docs(cli): CLI-005 harness-expansion spike — evaluation + proposed D-045

[joshuaboys]

Summary

CLI-005 (harness-expansion research spike). Evaluates candidate AI-coding
harnesses against the D-040 native-discovery gate — a harness is a zero-asset
add (like Grok) iff it both natively reads AGENTS.md and auto-discovers
.agents/skills/<name>/SKILL.md. Deliverable:
plans/research/2026-07-23-harness-expansion-spike.md.

Four parallel research streams with primary-source citations; the load-bearing
GO claims were spot-checked directly (Antigravity's .agents/skills default;
Amp + Windsurf skills discovery; the agents.md registry for AGENTS.md adoption).

Verdicts

Verdict	Harnesses
GO (zero-asset)	Antigravity, Amp, Gemini CLI, Windsurf, Roo Code¹, OpenClaw²
GO-WITH-ASSETS	Cursor³ (low), Hermes⁴
NO-GO / defer	Aider, Zed, Devin, Jules, Warp, Continue, Factory, Junie, goose

¹ AGENTS.md auto-load is a toggle · ² confirm hands-on before commit ·
³ reads AGENTS.md but per-tool skills dir; mirror cost is low & user base large ·
⁴ global ~/.hermes/skills/ path + severe name collision → deprioritise.

Proposed D-045 (needs your sign-off)

1. Add as zero-asset (CLI-006): Antigravity, Amp, Gemini CLI, Windsurf, Roo Code
2. Add after a 5-min hands-on confirm: OpenClaw
3. Separate GO-WITH-ASSETS decision: Cursor (low cost, large user base)
4. Deprioritise: Hermes · Not now: the AGENTS.md-only crowd

The spike does not self-authorise the set — D-045 is recorded as proposed.

Test plan

• aps lint plans/ clean (45 APS files); markdownlint clean on both files
• Flagship GO claims independently verified from primary vendor docs
  (antigravity.google/docs/skills, ampcode.com/manual, docs.devin.ai Cascade
  skills, agents.md registry)
• Recommendation marked proposed; every harness still gets a binding
  aps init --tools <x> smoke test in CLI-006 (Grok precedent, CLI-004)

APS Work Items

• CLI-005 → spike complete; evaluation linked from the module; D-045 proposed.
• Module stays Draft; CLI-006 (implement approved harnesses) unblocks once
  you approve the D-045 set.

Notes

Key nuance recorded in the doc: "zero-asset" bounds bespoke scaffold cost
only — each add still costs a tool key across Rust/bash/pwsh, wizard/install
dispatch, docs, and D-039 parity fixtures. Recommend the curated set, not
"everything that passes."