This matrix uses evidence tiers. It does not call a host “supported” merely because its directory name exists in legacy setup code.

Tiers are cumulative only when evidence says so. A portable host is not automatically verified. A legacy generated host output is not a native GStack 2 bundle.

npx skills add time-attack/gstack/skills

The standards installer owns host detection, destination paths, project/global scope, copy versus symlink, updates, removal, and selected-skill installation. Review its detected-host prompt; detection must never silently enroll a host.

Examples supported by the installer interface:

# One public skill only (point selection at the canonical skill directory)
npx skills add time-attack/gstack/skills --skill qa

# Installer-managed global scope
npx skills add time-attack/gstack/skills -g

Run npx skills add --help for the installed CLI version before scripting agent-selection flags. GStack deliberately does not reproduce those flags or host paths in ./setup.

The expected discoverable names are exactly:

plan
qa
debug
review
ship
make-pdf

Compatibility aliases live at skills/.compat/<name>/SKILL.md (routing-only, no judgment); the dot-prefixed directory keeps them out of discovery, so neither they nor references/legacy/*.md may appear as additional skills. Installing a subset must not pull the other public entries unless the user selected them.

The committed skills CLI 1.5.19 evidence run predates the design → make-pdf change, so its recorded discovery result is the six design-era names (debug, design, plan, qa, review, ship), not the current set. For selected installation, use the canonical skills/ source as in the example above. That CLI version's pre-filter display can count hidden compatibility aliases, but the committed actual-host artifact proves that only the selected qa directory was installed; the display count is not an installed-skill count.

The isolated standard-installer matrix passed 510/510 checks with skills CLI 1.5.19: 18 install cases and two removal cases. Default discovery projected only canonical skills/; separate explicit-selection cases covered a single canonical skill and an opt-in legacy alias. The matrix used symlinked and spaced source paths matching a clean checkout where ignored legacy host trees are absent. Every installed file was a physical copy and its hash matched the selected source. The standard skill installation remains Markdown-only and does not install the optional runtime. The committed evidence artifact is evals/installation/install-matrix.json. That artifact is historical evidence as of its 2026-07-20 generatedAt date: it was generated against the design-era skill set (design in place of make-pdf) by a matrix script that has since been removed. Its installer-mechanics results (copy fidelity, hashes, scope handling, removal) stand as of that date; it does not verify discovery of the current six-skill set.

The representative selection cases installed exactly qa, review, and ship, independently at project scope for Cursor and global scope for Codex, then removed them noninteractively. Separate cases installed only ship for OpenClaw and explicitly selected the office-hours compatibility alias for Codex; neither alias nor unselected canonical skill was silently enrolled. --copy was advertised and used.

The installer matrix itself is filesystem verification. Host UI execution is evidenced separately, per the Host UI/process column: Codex passed the full 4/4 live adversarial suite; Claude Code and Cursor passed scored one-shot UI-launch cells (single-fixture, recorded incomplete by construction — a UI-launch cell, not adversarial parity); Pi's cell is a retained honest failure; Kimi is blocked on an expired operator OAuth grant; OpenClaw and GitHub Copilot are installer-verified only, pending host availability. The activation token for the slash-capable hosts is /skill (probe-verified); Codex activates on $skill. A separate actual Codex invocation installed only qa from time-attack/gstack/skills, with the optional runtime absent, and passed the judgment/setup-gate behavior without changing its workspace or creating a runtime or browser. The Codex adversarial lane now has a passing live result: the 2026-07-22 paid live v3 one-shot passed 4/4 (Codex CLI 0.145.0, gpt-5.4, retry_count 0; artifact 2026-07-22T21-33-20-053Z-84fcb74b.json). The retained failed history is unchanged: v1 and immutable v2 failed, two earlier live v3 one-shots each failed 3/4 (review compound inspection, then ship git symbolic-ref under the pre-fix classifier), and a harness-4 gpt-5.6-sol one-shot (Codex CLI 0.144.4, 2026-07-21) failed 3/4 when the ship fixture logged one forbidden command attempt (the session ran npm run check during the report-only readiness assessment; artifact 2026-07-21-v4-live-gpt-5-6-sol.json). The 4/4 pass is a single-model result on the model recorded in its artifact and does not supersede the retained gpt-5.6-sol failure. None was retried or relabeled; the offline harness suite (now harness 4 with per-host adapters) is green at 32 tests / 183 assertions. The installer matrix used the current local canonical projection through the published npx skills CLI. The release branch remained unpushed; native CI used the temporary codex/gstack-2-ci-20260717-39bc307b ref only.

Hosts still without any live host-UI execution evidence:

• Kimi Code CLI — installer- and discovery-verified only; the scored cell is blocked on an expired operator OAuth grant (kimi login required).
• OpenClaw — installer-verified only; the CLI is not installed here.
• GitHub Copilot — installer-verified only; requires a paid Copilot seat.

Pi has live UI evidence, but its scored cell is a retained failure, not a pass.

Legacy generators currently know ten host names. That is historical breadth, not proof that all ten install correctly. Kiro's old Codex-rewrite behavior and other host-output transforms are outside the canonical 2.0 path.

