feat(planning): add research events and waiting-for-reply cues to the interview page

plugins/planning/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "planning",
-  "version": "0.53.0",
+  "version": "0.54.0",
   "userConfig": {
     "surface": {
       "type": "string",

plugins/planning/CHANGELOG.md

@@ -3,6 +3,21 @@
 All notable changes to the `planning` plugin are documented here. Format follows
 [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); this plugin uses semantic versioning.
 
+## [0.54.0] - 2026-09-30
+
+### Added
+
+- **The interview page has Research this and Cancel research buttons.** They post generic `research` and `cancel-research` events that name no skill or plugin; the interviewing session decides how to fulfill them. The `wait` op stamps `waitingSince`, and a held card reads "Research in progress, started <time>". While Claude holds a question, Accept, the alternatives and Own answer stay enabled (Answer anyway, as before); disabling them is not done and is left to the owner ([#5569](https://github.com/melodic-software/claude-code-plugins/issues/5569)).
+- **A question waiting for Claude's reply shows a chip, and Accept asks first.** An ask or rephrase with no later reply from Claude marks the question "Waiting for Claude's reply", and Accept then asks "Accept current recommendation anyway?" ([#5569](https://github.com/melodic-software/claude-code-plugins/issues/5569)).
+- **A bare `#N` in question text links to the issue in the new `meta.repo` key.** `round.py` warns, without blocking, when text carries one and `meta.repo` is unset ([#5569](https://github.com/melodic-software/claude-code-plugins/issues/5569)).
+
+### Fixed
+
+- **The revising chip reads "Answer not handled yet" and shows on a question only while Claude has not handled that question's own decision.** It no longer names an upstream question; a changed prerequisite stays in the Stale and Waiting-on chips ([#5569](https://github.com/melodic-software/claude-code-plugins/issues/5569)).
+- **The summary's Confirm and Something's off buttons sit in one row below the What-is-off box**, clear of it ([#5569](https://github.com/melodic-software/claude-code-plugins/issues/5569)).
+- **The decide area caps at 40% of the viewport height on short screens, and the phone layout wraps the status line and no longer scrolls the question rail inside itself** ([#5569](https://github.com/melodic-software/claude-code-plugins/issues/5569)).
+- **The wake contract says to re-read the events and restate after a stale-read refusal instead of forcing** ([#5569](https://github.com/melodic-software/claude-code-plugins/issues/5569)).
+
 ## [0.53.0] - 2026-09-30
 
