Planning poker

A poker session is a queue of work and a table of people. The facilitator puts one item on the table, everybody votes face-down, and the round opens when the facilitator calls it — or, if they opt in, when the last person votes.

The voting hand confirms that picks stay private until reveal, including when asynchronous voting is enabled. Until the facilitator reveals, the hand reads Hidden until reveal, and only the facilitator sees who the table is still waiting on. A revealed round keeps the votes on the table, and the estimate is only saved on the story when the facilitator presses Save.

Child-story foundation

The session API can retain one level of proposed child stories and explicitly adopt them as independent planning units. Drafting leaves the existing parent and round unchanged. Adoption records full coverage or names a remainder child, then keeps the parent as context with its Previous estimate. Child estimates stay independent, including an unestimated blank.

The implementation baseline allows one draft, at most 10 retained children (removed stories that were committed count; discarded drafts do not), and at least two active children for adoption and retention. The API also retains directed sibling planning dependencies, including proposed children. API → UI means API blocks UI; several siblings may block one child. Save checks the entire graph and reviewed content, and rejects stale revisions or a named cycle without saving part of the proposal. Dependency edits leave selection, ordering, points, and votes alone.

These links mean Planning dependency · Completion not checked. They do not track delivery or automatically unblock work. Removing a child requires reviewing its affected links; removed blockers remain Blocker removed—review. Undo validates the restored graph and keeps unsafe links marked Review needed. Retired links retain their local identifiers. No remote relationship is removed or verified by these local actions.

Continuation and shared screens

After adoption, the queue and space progress count retained planning children, excluding proposed drafts and parent context. The queue reports Unestimated and Needs review separately. A previous estimate needing review is retained, but does not count as settled. Removing a child never makes the parent count again, and a remainder is its own planning unit.

Next story reviews the exact child and parent before starting one round. Leave unestimated, in the room’s ⋯ menu, clears the table selection while retaining the child in the queue, its previous round and any unsaved votes. It neither accepts an estimate nor marks work complete. Selecting a child with prior votes still requires a fresh-round review; votes never migrate between stories.

The ended room groups unfinished children beneath their original ticket and retains named planning blockers, including unavailable blockers needing review. Reopen uses the same retained identities, notes, estimates and relationships. Presenter and Google Meet show the active child, parent context and compact accepted sibling estimates. Unrevealed votes stay hidden, and voting permissions are unchanged. Structural editing remains in the main room.

CSV remains the authorized copy/export handoff, with additive split identities, notes, roles, provenance and dependency JSON documented in the CSV reference. Flat historical exports retain six columns. There is no separate session archive/import file: ended sessions remain in Postgres under normal retention, and operators preserve them with the database backup and restore process. Local dependency intent does not confirm tracker links. Optional GitHub publication requires its own scoped connection, payload review and separate delivery receipts; see GitHub publication.

The facilitator can split the story on the table directly beneath its context. Type only a child title and press Enter, or tap Add story. Empty titles and IME composition do not submit. Parley keeps your input until the server acknowledges creation; an unknown response offers Check saved story before a retry using the same local identity. Quota and child-cap refusals keep the typed title. First-child drafts and unknown creation identities remain attached to their parent when another story is dealt.

Expand a child to read its notes and planning dependencies. Expansion is local; it never deals a story or changes anyone else’s round. Quick refine uses explicit Save and Cancel for title and optional multiline notes. Blur does not save, and Enter in notes starts a new line. Escape cancels the editor and returns focus to its trigger. A remote edit keeps your draft and shows Latest and Mine for deliberate conflict review. Saving changed scope marks previous accepted estimates Needs review; changing an active child also explicitly restarts its round.

Keyboard focus scrolls refinement fields and dependency choices clear of the fixed voting hand on narrow screens, including at 200% browser zoom.

