veche: unvalidated --option names make seats fail silently (COULD_NOT_DECIDE) unless they are short tokens #2136

What happened

--option accepts any non-empty text: the only checks are distinct names and at most 8 including do-nothing (.claude/skills/idea-critic/scripts/veche.py:170-172). The names are sent to every seat as one line, OPTIONS: a; b; c (:193), and each seat must echo every name verbatim as a key of its scores object (:195); scores_ok then looks the names up one by one (:133-137). A name containing ; makes that line ambiguous, and a sentence-length name (the docs suggest --option "<the claim>", references/council.md:22,34) is hard for a model to echo exactly. In our runs with claude -p seats, sentence-length option names came back keyed a and b; every seat was dropped as "unparseable or malformed answer" and the council ended COULD_NOT_DECIDE with only N of M seat answers parseable. --dry-run accepts the same arguments, so the mistake shows up only after the seats have run. Neither the --option help (:359) nor the docs state a naming rule. (Punctuation in --claim and --steelman is not affected; they are inserted as plain text.)

What you expected

A clear error at start-up, also under --dry-run, for example --option names must match [A-Za-z0-9][A-Za-z0-9_-]{0,39}; got 'ship it now; then measure', and help text that states the rule.

How to reproduce

Save as seat.py, a stand-in for a model that splits the OPTIONS: line on ; and scores every piece:

import json, sys
prompt = sys.stdin.read()
line = next(l for l in prompt.splitlines() if l.startswith("OPTIONS: "))[len("OPTIONS: "):]
scores = {name: {"impact": 3, "cost": 3} for name in line.split("; ")}
print(json.dumps({"verdict": "REVISE", "claims": [], "would_change_it": "x", "premortem": "x",
                  "kill_criterion": "x", "leading_metric": "x", "scores": scores}))

V=/path/to/perun/.claude/skills/idea-critic/scripts/veche.py
python3 $V --cmd "python3 seat.py" --claim c --steelman s --option "ship it now; then measure" --criteria impact:1,cost:1 --dry-run
python3 $V --cmd "python3 seat.py" --claim c --steelman s --option "ship it now; then measure" --criteria impact:1,cost:1

Observed: the dry run exits 0 with "verdict": "DRY_RUN". The real run exits 3 with "reason": "only 0 of 3 seat answers parseable", because the seat returned the keys ship it now and then measure, not ship it now; then measure.

Proposed fix (smallest)

Validate in the option handling at :170-172: reject names that do not match a short token pattern (the exact rule is the maintainers' call) with a Stop that names the offending option; this also covers --dry-run. Change the --option help (:359) and the examples in references/council.md (:22, :34) to short tokens, and say that the claim itself goes in --claim ("include the claim" currently invites a sentence). If free-text option names are wanted instead, send labelled options (option-1: <text>) and key the scores by label, as the critique round already does for answers (:281-285).

Environment

• deep-code-review version (from .claude/skills/deep-code-review/VERSION): 1.620.0
• Host agent (Claude Code / Cursor / Codex / other): Claude Code
• OS: Linux

Confidentiality

This report contains no secrets, real names, internal IDs, or private hostnames.