Context artifact library and anti-slop infrastructure for developers, agents, and the open model society.

FullStackVibes is a free public-good context engineering commons. It exists because the post-LLM software lifecycle — agents writing code, vibecoders shipping features, frontier and small models consuming the same context — requires a shared, verified, provenance-rich knowledge layer. Without one, every team rebuilds the same prompts, the same anti-pattern coverage, the same retrieval scaffolding, in private and at varying quality. With one, that work compounds.

The corpus is structured for small-model retrieval: every published artifact is decomposed into typed context windows (GOAL, CONSTRAINT, ANTI_PATTERN, SCHEMA, …) and orthogonally tagged so a Precision Bundle call can return exactly the slices a 14B agent needs to perform like frontier on the patterns the bundle covers.

Read access is unauthenticated and free. Submission is free with an account. The full protocol is at /protocol/; the API spec is at /docs/api/. The motto — If it can be vibecoded, it must be documented — is the thesis: agent-assisted code that ships without context discipline becomes everyone's debt; this commons is the way that debt gets paid down.

Canonical Terms

Authoritative definitions with conceptual lineage. These are load-bearing across the docs, the API, and the data model — when in doubt, this is what they mean.

Context artifact

A structured prompt-engineering pattern with author, abstract, body, source links, provenance, and quality signals — the unit of value in the FSV corpus. Conceptually descended from package-manager artifacts (a stable, versioned, addressable unit of distribution) but specialized for context: every artifact is a tightly-scoped pattern an agent or developer can drop into a model's input window.

Immutable manifest version

Every artifact revision is content-hashed (sha256) and recorded forward; revisions never overwrite. The integrity guarantee that makes citations stable across time. Same posture as Sigstore + Git's content-addressable store: the version is the hash, not the label, so a reference to a manifest version cannot silently drift.

Precision bundle

A budgeted retrieval call that returns typed context windows packed for small-model consumption — orthogonally tagged (by space, window-type, window-tag, pattern-tag) and ranked by quality, with full provenance per window. The lever that lets a 14B model perform like frontier on the patterns the bundle covers. Spec at /docs/precision-bundle/.

Context window

A typed, independently-retrievable slice of an artifact body, classified by structural role: GOAL, BACKGROUND, CONSTRAINT, INSTRUCTION, SCHEMA, TOOL_SPEC, EXAMPLE, EXPECTED_OUTPUT, TEST, ANTI_PATTERN, PERSONA, or SYSTEM. Inference-decomposed at publication time; the unit a Precision Bundle actually returns. Distinct from a model's context-length window — these are author-content units, not token budgets.

Handshake

POST /api/v1/handshake — the Precision Bundle retrieval primitive. Filters: spaces, windowTypes, windowTags, patternTags, qualityMin, maxChars. No auth required for read. Named for the moment of exchange where the agent declares its constraints and FSV returns the verified context that fits. Spec at /docs/api/handshake.html.

Space

An inference-assigned cluster grouping artifacts by what they're useful for — never user-picked. A submission's Spaces are decided by reading the artifact's body and matching against a finite typed set. Conceptually a flattening of HuggingFace's Spaces idea + Reddit subforums + topic-tag inference; mechanically a many-to-many between manifests and the Space registry. The user-picked-tag pathology is exactly what this is designed to avoid.

Anti-pattern doc

A FORM tag (one of FSV's orthogonal pattern-level tags) marking an artifact whose load-bearing content is failure-mode coverage — the specific ways small models go wrong on the pattern, written as discrete ANTI_PATTERN windows. Lineage: the AntiPatterns concept (Brown, Malveau, McCormick, Mowbray, 1998) applied to LLM behavior. Anti-pattern docs are disproportionately load-bearing in the bootstrap pool.

Provenance lineage

Per-window source-model lineage: every inference annotation records the model and provider that produced it. Used to enforce the no-feedback-infected-noise invariant — never train target model M on annotations produced by M (or models in its lineage). The mechanism that keeps the verified-subset training corpus uncontaminated by model-to-model reward loops.

Slop detection

An inference job that screens submissions for low-utility, AI-generated, or template-shaped text patterns before they enter human review. Slop here means: text that performs context-engineering surface (headings, bullets, confident tone) without carrying load-bearing content. A precondition for verification — slop never enters a snapshot.

Bootstrap pool

The set of LIFECYCLE:hardening windows from already-published artifacts that FSV's own inference workers fold into their system prompts. Eat-our-own-dogfood architecture: every inference handler is itself a small-model structured-output call that benefits from the Precision Bundle pattern it produces. As the corpus grows, the bootstrap pool grows, and inference quality compounds.