Plugin packaging and installation
Impeccable is packaged as one source skill that is rendered for several agent clients. The build turns that source into client-specific skill folders, agent definitions, hooks, launchers, and archives. The plugin/ tree is the distributable package for Claude Code and Grok marketplace installs, while dist/ holds provider builds and the separate OpenAI package. See Distribution and engineering tooling for the wider release model, Build, release, and validation for release checks, Agent design workflows for delegated roles, and Skill commands and references for the user-facing command surface.
Source package model
Source skill
The source skill is the single place where product behavior is defined. skill/SKILL.src.md holds the shared metadata and main instructions. Its references, scripts, and agent definitions add task guidance, runtime entry points, and delegated roles. Provider placeholders in this source are resolved during rendering, so the same behavior can use the path and command form required by each client. The .src.md suffix keeps a skills installer from treating the source as a ready-to-run SKILL.md.
The build reads these parts as one skill record. It excludes project state such as live-mode configuration and treats platform engine binaries as a separate release concern. An installed package therefore carries the launcher and shared runtime assets without copying one consuming project's state into another project.
References: scripts/…/utils.js• L1-16 scripts/…/utils.js• L195-239 skill/SKILL.src.md• L1-24
Provider rendering
Provider rendering converts the shared skill record into the file and metadata shape that each client can load. The renderer keeps the provider's instruction blocks, fills provider paths such as the scripts directory, copies references and scripts, and emits supported agent and hook forms.
The main output is under dist/<provider>/<config-dir>/. The provider table defines both parts of that path. It also defines special behavior such as native agent files, OpenAI metadata, hook manifests, and provider tags that remain in rendered text.
References: scripts/build.js• L696-715 scripts/…/factory.js• L253-327 scripts/…/providers.js• L12-40
Manifest contracts
Root manifest
The packaged Claude manifest at plugin/…/plugin.json is the source of truth for the plugin name, release version, description, author, and links. The installed SKILL.md now declares version 4.10.0 and a 24-command surface. The build uses release metadata for skill frontmatter and for the Claude, Grok, and OpenAI package manifests, so version-bearing outputs must stay aligned.
The marketplace manifest points the impeccable entry at ./plugin. Its version must agree with the packaged manifest and every generated version-bearing artifact. This keeps an installed package from reporting a version that does not match its content.
References: plugin/…/plugin.json• L1-12 .claude-plugin/marketplace.json• L11-31 plugin/…/SKILL.md• L1-8 scripts/…/validate-plugin-versions.js• L1-20
Claude and Grok manifests
The packaged plugin/…/plugin.json is the Claude Code marketplace entry point. It identifies the package as impeccable, reports release version 4.10.0, advertises the command surface, and points the loader to ./skills/. The trailing slash is required for skill registration. It does not declare an agents field because Claude Code discovers markdown files in plugin/agents/ itself.
The packaged plugin/…/plugin.json carries native Grok metadata for the same release. It adds the MIT license and search keywords while keeping the Claude-compatible skill, agent, and hook payload available to the Grok marketplace path.
References: scripts/build.js• L811-849 scripts/…/validate-plugin-manifest.js• L14-29 scripts/…/validate-plugin-manifest.js• L117-159 plugin/…/plugin.json• L1-12 plugin/…/plugin.json• L1-20
OpenAI manifest
The OpenAI package is assembled from the Codex-rendered skill rather than from the Claude marketplace tree. Its .codex-plugin/plugin.json contains public interface metadata. The package also names its skill folder, icon, and hook resources so the client can present and run the integration.
This separation keeps Claude paths and slash-command syntax out of the Codex package. The OpenAI stage fails when its root manifest, Codex skill payload, or icon is missing.