Split into N stories is available as soon as there are two stories, with nothing to answer first, and records that the stories cover the whole ticket. Ctrl+Enter does the same from the story field or when no text field has focus; in a card’s rename fields it does nothing to the split, and in the leftover field it adds the leftover. Work that is still left over can be named instead: Name what’s left over under the composer adds a story marked Leftover scope, and the split then records that scope remains in that story; … is not left over turns it back into an ordinary story. The designation is saved on the draft, so every participant sees it and it survives a reload or a second tab. Cancel discards the drafts; with two or more it asks Discard these N draft stories? first, with Keep drafting focused. Escape in the story field clears what was typed; with nothing typed it goes Back to the table, or leaves the split view when there are no drafts at all; it never discards a draft. Enter in a card’s rename title saves the rename, and Back to the table with a rename still open says Save or cancel the rename first. After a commit, the first story in the split’s order that is unestimated or needs review is dealt. If the first story of a new split still holds votes from an earlier round, nothing is dealt automatically and focus moves to that story’s Deal, which asks about a fresh round.

While one of a committed split’s stories is on the table, the split folds into a compact row under the ticket strip (two rows with nine or ten stories; one sideways-scrolling row on a phone) so the round has the screen: one compact item per story with its position, title and estimate, the story on the table marked, and a small marker on a story that waits on another. The facilitator can press an item to deal that story; Show the split on the strip brings the full cards back in place and Hide the split folds them again, for anyone, remembered in that browser tab for the session; a new tab starts folded. Removed stories are a quiet item at the end of the row. The full split stays open when nothing is on the table, while stories are being added, and while the ticket question is asked. The split does not fold while a rename holds typed text; Hide the split then answers Save or cancel the rename first.

The page only moves by itself within one rule: it never scrolls the round’s main control (Reveal, Save, Next story, Vote again) or whatever has focus out of the area above the hand tray, and it does not move at all while someone is typing or a dialog is open. After a reveal it scrolls toward the results as far as that allows, so on a small screen the results can stay partly behind the tray while Save stays in view. After a deal the round’s controls are brought into view, in rooms without splits too.

Focus is never moved onto a control that writes to the room. When the control that had focus goes away, or after an answer to the ticket question, focus lands on the on-table region, the ticket heading or the split itself; opening the removed stories from the folded row puts it on Hide the split. The one exception is after Undo split, where it goes to the ticket’s Deal in the queue. A card’s menus (Set/Change, ⋯ and Blocked by…) open with focus on the menu itself, not on an option; arrow keys or Tab reach the options, and the estimate menu marks the card that is currently set. A held Enter or Space acts once on any control in the room, while text fields keep their normal key repeat, and Reveal, Save and Reset are sent once however fast they are pressed. Leaving a split with no draft left — by Discard drafts, Cancel, Escape, removing each draft, or another tab discarding them — returns the facilitator to the round that is still on the table, with its votes untouched.

A ticket that has been split at any time and still holds votes from an earlier round asks Start a fresh round on …? before it is dealt, as its stories do.

Estimating a committed split goes story by story. After a round on one of its stories is closed — saved, set by the facilitator, left unestimated, or revealed with nothing to save — Next story: … deals the next story of the same split that has no current estimate, in the split’s order and wrapping past the estimated ones. It does not leave the split while one of its stories is open, and it asks nothing unless that story still holds votes from an earlier round, in which case Start a fresh round on …? says those votes are cleared. When every story has a current estimate the facilitator is asked once, in the page, All N stories are estimated. What about PAR-150?:

Use the total: 15 records the sum of the stories as the ticket’s estimate, shown everywhere as Story total. It is offered only when every story has a current numeric estimate, so not with T-shirt sizes.
Estimate PAR-150 with the team deals the ticket itself for an ordinary round; the stories and their estimates are untouched.
Leave it without an estimate closes the question, and Next story: … then offers the next unestimated ticket in the queue.

When the question appears on its own, focus goes to the question itself, not to any answer, so a stray Enter records nothing; Tab reaches the answers in order. If the facilitator is typing in the split, focus stays there and the question is only announced. Escape while focus is anywhere inside the question leaves the ticket without an estimate, and does nothing elsewhere. Asked again from Estimate PAR-150 on the strip, it opens with Leave it without an estimate focused, like the other questions the facilitator opens.

