LLP 0113: the classification ask is a menu, not an open prompt

[philcunliffe]

Summary

The session-start classification question (LLP 0106) is delivered to the assistant as SessionStart additionalContext, and the copy only said "ask the user" - so whether the user saw a real selectable menu (AskUserQuestion) or an open-ended prose question was left to the model's judgment that turn. On 2026-07-20 a fresh folder (~/workspace/bingo) got the prose form, the user interrupted the turn before any options appeared, and the whole exchange read as "the menu never showed". A consent surface should not render differently run to run.

• New decision llp/0113-classification-ask-menu-presentation.decision.md: the classification ask must render as a native selection menu when the client environment provides one (in Claude Code, the AskUserQuestion tool); plain text is a fallback for tool-less clients, never a choice. Numbered 0113 because 0111/0112 are claimed on the hyp-policy-verb branches.
• buildClassificationPrompt (src/core/usage-policy/classification.js) now ends with an explicit presentation instruction, tool-neutral with the Claude tool named. The mandate lives in the one shared copy, so the "hooks differ only in delivery, never in decision or copy" invariant is untouched and Codex degrades to the plain-text ask unchanged.
• Consent-copy test pins the mandate (selection menu, AskUserQuestion, the fallback clause) alongside the class vocabulary and verb lines, per LLP 0106's "the prompt copy is load-bearing" consequence.
• LLP 0106: editorial forward-ref only (Extended-by: LLP 0113 on the prompt-copy consequence).

Testing

• npm test: 2260 pass, 0 fail (1 skipped, pre-existing).
• ref-check: no new errors; the corpus's 101 pre-existing failures are identical with and without this change, and none touch 0106/0113/classification.

🤖 Generated with Claude Code

[philcunliffe]

Review round 1 — clean

Verdict: clean. No blocking, major, minor, or nit findings at head 18681b7.

Reviewed the full diff (LLP 0106 cross-ref, new LLP 0113, buildClassificationPrompt copy change, and the pinning test) in an isolated detached worktree.

Checks performed:

• Repo style — no statement semicolons; no em dashes (U+2014) anywhere in the four changed files. Runtime string uses - (e.g. local-only). The escaped environment\'s renders correctly as environment's.
• Consent copy — verified by executing buildClassificationPrompt: still names the folder, all three classes, and each hyp ignore verb line. The new menu paragraph is coherent and unambiguous (mandates a native selection menu, names AskUserQuestion for Claude Code, permits plain text only when no such tool exists).
• Tests — test/core/usage-policy-classification.test.js passes 10/10; the classify-hook consumer suites (claude-classify-hook, codex-classify-hook) pass 12/12, so the copy change causes no consumer regression. The only failing suite under npm test is test/core/leave-command.test.js (7 fail), confirmed pre-existing on origin/master and unrelated to this diff.
• @ref honesty — @ref LLP 0113 [implements] on buildClassificationPrompt, @ref LLP 0113 [tests] on the assertions, @ref LLP 0106 [constrained-by] inside 0113, and the Extended-by backlink in 0106 are all accurate and attached to the right constructs.
• LLP 0113 doc — correct NNNN-slug.type.md filename, complete metadata header, valid Systems values, internally consistent, all cross-referenced docs exist.
• No behavioral regression — buildClassificationPrompt({ cwd }) signature unchanged; the only other caller is unaffected.

Mergeable ∧ green ∧ reviewed-clean at this head.

[philcunliffe]

Approved — mergeable, CI green, and reviewed clean at head 18681b7 (review record above). Holding for a maintainer to merge; neutral does not merge a contributor's PR.