shrimpshack

Autonomous codebase research — discovers tunable parameters, runs experiments overnight in worktrees, delivers findings with KEEP/CHANGE/INVESTIGATE recommendations

7 skills · 7 commands · 10 agents

Modes as first-class Claude Code infrastructure — named, durable working stances that recompose the plugin catalog per mode via a cascading config system. Solves harness-bloat at scale.

3 skills · 13 commands · 1 agent

Run Claude on any model the local Superagent Gateway serves, across five surfaces. A headless lens returns a different-vendor answer as one JSON object on stdout, with failure classes distinguishable by exit code. An interactive launch materializes an attachable session on a named alias and prints a resume handle that never carries the token. A background agent runs unattended under a fixed repo-bounded permission ceiling and is judged against a contract written up front — deliverables measured against a pre-job baseline, an optional verify command the supervisor runs itself — never against the model's account of its own work. A team runs several named members at once, each in its own worktree on its own alias against its own contract, dispatched in bounded rounds from one record on disk that carries every member's probed state, its cause of failure and the model that actually answered it. The plugin owns its own gateway control layer (liveness by probing the model-list endpoint, idempotent locked start, append-only log, runtime install-dir resolution) and setup rewrites the existing gw wrapper to delegate to it and to read its token from the Keychain instead of carrying a hardcoded one.

5 skills · 6 commands

A config-driven bridge between one codebase's CSS design tokens and one Paper file. Sync custom-property tokens code → Paper (idempotent reconcile) and Paper → CSS (round-trip stable), and harvest rendered components code → Paper. Pointed at any codebase via --repo; theme scopes (data-attribute, media-query, class, file) and paths come from a codebase-local config, not hardcoded paths

5 skills · 5 commands

A registry-driven lint router: it picks which linters to run based on WHO the work is for — matched by git origin or marker files — then runs that profile's linters on just the files you changed. Profiles live in a routes.json registry (ordered, first match wins); add, configure, remove, or discover linters through the bundled skills — eslint, ruff, clippy, shellcheck, prettier, or any tool you register, across languages. A profile can layer a linter as an overlay on a repo's own config or run one standalone, and a profile that matches nothing runs nothing — so it never imposes a lint on code that shouldn't have one. Personal and local by design: it writes only a gitignored overlay or nothing at all, never a repo's committed config, deps, or CI. A SessionStart hook prepares the right profile automatically; invoke the skill to run the check before committing or opening a PR.

6 skills · 1 agent

Fork the current thread of work into its own place — takes the topic/plan/idea just discussed and moves it into a fresh git worktree with a new, already-briefed Claude session, so the current session stays focused and the new one picks up the context. Command-invoked only via /start-session, /start-split, /start-workspace, or the /start alias: branches a worktree (current HEAD or develop), writes a handoff doc linking back to the originating session's transcript, carries over the whole docs/ tree plus local config dotfiles (.env, .envrc, .tool-versions, .nvmrc), and launches Claude with the brief already attached as its opening prompt — in a new tab, in a split beside the pane you're in, or in a brand-new two-pane workspace with the handoff alongside. The mechanical work runs in a background agent so it doesn't consume the originating session's context. The receiving session is told to treat the handoff as directional author-intent — enough to orient and start, with the code and tests as the source of truth — both in the brief and via a self-declaring banner on the doc itself. Launches under herdr, cmux or Ghostty, auto-detected; the Ghostty path is AppleScript-driven, so it needs macOS Automation permission and cannot auto-answer Claude's MCP trust prompt.

1 skill · 4 commands · 1 agent

Ties each Claude session to its Linear issue through the herdr board. Agents write Linear with Linear's own MCP tools; the plugin's hook tells the board about every write, a SessionStart hook names the session's issue, column and marks, and start and layout create bound worktrees. A rules skill carries the write rules. Needs a herdr-board release with `board linear report`; run `/work setup` to install and connect everything, or see docs/cutover.md to move from 0.5.

4 skills · 1 command · 1 agent

A QMD-backed memory + document store for Claude Code. Keeps the auto-loaded memory index small and complete (a budgeted pointer index instead of one ever-growing file that silently truncates), surfaces the most relevant memory bodies on demand via fast local search, and captures the documents agentic work generates — brainstorms, handoffs, solutions — into a searchable store before worktrees are cleaned up. Hooks wire automatically when enabled; the invasive one-time steps (migrating an existing MEMORY.md, patching the Memory Protocol) are opt-in via /reflect-setup. Degrades gracefully when qmd is missing, slow, or wedged: the index still loads and bodies are read by file pointer, seeded recall self-suppresses after repeated failures (a cross-session cooldown) so a stalled qmd never taxes every prompt, and only search-based recall is dormant.

1 skill · 4 commands

Find and destroy zombie processes and repo slop from Claude Code sessions — kills by ownership and liveness, never age

4 commands

Cut comment bloat out of a branch without losing the comments that matter — four keep-rules, each needing a failure that nothing reports AND a tempting edit that would cause it, because a visible bug is a missing test rather than a comment; specs default to zero on the principle that a red test is not a silent failure. Reachable as /comment-cut:cut, and a PreToolUse gate stops a commit or a PR once when the change adds bloat on the lines it touches — opt-in per repo, fail-open on every path, and the retry always proceeds

1 skill · 1 command · 2 agents

Turn structured data an agent already holds into something a person can actually read in a chat transcript. Picks the form the data earns - a Markdown table by default, a line chart when the shape is worth a picture - validates the values, renders, checks its own output, and reports what it left out. It never invents a missing value and never claims more than the data supports. Saved reports re-run the exact calls a report was built from, check the data still has the same shape, and show what changed since the last run. It does not fetch data and does not interpret what the numbers mean.

2 skills · 1 command

Review a change too large for one reviewer — many files, several subsystems, or work built incrementally. A deterministic pre-pass measures the diff, a rubric the agent applies slices it by invariant and staffs the seams between slices, and a native Workflow fans out lens reviewers per slice, proves load-bearing assertions by mutation on an isolated worktree, and loops to an empty round with a three-rounds-same-class escalation. Companion to /ce-code-review and /code-review for large, seam-focused reviews — it reviews, it does not auto-fix.

1 skill · 2 commands

Claude Code plugin marketplace by @shawnroos

1 skill