Each answer, and Next story, is sent once: while a request is in flight the other answers and repeated presses are ignored. The question, like the undo, discard, estimate and remove questions, scrolls clear of the hand tray when it appears. A revealed round with nothing to save always leaves a way forward: Next story when there is one, otherwise Vote again. Committing added stories while the ticket itself is on the table first asks End the round on PAR-150?, with Keep the round focused.

The answer is recorded by the server, so a reload or a second facilitator tab does not ask again until a story is added, removed or brought back. Estimate PAR-150 on the ticket’s strip asks again later. Participants never see the question, only its outcome. If a story’s estimate changes afterwards, a Story total is marked Needs review and the strip offers Refresh the total; an estimate the team gave the ticket is not affected. Undoing the split keeps a Story total and marks it for review.

Removing or bringing back a committed story no longer leaves a question open. With no leftover named, the split’s coverage is confirmed again in the same step. With a leftover named, the strip says leftover needs review and offers Keep … as the leftover; publishing to GitHub waits for that, and nothing else does. Removed stories can be brought back with Undo removal at any time, including while a round is live; a ticket whose drafts were all discarded is an ordinary queue row with an N removed stories control for the facilitator.

Committing changes the parent to context and each active child to an independent planning unit atomically. It stops an active parent round without moving its votes or automatically dealing a child. Split saved acknowledges the change immediately. A previous parent 13 and children 3, 5, and Unestimated stay independent; there is no sum or required consensus. Drafting alone changes no planning counts. A new proposal beneath an already adopted parent stays separately labeled and excluded from planning and estimate counts until adoption is reviewed again. The queue provides a grouped summary and the active child’s name, with a link back to the single main tree.

Adopted children offer Deal through the normal round review and Set estimate with a deliberate Facilitator-set acceptance. Remove reviews affected local planning links and retains the child identity and any published key. Undo validates restoration; it cannot claim remote deletion or delivery. A displayed key means Remote status not checked. Blocked by indicators remain planning relationships, with completion unchecked.

A committed split keeps at least two stories and its leftover story, so the card menu explains that instead of offering Remove on them. To go back to one story, the facilitator chooses Undo split on the parent strip, or Undo split instead in that card menu, and confirms it in the page. The confirmation says the ticket becomes one story again and its stories are kept as drafts to come back to, and opens with Keep the split focused, so a repeated Enter changes nothing. Confirming leaves the split view: the ticket reads as one ordinary story again for everyone, nothing is labeled as being split, and the result is announced, for example Split undone. PAR-150 is one story again. 3 draft stories are kept. Focus moves to the ticket’s Deal control in the queue.

The stories are kept as drafts, set aside. In the queue the ticket is an ordinary row with a quiet 3 draft stories line; for the facilitator that line reopens the split, and for everyone else it is a count. Reopened, the drafts can be fixed and split again or discarded with Cancel; a story that was named as the leftover is an ordinary draft again. Estimates already set on the stories are kept and shown as Needs review if the split is made again; they count for nothing while the stories are drafts. Removed stories stay removed and links between stories are kept. Discarded drafts do not count toward the 10-story limit, so the ticket can be split again from scratch. If one of the stories is on the table, its round ends and its votes are kept.

The parent comes back as it was estimated: if it has an accepted estimate, the estimate and the votes behind it are kept; if it has none, the votes cast on it before the split are cleared so it has no round. Dealing it gives everyone an ordinary round with Reveal, Reset and Save, while the drafts wait under the title as a count such as 3 draft stories. The same applies to any story dealt while it already has drafts, and the presentation view shows that story’s title, seats and revealed result. A split the facilitator starts on the story that is already on the table is not set aside and takes the stage until it is committed or canceled, as before, and keeps its Splitting wording. The facilitator moves between the two with Continue splitting (or the queue’s draft line), which brings the split back for everyone, and Back to the table, which sets it aside again. Whether drafts are set aside is recorded by the server, so every participant, a second tab and a reloaded page show the same thing.