References: scripts/…/openai-plugin.js• L18-35 scripts/…/openai-plugin.js• L39-60 scripts/…/codex-plugin.js• L1-41
Agent definitions
Agent source
Agent definitions under skill/agents/ describe specialist roles that the main skill can use. Their frontmatter supplies a role name and description plus optional client names, tool limits, model hints, turn limits, and provider filters. The build reads the body as shared role guidance and renders it for each target.
The shipped roles divide delegated work into clear contracts. The asset producer creates raster plates from approved visual references without changing the direction. The documenter records the design system from the finished artifact. The finish reviewer checks a finished build against its direction contract, its approved comp when one exists, and the chosen world's quality bar. It reviews supplied evidence without rendering or editing, then returns a fixed disposition and an ordered list of material fixes. The manual edit applier changes only the source entries in a leased live edit event and returns a structured result to the parent live thread.
An absent providers field makes a role available to every client with an agent format. A provider list narrows delivery. This lets one source set support clients with different delegation features without copying role instructions by hand.
References: scripts/…/utils.js• L241-289 plugin/…/impeccable-asset-producer.md• L1-46 plugin/…/impeccable-manual-edit-applier.md• L1-97 plugin/…/impeccable-documenter.md• L1-28 plugin/…/impeccable-finish-reviewer.md• L1-43
Agent payloads
The same role reaches clients through different file forms:
| Client family | Packaged agent form | Delivery location |
|---|---|---|
| Claude Code and Grok | Markdown with YAML frontmatter | plugin/…/*.md |
| Codex skill installs | TOML role files | skills/…/*.toml |
| OpenAI skill metadata | YAML interface metadata | skills/…/openai.yaml in the Codex-rendered payload |
| Clients without native delegation | Degraded reference text | skills/…/*.md |
Degraded files keep role guidance available when a client cannot start a subagent. They state that the role runs inline and that the parent session keeps control of polling and protocol replies where needed. The generated files are derived from the source agent definitions and must not be edited as independent role sources. Native files use each client's naming rules. The marketplace package carries Claude-compatible roles, while Codex and OpenAI payloads carry provider-specific forms. The finish reviewer keeps its evidence-first review and fixed output contract in the native and degraded forms. The manual edit applier keeps source editing separate from live polling and protocol replies. The marketplace package also carries the motion reviewer, which judges captured normal and reduced-motion evidence and returns material fixes without editing the target.
References: scripts/…/factory.js• L108-137 scripts/…/factory.js• L348-395 plugin/…/impeccable-asset-producer.md• L1-19 plugin/…/impeccable-documenter.md• L1-29 plugin/…/impeccable-finish-reviewer.md• L1-43 plugin/…/impeccable-manual-edit-applier.md• L1-97 plugin/…/impeccable-motion-reviewer.md• L1-34 plugin/…/asset-producer.md• L1-6 plugin/…/documenter.md• L1-6 plugin/…/finish-reviewer.md• L1-10 plugin/…/manual-edit-applier.md• L1-24
Skill payload layout
Packaged skill tree
Each rendered skill is a self-contained folder. SKILL.md is the entry document. reference/ contains task guidance. scripts/ contains the launcher and browser-side assets needed by packaged workflows. Provider-specific agent files are placed beside or inside the skill according to the client contract.
The packaged skill exposes one user-invocable impeccable skill. Its command table names supported design workflows, and routing selects a workflow or asks the user when context does not decide it. The current entry advertises 24 commands, including direct live variants through generate. Setup and maintenance references cover durable product context, drift reports, and project-local detector hooks.
The Claude marketplace package has this high-level shape:
plugin/
├── .claude-plugin/plugin.json
├── .grok-plugin/plugin.json
├── skills/impeccable/
│ ├── SKILL.md
│ ├── reference/
│ └── scripts/
├── agents/*.md
└── hooks/hooks.json
Script references are rewritten to the installed plugin or skill root, so the skill can run from its actual installation path.
