Client compatibility evidence¶

This page records dated client checks. These are qualification observations, not minimum supported versions. Structural conformance is tracked separately in Standards and Forge policy.

1.1.0 validation (2026-09-22)¶

The project, shipped Forge plugin, marketplace, and lockfile are prepared for 1.1.0. The minor version adds forge update for reviewed existing-skill replacement, stricter YAML frontmatter validation, structured checkout diagnostics, and CI coverage of isolated wheel installation.

The macOS Python 3.12.13 suite passed 583 tests with three filesystem or platform-specific skips and 93.05% coverage. Ruff, formatting, ty, Forge validation, generation drift, the strict documentation build, and isolated wheel installation checks passed. Regression coverage includes MCP argument files and directories, path traversal, license evidence casing, shared provenance, rollback, YAML mappings, diagnostics outside a checkout, release source evidence, and preservation of unchanged catalog bytes. The wheel smoke imported two authored skills and updated one, ran their reviewed renderers only after package validation, and retained six verified Markdown, JSON, and SVG artifacts alongside the tested distributions. Peer review findings were corrected and verified.

With isolated local client homes, Codex CLI 0.155.1 and Copilot CLI 1.0.84-1 installed the local 1.1.0 marketplace package. Codex's cached package matched the source tree byte-for-byte; Copilot loaded the package live and listed package-agent-skill. These checks establish local installation and discovery. The isolated Codex home was not logged in, and Copilot reported no authentication information, so neither isolated installation performed an agent invocation. A fresh public Forge 1.1.0 marketplace installation in VS Code 1.138.0 remains pending merge.

For a separate end-to-end Codex check, Forge 1.1.0 imported the authored release-digest and status-chart fixtures into two disposable 0.1.0 plugins using separate reviewed plans. forge generate and forge check passed. Codex CLI 0.155.1 installed both from that local marketplace; both cached packages matched their portable sources byte-for-byte. Authenticated, ephemeral invocations read the installed skills and ran their reviewed renderers, producing Markdown, JSON, and SVG files that matched all four packaged expected artifacts. The JSON and SVG content was also checked semantically. Both temporary plugins and the marketplace were removed, and the client's registrations matched their pre-test snapshots. This verifies the Codex local-plugin user journey; the public 1.1.0 marketplace is still pending merge.

In a new Forge-1.1-E2E VS Code profile, the same two portable packages were loaded through chat.pluginLocations. Chat: Configure Skills... discovered both skills at their packaged paths. VS Code Chat then invoked each packaged renderer with its sample input; the Markdown, JSON, and SVG outputs matched all four packaged expected files by SHA-256, and the JSON and SVG content checks passed. This verifies the local-plugin skill journey in VS Code, not a fresh public marketplace installation.

The local release-status MCP server was discovered in VS Code 1.138.0 but did not start: VS Code passed ${PLUGIN_ROOT}/server.py literally, so no tool call or PLUGIN_ROOT/PLUGIN_DATA environment check passed. The portable configuration is valid under Agent Plugins 1.0. This matches the open VS Code placeholder-expansion bug. VS Code skill invocation is qualified separately; portable MCP runtime acceptance is blocked on this client version.

The PR must pass the protected branch's required quality and Linux, macOS, and Windows jobs, plus the project's Python 3.11 release check, before merge. Release qualification still requires the remaining client acceptance and public read-back steps in the release guide, including the blocked VS Code MCP check. These local results do not establish a fresh public marketplace installation or a Forge publication invocation from that public installation. Historical evidence below retains its original version and scope.

Previous release: 1.0.1¶

The project, shipped Forge plugin, marketplace, and lockfile use 1.0.1. This patch addresses two low-severity issues: import approval snapshot consistency and untrusted terminal diagnostics. The packaged skill instructions and bootstrap helper are unchanged from 1.0.0.

The regression suite covers source replacement between inspection and open, mutation during reads, bounded reads, byte-preserving copies, license snapshot hashing, exact file-map approval, CRLF metadata, all C0/DEL/C1 terminal controls, and POSIX surrogateescape output. It checks that ordinary Unicode paths and decoded JSON values remain unchanged. Import plans from older versions require fresh review after upgrade.

On macOS, the 1.0.1 suite passed 386 tests on both Python 3.11.16 and 3.12.13, with two platform-specific skips and 92% coverage. Ruff, ty, Forge validation, generation drift, actionlint, and the strict documentation build passed. Independent patch review identified a non-UTF-8 filename edge case; it was corrected and covered by a regression before the final full-suite runs.

The wheel installed into a fresh Python 3.12 environment and passed version, branch-name, and repository-validation smoke checks. Copilot CLI 1.0.84-1 enabled the 1.0.1 plugin with one skill from an isolated local marketplace. This is local installation evidence, not a fresh remote download or agent publication invocation. VS Code and Codex runtime acceptance was not repeated for this CLI-only patch; the dated 1.0.0 and earlier observations below retain their original scope. The GitHub release records final cross-platform CI and public marketplace read-back.