The server refuses Undo split once any of the stories has been created in a tracker, or its creation was sent without a definite answer: Parley never deletes a remote issue. A creation that was approved but never sent does not block. The action is hidden when the facilitator’s own receipts show a created or unknown story. Receipts are not always in view — after a hand-off, in an embedded meeting client, or once the connection is removed — so the action can still be offered there; the refusal is then explained in the same panel, and Undo split is withdrawn until the split changes.

Setting an estimate on the story that is on the table overrides its round. On a story whose round is live, choosing a card under Set first asks Set 5 while this round is live? with Keep voting focused: the estimate is recorded as Facilitator-set, the round ends and its hidden votes are discarded. Set estimate in the round controls says the same before Accept facilitator-set estimate. Once the round is revealed, setting an estimate leaves it on the table. Removing the story that is on the table says that its round ends and its votes are kept, with Keep focused, before anything is sent.

Expand a child and choose Blocked by to record several active siblings, including proposals. The checklist excludes the child itself and shows stable IDs beside names, so duplicate titles remain distinguishable. For example, UI is blocked by API and UI is blocked by Permissions preview the same intent shown as Blocks UI on each blocker. Showing a named sibling only expands it locally; it never deals it or moves another participant’s focus. Save dependencies acknowledges quietly with Dependencies saved, without changing votes, accepted points, parentage, or the shared active child. Reopening checks the saved selections. Deselecting previews removal; Cancel or Escape leaves saved intent intact and restores focus.

Direction reversal explicitly previews old-link removal and new-link addition and requires a separate review acknowledgment. Stale graph or content revisions show latest and mine, retaining unsent choices until you deliberately review the latest state. Server cycle refusals show the named chain. Removed or unavailable endpoints and unsafe Undo remain Needs review until explicitly retired; missing newly selected endpoints must be deselected before saving. No link count means No sibling dependencies recorded, never a claim of readiness or completion.

Connection loss pauses writes. Reconnect refreshes room authority and revisions; losing facilitation preserves readable local drafts, while reading revocation redacts room content. Ended trees remain readable and use the existing Reopen it room control for further edits. The tree uses native controls and semantic nested lists in both themes, without connector animation. See the API contract and CSV handoff for the currently available foundation.

Independent child estimates and rounds

An adopted child has its own estimate and one active voting round. Reveal opens the votes; it does not accept points. Save records a Poker estimate for the named active child and the reviewed round. Set estimate starts with no selected card and accepts a value from the room’s copied deck as Facilitator-set, without claiming consensus. Specials and numbers between cards cannot be accepted. Historical estimates keep their historical provenance.

Dealing a child with earlier votes offers Start fresh round. Keep working on this round cancels without changing anything. Switching away from a live or revealed-unsaved parent requires explicit review before parent voting stops. Votes stay with their story; they never move into children. Drafting, expanding and dependency-only changes do not deal a story or change its estimate.

For a child on the table, Change scope uses explicit Save/Cancel controls. Changing a title or notes restarts that child’s round after review and retains its previous acceptance as Needs review. Until deliberately reaffirmed with Set estimate or re-estimated through Poker, that value is historical and excluded from settled progress. Unchanged content and reorder/dependency edits do not restart voting. A stale save or vote is refused rather than assigned to a new child or a new scope.

The inline editor and scripts use the same guarded API. Follow the split rollout procedure before using split workflows: older binaries do not enforce these round/scope contracts. No tracker publication is enabled by these local acceptance fields.

Notification sounds

Notification sounds are off by default. Turn them on from Your profile or the Sounds off button in a poker room; both controls save the same account preference. A rising two-note cue marks a selected or reset story, and a higher two-note cue marks a manual or automatic reveal. The cue still plays while the tab is in the background.

Removing someone from the table thumps as the boot connects. Impact sounds follow the animation, so they are skipped under reduced motion and while the tab is in the background.

