Skip to content

Compatibility and evidence

Kiyo tracks six independent targets. A result on one never counts for another, not even between the CLI and IDE forms of the same product. Everything below comes from the framework’s live test matrix, checked 2026-09-29 on Windows 10.0.26200. That run used “native checks without quota”: no model turn, sign-in, global install, or publication was performed.

Target Observed environment Verified subset Remaining
Claude Code CLI 2.1.220 Native manifest validation; relocated directory and ZIP component inventory: eight Skills, no agents, hooks, MCP, or LSP Marketplace install and every model, activation, and lifecycle case NOT_TESTED
Claude Code VS Code Editor 1.139.1; extension metadata 2.1.283/2.1.284; active engine UNKNOWN Installed metadata only NOT_TESTED
Codex CLI 0.158.0 Disposable local catalog install and list; native cache byte-identical to the distribution; uninstall preserving synthetic project files Skill discovery and invocation, policy and Core loading, update, model behavior NOT_TESTED
Codex IDE extension Metadata 26.917.62051; active UNKNOWN None Native plugins UNSUPPORTED per official docs; no fallback adopted
GitHub Copilot CLI No executable found on the inspected PATH; version UNKNOWN None NOT_TESTED
GitHub Copilot VS Code Editor 1.139.1; Copilot extension metadata not found None NOT_TESTED

Key: V = VERIFIED/PASS for the bounded complete check · N = NOT_TESTED/NOT_RUN · U = UNSUPPORTED native plugin capability.

Check Claude CLI Claude VS Code Codex CLI Codex IDE Copilot CLI Copilot VS Code
01 Native installation and discovery N (a) N V (b) U N N
02 Eight Skills visible and loadable N (a) N N (c) U N N
03 Explicit Init writes only approved Memory/config N N N U N N
04 Simple Implement and real verification N N N U N N
05 Review-only writes zero files N N N U N N
06 Observation drift versus decision conflict N N N U N N
07 Sensitive approval and native restriction N N N U N N
08 Relocated or cached resource reads by agent N (a) N N (c) U N N
09 Fresh-session automatic activation N N N U N N
10 New session observes current policy/version N N N (d) U N N
11 Update preserves Memory and human bootstrap N N N U N N
12 Native uninstall preserves user state N N V (e) U N N
  • (a) Claude listed eight entries from a relocated directory and ZIP. This is component discovery, not installation or execution. Normal validation exits 0 with version/author warnings; strict validation exits 1.
  • (b) Codex registered a test-only local catalog and installed the unchanged candidate in a fresh child state directory.
  • (c) The Codex cache contains all 795 files unchanged, and a static checker passes eight entries and 4,956 contained links. No agent read trace exists.
  • (d) A new management process sees the installed and enabled state. This is not a new agent session.
  • (e) Uninstall removed the cache and entry; three synthetic files (human AGENTS.md, legacy Memory, project policy) stayed byte-identical.

Only two complete checks pass out of 72 records. There is no compliance score and no six-target PASS.

  • VERIFIED means only the stated observed property, and can describe a real failed check (for example Claude’s strict validation FAIL).
  • UNSUPPORTED describes a documented capability limit. It is not an executed failure.
  • DOCUMENTED_ONLY means documentation exists but there is no actual result.
  • None of these labels means “certified”.
Host Requirement recorded in the framework docs
Claude Code VS Code VS Code 1.94.0+; a paid Claude plan (Pro, Max, Team, Enterprise) or a Console account
Copilot CLI A Copilot plan; your organization’s CLI policy; PowerShell ≥ 6 on Windows; Node ≥ 22 for the npm install route
Copilot VS Code chat.plugins.enabled set to true
Codex CLI A new session after installing a plugin

These are vendor facts recorded as DOCUMENTED_ONLY. Check current vendor documentation before relying on them.

Owner-required tests (OA-05, OA-06) cover the 48-case behavioral suite and per-target discovery, invocation, Core loading, restart, update, managed-block, and state-survival cases. Until they run, compatibility claims stay limited to the table above. See Status and releases.