docs(contributing): reconcile Agent harnesses with the reference runtimes

[milamberspace]

Summary

• The ## Agent harnesses section of CONTRIBUTING.md named Claude Code as the framework's sole reference runtime and did not mention OpenCode at all, contradicting docs/prerequisites.md (OpenCode = reference implementation) and docs/vendor-neutrality.md (Claude Code; OpenCode; Codex experimental).
• The section was last touched 2026-05-26 (docs(contributing): rewrite for framework-as-a-whole audience #323), predating the vendor-neutrality work — OpenCode as a reference implementation and the first-class Codex adapter both landed later (feat(codex): first-class Codex CLI skill runtime #902). It was simply left behind.
• Reframed to two reference implementations — OpenCode (open source, model-agnostic) and Claude Code (Anthropic-powered) — with Codex as an experimental first-class adapter. Deliberately uses the softer, co-reference wording (consistent with the vendor-neutrality.md matrix) rather than crowning a single "the reference implementation". Added links to the two authoritative docs so the three don't redivergence.

Type of change

• Documentation (docs/, README.md, CONTRIBUTING.md)

Test plan

• prek run --files CONTRIBUTING.md passes (markdownlint, typos, lychee link check, SPDX, doctoc — all green). The section anchor is unchanged, so the TOC is unaffected.
• Doc-only change; no code paths affected.

RFC-AI-0004 compliance

• Vendor neutrality — this change improves alignment with RFC-AI-0004 §3 by removing the single-vendor "reference runtime" framing and pointing at the authoritative runtime matrix. No other principles touched.

Linked issues

Closes #1100

Notes for reviewers (optional)

Governance call to confirm: prerequisites.md says OpenCode is "the" reference implementation, while vendor-neutrality.md lists Claude Code and OpenCode side by side. This PR follows the softer vendor-neutrality.md framing (co-reference). If the PMC prefers the stronger "OpenCode is the reference implementation" wording, the opening sentence and the table's two Reference implementation rows are the only lines to adjust — happy to change it.

The path phrasing was also corrected: skills are authored under skills/<name>/SKILL.md and exposed to every harness through the committed .agents/skills/ symlink tree (per AGENTS.md) — CONTRIBUTING.md previously described only the Claude-Code view.

[dpol1]

Approving — the new wording matches the two docs it cites as authoritative, and the Codex paragraph matches the adapter guide. Two things worth recording, neither blocking.

Cross-checked against the authoritative docs

docs/prerequisites.md (line 57) does name OpenCode the reference implementation, docs/vendor-neutrality.md lists Claude Code / OpenCode / Codex in its runtime row, and the Codex paragraph here matches docs/adapters/codex.md: experimental, reads the .agents/skills/ tree natively, ships the in-tree sandbox profile and HITL exec-policy rules. The .agents/skills/ symlink tree is indeed committed, as the new text claims.

Smaller observations

• docs/adapters/registry.md ("Agentic runtime" row) still lists only "Claude Code, Codex experimental" — once this lands, that row becomes the next doc left behind on the same fact. A follow-up issue seems better than widening this PR.
• CONTRIBUTING.md (new text) — "powered by Anthropic subscriptions": Claude Code also runs on plain API keys; "powered by Anthropic models" would sidestep the inaccuracy.

---

This review was drafted by an AI-assisted tool and checked and
submitted by an Apache Magpie contributor — not a maintainer, so
a maintainer's review remains the authoritative next step. If
something feels off, please reply on the PR.

More on how Apache Magpie handles maintainer review:
CONTRIBUTING.md § Opening a pull request.