[ThinkOffApp]

What

petrus, 22 Sep 2026: "Stop these huge confirmation messages pls" — a Bash
confirmation's prompt (the question plus whatever the caller pasted after
it, often the full shell command or a heredoc) was posted verbatim into the
GroupMind room, running to hundreds of characters.

Adds shortRoomBody(prompt, {options}), a pure function in
src/confirmations.mjs:

• Keeps the prompt's first line (the question) verbatim.
• For a confirmation, reduces everything after the first line to the first
  non-plumbing line — skipping lines that are only a comment, a bare cd,
  a bare echo ..., or a bare VAR=value assignment — capped at
  COMMAND_SNIPPET_MAX (160 chars), followed by a fixed full command in CodeWatch note.
• For a choice card, the option labels (the buttons / typed-reply targets,
  /choose <id> <option>) are never shortened or truncated. If the options
  alone would already meet or exceed the budget, the description is dropped
  entirely rather than truncated.
• The whole shortened block is capped at ROOM_BODY_MAX (600 chars).

makeGroupmindAnnouncer now delegates to a new exported buildGroupmindBody()
helper so the assembled card is testable without a network round trip.

What is tested

• test/short-room-body.test.mjs (new, 10 cases):
  • A short single-line prompt passes through unchanged.
  • A prompt whose first line is a question keeps that question verbatim.
  • A 900-character command line is cut to the snippet cap with a pointer to
    CodeWatch, and the resulting body stays under ROOM_BODY_MAX.
  • Bare cd / echo / variable-assignment / comment lines are skipped to
    find the real command line.
  • A compound line that merely starts with cd/assignment but goes on to
    do real work (the actual shape from the reported card) is kept, not
    skipped.
  • Choice-card option labels are never touched by the shortening.
  • When options alone meet/exceed the budget, the description is dropped.
  • buildGroupmindBody() end-to-end: a 900-char shell command yields an
    assembled room body under 600 chars that opens with the verbatim
    question and still contains the intent id and every option label.
  • Negative control: the CodeWatch push still carries the full,
    untouched prompt — this is the test that would catch a future "fix" that
    truncates at the source instead of only in the room body.
• Verified each new behavioral test fails when the change is reverted
  (git checkout -- src/confirmations.mjs, temporarily): with the fix
  reverted, the whole test file fails to even load (missing exports), so
  every case fails; restored the fix and re-ran to confirm all 10 pass again.
• Full suite: npm test → 746 tests, 51 suites, 0 failures, exit code 0.

What is NOT touched / NOT tested here

• The stored intent's prompt field, the CodeWatch push payload, and the
  /intents HTTP listing all keep the full, unshortened text — only the
  text posted to the GroupMind room changes. This is the property the
  negative-control test guards.
• The GroupMind message metadata.intent_prompt field (used by richer
  clients, not the visible chat bubble) is left as the full prompt too —
  only the visible body is shortened.
• No changes to CodeWatch rendering, /intents HTTP handling, or the MCP
  tool schemas for request_confirmation / request_choice.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.