0.2.0

Two hundred and twenty commits since v0.1.62. The agent got smaller and more
honest, the transcript got readable, and updates now come from our own feed.

The tool surface: 118 became 16

A long tool list does not sit quietly in the model's context — it competes for
attention, and past about twenty definitions the model starts picking a
plausible wrong one or inventing a name. Thirty-four of ours were shell
wrappers, and a model drives a shell from training rather than from sixteen
descriptions it has to read first.

• The git, gh, npm, tsc, eslint, format, LSP, snapshot, memory and index tools
  are gone. execute_command replaced all of them.
• search_code works with no key of anybody's: the embedder now follows the
  keys that exist instead of defaulting to one nobody had.
• read_file returns a window with a map — 500 lines plus the file's symbol
  outline — instead of a 17,000-token dump.
• update_plan came back, because the harness could not see a plan written in
  prose. The transcript row is the tool call.
• sessions lets a chat read and message the app's other chats.
• A registry test fails above twenty tools, so adding one back is a decision
  rather than a side effect.

The transcript reads like a record of work

• One line per step, folded into a run with a sentence that says what happened:
  "Ran 3 commands (1 failed), read 2 files".
• Repeated calls collapse into one row with a count, and a repeat that was
  failing is marked — the loop is the finding.
• Short thoughts ride as a chip on the step they led to; long reasoning keeps
  its words.
• A run past eight steps windows down, and Inspect opens the whole truth in
  the rail.
• The plan docks over the composer; the answer stands before the receipts.

Updates come from our own feed

• The updater reads updates.vellumcode.com rather than a public GitHub repo,
  which keeps real client IPs in the logs and lets the feed file carry
  no-store while installers stay immutable.
• Release notes travel inside the feed, so the update modal is never blank.
• CI dropped macOS — ten times the minutes for an installer Gatekeeper blocks —
  and the Linux checks moved to a self-hosted runner.

The CLI is a real package

npm i -g vellum-code now works: the published package ships a bundle, because
Node refuses to strip types under node_modules. vellum task gained the
subagents it was missing, and background commands work without a terminal
emulator.

The model is told the truth more often

• One behaviour contract written three ways — terse for Claude, explicit for
  OpenAI, spelled out for open-weights models.
• The system prompt splits into a cacheable half and a per-turn half, so the
  ~4,100-token prefix stops being re-evaluated every turn.
• A tool call the model typed in its chat template's syntax now runs as the
  call it is.
• A SEARCH that matches twice edits nothing and says so; a failed SEARCH says
  where the file actually differs; CRLF checkouts match.
• A model repeating itself is stopped instead of streamed for an hour.
• Edit-time diagnostics for every language with a server, not just TypeScript.
• Compaction prunes on an absolute token floor, not only a fill ratio, and
  hands its casualties to the read-before-edit tracker.

Fixes worth naming

• A rebuilt agent came back with no memory of the conversation — a model switch
  made it answer as if the chat had just begun.
• Stopping a background session left its process running, so a dev server kept
  its port after the tool said it was stopped.
• A search engine's bot challenge was read as an empty web, and the agent gave
  up and guessed a URL.
• A project folder that is gone says so instead of throwing.
• Every project-less chat gets its own sandbox instead of sharing one folder.