Every version, and what it changed for you.

One call where the first pass over unfamiliar code used to take four.

Before you upgrade — Start a new agent session after upgrading. The tool set is announced when a session opens, so a session that is already running keeps the one it was given.

• NewA new tool, chi_context, answers a whole question about your repository in one call. An agent meeting an unfamiliar area used to work through a sequence — rank the candidates, outline the files, project the symbols, scan for the literal — and every one of those calls re-sent the entire conversation before it could ask anything, so the sequence was paid for three or four times over before any of its answers had been read. It now runs inside the server, where the passes cost time instead of turns. Measured on 0.32.0 across five real questions about a working repository, the same answers arrive for 2.4x to 4.4x fewer input tokens, and four of the five are as complete as the hand-written sequence was. The fifth is a ranking miss the older path had too, and the answer says so rather than dressing it up.
• NewIt takes an intent — understand, debug, edit, test or review — and that is an allocation rather than a hint. Explaining a subsystem and patching one function want opposite things from the same budget, and the caller is the only party that knows which one it is doing. edit returns the target function with its exact source, the callers it must not break, the signatures it has to keep honouring, the tests that cover it and the code either side of it: what a patch actually needs, and what otherwise cost four more calls. Each answer also carries how confident the selection is, and a low-confidence answer names the call that recovers it — because the recovery an agent invents for itself is reading files one at a time.
• ImprovedEvery turn now costs less before a word of it is useful. A model pays for the description of every tool on every turn, whether it calls one or not, so the announced set is smaller: compact_select and get_source came off the default roster — not because they stopped working, but because chi_context already runs the same ranking and returns more around it, and get_symbols already returns exact source when asked for it. Both remain callable and both are still on the full roster. Together with a shorter opening note, what a turn pays before it says anything fell from 3,513 tokens to 2,941, measured on 0.32.0.
• ImprovedAsking whether an exact string appears now searches the whole index rather than the best-ranked couple of dozen files. Scoping it to the ranking made the one question that exists to work around a ranking miss depend on that ranking — and widening it costs almost nothing. A literal you quoted that genuinely is not there now comes back as an answer, which usually means the text is assembled from a format string and no whole-string search will ever find it.
• FixedAn index that keeps itself current is no longer reported as a stale one. Archis opened every session by stating how long ago the graph was last written, and a bare timestamp reads as one thing: the files you edited since are missing from it. Agents acted on exactly that — announcing the index was hours old and going off to read the changed files by hand, which is the single outcome that line exists to prevent. Nothing was ever stale, because the tree is re-checked before every answer. Archis now says what the timestamp means, and where the age genuinely is the story — automatic indexing switched off, or a repository with no index yet — it says that instead, with the command that fixes it. The warning attached to an answer Archis could not check against your files changed for the same reason: an unreadable tree is something to repair, not a reason to go back to reading everything by hand.