RAC 2026.6.4 - the "unlock" release

[tcballard]

Your recorded knowledge stops living only in your repo. It travels out to RAG and graph backends, grounds five different coding agents, and rides a dated release line that finally says when a build was cut instead of making a compatibility promise it never kept.

pip install --upgrade requirements-as-code

Why "unlock"

For a long stretch RAC was a closed loop: author Markdown, validate it, enforce it in CI - all inside one repository. The knowledge was real and checked, but it was trapped. This release opens every door it was sitting behind. Your corpus exports into the systems your team already runs, your grounding works on whatever agent your team already uses, and releases resume on a scheme that doesn't lie about what they are. Same knowledge you already wrote - now it goes places.

Dates, not false promises

RAC now versions on CalVer (YYYY.MM.N, ADR-076). The number says when a build was cut. It was never a SemVer compatibility claim the project actually kept and compatibility lives where it always did, on schema_version (ADR-007), and the vX.Y.Z roadmap numbers continue as internal planning scope-fences. This is a one-way cutover: every prior vX.Y.Z tag is retained, and 2026.6.4 supersedes v0.19.0 as the latest release. A fail-closed publish gate (python -m rac.release) rejects any tag that isn't a well-formed YYYY.MM.N or that lacks a changelog entry, so a malformed or undocumented release can never reach PyPI.

Your knowledge, out of the repo

The corpus is no longer something you only validate in place. Two export projections hand it to the systems that consume knowledge:

rac export . --documents   # JSONL: one Markdown-bodied record per artifact
rac export . --graph       # the real, validated relationship topology

--documents emits an ingestion-ready JSONL projection any RAG or memory backend reads directly — one record per artifact, classified, with id/type/status stamped, no chunking to babysit. --graph hands a graph backend RAC's validated relationship topology (ADR-074) rather than edges inferred from prose: the connections you already checked in CI become the connections the graph stores. You re-author nothing, and you trust no heuristic to rebuild links you already declared.

Grounding on the agent you actually use

RAC ships worked, reproducible-from-the-file-alone setups for five coding agents. Each reads the generated agent-context file natively and connects to the read-only lore MCP server:

• Sourcegraph Amp Code — reads AGENTS.md, MCP via .amp/settings.json
• Claude Code — reads CLAUDE.md, MCP via claude mcp add / .mcp.json, plus the bundled rac-artifacts skill and optional pre-edit hook
• OpenAI Codex — reads AGENTS.md, MCP via ~/.codex/config.toml
• Cursor — reads AGENTS.md, MCP via .cursor/mcp.json
• GitHub Copilot — reads .github/copilot-instructions.md, MCP via .vscode/mcp.json

rac export rac/ --agent-rules     # write the context files
rac mcp --root .                  # serve the five read-only lore tools

One corpus, every major agent. The grounding demo (examples/guide/) runs the same task twice — once unconnected, once with lore — and the connected run cites a recorded decision the unconnected run violates. Enforcement stays separate and agent-agnostic: RAC supplies context and checks after the edit (ADR-067), so the post-edit gate protects you whichever agent wrote the diff.

See where the graph is thin, and how it's used

rac coverage .     # where the knowledge graph is incomplete
rac usage          # how agents actually consult Lore over time

rac coverage lists the gaps that matter — requirements no roadmap schedules, decisions nothing applies, roadmaps scoped to no requirement — advisory, always exit 0. rac usage is local, content-free telemetry: it records that a command ran and how it fared, never a path, query, argument, or byte of your repository, and only with your opt-in. And get_related now goes multi-hop: an agent can request a bounded neighbourhood (depth > 1) — cycle-safe, deterministic, capped to the same response budget — so it sees a decision's region of the graph, not just its nearest edges.

A place to read, not just lint

The Explorer TUI was reimagined:

pip install "requirements-as-code[explorer]"
rac explorer

Three themes — lantern (dark default), parchment (warm light), and high-contrast — switch in /settings and persist, with theme-aware, accessible type tags where every signal carries text, not just colour. A portfolio list (/list) gives a dense, sortable table of the whole corpus with live fuzzy search (ctrl+f), status filtering, and sort cycling. An optional split master-detail layout pairs that list with a persistent reading pane, so scanning and reading happen together — the pattern that suits review and cross-referencing.

Also in this release (the v0.23 hardening line)

The grounding claim moved from "trust us" to legible and proven:

• Explainable retrieval — a result shows why it matched (which field, which term, which edge) on search_artifacts / get_related and via rac find --explain.
• rac doctor — one command diagnoses a corpus and emits paste-ready fixes for malformed front matter, broken or cyclic relationships, orphans, duplicate ids, and injection-style content.
• Provenance on get_artifact — git-derived author and dates plus reconstructed status history, so an agent can cite who decided and when.
• A documented trust model (SECURITY.md) — artifact content is untrusted input made authoritative by human PR review (ADR-065); the read-only server protects the store, the PR gate protects the agent.
• Provably works, stays working — a gated, deterministic grounding benchmark (rac eval --check, with a committed baseline and a hard-negative check) turns the claim into a CI regression guard.

Upgrading

pip install --upgrade requirements-as-code

uv tool upgrade requirements-as-code

Everything is additive. Existing commands, JSON contracts, and exit codes work exactly as before; the new exports, the coverage and usage commands, multi-hop get_related, and the Explorer themes are all opt-in. The one change to know is the version scheme itself — releases are dated now, and 2026.6.4 is the latest, newer than every v0.x that came before it.

---

This discussion was created from the release RAC 2026.6.4 - the "unlock" release.