AW-12 — Chat, group conversations and the organization channel

Program: Agent Workspace · Epic ID: AW-12-chat-channels Feature ID: chat-channels · Branch: feat/chat-channels Status: Draft · Size: L · Depends on: — Created: 2026-09-06 · Last updated: 2026-09-06 Audience: Product, Engineering (backend + frontend), Design

Behaviour-first per Constitution IX. No class names, no file paths, no code in this document — those live in plan.md. Additive only (program rule #1): the single global assistant panel, its history list, its provider/model pins, its attachments, its dictation and its canvas all keep working exactly as they do today. Everything below is added on top.

Adjacent epics: AW-02 Mission board (owns the Chat about it card action; this epic owns what happens when it is clicked), AW-06 Knowledge library (owns the # document picker and reference resolution; this epic consumes it), AW-04 Live Feed and AW-13 Attention controls (own notification routing; this epic emits the events), AW-09 Runs & receipts (owns the receipt every agent reply links to), AW-18 Shared dashboards (owns teammate read access to surfaces this epic creates).

---

1. Overview

This epic turns Ever Works' single global assistant panel into a conversation system: many named Conversations instead of one running log, each one addressed at a specific Agent, a group of Agents, or the whole Organization.

Concretely, a person can: start a Conversation per piece of work and give it a name; keep that Conversation docked while they move around the product, resize it, navigate inside it back to the Agent's other Conversations and sideways to another Agent's; pull a second Agent in by mentioning them, which turns the Conversation into a group carrying its recent history with a link back to where it came from; post once to an organization channel that reaches every active Agent, or narrow that same post to two of them with a mention; read and join the Conversations Agents open with each other; see, while typing, exactly which @ words and # document references will actually resolve; attach files; and retry a message that failed to send instead of losing it.

The load-bearing idea is addressing. Ever Works can already run twenty-five Agent executions at once and parks the rest. The moment more than one Agent can hear a message, "who is expected to answer this?" stops being a UI nicety and becomes a spend-control question. This epic answers it explicitly, in the product, per message, and shows the answer back to the person who sent it.

2. Why now

2.1 The questions our users cannot answer today

"I have eleven Agents. Which one am I talking to, where did last Tuesday's conversation about the pricing page go, and how do I tell all of them at once that the launch slipped?"

They cannot answer any of the three. Here is what they do instead.

What the user wants	What Ever Works gives them today	What they do instead
"Talk to this Agent"	The assistant panel is not addressed at an Agent at all. A Conversation row records a user, a provider and a model — there is no Agent on it. The tool-calling assistant is a single generalist.	Open the Agent's detail page and read its run history, then come back to the panel and describe the Agent in prose.
"One conversation per job"	Conversations exist and persist, but a new one is created only by pressing New chat, which silently clears the panel. Nothing marks a Conversation as being about anything.	Let one Conversation run for weeks and scroll.
"Name this one, it matters"	Titles exist but are machine-written only — auto-derived from the first message, then rewritten once by a model at four messages. There is no way to set or clear a name.	Recognise Conversations by their first line in the history list.
"Keep it open while I work"	The panel does dock and does remember its width — but it holds exactly one live conversation, has no internal navigation, and reopening the history list replaces the whole panel body.	Keep two browser tabs open.
"Loop in the research Agent"	Impossible. A Conversation belongs to one user and has no participant concept of any kind — no second Agent, no second person.	Copy the transcript into a new Conversation and re-explain.
"Tell everyone the launch slipped"	Impossible. There is no broadcast of any kind.	Open eleven Conversations and paste eleven times.
"What are my Agents saying to each other?"	Agents can already delegate to one another through an explicit collaborator allow-list, and those child executions are real — but the coordination is only visible as separate rows in a run list.	Read two run transcripts side by side and infer the exchange.
"Did that reach the paused one?"	Nothing shows delivery. An Agent that is paused, in error, or parked behind the concurrency valve simply does not act, silently.	Check each Agent's status page one at a time.
"This @ — will it work?"	@-mention parsing exists and works server-side for Task chat, and @kb: document references exist and work server-side for the assistant panel — but neither has any composer affordance. Nothing lights up, nothing is picked, and an unresolvable token is dropped after send, silently.	Type the mention, send, wait, notice nothing happened, retype it.
"That send failed"	The panel shows a generic inline error banner and the text is gone from the composer.	Retype the message from memory.

