Elyra Desktop 0.4 adds a context inspector: see the context window broken down block by block, and exclude the boulders without deleting them.
0.3 adds line provenance — every agent-written line links back to the prompt that produced it, stored as a commit trailer rather than line ranges — plus a project switcher and a file explorer.
0.2.0 adds a searchable archive over every session you ever had, git worktrees so several agents can share one repository, and a tree-sitter symbols index so the agent reads ranges instead of whole files.
Elyra Desktop is the same agent as the terminal edition — same loop, same tools, same providers, same sessions — in a native window for macOS and Linux, sharing one directory.
Elyra has kept a local ledger of every model turn for a while. 0.9.37 turns it into recommendations — switch, pin, or keep for each routing tier — and stamps every agent commit with an Elyra-Cost git trailer showing what it took to produce.
Git records what changed. The reasoning lives in a Slack thread, a meeting nobody wrote up, or an AI session that gets compacted by Thursday. Elyra 0.9.36 gives architecture decisions a home in your repo — one ADR file per decision, versioned in git, and surfaced back to the agent every session.
Elyra 0.9.35 adds /thinking — a command that sets the reasoning effort level directly, listing only the levels your current model actually supports.
Claude Fable 5.1 lands in Elyra on day one — and the default still stays Sonnet 5. Here's why that's a philosophy, not an oversight.
Two quiet fixes in Elyra 0.9.33: mid-turn compaction that runs before the provider rejects your request, and manual overrides for terminal capability detection.
Five releases in under two weeks, none of them a headline on its own. Two-way extension management, visibility into forgotten sessions, a 400 error that taught us about prompt-cache anatomy, warnings that only fire once, and a model registry that kept pace with Grok 4.6, GLM 5.3 and Gemini 3.7 Flash.