Every version, and what it changed for you.

A crash you can report in one click, and a plan that says what it covers.

• NewA crash now leaves behind a report somebody can act on. Archis used to print the one standard line a panicking program prints and stop, so the two facts that would have made it fixable — what the machine was, and where in Archis it died — stayed on your screen, and what reached us was "it crashed". Archis now prints a GitHub issue link with the version, the platform and the failure location already filled in. Nothing is submitted for you: the link opens a form you read first and decide about, and if you never open it nothing leaves your machine.
• Newchi doctor --copy puts the report on your clipboard, with the version and platform on the front — the first two questions anybody asks of a pasted report, and neither was in the checks themselves. It works on either side of the area, so chi doctor --copy agents and chi doctor agents --copy both do what you meant, and with --json it copies the JSON alone. The confirmation goes to standard error, so chi doctor --copy > report.txt still writes just the report.
• ImprovedWhen your plan does not cover the repository you opened, your agent now hears about it. chi index explains that in your terminal, where you typed something and are waiting for an answer; over MCP there is no terminal to explain it in, and refusing the connection would have been the tidy answer and the wrong one — a server that quits during the handshake reaches you as "MCP server chi failed to start" in a log, which reads as a broken tool rather than as a plan. Archis now carries one line instead: which namespace, on which host, and what to run. Every other case stays silent, because a line on every turn costs tokens on every turn.
• ImprovedThe files Archis puts in front of your agent are now ranked by how much of the answer they contribute, not only by their single strongest match. One good symbol used to outrank a file that put five near-misses in the ranking, which is not how somebody deciding where to look reads the same evidence. Measured on 0.33.0 against issue-shaped questions from 63 repositories the ranking was never tuned on, the file that holds the answer reaches the shortlist more often and sits higher when it does; a file with one match keeps exactly the place it had, so the rule pays for evidence rather than taxing small files.