Capability map

Area	Maintained capability	Important boundary
Chat	Provider-qualified single-model sessions	Protocol/formatter isolation
AI Council	Teams, roles, runtime classes, workflows, human participation	Deterministic step state and bounded recovery
Providers	Ollama, OpenAI-compatible, OpenAI, Azure OpenAI routes	Endpoint-qualified identity and scoped credentials
Benchmarking	Bounded tasks, peer/self review, presets	Recommendations require user application
Projects	Versions, revisions, requirements, artifacts, topics	Database identity is separate from paths
Workspaces	Root resolution, policies, toolchains, build evidence	Fresh assessment before execution
Knowledge	Reviewed persistent entries and source refresh	Raw uploads/model output are not trusted automatically
DX functions	Typed callable application operations	Handler/service owns policy and execution
1-Wire	Peer discovery, identity, capabilities, work spooler	Discovery is not trust; remote flags are not approval
Embedded	Board/pin/transport plans and artifacts	Plan, compile, flash, and actuate are separate
GameDirector	Deterministic sessions with generative proposals	Models do not directly mutate state
Minecraft	Reviewable mod/datapack generation	Toolchain evidence required for build claims
Documentation	Conceptual docs, XML API, PDF, Pages	Publishes shipped static tree

Capability-gap contract

When LocalGPT cannot complete a requested task, it should report:

• the requested outcome;
• the missing dependency, source, API, framework knowledge, permission, or runtime capability;
• what evidence was inspected;
• what can still be produced safely;
• the next reviewable step;
• whether owner-side tooling, licensed dependencies, or human approval is required.

A gap report is not permission to self-expand or execute arbitrary installers.