Browsers may require one interaction before allowing audio. If that happens, the preference remains on and the room offers Enable audio in this tab. Muted rooms keep following live state without replaying missed cues when sound is enabled again. Signed-link guests do not have this account setting.

The story queue

Work goes in as either a ticket — with a reference like PLAT-412 — or an ad-hoc line item, which is what you want when someone says “and what about the migration script?” mid-session. Both take optional notes, and both can be reordered with the arrows. One + Add story control opens the dialog for both, with the ticket ID optional; Point something now inside it deals an ad-hoc round straight to the table. The queue header counts how far the room has come (“1 of 4 estimated · 8 pts”, with points left out when a deck is sizes rather than numbers), and a story that was dealt and then left unestimated is marked skipped.

The facilitator owns the queue: only they may create, edit, reorder, select, or delete a story. Ordinary members still cast and change their own votes. These permissions are enforced by the API as well as by the controls shown in the UI.

The reference is real data rather than decoration: it travels through the API and the WebSocket state and comes out in the CSV export as its own column, so a session’s output can be matched back to your tracker.

A ticket needs a reference or a title — either one on its own is enough, so “just estimate PLAT-412” is one field and a click. Titles are capped at 200 characters, notes at 2000, and a reference at 40.

A poker room with PLAT-412 on the table and its note beneath the title, and a queue of five items to the right — four ticket references and one ad-hoc line. — The item on the table carries its reference and its note; the queue behind it mixes tickets with the ad-hoc line somebody added mid-session.

The decks

Four built in, plus any a space has saved of its own, chosen when the session is created and fixed for its lifetime. Every deck also carries ? — “no idea” — and a coffee card. Both appear in the per-value counts, so you can see that three people said ?, but neither is ever fed into an average, a median, a mode or a range — nor into consensus: a round where everyone picked ?, or everyone picked coffee, is unanimous about nothing and the room is not told it agreed.

Fibonaccifibonacci

12358132134?☕

median, average, per-value counts

Modified Fibonaccimodified-fibonacci

0½12358132040100?☕

median, average, per-value counts

T-shirttshirt

XSSMLXL?☕

mode and range — no average, because there isn't one

Powers of twopowers-of-2

12481632?☕

median, average, per-value counts

A vote is validated against the session’s own deck on the way in, so a client cannot store a value the deck does not contain.

Decks of your own

A space owner can save its own decks under Settings → Decks: a name, between two and fifteen cards, and a switch for whether the cards are an order rather than numbers. They then sit beside the built-in four in the create dialog, and every member of the space can see the list even though only an owner can change it.

A saved deck is a template, not a reference. Starting a session copies the cards into that session, so renaming a deck, rewriting its cards or deleting it outright leaves every room already dealing them exactly as it was — no vote already cast can be invalidated by an edit here.

The New session dialog with Poker selected, the title Sprint 35 planning typed in, and six deck cards to choose from — Fibonacci, Modified Fibonacci, T-shirt and Powers of 2, then the space's own Platform sizing, selected, and Bug triage — above an unchecked Auto-reveal when everyone has voted. — The deck chooser at session create: the four built in, then whatever this space has saved. The choice is fixed for the life of the session.

Hidden votes and auto-reveal

Nobody sees a value before the round opens. This is enforced where it has to be: the server builds a redacted view of the session for each viewer, and the WebSocket broadcasts that view — an unrevealed vote is structurally absent from the payload rather than present-but-hidden by the client.

By default the round stays closed until the facilitator hits Reveal, even when everyone has voted. That beat is for discussion before the cards flip. Teams that want the old “last vote opens the round” behavior turn Auto-reveal on — at session create, or live in the room — and the check runs with the same rules as before.

When auto-reveal is on, the round opens once everyone who can vote has voted. Spectators are not in that denominator, and neither is anyone whose seat has gone away — otherwise “4 of 5” would hang forever because someone shut their laptop. Those seats stay on the table, dimmed and marked away, and the tally and the waiting line say how many (“0 / 2 voted · 1 away”, “Waiting on Dana (1 away)”).