References: scripts/…/factory.js• L253-381 scripts/build.js• L851-885 plugin/…/SKILL.md• L1-8 plugin/…/SKILL.md• L43-88 plugin/…/command-metadata.json• L1-98
Workflow references
The packaged reference/ tree supplies the task-level contracts below the entry skill. It covers routing, setup and maintenance, new-work decisions, platform adaptation, technical audits, hardening, critique and polish, layout and typography, hooks, live setup, and visual asset production. These workflows remain separate from the launcher and runtime files so the entry skill can route a request without embedding every task procedure.
The document workflow records an existing visual system in the standard DESIGN.md shape and keeps extra design data in its sidecar. The new-work and visualization contracts separate comp-led work, which carries an approved visual reference through measured asset and review stages, from code-led work, which carries its direction and interaction promises into the finish review. The packaged agent and degraded references preserve the same handoff rules when native delegation is unavailable.
The hook reference defines a project-local detector service. Supported clients install provider-specific hook manifests or settings entries, then use the packaged launcher to scan edits and report findings through each client's event model. Per-edit checks focus on immediate findings. A deeper stop-time pass runs where the client can return that context, while clients without that support keep the full detector pass per edit. Projects can add server-side template extensions to the detector, and manual scans still use detector filters even when automatic hooks are off.
The live reference defines an ordered session contract. It covers boot, page selection, polling, variant delivery, accept or discard, recovery, and cleanup. Boot resolves an app root for the session, which keeps later helpers attached to the same project in a monorepo. One-time live setup defines the project file targets, framework adapters, and development-only CSP allowance before injection. A command can be present in the entry table and still be unusable if its workflow reference or required launcher path is absent from the installed payload.
References: plugin/…/audit.md• L1-9 plugin/…/craft-floor.md• L1-18 plugin/…/distill.md• L1-26 plugin/…/doctor.md• L1-11 plugin/…/document.md• L1-7 plugin/…/live-setup.md• L1-37 plugin/…/hooks.md• L3-21 plugin/…/live.md• L9-20 plugin/…/live.md• L32-50 plugin/…/new-work.md• L1-16 plugin/…/new-work.md• L91-118 plugin/…/new-work.md• L136-154 plugin/…/optimize.md• L1-20 plugin/…/visualize.md• L1-16 plugin/…/android.md• L1-5 plugin/…/finish-reviewer.md• L16-34
Runtime assets
The packaged runtime uses a native engine launcher instead of shipping detector and command implementation as a large set of JavaScript files. plugin/skills/impeccable/scripts/impeccable selects an explicitly requested binary, a matching sibling platform binary, a trusted home or version-pinned cache binary, or a compatible binary on PATH. It uses an engine probe to reject old or mismatched candidates. When none exists, it downloads the version named by VERSION. A fresh download and an existing cache must pass a SHA-256 check from engine-checksums.txt before use; a verified size-and-digest stamp avoids rehashing an unchanged cache. The launcher needs no Node runtime.
The launcher exports the installed skill directory before it selects any engine. This lets every accepted engine find the packaged references and command metadata. An explicit IMPECCABLE_BIN choice bypasses candidate probing, while sibling and unversioned candidates must answer the engine probe and match the pinned version where required.
The launcher currently pins engine version 0.1.23 in plugin/skills/impeccable/scripts/VERSION. This engine version is separate from the plugin release version. The Windows companion impeccable.cmd follows the same candidate checks, platform fallback, checksum rules, and version-pinned cache behavior. Its download path stages each run separately before placing the verified binary, so concurrent first use cannot consume another run's partial file. The checksum manifest records trusted digests for the supported Darwin, Linux, and Windows engine assets. Release builds can stage platform binaries into dist when IMPECCABLE_BUNDLE_ENGINE=1 is set. The normal release shape remains launcher-only, so an installed client can fetch its matching engine on first use.
