Changelog

Language / Idioma: English · Español

This project uses SemVer for plugin/runtime metadata. Release candidates are not public releases and may change before general availability.

Git source runtime vendoring (2026-10-18)

• Windows x64 Git marketplace now launches ./runtime/node.exe included in the repository; no external Node/npm or runtime download on launch.
• Pin runtime provenance to immutable 0.6.0-rc.4 release SHA-256, check the Node license and exact binary hash on CI, and explicitly allow just this oversized binary in history scans.
• Correct portable-package MCP commands to plugin-relative paths required by Agent Plugins 1.0.0.
• Other platforms continue to use their own OS-specific builds; Desktop host support must be tested separately.

0.6.0-rc.4 — Windows Git runtime and marketplace refresh

• Ship the Windows x64 executable in the Git source and launch it through a plugin-relative path, without requiring system Node.
• Increase the manifest version to distinguish the pre-runtime rc.3 package from the current source.
• A newly installed plugin still requires a local ChatGPT execution session and a valid MCP tool connection; versioning alone does not guarantee a cache refresh.

0.6.0-rc.3 — Read-only ChatGPT onboarding and explicit host gates

• ChatGPT's default onboarding skill now verifies already connected read-only MCP tools instead of invoking native Codex/Claude installers or requesting combined network/filesystem permissions.
• The Bash and PowerShell agent bootstraps refuse execution without an explicitly selected Codex or Claude host; ChatGPT is not an allowed bootstrap target.
• Document that the first-party ChatGPT Plugins manager may store enable/disable preferences in Codex's config.toml, which is not authorization for QuotaPilot to install a native Codex footer.
• Add host-scoping regression tests and package the read-only ChatGPT skill with the self-contained distributions. ChatGPT local MCP compatibility still depends on the desktop host.

0.6.0-rc.2 — Host-scoped permissions and safe OpenAI hooks

• OpenAI's shared ChatGPT/Codex manifest now loads an explicitly empty hook, preventing plugin installation from automatically editing Codex config.toml in ChatGPT Desktop/Work.
• Codex CLI native footer changes require a separate explicit request/approval; Claude Code retains its own scoped hook.
• Agent setup now verifies local-versus-cloud execution before requesting write/network permissions and refuses broad ChatGPT permission prompts.
• Refresh pinned Windows/Linux/macOS agent bootstrap checksums for the new verified release; preserve English/Spanish setup guidance and tests.

Desktop MCP compatibility patch (2026-10-18)

• Accept standard MCP request params._meta on tool and UI resource calls. Keep strict tool argument schemas and reject unexpected parameters or invalid metadata types.
• Add host-shaped protocol regression tests for get_usage, open_hud, metadata redaction and the UI resource; separate protocol errors from provider-unavailable messages.
• Document upgrade/restart steps in English and Spanish for affected ChatGPT Desktop plugin installations.

Agent-assisted GitHub URL installation (2026-10-18)

• Add local-agent onboarding: paste repository URL and the agent downloads and SHA-256 verifies the matching self-contained GitHub release without requiring manual ZIP/Node/npm installation.
• Add reviewed Bash and PowerShell bootstraps, strict pinned hashes, private per-user install locations, idempotent behavior, security tests, and host-aware setup instructions.
• Keep ChatGPT Desktop local bridge and provider telemetry limitations explicit; host trust consent remains mandatory.

0.6.0-rc.1 — Portable, dependency-free runtime

• Bundle official Node.js 24 LTS, its license notices, MCP server and hooks into OS-specific release archives: Windows x64, Linux x64, macOS ARM64/Intel.
• Build native release bundles in four pinned GitHub Actions environments and test with a completely empty child PATH.
• Fix Codex-vs-Claude lifecycle hook ambiguity with an explicit host argument and preserve Claude's bundled runtime in its private status-line directory.
• Document honest host/provider limitations and release verification in English and Spanish. No credentials, cloud service or automatic network downloader is introduced.

0.5.0-rc.1 — Public source, host-validation release candidate

• Show advisory consumption rate, estimated runway and possible quota gap within the local HUD.
• Add visual low-quota/critical and pacing alerts; missing provider metrics remain unavailable, never fabricated.
• Document local Codex and Claude installation, constrained ChatGPT host compatibility, privacy and uninstallation.
• Add offline npm run doctor diagnostics with no secret or filesystem path output.
• Add full reachable-history secret-pattern audit (npm run security:history) and Linux, macOS, Windows CI coverage.
• Avoid side effects in Claude installer when the user has a different custom status line.
• Normalize missing Claude percentages correctly and limit status-line input accumulation before appending data.
• Add MIT licensing metadata, CODEOWNERS policy, GitHub Actions Dependabot monitoring and a step-by-step administrative security checklist. Public GitHub source now supports admin-side protection; connected tooling cannot change branch settings.
• Public marketplace naming (ai-usage), HTTPS installation documentation, sanitized GitHub issue templates and a host acceptance guide.
• Add English/Spanish documentation, bidirectional language links, EN/ES widget toggle and localized Claude status line/doctor, without changing machine-readable data.
• Retain all existing 0.4.x filesystem/protocol hardening and read-only/no-network invariants.

0.4.0 — Security hardening

• Bounded MCP/provider input, private file handling, stronger CSP, generic errors and minimized Claude snapshots.
• Read-only MCP tools and host-dependent branding with CI safety gates.

See release readiness for remaining real-host/admin validation evidence.