# Hawk DRMP roadmap

Evidence date: 2026-10-25
The private runtime is deployed and verified. General public distribution, signed installers, and the full quality and hardware certification remain unfinished.

## Existing capability evidence
- Resources and workflows: Local authority, versioned resources, catalog and search, review and rights, planning, workflow execution, original-job recovery, and receipts. Limit: Full production-scale and failure certification remains separate.
- DaVinci Resolve: One shared integration exposes the recorded 458 native tools and 1,357 action entries across three MCP servers, including editing and media management. Limit: Exact schemas, installed SDK support, context, and job-completion requirements apply. Every upstream action was not individually exercised.
- Media generation and finishing: Local finishing and registered Grok, OpenAI, BFL, Runway, Google, Fish Audio, and Descript paths have operational evidence. Limit: Availability depends on the installed account, model, mode, and accepted configuration. Firefly and ElevenLabs are not configured in the reference installation.
- Native provider APIs: Agents can use 881 non-deprecated OpenAI, xAI, Runway, Gemini, BFL, Fish Audio and TwelveLabs APIs through one managed execution path, including native parameters, uploads, streaming and original-job polling. Limit: The count includes API versions. Normal discovery excludes 52 deprecated contracts retained only for explicit legacy recovery or migration. Provider permissions still apply: the reference xAI key permits image/video only; administrative operations need appropriate credentials. Signed URL transfers add 14 helpers.
- Local analysis: Tesseract OCR, DINOv2 and SigLIP embeddings, SigLIP semantic annotation, SAM 2.1 masks, OpenCV face observations, and MediaPipe 33-landmark body pose run through explicit workflows. Limit: Curated accuracy benchmarks and strict multi-character continuity remain open. Analysis does not approve identity or rights.
- Platform intake: Local file intake, a real Apple Photos export, the native Takeout matcher, and Android debug/emulator checks have evidence. Limit: Google Picker OAuth, selected real Takeout archives, physical-phone permissions, reboot, transfer, and release signing remain separate.
- Local model training: Canonical model.train admission connects frozen packages to the pinned local SimpleTuner CLI. Native training and installed background-worker collection returned the exact verified artifact. Limit: A one-step SDXL LoRA run verifies the connection and artifact return. Training-input access must be granted by the host. Model quality remains separate, and active training jobs take priority.
- Local rendering and conditioning: Agent-driven media.render uses installed Draw Things and ComfyUI with exact managed models, source policies, receipts and original-job recovery. Native evidence covers Draw Things Canny plus ComfyUI pose, IP-Adapter and the returned SimpleTuner LoRA. The LoRA loader records 560 applied layers at strength 0.8. Limit: The pinned Draw Things CLI cannot pass IP-Adapter reference hints; use ComfyUI. The fixed images showed subject-count and pose errors. This verifies execution, not strict identity continuity or production model quality.
- Usage reporting: The independent reporter collects scheduled usage for four configured billing accounts and delivers exact report revisions to a persistent MCP inbox with explicit acknowledgements. Limit: The inbox works without a UI or Events. Automatic chat notifications require host Events support; additional billing coverage needs real account access.

## Remaining work and why
- Additional model and control combinations: Accept any further model family, custom node, video mode, or renderer/control combination against its exact installation before use. Why: The installed SDXL and conditioning cases are verified; that evidence cannot certify every upstream model or extension.
- Account and delivery connections: Configure Firefly/ElevenLabs credentials, wider xAI key permissions when needed, Google Photos Picker OAuth, and any additional billing accounts. Why: These paths need real credentials or account consent; they cannot be supplied by documentation or tests.
- Quality and continuity: Use representative benchmarks with hard negatives, multiple characters, and clothing, pose, lighting, location, and style variation. The first fixed render set is recorded, including its subject-count and pose failures. Why: Successful execution proves the path works. It does not establish accuracy or safe strict-continuity thresholds.
- Real devices, recovery, and scale: Complete missing real archive/phone, separate-machine, reboot/failure, and large-library scenarios. Why: Component tests and emulator checks do not certify every production environment or workload.
- Public distribution: Finish signed/notarized packaging, Android release signing, public privacy/provenance review, and release certification. Why: The working private installation is a different milestone from a supported public download.

## Optional and separate areas
The UI remains optional; focused native-view and accessibility checks do not block agent operation. Peer delivery through HawkSpan is a separate integration and is not accepted in the reference installation. Automatic OpenAI Batch scheduling remains disabled; explicit native Batch requests require the task’s authorization. Concurrent provider rendering and coordinated parallel creative agents remain future options.

Do not interpret an old missing-capability note as current state. The permanent runtime-directory fix, broad Resolve integration, current Descript connection, installed local analysis providers, and the accepted local renderers and conditioning cases are completed work, not items to reimplement.