Harness landscape — format snapshot

Measured 2026-08-14 from vendor documentation. This is a fast-moving surface; every claim below is dated and sourced. Authoritative for: what each harness reads and which frontmatter fields it accepts. Not authoritative for: what rungs does about it — that is ADR-0001.

The four source repos in repos/ were built between 2026-03 and 2026-08 and do not reflect the current landscape. Two things changed underneath them, and both shrink the problem they were solving:

1. Agent Skills became an open standard (agentskills.io, from Anthropic, adopted by 45+ clients). Procedures are now portable by default.
2. Every major harness gained path-scoped rules — including Claude Code, via .claude/rules/ with paths: frontmatter. rift-forge’s 1513-line CLAUDE.md now has a native fix.

1. The four primitives

Agentic configuration is not one thing. It is four, with very different portability:

2. P3 — Agent Skills (the standardized one)

Format: a folder containing SKILL.md (YAML frontmatter + markdown body), optionally with scripts/, references/, assets/. Loaded by progressive disclosure — name + description at startup (~30–50 tokens per skill), full body on activation, referenced files on demand.

Spec frontmatter fields — the portable set (6):

Claude Code extensions — non-spec: disable-model-invocation, user-invocable, disallowed-tools, context, arguments, and others. Claude Code accepts all of them. Outside Claude Code, claude.ai skill uploads / the Skills API / package_skill.py accept only the six spec fields and fail with a hard error on anything else:

Unexpected key(s) in SKILL.md frontmatter: argument-hint. Allowed properties are: allowed-tools, compatibility, description, license, metadata, name

Cursor documents its own optional paths (globs limiting availability) and disable-model-invocation — overlapping but not identical to Claude Code’s extension set.

Discovery directories — this is the part that still differs:

Consequences:

• .claude/skills/ has the widest coverage today (3/3) but is vendor-named and Cursor labels its support legacy.
• .agents/skills/ is the neutral direction (2/3) and Claude Code does not read it.
• There is no directory all three read that is not vendor-named.
• Claude Code also loads nested .claude/skills/ below the working directory, on demand — a monorepo package can carry its own.

3. P1 — Always-on context

The Claude bridge, from the vendor’s own docs: a CLAUDE.md whose first line is @AGENTS.md (an import, expanded at launch), with Claude-specific content appended below. A symlink also works, but on Windows a symlink needs Administrator or Developer Mode, so the import is the portable form.

Vendor size guidance: “target under 200 lines per CLAUDE.md file. Longer files consume more context and reduce adherence.” This is direct vendor confirmation of the core-size-budget pattern that repos/rift-forge.md derives from a 1513-line counter-example.

@path imports load at launch — they organize, they do not reduce context.

4. P2 — Path-scoped context (the fragmented one)

Same idea in every harness, three incompatible dialects:

Notes that matter for rendering:

• Cursor requires the .mdc extension; a plain .md in .cursor/rules/ is ignored.
• Claude Code path-scoped rules trigger when Claude reads a matching file, not on every tool use, and are not re-injected after /compact — they reload on the next matching read.
• Claude Code’s paths brace expansion is budgeted (1,000 expanded patterns / 4 MiB per rule).
• Cursor has four activation modes (always / auto-attached by glob / agent-requested by description / manual); Claude Code’s rules have two (always / path-matched) and push intent-based activation into skills instead.

5. P4 — Enforcement hooks

Claude Code: .claude/settings.json with PreToolUse (and other lifecycle events) running shell commands, plus permissions.deny for hard blocks. The vendor states the distinction rungs’ enforcement-declaration pattern depends on:

“Claude treats them as context, not enforced configuration. To block an action regardless of what Claude decides, use a PreToolUse hook instead.”

No cross-vendor equivalent is documented. rift-forge’s shell-backtick guard is Claude-only.

6. Migration surfaces already shipped by vendors

Relevant because they set expectations for what rungs’ own retrofit path should look like:

• Claude Code /init reads .cursor/rules/, .cursorrules, .github/copilot-instructions.md; with CLAUDE_CODE_NEW_INIT=1 also AGENTS.md, .devin/rules/, .windsurf/rules/, .windsurfrules, .clinerules.
• Claude Code /import copies another agent’s config in, including MCP servers, commands, subagents, and skills.
• Cursor converts eligible dynamic rules to skills.

Every vendor is building a one-shot importer. None of them maintains a repo across harnesses over time, which is the gap ADR-0001 targets.

7. What this changes about the Phase 2 findings