Sitting out ends the wait. Switching yourself to spectator mid-round takes you out of the set the round is waiting for, and if you were the last person it was waiting on, the round opens there and then — with open voting on, against the recorded list just the same. Sitting back down before the reveal puts you back in the set, and the round waits for your vote again — a vote you had already cast still counts. A vote you cast before sitting out is set aside while you are out, so it never holds the round open behind you: somebody else finishing, or sitting out themselves, still ends the round. Anyone who sits out rather than voting sees the estimates without having committed one, so a facilitator watching the room may want to say something about that.

The check otherwise runs only when a vote lands, or when the facilitator turns auto-reveal on mid-round with a full table — never on connect or disconnect. A disconnect can shrink the denominator; it can never open the round. Otherwise closing a laptop would be a way to reveal everyone else’s votes. Turning auto-reveal off mid-round does not close an already-open round.

Open voting, for a team that is not in the room

A distributed team cannot all be at the table at once, and the denominator above is unforgiving about that: it counts who is connected, so a round cannot complete while half the team is asleep.

Wait for absent voters changes what the round waits for. When a story goes on the table with open voting on, Parley records who is expected to vote — everyone who has been in this room, whether or not they are here now, minus the spectators — and the round waits for exactly those people. Opening the room once is enough to be counted; you do not have to still be connected, and you do not have to have voted. Somebody who comes back hours later still finds the round waiting, and their vote is the one that completes it.

The list is the room, not the space. A forty-person space that runs five-person rounds gets a five-person round: the thirty-five who never opened this room are not in it and never hold it up. A link guest who has been in the room is in it on the same terms as anyone else — no members row required.

Because the list is a snapshot, it does not move under the round. Someone opening this room for the first time after the story went on the table is not in it: they are very welcome to vote, and their vote is a bonus that can never hold the round shut. Somebody who merely closed their tab is in it and is still waited for. And if a person in the list is removed from the space, or their guest link is revoked or expires, the round stops waiting for them — their vote could never arrive, and a round wedged on it would never complete.

Turning open voting on mid-round records the list for the story already on the table, so that round can complete too. If the story has already been revealed there is no round to record, so Reset takes the list instead — a re-vote on a revealed story waits for whoever is in the room at that point, latecomers included.

Open voting reveals nothing by itself. It only changes who the round is waiting for; whether the last vote opens the round is still auto-reveal’s question, and with auto-reveal off the facilitator still hits Reveal. The two are independent switches and all four combinations do what they say.

Either way the room names the seats it is still waiting for, so an open round that has stalled shows who it has stalled on.

Open voting sits beside auto-reveal on both surfaces: it is a checkbox in the create dialog — Wait for everyone who has joined, even if they have dropped off — and a switch labeled Wait for absent voters under Session actions → Round settings, next to auto-reveal, that only the facilitator sees. Turning it on or off there takes effect for the round on the table, without resetting it: switching it on widens the round to the people who are not in the room, and switching it off narrows it back to whoever is present, which is how a round stalled on somebody who has gone home gets moving again. The same change is available to scripts as PATCH /api/sessions/{id}/actions/config; see the API reference.

Four face-down cards at the table and two empty seats, with the line '4 / 6 voted' above them and the viewer's own 5 raised in their hand. — Mid-round: four cards are down and the count reads 4 of 6. The other values are not hidden by the browser — the server strips them out of the payload, so there is nothing to peek at until the round opens.

Statistics that suit the deck

Numeric decks report a median up front, an average underneath, and a count for each value. A median can land between two cards — 3 and 5 give 4, which is not a Fibonacci card — and an estimate has to be a card from the session’s deck, so the round shows that median without offering to save it. Re-run the round, or talk it through and vote again until the room lands on a card.

Ordinal decks do not. A T-shirt round reports the mode and the range — “mostly M, spread S to L” — because the average of S, M and L is not a size. An “average: 4.5” over T-shirt sizes is a correctness bug displayed to the whole team on a shared screen, and Parley declines to render one.