Live mode carries a matched browser runtime group. live-browser.js owns the injected live UI, selection flow, annotations, variant progress, and browser-to-server session behavior. The browser script receives the live command vocabulary and UI mount contract from the server, so those values are part of the server-to-browser contract rather than a second hand-maintained command list. live-browser-session.js preserves session and scroll state. live-browser-ignores.js applies project detector waivers to live scans. live-browser-dom.js supplies shared page selection, DOM lookup, focus, mounting, and modal top-layer helpers. The browser script refuses to start safely when the session or DOM helper is missing, so these assets must be shipped and loaded together.
Project-local mirrors preserve selected project state during a rebuild. The sync replaces generated skill content while keeping live-mode config.json that belongs to the consuming project. This prevents a release build from carrying one project's state into another project.
References: plugin/…/impeccable• L1-18 plugin/…/impeccable• L21-33 plugin/…/impeccable• L82-110 plugin/…/impeccable• L112-180 plugin/…/impeccable• L182-299 plugin/…/impeccable• L300-319 plugin/…/impeccable.cmd• L1-23 plugin/…/impeccable.cmd• L35-101 plugin/…/impeccable.cmd• L103-173 plugin/…/impeccable.cmd• L229-251 plugin/…/VERSION• L1 plugin/…/engine-checksums.txt• L1-9 plugin/…/live-browser.js• L16-30 plugin/…/live-browser.js• L78-91 plugin/…/live-browser.js• L105-132 plugin/…/live-browser.js• L292-300 plugin/…/live-browser-ignores.js• L1-9 plugin/…/live-browser-session.js• L1-7 plugin/…/live-browser-dom.js• L1-7 scripts/build.js• L548-584 scripts/…/utils.js• L1-16 scripts/build.js• L717-741
Client assembly
Provider outputs
The build creates a provider output for each supported client. Most clients use an Agent Skills-style folder under a dot-directory. Clients with extra features receive native agent or hook files.
This diagram shows how one source skill becomes project mirrors, the marketplace package, and the separate OpenAI package.
flowchart TD
source["skill/SKILL.src.md<br/>references, scripts, agents"] --> render["Provider rendering"]
render --> mirrors["Project mirrors<br/>.claude, .cursor, .agents, and others"]
render --> claude["Claude and Grok plugin<br/>plugin/"]
render --> codex["Codex output<br/>dist/codex/"]
codex --> openai["OpenAI plugin staging<br/>dist/openai/"]
mirrors --> universal["Universal archive"]
claude --> universal
| Client | Project folder | Notable native output |
|---|---|---|
| Cursor | .cursor | Agents and hooks |
| Claude Code | .claude | Agents and settings hooks |
| Gemini CLI | .gemini | Shared skill files |
| Codex | .codex | Nested skill agents and hooks |
| Codex repo skills | .agents | Nested skill agents |
| GitHub Copilot | .github | Custom agents and hooks |
| Kiro | .kiro | Shared skill files |
| OpenCode | .opencode | Shared skill files |
| Pi | .pi | Shared skill files |
| Qoder | .qoder | Shared skill files |
| Trae | .trae or .trae-cn | Shared skill files |
| Rovo Dev | .rovodev | Shared skill files |
| Mistral Vibe | .vibe | Shared skill files |
| Grok Build | .grok | Agents and hooks |
| Antigravity | .agent | Shared skill files |
| Hermes Agent | .hermes | Agent Skills fields only |
The root sync copies generated folders for direct project or repository installs. It excludes the .codex root mirror because Codex's current repository and user install path is .agents/skills; the Codex preview remains under dist/ for package assembly.
References: scripts/…/providers.js• L12-76 scripts/…/providers.js• L88-206 scripts/build.js• L717-764
Marketplace package
The Claude and Grok marketplace package is built from the Claude provider output. The build writes the Claude and native Grok manifests, copies the rendered skill and Claude-compatible agents, and writes a plugin hook manifest. The package is intentionally small so a marketplace cache does not need the full repository.
