v0.56.0

Release Notes — v0.56.0

Major: Find the conversation you half remember

The session list could only filter titles, so "which conversation was that about caching?"
had no answer unless you remembered what you had called it. Every turn already writes a
summary with its topics into .ogcode/memory, and that text was sitting there unread.
Press ⌘K and it is a search box: type a few words and the palette answers with the
conversations whose memory is about them, each with its matched turns, their dates and
the labels those turns were filed under.

The index is a new full-text table (migration 073, fts5 over title, labels and body),
written by the server rather than by a trigger, so a document and its search row land
in the same transaction. It backfills itself at boot from the project's memory
directory and again on the first query of a fresh process, and it stats the files it
holds before answering, so notes the agent wrote earlier in the same turn are
searchable immediately. Ranking is not one bm25 pass. Each query term is matched
separately and the results are unioned per conversation, so a question whose words
are spread across different turns still finds the conversation that answers it. A
best-turn score is multiplied by how much of the query the conversation covers, and
label matches, title matches, breadth across turns and a 30-day recency tail sit on
top. Two characters is the shortest query, a page is twenty conversations, and the
answer is always a list, never null.

As you type, the palette also fuzzy-matches titles locally, so picking up a
conversation you can name but cannot spell stays instant, and your pinned sessions
now read the same list from the sidebar and the palette. GET /api/session/search
is the same search over HTTP for anything that is not the palette.

Minor: The agent board tells idle apart from working

The board notice counted every agent that was not finished as still working, so a
session parked between turns read as one that was busy. The count is now split three
ways: running, idle between turns, finished. The tree also draws its newest rows
first, because the notice's cap keeps the head of the list and the agents you just
spawned are the ones you are waiting on.

Minor: The agent's housekeeping rows are records, not disclosures

codebase_map, file_map, compact_context and update_intent maintain the agent's
own state; none of them hands back something the user asked for. Their rows in the
chat used to look like every other tool card and open on click, which promised a
result that was not there. They are now flat one-line records: no chevron, not
clickable, not focusable, and they stay shut even while the tool runs. Every real
tool keeps its disclosure.