[MaggieAppleton]

Maggie answered one decision in a blank document and was left with a lone "⤭ Decided: … · @MaggieAppleton / No related text" line. While the Planner writes up several answers, the document showed a stack of these lines too.

Why it said "No related text"

When the Planner asks a question in a blank document, the question has no blocks to point at. The empty anchor set gets stored while the question is open (pending: false). Any commit's rebase saves it as it reads. Once the question is answered, that set reads as reviewed and deliberately empty. So the card says "No related text", and the question never appears in anchors_pending, which means nothing asks the Planner to write or link its prose. The fake Planner's ASK-DECISIONS reproduces this exactly (screenshot below).

Fix (server): answering a decision now marks each anchor set that points at nothing as owed (pending, missing) and sends plan:anchors to the room. A decision that already points at prose is unchanged.

flowchart LR
  A[ask in blank doc<br/>blocks: none] --> B[open: anchors [] pending:false]
  B -->|answer, before| C[reads as reviewed-empty<br/>'No related text', no review owed]
  B -->|answer, now| D[pending: missing<br/>on anchors_pending, relayed]
  D -->|Planner edit_plan + anchor_plan| E[linked: marker + wash]

Loading

What the document shows now

• Display only. A decided card shows in the document only through the prose it lives in, as the margin marker and wash. If it is pending, deliberately empty or orphaned, its decorator draws nothing. The projection node stays and records are untouched. The full card, its counts, Reopen and Discard stay in Decisions.
• Real in-progress work only. While a Planner turn is running, or a conversation write-up job is writing, for decisions that have no prose yet, the first of them shows one quiet line: "Writing up N decisions…". Decided cards that a deck still draws are not counted. The line goes away when the prose lands.
• Blank stays blank. When hidden decisions are all a document holds, the usual "Start writing, or ask Chopin to plan" placeholder shows in the first empty paragraph, at the same spot as in an empty document, and typing goes there. Lexical shows no placeholder next to a decorator node, so the page sets it. Backspace and Delete do nothing while the document is blank, so they can't strand the caret beside a hidden decision.
• Decisions tab. The write-up status and Retry for a conversation decision move to its Decisions card. "Linking…" now shows only while a Planner turn runs. Otherwise it says "Not linked yet".
• Open decisions are unchanged. Inside a deck, decided members use Fix the deck's card-to-card motion, keep decided cards full size, and bound the pager #540's full-size card. The settled line is gone: nothing draws it any more.

Before / after

Maggie's case: a blank document with one answered decision.

Before	After

It stays fully visible in Decisions:

Several decisions waiting on prose:

Before	After, Planner writing	After, prose landed

The same flow when the Planner ends its turn without writing: the document is blank (placeholder) and Decisions says the answers are not linked yet.

Document	Decisions

A linked decision still shows inline, as its marker:

Tests

• Unit: cardPresentation for inline, deck and list. Document visibility for each relationship state (pending, empty, orphaned, linked). The writingProse predicate. The write-up Retry on the Decisions card. Anchors.answer. A durability test that a question asked into a blank document still owes a review once answered, and that the change is relayed.
• E2E: the harness covers a blank document with answered decisions (placeholder shows, typing works, Decisions keeps the cards). The new ASK-DECISIONS THEN-WRITE fake Planner shows "Writing up 3 decisions…" and then the linked prose. Settled-line expectations are updated in the decision-marker, icon-size, inline-deck, sidecar-card-states and decision-prose suites. I ran the harness, decision-marker, icon-size, inline-deck and sidecar-card-states suites locally against my own servers. decision-prose.e2e.ts passed all 12 cases in the manual conversation-plan configuration after synchronizing Document navigation, persisted history expansion, and the collapse/reopen test with real browser animation.

Notes

• Rebased onto Fix the deck's card-to-card motion, keep decided cards full size, and bound the pager #540. A decided card is hidden inline and uses Fix the deck's card-to-card motion, keep decided cards full size, and bound the pager #540's full-size card in a deck (InlineCard's deck prop). The settled line and its tests are removed.
• When the last open card in a deck is answered and the deck dissolves into hidden decisions, the caret moves to the start of the block after the run (or to the first paragraph of a blank document), and a page-wide polite status says "Decision saved". Covered by e2e.
• Design-contract hashes were renewed in their own commit. The dynamic class owners did not change.