The plugin hook prepares a session, runs the detector after supported file edits, and runs a deeper pass at stop when the client can return stop-time context. Each event resolves the launcher through the installed plugin root. This makes the hook work after the marketplace unpacks the package into an arbitrary cache path. The edit and session checks use a short timeout, while the stop pass has a longer timeout for the deeper scan. The installed reference also carries the detector extension and manual-scan rules that keep project-specific templates and waivers consistent.
The marketplace hook contract now identifies the VS Code harness and provides both Unix and Windows command forms. Hook events use CLAUDE_PLUGIN_ROOT to reach the installed launcher, so the same package remains usable after extraction to a different cache path.
Hook behavior varies by client. Claude Code, Codex, Cursor, and GitHub Copilot use post-edit reporting. Grok Build reports findings at stop, and Hermes returns per-edit findings in the tool result while using its plugin for the stop pass. Clients that cannot feed stop-time context back to the agent keep the full detector pass per edit. The package therefore delivers one detector contract through provider-specific hook events rather than assuming one shared event shape.
References: scripts/build.js• L786-887 scripts/…/hooks.js• L119-145 plugin/…/hooks.json• L1-56 plugin/…/hooks.md• L3-21
OpenAI package
The OpenAI package starts with the Codex output in dist/codex/.codex/skills/impeccable. The stage creates dist/openai/impeccable, writes the Codex manifest, copies the icon, copies the Codex skill and nested agents, and writes the OpenAI hook manifest. A ZIP is then created from that staged directory.
The package uses ${PLUGIN_ROOT} for its hook path. This keeps the hook tied to the installed OpenAI package rather than to Claude or project skill paths.
References: scripts/…/openai-plugin.js• L26-60 scripts/…/hooks.js• L138-145 scripts/build.js• L892-896 scripts/…/zip.js• L20-61
Universal archive
The universal archive combines provider output folders into one directory. Its README tells an installer which hidden folder belongs to each client. An install copies only the relevant folder into the consumer's project root, while Codex uses .agents/skills rather than .codex/.
The archive is created only after provider rendering and universal assembly complete. A missing source directory, empty archive, or zero-byte archive fails the build instead of producing an installable-looking download.
References: scripts/build.js• L613-667 scripts/…/zip.js• L24-56
Installation paths
Marketplace installation
Marketplace installation uses the committed plugin/ subtree as the source. Claude Code reads its .claude-plugin/plugin.json, then discovers the skill, agents, and hooks from the package. Grok can use its native manifest or its Claude-compatible plugin path.
The end-to-end test copies this subtree into an isolated temporary marketplace, installs it through the real Claude CLI, and checks the component inventory. This tests loader behavior that static manifest validation cannot prove.
References: scripts/build.js• L786-792 tests/plugin-e2e.test.mjs• L58-101 tests/plugin-e2e.test.mjs• L115-142
Project installation
Project installation uses generated dot-directories or the universal archive. The build mirrors provider output into the project root and writes client-specific hook locations when the client supports hooks. The consuming client then loads the skill from its supported folder.
This path differs from marketplace installation because hooks are project-scoped. Plugin hooks resolve from the installed plugin root, while project hooks resolve from the project's client folder. Some clients also require a trust or enable step: Hermes uses a user plugin and project skill trust, while Copilot reads its repository hook after the file reaches the repository's default branch. A package change must not assume that one path serves both modes.
References: scripts/build.js• L717-764 scripts/…/hooks.js• L5-18 scripts/…/hooks.js• L147-190 plugin/…/hooks.md• L17-27 plugin/…/hooks.md• L103-115
Change checklist
When these files or sections are changed, remember to consider these:
- A version change in plugin/…/plugin.json or packaged SKILL.md must also reach .claude-plugin/marketplace.json, generated skill frontmatter, the Grok manifest, and existing OpenAI metadata.
