Quartermaster setup

Quartermaster looks after your Claude Code setup the way a quartermaster looks after gear: it outfits new workspaces, watches what actually gets used, and retires or adds equipment based on evidence. The evidence is your own session history, mined locally by a script and reduced to counts before any model sees it.

Install

Quartermaster works across every project, so install it once at user scope:

/plugin marketplace add Eigenwise/eigenwise-toolshed
/plugin install quartermaster@eigenwise-toolshed --scope user

Reload plugins or start a new Claude Code session.

Set up a workspace

In any project, new or existing:

/quartermaster:setup

It reads the project, mines your history across all projects (which plugins you lean on, which permission denials repeat, which corrections you keep giving), interviews you briefly, and proposes a plan: Toolshed core (codebase-mapper, live-rules, sidequest where it fits), stack plugins, starter rules, and permission allowlist entries. Every item is approved individually before anything is installed or written, and the result is verified against a really-loaded session rather than assumed.

The in-the-moment loop

Most self-improvement happens without any pass at all. A SessionStart line keeps one question in front of Claude while it works: is this the thing being done for the third time that should become a skill, a codebase-map entry, a rule, or a committed measurement? When it notices one, it says so and offers to capture it right then; if nothing was missing, it stays silent. Workspaces that ran setup get the stronger version, a live rule re-injected on every prompt, and the session line steps aside for it.

Resupply an existing workspace

After a stretch of real work, or automatically when the SessionStart nudge fires:

/quartermaster:resupply

The miner streams your recent transcripts (subagents included) and emits a bounded aggregate. Every session carries what it was for: its own title, its first real prompt, any explicit /goal and whether that goal was ever met. An explicit goal is the strongest signal and also the rarest, so the title and the opening ask carry most sessions. Sessions nobody typed into twice are flagged, because a hook that spawns fifty review sessions would otherwise report its purpose back to you as yours.

Alongside that: the areas of the tree your work landed in, repeated commands, per-plugin and per-MCP attribution, documentation lookups, per-session cost, and friction (denials with their targets, interrupts, corrections with themes).

The skill looks for what is missing against that purpose, in value order: something you have no way to measure, work you keep doing by hand, knowledge you keep re-deriving, and only then the setup pushing back. That is the order it looks in. What it leads with is whatever your history actually attests, so a missing measurement comes first when the evidence carries it (a goal restated and never met, a check improvised dozens of times, one question answered two different ways) and gets labelled as inference when it is a guess, which puts it below the cheap fixes it is sure about. At most seven findings, each proposed one at a time with evidence and the exact change. Say no and that recommendation never comes back; say yes and the next pass reports whether it helped.

Friction is last on purpose. Fixing what went wrong returns you to the speed you already expected; adding a capability you never had moves that baseline, and the best capabilities leave no friction trace at all. A measurement you are missing throws no errors, gets nothing denied, and happens exactly once, so anything that watches for repeated pain will never find it.

Findings route to the cheapest durable fix: a measurement built as a committed skill, a plugin install from the catalog (2,500+ plugins searchable locally, install counts included), a workflow skill, a codebase-map or CLAUDE.md entry, a live rule, a permissions.allow entry, or a plugin disable.

How the loop closes

A SessionEnd hook tallies each session locally: no LLM, no network, one streamed pass. Applied recommendations record what they target, and verify compares that signal per session before and after. For a capability no counter tracks, the check is whether the new skill or plugin shows up in attribution at all. A SessionStart nudge (one line, 24-hour cooldown) fires only when enough unreviewed sessions or friction pile up, and has Claude run a pass at the next natural pause without being asked. The pass only reads and proposes; nothing is ever changed without your per-item approval.

What it stores

Tallies and decisions live under ~/.claude/quartermaster-state/: per-session counters and a decision ledger with fingerprints. Raw transcripts are never loaded into model context; the resupply skill is explicitly forbidden from opening them. Everything taken from a transcript is clipped: session titles to 120 characters, opening asks to 240, goal conditions and evidence quotes to 300. Paths are reduced to the two directory segments nearest the file, and scratch directories are dropped entirely.