🤖 Generated with Claude Code

Landing verification

The author's newer caret handoff and settled-line removal are preserved. A regression fix now skips hidden accepted-comment projections when finding the editable block after a dissolving deck; the new headless test failed before the fix and passes after it. Answered empty anchors remain pending after storage restart.

Fresh validation: CI, types, production build, explicit changed-source lint, 102 focused units, and both browser caret handoff cases pass on the combined UI base. The earlier full 21-case deck suite and all 12 manual decision-prose cases passed. Assertions and committed timeouts remain enforced. Full GitHub CI is running on the final head.

Final landing: full GitHub validation, browser tests, and both CLI checks passed with the caret regression fix. The shared container size failure is repaired by #553. All twelve source patches remain identical on current main 0c525fed; fresh local project checks and types pass, and lane audit confirms the existing option-lifetime guard. Admin rebase merge skips repeating full CI after this patch-preserving base update.

[MaggieAppleton]

Review: Looks good. Two small follow-ups below, and it needs a rebase now that #540 has merged.

I checked it in the browser on a fake-Planner server (desktop 1440×900). In a blank document with three answered ASK-DECISIONS, the document stays blank and typing works. THEN-WRITE shows one "Writing up 3 decisions…" line, then the prose and markers. I also probed the keyboard: Backspace and Delete at the start, at the end and at the block cursor never remove a hidden projection, the server logged no CRDT rejection, and the projections survive a reload.

Server

• Anchors.answer() does not break "deliberately empty". The Planner's real empty review comes from anchor_plan/relate() after the answer, and answer() runs only at submit, so that review survives. Before the answer, an empty set is either an ask-time placement or a rebase write-through, so it was never a review. A Reopen followed by a re-answer now owes a review again, which seems right because it is a new answer.
• No loop. anchor_plan with [] clears the set (pending: false), and answer() never runs again. One side effect is new but bounded: relinkDecisions (on editing unlock under liveBuild, and when a remote Planner attaches) will now also start a relink turn for answered blank-document decisions the Planner never wrote up. That is one turn, and it clears them. It is worth knowing, not blocking.
• Persistence comes before publication: the new anchors are part of the answered record inside publishStaged, and Service.anchors runs only in the post-finish() notify list. One nit: Service.anchors already re-announces each prose widget, so the announce just before it now sends a duplicate frame. Harmless.

Client

1. styles.css (the [data-plan-blank] rules). The placeholder sits about 12px higher than the real empty-document placeholder: 164px against 176px at 1440×900. The first empty caret paragraph is collapsed by the existing rule .plan-content > p:has(> br:only-child):first-child:not(:last-child) { block-size: 0; margin-block: 0 }, so the ::before and the caret sit in a paragraph with zero height. Also, one Backspace in the blank document deletes that leading paragraph. The caret then becomes a Lexical block cursor before the invisible decorators, and the placeholder jumps down to the trailing paragraph (p:first-of-type) where the caret isn't. Typing recovers it. Suggested fix: under [data-plan-blank], give the first paragraph its normal size and margins back (block-size: auto; margin-block: <paragraph default>), so the placeholder and caret match the true empty state.
2. writingEntries / useWritingUp (questionnaire.tsx). writingProse uses the "inline" presentation, so decided deck members count too, even though their deck still draws them. If the first writing entry is a deck member, its InlineCard passes writing={undefined} and no line shows anywhere. If it isn't, the count includes cards that are already visible. Suggested fix: filter out entries the deck shows, using the same state inlineMemberState uses, before counting.

#540 conflict. #540 is already merged, so this PR goes second. git merge-tree shows conflicts in questionnaire.tsx and dynamic-editor.json. Resolve it as the PR notes describe: keep where === "inline" → hidden, use #540's if (where === "deck") return "resolved", and settle on one prop shape (#540's deck boolean or this PR's presentation). Then renew the hashes. inline-decision-deck.e2e.ts and harness-decisions-deck.e2e.ts auto-merge but may still conflict in meaning, because #540 shows decided deck members at full size where this PR expects settled lines. Re-run e2e after the rebase.

The hash renewals are fine. relationNote is an explicit ReactNode prop, not part of the pointing spread, and the plan-editor and question-view class boundaries are unchanged.

CI: all 5 jobs pass on 26e264b (format/lint/types/tests, e2e, container, CLI, CLI Windows).