 ### Added

plugins/planning/skills/interview/context/surface.md

@@ -25,7 +25,7 @@ The page is the input surface SKILL.md "Question surface: the page" selects. The
 - **Start** with the command in SKILL.md (it carries the configured emoji setting). It prints the page URL; give that URL to the user. A missing prerequisite exits non-zero with its name: take the degrade below.
 - **Ids.** Page question ids are `Q<N>` on the session's one continuous counter, so the register rows written at ask-time match what `export-ledger` emits (it renumbers any other id set). When the register already has rows as the page starts (earlier terminal rounds, or a resumed topic whose data dir was discarded), seed the still-empty data dir with `round.sh import-ledger --ledger '<memory_dir>/<topic-slug>/interview-checklist.md'` before the first `add-round`; it keeps each row's `Q<N>` and decision. `export-ledger` writes each row's resolution as escaped named fields in a fixed order (`hold`, `proposal`, `was`, `answer`, `note`, `aside`, `commitments`; the grammar is in `surface/exporters.py`), and import restores them: the hold, a superseded-by-plan row's proposal and displaced answer, the decision on any status (an accept or a defer on a superseded-by-plan row too), the note, a decision a user hold set aside (restored still set aside), and every commitment with its confirmed or unconfirmed mark. An unknown field or a contradictory pair is refused, naming the field. Every older grammar still imports, read as it always was: `waits on::`, `awaiting user::`, `confirmed::`, `plan proposes::` and `answer::` rows, a settled row's `; confirmed::` tail, and the unescaped `waits on:`, `confirmed:` and `; confirmed:` forms. A hand-written resolution imports as the row's own text.
 - **A round.** Write the frontier to `'<data_dir>/round-<n>.json'` (`{"meta": {...}, "groups": [...], "questions": [...], "visuals": [...]}`), run `round.sh add-round --file '<data_dir>/round-<n>.json' --round <n>`, and write the register's `open` rows in the same step. Each question carries `recommendation` (one line), `basis` (2-3 sentences, shown behind Why), at least two `alternatives` (`{key, text}`), `commits` (what accepting commits the user to, or `[]`), and `dependsOn` for its prerequisites. Send the round's closing constraint probe with a `note-reply` op (no `seq`) so it lands in Notes to Claude, or as a Claude thread line on the round's first question. `add` and `add-round` refuse a question without `commits` or with fewer than two alternatives.
-- **Meta.** `meta` takes `title`, `eyebrow`, `stages` and `next`. `stages` is an object mapping each stage key to its label, for example `"stages": {"frame": "Frame", "decide": "Decide"}`; a list is refused (`$.meta.stages: expected object, got list`). Set with `add-round`'s `meta` object or the `meta` op; any other key is refused. `meta.next` is what Claude does after wrap-up, shown on the finished screen. Keep round numbers out of `meta.eyebrow`: the page derives the round label (`<stage> round N`) from the questions and shows it beside the eyebrow, so a number in the eyebrow goes stale at the next round.
+- **Meta.** `meta` takes `title`, `eyebrow`, `stages`, `next` and `repo`. `stages` is an object mapping each stage key to its label, for example `"stages": {"frame": "Frame", "decide": "Decide"}`; a list is refused (`$.meta.stages: expected object, got list`). Set with `add-round`'s `meta` object or the `meta` op; any other key is refused. `meta.next` is what Claude does after wrap-up, shown on the finished screen. `meta.repo` (`owner/repo`) makes the page link a bare `#N` in any markdown field to that repo's issue; `owner/repo#N` always links to its own repo, and `[text](url)` links, `Qn` refs and code spans are left as written. A bare `#N` with `meta.repo` unset stays plain text and `add`, `add-round` and `apply` warn (non-blocking). Keep round numbers out of `meta.eyebrow`: the page derives the round label (`<stage> round N`) from the questions and shows it beside the eyebrow, so a number in the eyebrow goes stale at the next round.
 - **Visuals** are declared by format (`svg`, `mermaid`, `image`, `markdown`, `html`, `chart`) with a `scope` (`question:<id>`, `group:<id>`, `round:<stage>:<n>`, `all`) and inline `content` or a data-dir-relative `file`. `label` is its tab name (short; the page falls back to `title`, then `id`). `group`, `order` and `primary` arrange visuals: those sharing a `group` are versions or alternatives of one another, `order` sorts them, and at most one live `primary` is allowed per `group` within a `scope`. Change a visual with `replace-visual` (a full object with the same id) or retire it with `archive-visual`; an archived visual stays in `questions.json` and the page and report never show it. `add-round` refuses an id that already exists. Describe what a visual shows; leave out the tool or skill that made it. An `html` visual runs its scripts on the page, in an opaque-origin sandbox that cannot reach the page, so an interactive prototype or an inlined chart library works; the page CSP still blocks remote content. The exported report runs no scripts: when the decision rests on what a scripted visual shows, attach an `image` of it as well. Two or more `image` visuals on a question also get a Gallery tab (one more per group holding a smaller set) with a thumbnail strip, arrow-key flip and a side-by-side compare, in full screen too. Any visual opens in a new tab from the panel or full screen, still sandboxed.
 - **Arm** the watcher as a background Bash task (`run_in_background`): `bash '<surface_dir>/watch.sh' '<data_dir>'`.
 - **One watcher.** One session watches an interview at a time: the first watcher holds the server's lease, and `watch.sh` exits 3 naming the holder, since when and its last poll when another session holds it. Do not re-arm. When the holder is another Claude session, coordinate with it through the cross-session messaging tooling this session provides (discover what is available; assume no particular tool) and agree which session runs the interview, or ask it to hand over with `round.sh lease --release`. When it cannot be reached, tell the user which session holds the lease and since when; the lease frees itself once the holder stops polling for `leaseTimeout` seconds (default 600). `round.sh lease` prints the current holder. `watch.sh` also exits 3 when this session's lease was released while it waited: run `round.sh lease`, and re-arm only when this session should still watch. When it exits 2 with "no watcher id" (no session id is exported and the parent pid is 1), export `WATCH_ID` with a name for this session and re-arm.
@@ -41,7 +41,7 @@ The watcher exits with one JSON line: `{"seq", "timedOut", "events": [...], "not
 
