
- Basic Memory (MCP) — the maintainer's Markdown knowledge base (Obsidian vault): design notes,
decisions, tooling procedures (e.g.
notes/tooling/repowise-local-indexing). search_notes /
read_note / recent_activity. Personal + per-user (each dev brings their own vault); complements the
file-based auto-memory. write_note is mutating.
Quick routing
| Question | Tool |
|---|
"Who calls build_note?" / "where is X defined?" | LSP (findReferences / goToDefinition) |
| "What's the riskiest file to touch here?" | repowise get_risk / get_overview |
| "Why is the tokenizer matching shaped this way?" | repowise get_why |
| "What did we decide about dict tabs / the port?" | repowise get_why or Basic Memory search_notes |
| "How do I set up / reproduce this stack?" | the agent-setup skill · docs contributing/agent-tooling.md |
Repowise readiness and fallback
Treat every response as usable only when _meta.indexed_commit equals the worktree HEAD. For
get_answer, also require a non-empty answer and no degraded marker. If either check fails, do not
retry synthesis in a loop or let it steer selection: use bounded get_context / get_symbol, then
confirm correctness-critical claims against live code. Index freshness on disk is not sufficient to
prove a long-lived MCP process has reloaded it, so a shell readiness helper cannot replace these response
checks. Raising REPOWISE_ANSWER_TIMEOUT_S trades more latency for synthesis headroom; it is a per-user
choice, not a repository default.
Enabling these
Not wired yet? → the agent-setup skill (renders MCP config, registers the LSP, symlinks skills).