A revealed T-shirt round showing a large M under the heading MODE, the line 'S to L · 6 votes · sizes have no average', and per-value counts of S ×1, M ×4 and L ×1. — An ordinal deck reports the mode and the range, spells out that it will not average, and offers to save M onto the story.

When everyone agrees

A unanimous reveal is celebrated: seats pair off across each row, crouch, jump and high-five at the top, with confetti at the point of contact. Anyone left without a partner jumps solo on the same beat. It waits for the last card to finish turning over, then runs for a little over a second — under three seconds all told, even at a full table — so it never delays the next round.

Everything here is decoration. prefers-reduced-motion: reduce skips it entirely — not just the movement, the confetti is never drawn — and the reveal, the statistics and the saved estimate are identical either way.

Discussing different estimates

Different estimates are an invitation to discuss assumptions. Revealed rounds show the spread without automatically directing negative reactions at a voter. Consensus celebrations remain decorative and respect reduced-motion preferences.

When someone joins

A seat that appears mid-meeting drops into its slot and takes one light bounce. The row opens first — the seats already on the table slide to their new positions, and only then does the new one fall into the gap they made, rather than landing on a row still rearranging itself under it.

Several people arriving at once — the usual start of a standup — are staggered across a fixed budget rather than a fixed gap each, so a dozen joiners are all on the table inside half a second instead of queueing up one after another. A reconnection is not an arrival: coming back from a dropped network does not replay the whole room as though everybody had just walked in. Someone who genuinely leaves and comes back does drop in again.

prefers-reduced-motion: reduce skips it entirely — the seat is simply there, with no movement and nothing scheduled.

Handing the room over

The facilitator has a Hand the room to someone… item in the room’s ⋯ menu. It opens the roster; picking a name moves the role straight away — no grace period, no accept step — and the room is told who is driving now.

That is the deliberate path. The other one is the claim: if the facilitator loses their connection, a banner counts out a 60-second grace period, after which anyone at the table can take the chair. A guest joining through a signed link is offered neither, and cannot be handed the role.

Removing someone from the room

The facilitator can remove a participant from the room they are running: hover a seat at the table and a small × appears on it. Only the facilitator sees it, it never appears on their own seat, and it goes away once the session has ended. Confirming asks for an optional message before it fires. It ends that person’s seat in this meeting and nothing else: their sockets for the room close and the seat they were occupying at the table empties on the next update. They are still a member of the space, so they still appear wherever the space’s people are listed, and they can walk back in through the same link or room code.

If a round is open when they go, it stops waiting for them: a round that needed their estimate to auto-reveal reveals as soon as everyone still at the table has voted. A vote they had already cast stays on the table and still counts towards the result — removing someone ends their seat, it does not retract what they had already said.

The removal can carry a short message — “we will pick this up tomorrow” — which the person sees on the way out. The box stops at 80 characters, which is what fits the connection’s closing frame once accents and emoji are counted.

The person removed lands on a screen of their own, with the message if there was one. It is not the “you no longer have access to this space” screen, because that is not what happened — they still belong to the space, and the way back in is the same link they came through.

Everyone still at the table watches a boot swing in from beside the seat, connect, and launch it off the screen; the row closes up only once the seat has fully gone. The boot is a pendulum rather than a canned slide, so it dips on the way across and is on its way back up when it lands — which is why the seat leaves sideways instead of straight up, and why a seat on the far side of the table takes a little longer to see off. If you have asked your system for reduced motion, the seat simply goes.

It is deliberately the small hammer. Removing somebody from the space itself is a separate, owner-only action on the space settings page, and there is no per-room block list: a removal is reversible, and stopping someone coming back means taking away their way in.

Ending the round

After the reveal the facilitator can save the agreed value onto the story with one click, which writes the estimate and moves the queue along. If a round has no saveable result — everyone voted ?, or the deck is ordinal and there is no single winner — the save control does not appear rather than inventing something to write.

Reset, in the ⋯ menu, clears the votes on the current story and starts it again.

