Primary prose is too long to read: [agree:] and [observation] blocks run to paragraphs, and pr-watch publishes them verbatim #141

What

The step-3 [observation] and the [agree:] verdicts have no length guidance, and in a long review the primary writes each one as several hundred words of standing bounds, mechanics notes and meta-commentary about the protocol itself. In PR mode those are published to the author verbatim — pr-watch's _render_block copies reviewer and adjudicator words exactly on purpose ("a daemon that paraphrases a finding is one nobody can trust"), so it cannot trim them without breaking the property it exists to hold.

The result on a 14-round review: a single [agree:] response running ~350 words, of which the part the author can act on is one clause; observations that are almost entirely notes to the gate about interpreter availability, tee seeding, quarantine bookkeeping and round-bound edits. A human reading the PR comment does not read any of it.

Example

From a round-14 comment (one [agree:] verdict, abridged):

Confirmed — the same defect as codex-rd14-r1, reached independently by the other vendor from the same direction, and high is the better of the two severities: severity is consequence-if-true, not likelihood... Two of the three examples hold exactly as traced... The third does not — north.farm//promo blanks to //promo, and _LINK_RE's (?<!:)//[\w] alternative matches... The half of the evidence line to act on is the last clause: the push adds four acceptance cases and no refusal case...

Everything the author needs is "confirmed, high, and the push added no refusal case for a separator-then-separator continuation". The rest is adjudication reasoning that belongs in the artifact, not in the author's inbox.

And a representative observation, in full, is ~200 words of which the operative sentence is "codex contributed nothing this round".

Suggested direction

Not a formatting fix in pr-watch — verbatim is deliberate there. The prose has to be shorter at the point it is written:

• commands/multi-review.md step 3 ([observation], ~line 1038) and the [agree:] grammar: state a length bound, and say what belongs in a published block versus what belongs in the round's own record. Standing session bounds (no interpreter, blocked redirection, tee seeding), round-bound edits and quarantine bookkeeping are gate material and repeat verbatim every round; the author never needs them.
• Consider a distinct marker for gate-only notes so they stay in the artifact and are never composed into the PR comment — that would cover most of the volume without asking the primary to say less.

Related

pr-watch separately republished every round's observations on every comment, since the _ROUND fold covered findings only. Fixed on the pr-watch side (fold-stale-observations); that removes the duplication but not the per-block length, which is this issue.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MBFDiJrnWDv8JLda56bHiB