- The engine version in
plugin/skills/impeccable/scripts/VERSIONis separate from the plugin release version. A change must stay aligned with launcher candidate checks, bundled binaries, and every matching entry in engine-checksums.txt. - A command addition, removal, or rename must stay aligned across packaged SKILL.md, command-metadata.json, its reference playbook, and provider-specific pin behavior.
- A change to the document workflow must keep the canonical DESIGN.md sections, token frontmatter, and .impeccable/design.json sidecar contract aligned with the packaged command guidance.
- A change to hook event behavior or provider support must stay aligned across the hook reference, provider manifests, event-specific output behavior, the session-start contract, the project-local config contract, and the Windows and Unix command forms. Do not treat
hook.enabledas the switch for manual detector scans. - A new agent or an agent rename must update provider filters, emitted filenames, native agent payloads, degraded references, and the Claude plugin's auto-discovered
plugin/agents/copy. - A launcher or engine change must be checked in provider output, plugin hooks, optional bundled binaries, checksum entries, and the OpenAI
${PLUGIN_ROOT}hook path. Keep the shell and Windows launchers aligned, including probe checks and safe concurrent download staging. - A change to live browser state, injected globals, UI surfaces, event names, or live detector filtering must keep live-browser.js, live-browser-session.js, live-browser-ignores.js, live-browser-dom.js, local server assembly, and the installed runtime asset list aligned.
- A change to live setup targets, framework adapters, or CSP guidance must stay aligned with the installed live reference and the project-local configuration contract.
- A change to Codex skill layout must be checked in both
.agents/skillsproject installs and the separately staged OpenAI package. - A new manifest field must be verified against the real client loader before it is added to the Claude plugin manifest allow-list.
Important failure modes
| Trigger | Consequence | Guard |
|---|---|---|
The Claude manifest declares an agents field | Claude Code loads no agents or rejects the plugin | Manifest shape validation and the real Claude loader test |
The Claude skills value loses its trailing slash | Slash commands do not register | Manifest shape validation requires ./skills/ |
| A version-bearing manifest or packaged SKILL.md is stale | An installed package reports a version that does not match its content | Version collection compares generated files with the package manifest |
| The launcher accepts an old sibling, PATH binary, or retired CLI | Commands or hooks fail with the wrong engine | Engine-probe checks and sibling version matching before execution |
The launcher VERSION does not match an available engine or checksum entry | Commands and hooks cannot start after fallback or offline installation | Keep the pinned version, launcher logic, bundled binaries, and engine-checksums.txt aligned |
A source agent is not copied to plugin/agents/ | Marketplace installs cannot discover that role | Agent parity validation compares source and emitted names |
| The motion reviewer is missing from the packaged agent payload | Motion work loses its evidence-based review step | Keep the packaged agent set aligned with the supported delegation roles |
| A generated degraded agent is edited instead of its source definition | The next render removes the fix or leaves provider payloads out of sync | Treat degraded files as generated output and change the source agent |
| A workflow reference, command metadata entry, or required launcher asset is absent | A command appears in the skill but fails after installation or shows incomplete help | Package assembly and provider payload checks |
| Live browser helpers, injected vocabulary, or the UI mount contract do not match the local server contract | Live mode cannot start safely or cannot show the expected controls | Browser-part assembly checks and live runtime validation |
| live-browser-session.js or live-browser-dom.js is missing or loads after live-browser.js | Live mode refuses to start | Runtime asset assembly and browser dependency guards |
| The launcher cannot find or download a matching engine | Commands and hooks stop before they can run | Explicit binary, sibling, cache, PATH probe, and verified download fallback |
