ai-progress-workspace

# AI progress workspace Use this skill when the product needs a central editable artifact and a truthful view of long-running AI work. Do not use it for decorative loaders or ordinary request/response chat. ## Route - For event names, persistence, replay, cancellation, and failure semantics: read [references/event-contract.md](references/event-contract.md). - For choosing an editor or canvas and wiring tools to artifacts: read [references/workspace-implementation.md](references/workspace-implementation.md). Load both only when implementing the end-to-end system. ## Invariants - Display only observable work: a job, tool call, API operation, artifact mutation, failure, retry, approval, or completion event. - Never present private model reasoning or invented exact percentages as telemetry. - Give jobs, events, artifacts, and tool calls stable IDs. Persist enough state to restore after refresh and replay missed events. - Insert model output into the workspace through validated structured data, not unparsed prose. - Long-running work needs cancellation, recoverable failure states, and idempotent retry behavior appropriate to its side effects. - Prefer the existing application stack and the simplest orchestration model that satisfies the workflow. Completion means a real job updates the progress surface and editable artifact, refresh/replay works, cancellation and a forced failure behave safely, invalid artifacts are rejected, and all affected paths pass after fixes and rechecks.