Every version, and what it changed for you.

Archis connects the coding agents you install as they appear, shows you every file it has ever written on your machine, and names 96% of declarations in 26 of its 29 languages.

• Breakingchi support is gone. Hidden and deprecated since 0.29, it did exactly what chi status --languages does and said so on every run. chi status --support still works as a spelling of --languages.
• LanguagesPer-language coverage is measured over the whole benchmark corpus and published beside the declared score. Measured with 0.39.0 across 123 repositories, 26 of the 29 languages name at or above 96% of the declarations they see; the three below that line are named as such — GraphQL 94.3%, Zig 90.2% and PHP 89.0%. chi status --languages prints that figure instead of quoting a note about the single repository each language was first tried on. Support levels did not move: Stable still asks for the agent-level benchmark, which coverage alone does not meet.
• LanguagesDeclarations five language packs were counting but could not name. Dockerfile ARG and ENV names, Zig tests named by a string, GraphQL's anonymous operation, PHP properties and class constants, and proto map fields and oneof variants all reach the graph now — Dockerfile alone went from 103 symbols to 186 on its fixture. Proto and Ruby were also counting the keyword that introduces a declaration as a second declaration, which surfaced as coverage they had never actually missed; both are at 100%. A pack that counts a declaration it cannot name now fails a test rather than reporting itself as broken.
• NewA coding agent you install later gets connected on its own. chi init connected whatever was on the machine the day it ran and never looked again, so installing Codex or Gemini the following week left that agent with no registration, no briefing and no skill — and nobody told you. Archis now looks again wherever it already runs, gives an agent that appeared exactly what chi agents on would have given it, and tells you once what it wrote and the command that undoes it. chi agents off <id> is sticky: nothing reconnects an agent you took out, and chi agents auto off turns the whole behaviour off.
• Newchi config lists every file Archis has touched — on this machine, in this repository, or across every repository it knows — and how to undo each one. One row per file: where it is, which coding agent it serves, what Archis put there, whether it is current, stale or absent, and the command that takes it back out. A registration you made through a vendor's own tooling is listed at the location that vendor documents and marked as theirs. It reads and never writes, needs no licence, and --json returns the same document as data.
• FixedAn organization's repository on a Free plan is refused over MCP instead of served with a caveat. The server used to say the plan did not cover the repository and then answer anyway: every agent took that at its word and kept asking for the rest of the session, so the repository was served in full and nobody ever heard about a plan. The tools now refuse, tell the model to stop calling them and to tell you a Team plan is needed, and nothing is opened, indexed or read.
• FixedThe skill is written for every agent that reads one. Setup picked the first agent with a skills directory, so a machine running Antigravity or Pi was told no agent here reads skills — and the copy it wrote instead was unstamped, which made chi skill status report it as somebody else's and left the session hook refusing to refresh it.
• FixedThe closing summary says why an accounting is missing instead of printing a zero for the model to explain. Sessions were ending with lines like "3 tool calls with zero unmeasured query tokens" because the report offered a bare zero and the model reached for the one number it recognised. It now names which case it was — Archis was never called, the index was missing or could not be verified, or every call was one that replaces no file read — and a lapsed licence, which used to end a session in silence, now says so once.
• FixedA file that declares nothing no longer scores 0% coverage. A bash script of plain commands, a Groovy build file and a PowerShell script legitimately declare nothing, and scoring them as a total miss dragged their language's published mean down by their share of the corpus: bash read 47.1% and is 100%, PowerShell 55.5% and is 100%, Groovy 69.1% and is 99.8%. The extractors were never the problem.
• FixedCompression figures describe your repository rather than everything sitting beside it. chi measure walked the tree its own way instead of the way chi index does, so .git internals, node_modules, vendor/ and build output were measured too — 10.6% of every file it counted. Both walk the same way now, and the corpus-wide reduction measured with 0.39.0 is 80.25% of the tree the index actually holds.