| A downloaded engine has no usable checksum entry or its hash differs | The launcher refuses to run the new binary | SHA-256 manifest verification in both launchers; preinstall a trusted binary when offline |
| Two first-use launches share a partial download | One hook can read another hook's incomplete engine | Per-run staging and atomic placement before cache use |
| The OpenAI stage reads the Claude payload | Codex receives Claude paths or slash-command syntax | OpenAI staging requires and copies only the Codex payload |
| A hook points at a project path inside a plugin package | The hook misses its launcher after installation | Plugin hook manifests resolve through the installed plugin root |
| A client receives the wrong hook event form | Findings are hidden, edits are blocked unexpectedly, or the deep pass never reaches the agent | Provider-specific hook manifests and the packaged hook reference |
| Automatic hooks are disabled and a manual scan is treated as disabled too | A requested detector scan silently skips configured filters or uses the wrong scope | Keep hook.enabled separate from detector.* settings |
| A generated or universal archive is empty | Users download an archive that cannot install the skill | ZIP creation rejects zero entries and zero-byte output |
Key files
| File | Why to read it |
|---|---|
| skill/SKILL.src.md | Canonical skill metadata and shared instructions |
| skill/…/*.md | Source roles that become native agents or degraded references |
| plugin/…/plugin.json | Manifest consumed by the committed Claude marketplace package |
| plugin/…/plugin.json | Native Grok manifest in the same distributable subtree |
| plugin/…/SKILL.md | Installed command surface, routing rules, and launcher contract |
| plugin/…/document.md | Canonical design-system document and sidecar contract |
| plugin/…/hooks.md | Project hook behavior, detector tiers, template extensions, provider limits, and config contract |
| plugin/…/live-setup.md | One-time live injection, framework adapter, and CSP setup contract |
| plugin/…/live.md | Live session flow, app-root resolution, and browser runtime contract |
| plugin/…/new-work.md | New-surface build paths, direction contract, and finish handoff |
| plugin/…/visualize.md | Comp approval, measured asset flow, and raster provenance |
| scripts/…/utils.js | Source collection, project-state exclusion, and script metadata rules |
| scripts/…/providers.js | Supported client folders and provider capabilities |
| scripts/…/factory.js | Shared rendering model for skills, references, scripts, and agents |
| scripts/build.js | Assembly of provider outputs, root mirrors, plugin trees, and archives |
| scripts/…/openai-plugin.js | Separate OpenAI package staging from the Codex payload |
| scripts/…/hooks.js | Hook locations and installed-root path contracts |
| plugin/…/hooks.json | Marketplace session, edit, and stop hook events plus plugin-root launcher paths |
| plugin/…/impeccable-asset-producer.md | Claude marketplace asset-production payload and its approved-comp handoff contract |
| plugin/…/impeccable-documenter.md | Claude marketplace design-system recording payload |
| plugin/…/impeccable-finish-reviewer.md | Claude marketplace finish-reviewer payload and its evidence-first review contract |
| plugin/…/impeccable-manual-edit-applier.md | Claude marketplace live source-edit payload and atomic edit contract |
| plugin/…/impeccable-motion-reviewer.md | Claude marketplace motion-review payload and its evidence-based verdict contract |
plugin/skills/impeccable/scripts/impeccable | Unix launcher, engine probe, fallback path, and verified download |
plugin/skills/impeccable/scripts/impeccable.cmd | Windows launcher, engine probe, fallback path, and verified download |
plugin/skills/impeccable/scripts/VERSION | Pinned engine version used by the launcher |
| plugin/…/engine-checksums.txt | Trusted SHA-256 digests for supported engine downloads |
| plugin/…/live-browser.js | Installed browser-side live interaction, session startup, and chrome lifecycle |
| plugin/…/live-browser-ignores.js | Browser-side project-scope resolution for live detector ignores |
| plugin/…/live-browser-session.js | Live session and scroll recovery state |
| plugin/…/live-browser-dom.js | Shared live page selection, mounting, focus, and DOM helpers |
| tests/plugin-e2e.test.mjs | Real Claude installation check for skills, agents, and hooks |
| tests/validate-plugin-manifest.test.js | Static checks for the Claude loader contract |