ACRI — Agent-Compiler Repair Interface

This repository is the home of the ACRI specification — the standalone, implementation-neutral "spec house", separate from any implementing project.

What is ACRI

ACRI defines the interface between a compiler/toolchain and the AI agents that consume its diagnostics and act on its repair suggestions. It is a thin, deliberate contract: structured diagnostic objects, confidence-graded repair suggestions governed by a safety axiom (auto-apply only high-confidence, non-hint actions), a minimal CLI protocol with an exit-code matrix, and a methodology for contamination-safe repair-loop benchmarks. The reference implementation is the Lom language and toolchain; conformance requires no adoption of Lom or of any particular language, runtime, or vendor.

Three layers

Layer	Content
L0 — Object model (normative)	The Diagnostic and RepairSuggestion objects (§1.1/§1.2) and the three safety axioms (§1.3): AXIOM 1 auto-apply gate, AXIOM 2 cognitive-suggestion rule, AXIOM 3 plan/apply separation.
L1 — CLI protocol (normative)	The command set check --json / fix --plan / fix --apply / codes (plus OPTIONAL explain), the check --json envelope, the exit-code matrix, and the conformance levels (L0-conformant / L1-conformant / L2-documented).
L2 — Repair-benchmark methodology (informative)	Six clauses for a contamination-safe repair-loop benchmark: dual verbatim capture, synthesis declaration, multi-chain reference validation, black-box scoring, machine gate on prompts, scope honesty. Reference instance: RepairLoopBench.

Versioning

• 0.x drafts may revise freely between cuts; 1.0 will be the first stable cut.
• Additive field/enum changes are minor; removal or semantic change of a REQUIRED field, or an axiom change, is major.
• A numbered cut is immutable: once a version is cut, its content never changes in place — any textual change, down to a typo, requires a new version number (errata ship as a new draft/cut; see spec/).
• Every cut is anchored by two carriers: a git tag for that version in this repository, and an entry in CHANGELOG.md. A version lacking either is not cut.

Governance

• Current steward: the Lom project & wytyKen.
• The specification is implementation-neutral (§0 of the spec); stewardship of the Lom language and stewardship of the ACRI specification are separable, and a future transfer to a neutral foundation is explicitly not precluded (precedents: MCP and AGENTS.md → Linux Foundation-hosted Agentic AI Foundation, 2025-12).

Repository layout

spec/                         # The specification itself
  acri-v0.1.md                # v0.1 — numbered cut, immutable (byte-identical to the main repo's cut)
  acri-v0.2-draft.md          # v0.2 errata draft (undecided; three [v0.2-draft]-tagged revisions)
conformance/                  # Conformance suite (runner delivered 2026-10-28)
  README.md                   # Scope, three-state reporting, probing method, batch plan
  suite-design.md             # Probe inventory (C-01…C-14)
  suite/                      # Executable runner (Python, standard library only)
  adapters/                   # Adapter manifests (lom.json = first/reference)
  probes/                     # Implementation-owned probe files (probes/lom/ = first set)
implementations/              # Implementation registrations
  lom.md                      # Lom (reference implementation) — L0/L1/L2 conformant since v1.15.0
mappings/                     # Adapter-source mappings (informative; never modify spec/)
  sarif.md                    # SARIF v2.1.0 → ACRI L0/L1 — adapter-author guide, field-checked against the OASIS spec
adopt/                        # Adoption aids for toolchain repos
  AGENTS.md.snippet           # Paste-ready AGENTS.md block (commands, exit codes, safety rules)
  README.md                   # What to paste where
CHANGELOG.md                  # Version carriers (one entry per cut)
LICENSE-SPEC                  # CC-BY-4.0 (spec text)
LICENSE-CODE                  # Apache-2.0 (code)

How to conform

1. Read the specification — spec/acri-v0.1.md is the operative version. Map your toolchain's diagnostics, suggestions, CLI, and exit codes onto L0/L1; document any enum-name or field-shape mappings. If your tool already emits SARIF v2.1.0 (GCC 13+, Clang, most static analyzers), start from the verified mapping in mappings/sarif.md; to teach agents in your repo how to consume your interface, paste adopt/AGENTS.md.snippet into your AGENTS.md.
2. Run the conformance suite — write an adapter manifest for your CLI (see conformance/adapters/lom.json for the shape) plus your own probe files, then run python conformance/suite/runner.py <manifest.json> (delivered 2026-10-28; see conformance/README.md).
3. Register: open a PR adding implementations/<your-tool>.md (use implementations/lom.md as the template) stating your conformance levels, registered deviations, and evidence pointers.

License

Dual-licensed by artifact type:

• Specification text (spec/, plus the descriptive documents in this repository): Creative Commons Attribution 4.0 International (CC-BY-4.0) — see LICENSE-SPEC.
• Code (the conformance runner and any scripts, when they land): Apache License 2.0 — see LICENSE-CODE.