fix(pstack): list every waiting item so each ketchup call is decidable

[poteto]

Why

/ketchup capped the whole reply at 150 words and gave each action item two lines. When many items waited, the model folded a set of PRs into a count, such as "Merge the 12?", and moved the other calls to a linked file. The user could not decide from the reply.

What changed

• Cap only the story at 100 words, shorter when less happened. Give each waiting item about 30 words, so the reply grows with the item count: two items stay well under 150 words, and thirty need far more. When many items wait, the reply runs long. Never cut or shrink an item to stay under a word limit: no count in place of a list, and no link in place of an item.
• Fit the shape to the call. A call about one item is one bullet, with the item named once. When items share a call, ask it once, then give each item its own bullet: link, what it changes for users, evidence, and risk, even when the risk is low. Name the PR in every call.
• A call between options gives each option's gain and cost, and those costs are its risk. Say what waiting costs only when it is more than a delay. Write "unknown" only for a fact the call needs, and leave out a link you do not have.
• Ban drawings of action items and status lists. Keep stack, flow, and before-and-after diagrams.

Scope

• pstack/skills/ketchup, the matching poteto-help line and README row, and their regenerated Claude Code copies.
• Does not change /bro or the screenshot rules.

Blast Radius

Only /ketchup replies change. A large batch now makes a long reply, by design. On a small chat with two calls, replies stay near the old size, about 20% longer. The extra words are the evidence and risk that the old text left out.

Verification

• Three fixture chats, a frozen grader, and 50 runs on two models. Share of waiting items decidable from the reply alone, on the batch chat: 0.00 → 0.93 on both models. Every reply with the final text listed every waiting item.
• Blinded judges from a third vendor. On the two-call chat, all six final-text replies ranked above all five old-text replies. On the batch chat, the final text ranked above the previous revision, with higher brevity and accuracy scores.
• The README row changed after these runs. The skill text the runs measured did not.
• node scripts/validate-plugins.mjs and node scripts/check-pstack-docs.mjs pass, and node scripts/generate-pstack-harness-trees.mjs --check reports the harness tree up to date.

---

Note

Low Risk
Prompt-only changes to the /ketchup skill and help copy; behavior shifts to longer replies on large batches but no runtime code or security paths.

Overview
/ketchup no longer squeezes the whole catch-up into ~150 words. The narrative recap is capped around 100 words (shorter when little happened), while each thing waiting on the user gets ~30 words, so replies grow with backlog size instead of collapsing many PRs into “merge the 12?” or pushing detail behind links.

Action items are reframed as explicit calls with a shape that matches the decision: one bullet per single-item call (call, recommendation, user impact, evidence, risk); shared calls asked once, then one bullet per item (link/name, impact, evidence, risk, riskiest first). Counts can’t replace lists and links can’t replace items—the reply alone must be enough to decide. Option-style calls weigh gain/cost per option (that tradeoff is the risk); separate risk lines drop in that case.

Docs stay aligned: ketchup/SKILL.md frontmatter and rules, plus the /ketchup blurbs in README.md and poteto-help. Visual rules move slightly—no diagrams of action-item lists or status lists; stack/flow/before-after diagrams remain.

^{Reviewed by Cursor Bugbot for commit 55c6ce9. Bugbot is set up for automated code reviews on this repo. Configure here.}