Okf Project Knowledge Base

OKF Project Knowledge Base

Durable project knowledge - decisions with rationale, findings with evidence, rules with what they protect against - lives in the repository as an Open Knowledge Format (OKF) v0.2 bundle: plain Markdown concepts with YAML frontmatter, readable by humans, parseable by any agent, diffable in git. This skill is the discipline for creating, writing, reading, and maintaining such bundles. It standardizes the mechanics; without it every model invents its own incompatible structure (measured in evals/runs.md).

Finding the bundle

A bundle root is any directory whose index.md declares okf_version in its frontmatter. That marker, not the path, is authoritative: to discover bundles, search the repository's index.md files for okf_version. The canonical location for a NEW bundle is docs/knowledge/ at the scope it serves - the repo root for project knowledge, <subproject>/docs/knowledge/ for a subproject - but respect existing bundles wherever they live (knowledge/ at the root is a common alternative).

The local law

Before reading or writing anything else, read the bundle's root index.md. Its preamble states what belongs in this bundle and its local conventions (type vocabulary, directory layout). The preamble may tighten the rules below; it never loosens them.

Three fences

Route by what the information is about, never by how you discovered it. A finding about a tool this project depends on is project knowledge even when a session-local tool surfaced it; the admission test is whether it is true and useful in a fresh clone, to a reader who never used this session's tools.

1. Durability fence (knowledge vs state). State - counts, versions, hashes, in-flight work, "where we left off" - never enters the bundle; it lives in its authoritative source (ledger, git, the tracker for work items). A number belongs in a concept only when the number IS the finding. Eventual mutability is NOT state: the test is what invalidates it - the next commit or session means state; an upstream release or vendor decision means knowledge, captured with stale_after to absorb the decay.
2. Publicity fence. Write every concept as if the repository goes public tomorrow: no secrets, no personal or sensitive data, no host-local context (absolute paths, machine names, internal hostnames). This holds even in private repositories - bundles travel with repos.
3. Instructions fence (knowledge vs AGENTS.md). Agent instruction files hold standing orders obeyed every session; the bundle holds facts consulted when relevant. When a decision produces both, the instruction file gets one line stating the law with a link to the concept; the rationale, history, and evidence live in the concept. Never grow an instruction file with a rationale paragraph.

Concept format

Every concept is one Markdown file: YAML frontmatter, then a body. Minimal conformance is a parseable frontmatter block with a non-empty type; everything else below strengthens trust and should be present.

---
type: Decision
title: OAuth2 flow standardized on PKCE
description: One sentence a reader or index can reuse verbatim.
tags: [auth, security]
status: stable
generated: { by: claude-code/fable-5, at: "2026-09-07T14:00:00Z" }
sources:
  - id: rfc
    resource: https://datatracker.ietf.org/doc/html/rfc7636
    title: RFC 7636 (PKCE)
---

# Decision

Chosen because public clients cannot hold a secret.[^rfc]

[^rfc]: RFC 7636 (PKCE)

Field reference (all optional except type):

Attribute specific claims with Markdown footnotes whose labels are sources[].id values - keyed, never positional, so reordering the list cannot misattribute. Link concepts to each other with normal Markdown links, bundle-relative (starting with /) preferred. All timestamps are ISO 8601 with explicit UTC offset.

Actors

Identity is derived from the environment, never hardcoded:

• Humans: human:<id> where <id> is the git user.name configured in the repository holding the bundle.
• Agents: <harness>/<model> as the session knows itself, e.g. claude-code/fable-5, codex-cli/gpt-5.6-sol.
• Automated processes: process:<id>.

generated is written on every create or content update. verified with a human: actor is written ONLY when the human explicitly instructs it or made the edit themselves - an agent never volunteers human verification. An unverified concept is honest, not deficient; consumers derive trust tiers (unverified, machine-confirmed, human-reviewed) from what is actually there.

Invariants

1. Search before write. Query the index and existing concepts before authoring. Prefer updating an existing concept over creating a near duplicate.
2. A bundle is never created empty. Create it at the moment of first capture, with that real concept as its first entry - never scaffolding, never a placeholder. The first concept is the convention every later writer imitates, so give it full frontmatter and real sources.
3. Every claim is traceable. Record sources for whatever the concept derives from; when no durable link exists, write an honest scope descriptor rather than dropping the source or inventing a URL.
4. Bookkeeping travels with the change. Adding, moving, or removing a concept updates the nearest index.md (a bullet with title, link, and the concept's description) and adds a dated entry to the bundle's log.md (newest date first, ## YYYY-MM-DD headings), in the same change.
5. Deprecate, never delete. A concept that stops being true gets status: deprecated and, when replaced, a link to its successor. History and inbound links survive.
6. Progressive disclosure on read. Enter through index.md and open only the concepts the task needs; never bulk-dump a bundle into context.
7. Verify before claiming conformance. After writing, confirm what the format requires: the frontmatter parses, type is non-empty, links you added resolve, and the index matches the directory. Use whatever tools the session has; the checks, not the commands, are the contract.

Workflows

Cold start (a repo with knowledge to keep and no bundle): create docs/knowledge/ with a root index.md whose frontmatter declares okf_version: "0.2" and whose preamble states, in a few sentences, what this bundle holds and what it refuses (the three fences, localized); a log.md; and the first real concept (invariant 2). Then add one line to the repo's agent instruction file naming the location, telling agents to load this skill before reading or writing the bundle, and asking for an end-of-task capture review.

Capture (mid-work or on request): decide with the fences whether it is knowledge; search first (invariant 1); write or update the concept with full frontmatter; update index and log (invariant 4); verify (invariant 7). If the knowledge also implies a standing order, apply the instructions fence: one line in the instruction file, linking here.

Recall ("why did we choose X", "have we established Y"): find the bundle by its marker, read the index, open only matching concepts. Treat status, stale_after, and the trust tier as part of the answer - a deprecated or stale concept is reported as such, not as current truth.

End-of-task review (after substantial work, when the repo's instruction file asks for it): scan the work for decisions taken, findings established, or rules adopted; capture what passes the fences; say plainly when nothing does.