2.2 Why this is the right moment

1. The addressing machinery already exists — in the wrong room. Task chat already parses @<slug> and [[document]] tokens server-side, validates each against the Agents and documents the caller can actually reach, strips unresolved tokens so a model never sees a hallucinated reference, and dispatches a reply execution per mentioned Agent — deduplicating against an Agent's already-live execution instead of double-spawning. That is the hard half of this epic, already built, tested, and running. It is simply confined to one object type.
2. Multi-Agent delegation already ships. Every Agent carries an explicit allow-list of the other Agents it may hand work to. Agent-to-agent coordination is therefore already a real, permissioned, bounded thing in this product — it just has no readable surface. This epic gives it one instead of inventing a new capability.
3. Spend control already exists and is about to be stressed. The platform admits Agent executions through a concurrency valve (default 10 concurrent per Work, 25 per Organization) and parks the overflow with a recorded reason. A broadcast that starts one execution per Agent is exactly the traffic shape that valve was built for — but only if the broadcast surface reports what the valve did. Building the broadcast without the reach receipt would be building a way to silently exhaust the org's concurrency budget.
4. The panel is already the right shape. It already docks, already resizes by pointer drag, already persists its width, already has a distinct mobile full-screen mode, and already restores the active Conversation across reloads. This epic adds navigation inside it, not a new panel.
5. There is dead weight to reclaim. Message-edit copy already exists in the message catalogue with no component rendering it. This epic either uses those strings or leaves them alone — it does not add a second vocabulary next to them.

2.3 The problem this epic must not create

The defining failure of a multi-Agent room is that every Agent answers every message. In this product that failure is not merely noisy — it is metered. Eleven Agents replying to one "thanks, looks good" is eleven executions, eleven model bills, and, past the concurrency valve, a queue of parked executions that delays real work.

So the addressing rule is a first-class functional requirement, not a prompt-engineering detail (§4.8), and it has two halves that must ship together:

• The mechanism — mentions decide who answers; unmentioned Agents may answer only under a stated, bounded condition.
• The expectation — the product says out loud, in the interface, that an Agent staying quiet is correct behaviour and not a fault. Without the second half the first half reads as a bug.

3. User scenarios

3.1 Primary scenarios

S-1 — Starting a Conversation with a specific Agent. Given a person on any dashboard screen with at least one active Agent, when they open the conversation panel, choose an Agent, and send a first message, then a new Conversation is created addressed at that Agent, the panel shows it immediately, the Agent's reply arrives in the same Conversation, and the Conversation appears in that Agent's Conversation list titled by its first message.

S-2 — Naming a Conversation. Given an open Conversation whose title was written automatically, when the person clicks the name control in the header, types Q4 pricing page, and confirms, then the Conversation is listed under that name in bold with its first-message preview beneath it, the name survives reload, and no further automatic re-titling ever overwrites it.

S-3 — Clearing a name. Given a Conversation the person named, when they open the name control, delete the text and confirm, then the Conversation reverts to showing its first-message preview as its identity, and automatic titling is re-enabled for it.

S-4 — The panel follows the person. Given an open Conversation docked on the right, when the person navigates from the Mission board to a Work's items page to Settings, then the Conversation stays docked, keeps its scroll position and its in-flight reply, and does not close. Only the panel's close control closes it.

S-5 — Navigating inside the panel. Given an open Conversation in the docked panel, when the person uses the panel's back control, then the panel body becomes that Agent's Conversation list without leaving the current screen; and when they then click the participant name at the top, then the panel offers the other Agents, the group Conversations and the organization channel, and choosing one loads it in place.

S-6 — Resizing. Given the docked panel, when the person drags its left edge, then it resizes live between 350 px and half the viewport and remembers the width per browser; and when they double-click that same edge, then it snaps back to the default width of 420 px.

S-7 — Pulling a second Agent in. Given a Conversation addressed at one Agent, when the person mentions a second Agent and sends, then a group Conversation is created containing both Agents and the person, carrying the last 20 messages of the original, both Conversations show a link to each other, the mentioned Agent replies in the group, and the original Conversation stays intact with its own history.

S-8 — Broadcasting to the Organization. Given a person in an Organization with nine active Agents, when they post to the organization channel without mentioning anyone, then the message is delivered to all nine, each one decides for itself whether it has something to add, and the message carries a receipt the person can open to see, per Agent, whether the message was delivered, queued behind the concurrency valve, or skipped because the Agent was paused.