 1. For an `ask`, `own` or `rephrase`, open the turn with a one-line status (which question, what you are doing) before the reply (R10).
 2. Answer every `ask`. For decisions on one question, the latest live event wins; mark the earlier ones handled with it (R7).
-3. Write `'<data_dir>/ops.json'` fresh with the Write tool on every wake; a stale file re-applies old replies.
+3. Write `'<data_dir>/ops.json'` fresh with the Write tool on every wake; a stale file re-applies old replies. When `apply` refuses a `rec` because the question has a newer live user event (the stale-read refusal, R2), re-read the events, restate the reply or revision against what the user just did, and apply again instead of passing `"force": true`.
 4. Run exactly one background Bash call that records and re-arms (R8):
 
 <!-- wake-command: surface/watch.test.sh runs the fenced command below -->
@@ -67,7 +67,7 @@ When the first wake prompts for permission, offer the user one allow rule, `Bash
 | `revise` | `id`, `title`, `short`, `facts`, `basis`, `rec`, `why`, `text`, `alternatives`, `commits`, `seq`, `affects`, `force` | Reword a question; `commits` replaces the list and resets its confirmations |
 | `note-reply` | `text`, `seq` | Answer a note in Notes to Claude; with no `seq`, post a closing probe there |
 | `add`, `add-round`, `group` | `question`; `round`, `meta`, `groups`, `questions`, `visuals`; `id`, `title`, `summary`, `dependsOn` | New questions and groups; writing a `summary` records the group's current question ids as `summaryOf`, and the page marks the summary Stale once the members differ, so rewrite the summary after adding questions |
-| `meta` | `set` (`title`, `eyebrow`, `stages`, `next`) | Merge into `meta`; other meta keys stay |
+| `meta` | `set` (`title`, `eyebrow`, `stages`, `next`, `repo`) | Merge into `meta`; other meta keys stay |
 | `archive` | `ids`, `why` | Take off-path questions out of the open count |
 | `replace-visual` | `visual` | Swap in a full visual object for the top-level visual with the same id; an unknown id is refused |
 | `archive-visual` | `ids`, `why` | Hide top-level visuals from the page and report; they stay in `questions.json` |
@@ -93,7 +93,7 @@ The page header shows a Claude line: the `set-status` text with its age while on
 
 A `wait` holds a question in one of two ways. The page labels them as follows:
 
-- **Pending research** (`by: claude`, the default): Claude is working something out. The question shows `Pending research: <waitsOn>`, counts in the header's pending-research count, and is listed under Show: Pending. The user can still answer it (Answer anyway).
+- **Pending research** (`by: claude`, the default): Claude is working something out. The question shows `Pending research: <waitsOn>`, counts in the header's pending-research count, and is listed under Show: Pending. The card reads `Research in progress, started <time>` (the time the `wait` landed). The user can still answer it (Answer anyway) and can post `cancel-research` (Cancel research) to release it; a question with no hold offers `research` (Research this) instead, and one waiting on the user offers neither.
 - **Needs your answer** (`by: user`): the question needs the user again, even though a decision is recorded. The question shows `Needs your answer: <waitsOn>` and counts as open and in the needs-you navigation. The `wait` also stamps `setAsideAt` and `setAsideSeq`: a page or terminal decision recorded before it no longer counts as an answer anywhere, even after the hold clears; the user's next decision counts. A recommendation revision (`reply --rec`, `revise --rec`) stamps the same fields on a counted `own` answer without a hold: the question returns to open until a new decision counts.
 
 Both count as not answered in the page meter, `round.sh status` and `export-ledger`.
@@ -127,6 +127,8 @@ When the work returns, clear both (`wait` with `"clear": true`, `set-status` wit
 | `reopen` | question | clears it, keeps the note | `handle` |
 | `ask` | question | no | `reply` with its `seq` |
 | `rephrase` | question | no | `reply` with `"kind": "rephrase"` and its `seq` |
+| `research` | question, optional `text` | no | `wait` (`by: claude`) plus `set-status`, then `reply` or `revise` when the lookup returns. Run `/discovery:research` when that skill resolves in this session, else look it up inline. See [Status, activity and holds](#status-activity-and-holds) |
+| `cancel-research` | question | no | `wait` with `clear` and `set-status` with `clear`; ignore the pending result, and `handle` both seqs |
 | `note` | none | no | `note-reply` with its `seq`, or `reply` on a question |
 | `undo` | question, `undoSeq` | withdraws `undoSeq` | Drop that decision from the ledger; `handle` both seqs |
 | `wrapup` | none | no | Run [Wrap-up](#wrap-up), then `handle` |