Roadmap
This page separates completed repository work from work in progress, proposals still being evaluated, and directions intentionally kept out of scope. Roadmap Shipped can include work implemented in a source candidate; the install page and homepage separately identify the latest published package. Release records and GitHub issues refresh these categories when available.
Work grouped by status
Shipped
v0.8.45
Moonshot/Kimi provider support, API-key setup guidance, provider-surface sync, and current Windows install/runtime guidance
Underway
VS Code extension
The repository ships a Phase 0 local-runtime companion: terminal launch, health checks, read-only thread summaries, and restore-point browsing. Full chat and editor actions are not part of this slice.
Memory typed store
The native store has shipped: Markdown is the durable source of truth, with a rebuildable SQLite FTS5 index over it. Graph-structured agent memory and multi-signal recall are still ahead (#534–#536)
Feishu / Lark bot
First long-connection bridge over the runtime API shipped; richer chat features underway (#757)
Chinese-market & i18n
Locale-aware UI, platform refinements, region-specific search backends (#755)
Hugging Face model discovery + Model Lab
Browse, download, and manage models from Hugging Face Hub directly in the TUI
Considered
Workrooms
Durable, addressable agent-work threads over the Runtime API and user surfaces (#3209, docs/WORKROOM_ARCHITECTURE.md)
Exa web-search backend
Bundled alternative to the existing DDG + Bing path (#431)
Homebrew core formula
Tap exists; pursuing homebrew-core inclusion
Native Windows installer
MSI / WinGet; Scoop manifest already ships
Unsloth / NeMo / Arcee fine-tune integration
One-click fine-tuning workflows backed by Unsloth, NVIDIA NeMo, and Arcee toolkits
Ruled out
Silent or content-level telemetry
Anonymous usage counting is on by default, but the first-run disclosure and every opt-out are mandatory. What stays ruled out: conversations, code, prompts, files, model content, credentials, per-turn or per-tool timelines, collection hidden from the user, and any third-party ad or analytics SDK in the runtime binary. A selected hosted provider still receives the context required for its model turn; loopback routes can keep inference local.
Mandatory hosted relay for local sessions
The local runtime and bring-your-own-provider routes continue to work without sending sessions through a Codewhale service
Required account for the local runtime
Installing and running Codewhale locally requires no account
Sponsored model promotion
Model picker stays neutral — no paid placement
Public share links for local sessions
The retired share-link direction (#471/#481) is not coming back; any future sharing design starts fresh.
Open model platform
Community model registry
Discover, share, and rate community fine-tuned models with reproducible recipes
One-click deploy
Deploy any model to RunPod, Replicate, or your own infra with a single command
Model evaluation dashboard
Transparent, reproducible comparisons across providers, quantization levels, and hardware
Keep roadmap decisions in the open.
Use issues for bugs and well-scoped feature requests, Discussions for ideas that need shaping, and pull requests for concrete changes with tests or documentation. Verification across languages, platforms, and providers helps maintainers judge priority.