Use a clean temporary home and project for every cell; never test against an operator's live skill directory.

1. Record OS, host version, Node/npm version, and skills CLI version.
2. List the source and assert exactly five default entries.
3. Install all five at project scope and verify each host discovers only those five GStack public skills.
4. Remove them through the standard installer.
5. Repeat at global scope.
6. Install a selected subset; verify unselected skills were not enrolled.
7. Reinstall/update without re-detecting or enrolling an unselected host.
8. Invoke a pure judgment mode with gstack absent from PATH.
9. Invoke a capability-dependent mode and verify one actionable runtime offer, without breaking judgment.
10. Exercise a path containing spaces and, where supported, symlink and read-only failure behavior.
11. Record exact command, exit status, output artifact, and cleanup result in TEST-EVIDENCE.md.

The scripted matrix runner (scripts/gstack2/test-install-matrix.ts) and its test file were removed with the rest of scripts/gstack2/, so the committed matrix JSON cannot be regenerated and stands as historical evidence at its recorded date. To re-verify installation against the current tree, run the standard installer manually from a clean temporary home and project:

# From a clean temp project (HOME pointed at a temp dir):
npx skills add time-attack/gstack/skills            # all six, project scope
npx skills add time-attack/gstack/skills -g         # global scope
npx skills add time-attack/gstack/skills --skill qa # selected subset

Then assert the discovered set is exactly the six canonical names, compare each installed file's hash against its source, verify the subset case installed only the selected skill, and remove everything through the same installer.

The committed historical matrix (2026-07-20, design-era skill set) passed 510/510 installer CLI checks across 18 installs and two removals; its JSON artifact is committed at evals/installation/install-matrix.json. Steps 8–9 passed for the recorded Codex runtime-absent invocation; UI-launch cells now also cover Claude Code, Cursor, and (as a retained failure) Pi per the matrix above. The paid live v3 adversarial gate passed 4/4 one-shot on 2026-07-22; the 3/4 one-shots are retained as immutable failed runs, not pending runs. Evidence: standard-codex-runtime-absent-2026-07-17.json, 2026-07-22T21-33-20-053Z-84fcb74b.json (passed), and the retained failed one-shots 2026-07-17T19-48-45Z-v3-live-gpt-5-4.json, 2026-07-21-v4-live-gpt-5-6-sol.json, and 2026-07-22T20-57-26-117Z-b39a6a57.json.

The optional runtime is separate from skill placement. From a repository checkout, the one host-neutral setup entrypoint is:

./setup
~/.gstack/bin/gstack doctor --json

./setup resolves a symlinked checkout, reinstalls the frozen production dependency set on every run, builds missing allowlisted capabilities through the runtime-only build target, validates and hashes every staged file, smoke-tests the CLI, atomically activates the version, and writes stable POSIX and Windows launchers under $GSTACK_HOME/bin (default ~/.gstack/bin). The bundle uses .exe targets on Windows and includes the CoreDevice/iOS bundle only on Darwin. Add the bin directory to PATH if the short gstack command is desired.

Official bundles capture the pinned Bun 1.3.14 executable inside the immutable runtime and expose it as $GSTACK_HOME/bin/bun; browser launchers do not use host-global Bun. Node remains the bootstrap/launcher floor. On Windows, Git for Windows Bash is required only for retained shell helpers and is a separate doctor check. Python 3 is an optional prerequisite only for specialist flows that name it. Doctor reports both tools independently of native browser readiness.

Twenty-five focused installer tests pass with 341 assertions. They cover manifests, paths with spaces, source-root symlinks, internal-link/path-escape rejection, failed build/ validation/smoke rollback including native-load rollback smoke, interrupted-pointer recovery, stable POSIX/Windows launchers, runtime-only builder selection, deterministic exact Sharp/ngrok platform closure, managed uninstall, and the host-neutral wrapper.

The deterministic clean macOS arm64 managed-bundle audit records 110 components, 1,829 files, 450,044,315 bytes, and 50 capability launchers. This is a platform-specific bundle measurement, not a universal byte count; platform-native package payloads differ. Setup installs frozen production-only dependencies and excludes the development-only Claude Agent SDK. The Sharp/ngrok closure is included. The Hugging Face sidecar is excluded and its package is development-only, so setup installs neither its inference runtime nor model weights and reports the L4 capability unavailable.

Native CI run 29615621805 passed every job at commit a8a5fa1aa381f9b948dfc57af26016092fc33277: macOS, Ubuntu, native Windows, the 470-check installer matrix, and the Dev Container. The sanitized committed record is evals/ci/native-2026-07-17.json. Runs 29608904265, 29611504979, 29611757175, 29612056517 (cancelled), 29613668419, 29614448170, and 29614899434 are retained as superseded diagnostic evidence, not substituted for the passing run.

The six portable skills remain useful when runtime installation fails. A runtime failure must not remove or corrupt their standard-installer placement.

No GStack 2 native tier is currently required or awarded. A future native bundle needs an accepted issue showing that a host API is necessary, must load the same canonical judgment modules, and must pass the portable parity suite. Host-specific presentation metadata such as agents/openai.yaml does not by itself create a second judgment source or a Native claim.