Add specialist roles and original code intelligence, media and experiment skills

[jdavis-software]

Delivered expansion

Implement the selected reusable ideas from the Graft, Codebase Memory MCP, OpenMontage, Agency Agents and Scientific Agent Skills review without vendoring their engines or skill bodies.

Eight original skills: Codebase Orientation; Code Index Qualification; Reference Video Analysis; Media Timeline Assembly; Shot Continuity Review; Specialist Role Composition; Engineering Experiment Analysis; Dataset Readiness Review. Each has task-specific procedures, original examples, limitations and three explicitly not-run agent-host scenarios.

Eight original role presets: Frontend Engineer, Go Backend Engineer, Architecture Reviewer, QA and Accessibility Reviewer, Security Reviewer, Research Analyst, Video Producer, Release Verifier. Roles reference canonical skills, list deliverables/context/non-goals and requested mode, and do not grant permissions or launch workers. Reviewer/analyst presets request read-only operation; actual authority remains with the host.

User-facing and executable changes

• Searchable Roles directory and individual pages; keyboard, query restoration, reset, light/dark and mobile navigation; skill-to-role backlinks.
• catalog/roles.json is canonical, with strict role validation and generated roles.json/sitemap/build metadata.
• Existing Bundle Resolver gains roles and role ROLE commands using bounded source hashing, provenance and no-execution behavior. No new tool entry or installer.
• Two bundles: Code Intelligence and Orientation and Engineering Experiments and Data. Existing media/orchestration bundles gain relevant skills.
• Six focused refinements to existing context, canary, storyboarding, provider, lineage and metering procedures.
• Existing Publicationcheck gains local index, timeline, dataset, and descriptive experiment commands. They check supplied records against selected local hashes and preserve incomplete evidence; they do not call index engines, render media, authenticate approvals or compute inferential statistics.
• Original synthetic source/timeline/data fixtures demonstrate changed source and narration detection, complete task accounting, retries and missing metrics. Role resolution references existing canonical skills and their companions without installation.

Inventory: 121 entries / 112 original skills / 9 tools / 18 bundles / 15 categories, plus 8 separately counted roles. The 24 new scenario inputs bring the domain-input count to 318; they are not completed host evaluations.

Verification — passed

Final head f7fb788a45739383179affdf9e1a2ffcda2306f2, tree 50494e3e29e6556e78ecddc61bf21316e8bc0c85, tested as combined candidate 0830020ec3d1975e0c721793ca2925ab1738e00d against main 828f1101ae8ac6a0e91eaa7f08cd7c42d8ef9e61.

Final toolkit validation and existing React fixture qualification completed successfully.

Check	Result
Frozen root installation; catalog/category/role/bundle/source/README checks	Passed
Root Node suite	291 passed, including 29 new role tests
Sourcekit Python regression suite	80 passed
Publicationcheck Python suite	117 passed, including 47 new specialist cases
Astro/TypeScript and production build	Passed
Existing static-site/media checks and public HTTPS reader checks	Passed
Website Chromium browser/request cases	70 passed, zero unexpected/flaky/skipped
Existing independently locked React fixture qualification	Passed: install, type/build, unit/browser checks and source/lock preservation

The first full validation exposed an unclosed Base layout in the new Roles directory. The markup was repaired, then the full canonical pipeline ran again successfully. No failing tests were removed or skipped. No dependencies or lockfiles were changed.

Actual demonstrations and limits

python3 examples/specialist-collection/demo.py reads and modifies only owned disposable fixture files. It detects source-byte changes against a supplied index receipt and changed narration against timeline dependencies. It reports the fixed synthetic workload with baseline 3/3 accepted versus candidate 2/3, two complete timing pairs, explicit excluded task c, retry-inclusive paired differences of -4 and -12 milliseconds, and one unknown output-token cell. Those are hand-checkable fixture values, not evidence that any real agent or engine is faster.

The 29 role tests exercise registry identity, canonical selections, modes, unsupported fields, path/symlink bounds, missing helpers, content hashes, Git provenance and actual CLI behavior. The 47 specialist tests cover supplied index freshness/coverage/expected probes, frame timeline coverage/trims/assets/dependencies, exact CSV units/schema/workload/attempt identity, retries, failures, null metrics, unchanged source and CLI statuses.

Browser and visual evidence

Existing Playwright automation ran desktop Chromium 1440×1000 and iPhone 13-sized Chromium emulation. The final journey clicks home → Roles → Architecture Reviewer → Code Index Qualification → Code Intelligence and Orientation, with canonical source links and skill-to-role backlinks. Role search, URL restoration, empty/reset states, keyboard focus/activation, light theme, no-JavaScript discovery, all eight role routes, sitemap, new skill pages and extended bundle selections passed alongside existing regressions.

The final production-build screenshots were inspected for desktop/mobile Roles, light-theme Roles, and the mobile Architecture Reviewer contract. The existing dark/light lime design is retained; the header wraps on mobile, cards and role-contract columns stack, and inspected views showed no clipping or overlap. Document-overflow assertions passed. This is not Safari, physical-device, assistive-technology or actual agent-host certification.

Final build/browser artifact was downloaded and ZIP SHA-256 verified as 4357174ddf24b50d29f53a188b4de2a97bc0450ec7a2edc1c3a56eb59a8bc12d before inspection. The embedded Playwright report confirms 70 expected results, zero unexpected/flaky/skipped; build-info identifies the exact combined candidate and complete inventory above.

Preserved boundaries

No Graft, Codebase Memory MCP, OpenMontage, Agency Agents, K-Dense, new renderer or statistics package was installed or copied. The research record links pinned sources. There is no new scheduler or human-authored KB duplicate, root dependency/lock change or original-content license decision.

Index checking validates a supplied receipt and known expected probe results, not runtime index authenticity or exhaustive coverage. Timeline checking is a bounded local cut-only contract, not a render, decoded-frame analysis, visual continuity check or permission verification. The skills describe additional actual-host work required for those claims. Experiment summaries retain all tasks and explicit excluded pairs but calculate descriptive paired execution-time aggregates only: no p-values, population confidence, causal claim or adoption decision. Requested role mode does not sandbox a worker.

No private DriftGate/AvatarOps repository, configuration, data or internal MCP changed. No paid service, credentials or live agent execution. Temporary transfer/export workflows and patch fragments are absent from the final tree.

Contracts: docs/ROLES.md, docs/SPECIALIST_COLLECTION.md. Research: docs/research/CODE_ROLES_MEDIA_METHODS.md.

After the approved merge, main validation and Pages deployment must pass; the deployment verifies the exact public revision, role inventory, category/bundle membership and representative new/existing pages before publication is reported complete.