v0.50.2

Release Notes — v0.50.2

Patch: A search that skipped the map gets one reminder

The prompt makes codebase_map the first exploration step, but nothing below it
noticed when a turn skipped the map and grepped the whole tree instead. A search
returns matches in one call and reads as a successful step, so a turn could work
from match lists without ever learning which branch a file lives in — the
question the map's folder labels answer.

Searches are watched now. When the project index holds files, the agent holds
codebase_map, and a grep or glob result lands before anything has located
files — no codebase_map call with subdir, no read, no file_map naming a
file — that result carries one reminder to descend. It rides the stored tool
output, so the model reads it on its next step, and it is sent once per turn: a
second copy would add context without adding a decision. The search still runs.
The matches were true when it returned; what the model is missing is the map's
answer to a different question.

Patch: The exploration rules are one section instead of four

"Map before you read" was stated three times under three Mandatory headings,
and a fourth mandatory heading covered reading files with the tool rather than
the shell. On one subject, five headings competed for attention and the longest
read like a policy rather than a procedure.

They are one section now, Explore before you read, built around the three
moves it is actually asking for: find the files (the map when an index exists,
glob and grep when it does not), map every file the turn expects to touch in one
block, then read the ranges in the next. Map a file before reading it survives
as a subsection, the read-not-the-shell rule loses its top-level heading, and
the workflow diagram gained the linter.

Patch: HISTORY.md joins MEMORY.md, HOWTO.md and AGENT.md

A project's memory was outgrowing the file that carries it. MEMORY.md is
loaded into every turn's cached prefix, which makes a dated measurement or the
record of an experiment expensive to keep and useless in most turns, and the
prompt said nowhere else to put one.

HISTORY.md is that place: dated changes, measurements with their numbers,
experiments and what they ruled out. Like HOWTO.md it is never loaded — the
prompt carries a constant pointer and the agent opens it when a task touches
past work or before re-running an experiment — so the file can grow without
rewriting entry [0] of the prompt. The MEMORY.md section was rewritten around
the boundary, with a size budget and one instruction that matters more than the
rest: when a line is a story about what happened rather than a fact to act on,
move it out — do not delete facts to hit a size.

Patch: The plan pill moves only when the statuses move

update_intent says the statuses are what the user reads, and a plan re-sent
with every status unchanged has updated nothing — but the tool's own description
and its status schema field said only what the four values meant, never that
they were expected to change. Both say it now, and so does the prompt, which
asks for the closing call to carry every step's real status instead of a
sentence about where the turn got to.

Other changes

• The build agent's workflow step 5 is three lines instead of seven, the edit
  contract points at Parallel tool calls rather than restating it, and the
  clarifying-question rule names ask_user.
• Public copy across the homepage, the OGX page, the docs site and the blog no
  longer uses em dashes (colons, semicolons and parentheses carry the break),
  and the docs pages were refreshed alongside: MEMORY.md, HISTORY.md and
  HOWTO.md are described as they now behave.
• HOWTO.md's Install ogcode on a host gained the post-install boot check and
  the reminder that replacing the binary restarts nothing.