When the session ends

Ending the session keeps the results in the space. Anything in the queue that never received an estimate — including a story left mid-vote — is listed on the ended panel, in queue order, so the leftover is a backlog rather than a memory. A session that estimated every story says so, instead of showing an empty list.

What is not here

Capability	Status	What to do instead
Vote timer	Not built	Use a facilitator's own timer. Opt-in auto-reveal can end a round without one when everyone has voted.
Estimate history across sessions	Not built	Export each session to CSV and keep them; cross-session analytics is on the roadmap.
Sharing a deck between spaces	Not built	A saved deck belongs to one space. Copy the cards into a deck of the same name in the other space; an org-owned deck is on the roadmap.
Editing a revealed vote	Out of scope	Reset the round instead — silently changing a revealed vote would make the reveal meaningless.

Optional GitHub publication

The current facilitator can choose Publish splits to GitHub in the ordinary room. The operator must configure PLUGIN_SECRET_KEY; no plugin installation is required. Without it the room does not show the action. Use a dedicated PAT scoped to the chosen repository and Project. Metadata reads are separately consented from write permission. The encrypted connection belongs to this org, space, room and account. Replacement invalidates earlier approval; revocation removes the credential and preserves issued keys. Link guests, embedded meeting clients and org custody cannot use this surface.

Choose one GitHub.com repository, an ordinary native parent issue, a Project owned by the same repository owner, and its NUMBER field. The host verifies current local access and bounded provider metadata. The Project must contain at most 100 complete items/fields, only issues from this repository, and no parent item. Exclusion means absence from Project membership, not hiding a view or rollup. Automation can change this later; detected drift pauses delivery.

All active adopted children start selected. Review actual titles and refinement, including the reconciliation marker, destination/type, native parent, directed blockers, Project membership, numeric mapping and counting before payload consent. A UI-only selection offers including its API sibling or leaving that dependency pending. Accepted numeric cards 1/2/3/5/8/13/21/34 map exactly. Blank confirms actual unset only; unsupported cards, changed values and unexpected defaults pause or fail rather than become zero or be cleared. Parent context is excluded; its historical estimate is never restored or cleared by publication.

Create, parent, Project membership, each dependency/removal and estimate have separate receipts. Created or linked never means every field reached GitHub. Refresh unknown receipts deliberately reads the same identity without another write. A complete current readback of an unset numeric field or missing Project/native relation can record Failed, allowing an explicitly reviewed retry. An inaccessible or incomplete readback, unexpected numeric default, or uncertain creation remains verification-needed. Review the current split again to retry pending or failed steps. A story GitHub cannot take as written — a title over 256 bytes, which a title within Parley’s 200-character limit can exceed in non-Latin scripts, a body over 60,000 bytes, or a refinement that already contains the reconciliation marker — is recorded Failed before any request is made and without using an attempt, so nothing is left unknown; shorten it and review again. A creation that is already unknown stays unknown. Confirmed is a historical readback, not a promise of current upstream access. Confirmed creations reuse their issue without rewriting title or refinement. For cleanup, change the local graph and review the explicit removal. There is no issue deletion, completion retrieval, automatic unblocking or general sync.

GitHub has no conditional field update. Fresh preflight and readback bound the check, but another writer can race between them. Each request has a 60-second cumulative deadline; durable receipts survive interruption, room closure, kind retirement and credential revoke. Saved review IDs remain bound to their approving account. A replacement facilitator connects with their own credential and makes a fresh review while retaining issued child identities. Revoked credentials cannot refresh provider receipts. Local access denial clears the open payload, receipts and connection view. Parent text and votes are omitted.

Planning poker

Child-story foundation

Continuation and shared screens

Inline splitting and refinement

Independent child estimates and rounds

Notification sounds

The story queue

The decks

Decks of your own

Hidden votes and auto-reveal

Open voting, for a team that is not in the room

Statistics that suit the deck

When everyone agrees

Discussing different estimates

When someone joins

Handing the room over

Removing someone from the room

Ending the round

When the session ends

What is not here

Optional GitHub publication