🪨 benitogf.github.io/detritus — what detritus does, at a glance: the knowledge base, code intelligence, candyland, and the delivery pipeline.
Detritus is an MCP server for AI coding assistants. It exposes a curated
knowledge base of engineering doctrine (kb_* tools) and live codebase
exploration (code_*) — and layers
on slash commands that share one delivery pipeline: plan, build,
review-with-rework, deliver. Those commands run either in your session (/plan,
/smith, /forge, /janitor, /gh-*, /todo, …) or hand off to the
out-of-process candyland sidecar for multi-agent, dashboard-observable
delivery (/candyland, /quest). No flow merges on
its own — the human merge stays the gate.
Ask your AI coding tool to install Detritus from this repository:
please install detritus mcp https://github.com/benitogf/detritus
Or, if your tool supports Markdown links:
please install detritus mcp [benitogf/detritus](https://github.com/benitogf/detritus)
That is the human install path. You should not need to run Detritus as a CLI
tool yourself. Manual install details for an assistant or maintainer live in
INSTALL.md.
Use Detritus by command name. This table is generated from docs/flows/ (run
detritus --readme); the folder a doc lives in is its category. Internal
building blocks (docs/core/), agent roles (docs/roles/), and reference
material (docs/ooo/, docs/patterns/) are kb_get-only and intentionally
absent — they are not commands.
| Command |
Use it for |
/plan |
Analyze requirements/feedback, create implementation plan, provide insights and questions |
/vibe |
Executive intake for autonomous delivery. |
| Command |
Use it for |
/candyland |
Launch ONE candyland run in the sidecar — a tech-lead agent partitions the work, coders build it concurrently in worktrees, a reviewer loops fix→re-review until clean, then the run delivers. |
/forge |
Drive a settled plan to a PR with a parallel tech-lead + coders implementation loop, in-process. |
/janitor |
Create a recurring proactive code-maintenance worker that audits, safely fixes, verifies, self-reviews, and routes delivery through /gh. |
/quest |
The one flexible, persistent iterative loop in the candyland sidecar — a quest-lead ticks discover→triage→launch serial child runs against an objective + scope lens, with two delivery modes (converge-until-clean → one PR per impacted repo; per-finding → one PR per accepted finding). |
/smith |
Recurring loop that takes a feature from /plan all the way to an open PR — the in-session single-agent instantiation of the universal pipeline. |
| Command |
Use it for |
/babysit |
Watch a single PR on an interval — fix any review feedback via /gh-feedback-work, and merge once an approval lands on the latest commit. |
/caregiver |
Watch a single PR from the reviewer's seat — re-review on each push or discussion by dispatching the one-shot /gh-pr, re-posting a commit-pinned verdict only when it materially changes, until the PR is merged or closed. |
/gh |
Router for GitHub issue/PR workflows — reads conversation context and dispatches to gh-issue-create, gh-issue-work, gh-feedback-work, gh-self-review, gh-pr, and the PR-watch loops babysit and caregiver. |
/gh-feedback-work |
Address open review feedback on a PR, push fixes, and update the PR body in place — never posts issue/PR comments. |
/gh-issue-create |
Draft a GitHub issue from the current conversation, confirm with the user unless the post was already directed, post it with the Claude Code attribution footer, then offer next steps (/gh-issue-work, refine, or leave). |
/gh-issue-work |
Take a GitHub issue end-to-end — branch, fix, test, commit, push, self-review the diff, confirm with the user unless opening the PR was already directed, then open PR with a product-focused summary and the Claude Code attribution footer. |
/gh-pr |
Hard-review a GitHub PR under truthseeker rigor — verify the PR's own claims, hunt for fragility, demand evidence before flagging OR approving, and post an APPROVE or REQUEST_CHANGES review via gh api. |
/gh-pr-safe |
Hard-review a posted GitHub PR with the same truthseeker rigor as /gh-pr and auto-post an APPROVE or REQUEST_CHANGES review, but never mutate the working tree of the clone it runs from. |
/gh-self-review |
Deep self-audit of pending local changes (committed + uncommitted) under truthseeker rigor before commit/push/PR. |
| Command |
Use it for |
/checkpoint |
Durable per-topic session context — serialize load-bearing state to .checkpoint/.md at a boundary so a /clear is lossless and a later session (hours or weeks on) resumes the topic by name. |
/code |
Explicit entry point for code exploration backed by the zero-setup code_* tools (code_map, code_outline, code_graph) plus native Grep. |
/settings |
Show or change global detritus settings — the model and effort the generated reviewer and coder agents run on. |
/todo |
Cross-session todo management — router, conventions, and ALL everyday item operations (view, add, done, edit, defer, clear, file) in one doc. |
| Command |
Use it for |
/flaky-check |
Flaky test detection - manually-driven batches of -race runs that PRESERVE full failure logs, reproduce faithfully, and account for cross-test/ordering flakes |
/testing |
Testing index - entry point for all testing workflows |
/testing-go-backend-async |
Async testing - deterministic async event synchronization patterns |
/testing-go-backend-e2e |
E2E testing - consolidated tests covering full state lifecycles |
/testing-go-backend-mock |
Mock testing - minimal mocking at boundaries, simple state toggles |
| Command |
Use it for |
/coding-style |
Self-documenting code - naming, extraction, readability rules for AI |
/go-modern |
Modern Go patterns - auto-fix with gopls modernize after Go edits |
/line-of-sight |
Line-of-sight code style - flat code, early returns, separate error handling from business logic |
/truthseeker |
Foundational principles - ALWAYS ACTIVE, do not invoke |
| Command |
Use it for |
/absorb |
Close the learning loop on a shipped PR - resolve its candyland unit, fix outstanding blockers, and distill the review outcome into the KB |
/cleanup-extra-rules |
Remove all detritus-generated rule files, hook scripts, and the matching settings.json hook entries. |
/detritus-release |
Cut a release of detritus and/or candyland (bump → annotated tag → CI build), then update the local installation. |
/detritus-update |
Update detritus to the latest released version by running detritus --update. |
/grow |
Learn from conversation corrections - distill manual fixes into KB updates |
/learn |
Learn from candyland telemetry - mine failure signatures across runs/quests into audited KB updates |
/optimize |
Re-index and optimize KB docs for agent retrieval efficiency |
/setup-extra-rules |
Generate personalized Claude Code rule files and hook scripts based on the user's actual environment. |
/setup-superpowers |
Apply baseline Claude Code settings (deny list, status line, effort/thinking, autoMode environment). |
| Command |
Use it for |
/pdf |
Generate a PDF to whatever format and style the user describes — general-purpose, multi-page, diagrams when useful. |
/pdf-management |
Turn the issue under discussion into a single-page decision PDF for a non-technical stakeholder — premise-only, plain real-world terms, no jargon, no component names or technical mechanisms. |
/pdf-tech |
Turn the issue under discussion into a single-page decision PDF at a bare-minimum, high-level technical register — name the component and mechanism conceptually, no code or deep internals. |