detritus

One binary. Three seams of value.

Detritus is a single, embeddable Go server that speaks the Model Context Protocol. It carries a curated knowledge base of engineering doctrine, a set of zero-setup tools for reading a codebase, and a layer of slash commands that move real work through one pipeline — plan → build → review → deliver — either inside your session or handed off to a separate multi-agent sidecar.

The genuinely unusual bet: the docs are the product, and the Go is the delivery mechanism. Everything a human touches installs itself into whatever coding tool you already use — and no flow ever merges on its own. The human merge stays the gate.

// the kb_* seam

A knowledge base tuned for a machine reader

The corpus is markdown, split one chunk per ## section, enriched at build time with TF-IDF keyword triggers, and frozen into a deterministic data.gob — sorted and tie-broken so the same docs compile to the same bytes on any machine. At runtime it's a title-weighted bleve index reranked with MMR.

Every response is shaped for the consumer that reads it: an AI, not a person. Search punts identifier-shaped queries straight to code_graph rather than answering them badly. That restraint is the design.

// the code_* seam · zero-setup

Live code intelligence, no index to build

No pack, no config, no stored index — the tools read the repo live and auto-discover the project root (go.mod > .git > language marker). .gitignore honoured, binaries sniffed, and a per-file, mtime-keyed, atomically-written parse cache means the first pass pays and the rest is fast.

// learned long-term memory

A lesson is a pull request

The learned-memory subsystem was deleted and redesigned into nothing — and that's the clever part. There is no local store, no agent-write API, no LRU, no dedup pass, no recall-miss counter. A lesson is just a markdown doc under docs/lessons/, shipped by a reviewed PR, distributed by detritus --update, and retrieved by the same kb_search as everything else.

// candyland · out-of-process delivery

When one session isn't enough: candyland

For work too big for a single agent, detritus hands off to candyland — a long-running REST sidecar it launches and correctly detaches (new session on Unix, new process group on Windows) so it outlives the session that started it. A tech-lead partitions the work, coders build concurrently in git worktrees, a reviewer loops fix→re-review until clean, then it delivers. (That engine lives in its own repo — what ships here is the launcher and the classifier.)

// docs/flows · one source, many surfaces

Doctrine that installs itself

A doc is a command if and only if it lives under docs/flows/. From that single rule, one binary generates skills, prompt files, sub-agents, router instructions, the README table and marketplace shims — then registers itself as an MCP server across Claude Code, VS Code/Copilot, Cursor, Windsurf, Codex and Verdent. The generated shim is a stable kb_get indirection, so editing a doc ships new behaviour with nothing to regenerate client-side. Drift tests fail the build if the surface and the docs ever disagree.

• truthseeker — always active. Prove before acting; reject fragility; make the invisible visible; push back on the user when the evidence demands it.
• review-rigor — forbidden-action tables with IDs, mandatory checks R1–R11, closed-enum verdicts (approve|changes), and hedge-words banned outright. No "approve with reservations."
• the universal pipeline — every flow runs plan → build → review-with-rework → deliver, and no flow may skip a stage. /babysit merges only when an approval's commit matches HEAD — a push invalidates it.
• todoguard — a PreToolUse hook that trusts Claude's agent_id, deliberately distrusts the spoofable agent_type, and fails open. A small, honest threat model.