v0.45.2

Changed

• The MCP server instructions are one short text, whole on every
  client. Claude Code shows an agent only the first 2048 characters of
  them; 0.45.1 put a must-know digest in front of a long form that ran
  to ~5.5k characters, most of it never seen there and all of it to be
  kept in step with the digest and doctor. The long form is gone: the
  instructions are one rule per line, at most ~1.7k characters on any
  combination of grant and client, and now include the three rules only
  the long form had — surface a plan in a line rather than re-arguing
  it, a granted verb you cannot see usually means the client needs a
  restart, and read the tool CAVEATS. doctor still carries the full
  host-denial remedy, uncut. A test fails the build if the text outgrows
  the cut.

Release review

Two review lenses (architecture and correctness), with two re-review rounds:

• Found and fixed: dropping the long form had lost four rules an agent acts on:

  • naming the write verbs on a read-only surface (without it, an agent can't tell ebman has DLQ tools and reaches for the AWS CLI);
  • that each write is still put to the operator to accept or decline;
  • that confirm_action may take as long as a person takes;
  • that a denied undo is urgent.

  All four are back, each pinned by the size guard, and the longest text is 1,693 of 2,048 characters. A test that could no longer fail (it checked for an upper-case word the text never used) now checks the real wording.

• Docs: headless.md listed queue peek and :why as missing from the MCP surface, but both are tools. commands.md counted five write verbs (there are eight). Four remaining places in the docs still claimed a person answers every confirmation, and now say the client is asked.

• Decided in this cycle: confirm_action will not carry requiresUserInteraction (two prompts per write, to save one setup step that is already verified).

• cargo semver-checks against 0.45.1 found no breaking changes, and no dependency changed.

Full Changelog: v0.45.1...v0.45.2