Initial Windows CI exposed a cross-API file-metadata mismatch. The correction preserves file identity and read-change checks while accounting for CPython's Windows path-stat timestamp and executable-suffix behavior. The regression suite includes modeled Windows metadata and executable suffixes.

Previous release: 1.0.0¶

The project, shipped Forge plugin, marketplace, and lockfile use 1.0.0. This release removes the built-in publication repository and adds a confirmed destination saved across projects. The publication guide and destination settings reference describe first use, reuse, overrides, and replacement of the saved default.

The destination regression suite covers missing configuration before any checkout, read-only inspection, persistence across processes and projects, override precedence, protected replacement, competing saves, invalid settings, failed-write recovery, and lock release after a client exits unexpectedly. The real CLI tutorial also runs from a remembered destination through plan, approved apply, generation, and validation, checking the repository shown in the plan and manifest.

On macOS, the 1.0.0 suite passed 288 tests on both Python 3.11.16 and 3.12.13, with two platform-specific skips and 92% coverage. Ruff, ty, Forge validation, generation drift checks, and the strict documentation build passed. The built wheel installed into a fresh environment and passed forge --version, branch-name validation, and forge check. A one-file benchmark smoke run completed the documented workflow; it is not a new performance benchmark.

Copilot CLI 1.0.84-1 registered a local marketplace in an isolated configuration and enabled the 1.0.0 Forge plugin with one skill. That client loads local marketplace packages live from the source directory, so this is local installation evidence, not a remote download or an agent publication invocation.

The client commands were checked on 2026-09-12 against local Codex CLI 0.153.4, GitHub Copilot CLI 1.0.84-1, and the upstream Codex, Copilot CLI, and VS Code references. Historical client experiments below retain their actual versions and scope; they are not fresh 1.0.0 installation or runtime evidence. GitHub Enterprise Server authentication still requires separate qualification.

Previous release: 0.3.0¶

The 0.3.0 release passed 257 tests on Python 3.11 and 3.12. A fresh public Copilot CLI 1.0.84-1 installation matched all nine package files from merged revision 2a97c34889aa9fa035b6e1397e799df491cb7562, and the installed skill completed its check-only workflow without changing files or branches. That evidence is recorded in the 0.3.0 release.

Earlier development snapshots also used 0.3.0 and 0.3.1, as recorded below. After an explicit refresh and update, verify the intended installed version and package content.

Evidence recorded on 2026-09-11–12¶

The local 0.0.2 candidate passed the full suite on macOS with Python 3.11.16 and 3.12.13: 257 tests passed, two platform-specific tests skipped, and coverage reached 92%. Ruff, ty, Forge validation, the strict documentation build, and the built wheel's version and branch-name commands passed.

Forge 0.0.2 created the fixture's skill branch, applied the reviewed import hash, generated both marketplaces, and validated the package. The installed cache matched fixture revision a170e50d5417695c6d0780276ee2aa664adfc8fd; its skill bytes matched the reviewed source.

The multi-skill installation used fixture revision 796fb8f149f0cbc045f762235206355809e9cffd; the update used 8985331bed92c8de7de9073476f4f14a8ffb9342. Private visibility and the final remote revision were read back. The original forge-smoke package stayed unchanged.

The shipped skill also passed five local conformance scenarios and an independent existing-skill maintenance walkthrough, including the documented provenance example and full local gates. The benchmark guide describes methodology, measured results, and boundaries.

This qualifies private GitHub.com fixture installation and skill invocation, plus the recorded local contributor workflows. It does not qualify a fresh public 0.0.2 release installation, the complete installed Forge publication journey, GitHub Enterprise Server authentication, executable MCP or hook behavior, or Linux/Windows execution for this candidate.

Evidence recorded on 2026-09-07¶

The project, shipped Forge plugin, and marketplace version are reset to 0.0.1. This is an intentional version reset from the earlier 0.3.x numbering. The historical installation evidence below retains its original versions; it does not qualify a fresh 0.0.1 client installation. Clients may not offer a lower version as an automatic update, so verify the installed version after explicitly selecting the reset release.

The command review used the current Codex reference, Copilot CLI reference, and VS Code plugin documentation. This review did not qualify private-host authentication or an end-to-end publication.

Evidence recorded on 2026-08-22¶

At that time, the working-tree candidate was plugin version 0.3.1; its local package, provenance, generated marketplaces, tests, and strict documentation build passed. The public-repository installation above qualifies the 0.3.0 baseline. It does not qualify a 0.3.1 remote installation; a fresh release install must be recorded separately.

The 2026-08-22 documentation change did not qualify authentication against a live private GitHub or GitHub Enterprise Server host. Private-host release acceptance remains a separate required test.

Record a new qualification¶

Update this page when a release changes client installation, marketplace generation, or portable runtime behavior. Record exact versions and distinguish a live install or runtime test from command-surface and schema validation.