token-master

# /token-master This skill turns on TokenMaster's token-efficient routing for the repository the user is working in. It is the on-switch: it builds the code graph and installs the routing engine. ## What it does 1. Builds a structural **code graph** of the repo with `graphify` (no LLM, fast) — the default, cheapest backend. 2. Stores it at `.token-master/graph.json` and adds `.token-master/` to the repo's `.gitignore`. 3. Best-effort installs a second **`codegraph`** backend (AST-resolved call edges) for precision escalation, and indexes the repo with it. This needs `node` + `npm`; if they are missing the plugin still works graphify-only (the agent is written without the codegraph server). 4. Installs/refreshes the **`token-master` routing agent** (user scope) and its graph MCP servers. The two backends are not interchangeable: graphify is the cheap default; codegraph is a targeted escalation for precision-critical or sparse-call-graph cases (see "After setup"). If graphify's call graph is **sparse** for the repo (common in JavaScript/TypeScript), setup prints a warning — that is the signal codegraph will do the call-resolution work on that repo. ## How to run it The setup script lives in this skill's own directory. Locate it reliably, then run it: 1. Run `/skills info token-master` and read the skill's directory path from the output. 2. Run the script, passing that directory and the repo root (normally the current working directory): ``` python "<skill-dir>/setup.py" "<repo-root>" ``` Then relay the script's summary output to the user verbatim. - If the script reports that `graphify` (or `uv`) is not installed, tell the user to install the missing tool — `uv tool install graphify` (see https://github.com/safishamsi/graphify) or `uv` itself (https://docs.astral.sh/uv/) — then run `/token-master` again. - The routing agent is loaded at CLI startup, so the user must **restart their CLI** before routing takes effect (unless it is already active). The setup script prints the exact host-specific restart command in its summary — relay that line as-is rather than guessing the host. ## After setup Once the agent is active, the user just asks structural questions normally ("who calls X?", "what breaks if I change Y?") and they are answered from the graph. To confirm it is working, ask a known structural question and check that the answer comes from a `graphify-nav` MCP tool call. The agent prefers `graphify-nav` (cheap) and escalates to `codegraph` (precise, ~2x tokens, AST- resolved) only when it matters: graphify returned nothing for a symbol that exists, name collisions where text matching over-reaches, or when exact `file:line` call sites are load-bearing. On a repo where setup warned the call graph is sparse, expect that escalation to fire more often. For **cross-session** continuity, the agent uses the host CLI's **native** session memory (no extra server): keyword recall over past turns and the host's `--resume`/`--continue`. This is lexical and opt-in, not automatic semantic memory — resume replays the full prior transcript (re-billed as input), so it is only worth it after a large investigation, not for short sessions. Re-run `/token-master` whenever the code has changed enough that the graph is stale.