S-9 — Narrowing a broadcast. Given the same channel, when the person posts and mentions exactly two Agents, then only those two are addressed, the other seven receive nothing at all, and the receipt says so explicitly rather than showing seven silent deliveries.

S-10 — Reading and joining an Agent-to-Agent Conversation. Given two Agents that have coordinated with each other, when the person opens the Agent conversations list and selects that pair, then they see the full exchange in order; and when they type into it, then their message joins the exchange attributed to them by name, and both Agents receive it.

S-11 — Live mention and reference highlighting. Given a person typing in any conversation composer, when they type @ and two more characters, then a picker appears listing at most 8 matching Agents and Organization members, and the inserted mention renders as a highlighted chip; and when they type an @ word that matches nothing, then it stays plain text, so a highlight always means the mention will land.

S-12 — Referencing a document. Given a person composing a message, when they type # and pick a Knowledge Base document, then the reference renders as the document title as a link, and when the message is sent the addressed Agent receives that document's content in its execution context.

S-13 — Attaching a file. Given a person composing a message, when they drop three files onto the composer, then each uploads immediately with a visible chip and progress, the send control stays disabled until all uploads settle, and the addressed Agent receives the files as reference material.

S-14 — Retrying a failed send. Given a message that failed to send because the account hit a rate limit, when the person looks at the Conversation, then the message is still there, marked failed, with the reason in plain language and a Retry control; and when they press Retry, then the same message is sent once — never twice — and on success it becomes an ordinary message in place.

S-15 — Archiving and restoring a group. Given a finished group Conversation, when the person archives it from its list row, then it leaves the active list, its transcript is preserved in full, and it appears under an Archived entry at the bottom of the section; and when they restore it, then it returns to the position its last activity earns it, not to the top.

S-16 — Chatting about a Mission. Given a Mission card, when the person chooses Chat about it, then the docked panel opens a Conversation carrying that Mission as attached context, the Mission is shown as a chip in the Conversation header, and the addressed Agent receives the Mission's identity without the person pasting anything.

3.2 Edge cases, failures, races and denials

S-17 — Mentioning an Agent the person cannot reach. Given a person composing a message, when they type the name of an Agent that belongs to a different Organization or has been archived, then the picker does not offer it, a hand-typed match does not resolve, the token stays plain text, and the message sends as ordinary prose. The response is identical to mentioning a name that does not exist anywhere — the product never reveals that a hidden Agent exists.

S-18 — Two mentions of the same Agent in one message. Given a message that names the same Agent twice, when it is sent, then exactly one reply execution is started for that Agent.

S-19 — Mentioning an Agent that is already working on this Conversation. Given an Agent with a live execution started from this Conversation, when the person mentions it again with a follow-up, then the follow-up is delivered into the running execution rather than starting a second one, and the Conversation shows that it was delivered to the run in progress.

S-20 — Broadcast over the concurrency valve. Given an Organization already at its concurrent-execution ceiling, when a person broadcasts to the organization channel, then the post still succeeds, every Agent that could not start immediately is shown as queued in the receipt with the reason, those replies arrive later as capacity frees, and no message is lost.

S-21 — Broadcast to an Organization with no active Agents. Given an Organization whose Agents are all paused, draft or archived, when a person posts to the organization channel, then the post is stored and visible, the receipt reads Reached 0 of 4 — all Agents are paused, and the channel offers a link to the Agents list rather than failing the send.

S-22 — Group at capacity. Given a group Conversation that already contains 8 Agents, when the person mentions a ninth, then the message sends as prose, the mention does not resolve, and an inline notice explains that a group holds at most 8 Agents and offers to start a new group instead.

S-23 — Promotion race. Given two browser tabs open on the same one-to-one Conversation, when both send a message mentioning the same second Agent within the same second, then exactly one group Conversation is created, both messages land in it in send order, and the second tab follows the first tab's group rather than creating a duplicate.

S-24 — Archiving something that cannot be archived. Given the organization channel or an Agent-to-Agent Conversation, when the person opens its row menu, then no archive control is offered, and a direct attempt through any other route is refused with a plain explanation.

S-25 — Losing Organization membership mid-session. Given a person reading the organization channel who is removed from the Organization, when they send their next message, then the send is refused with "You no longer have access to this Organization's channel", the channel is removed from their sidebar on the next load, and the message text is preserved in the composer so nothing they typed is lost.

S-26 — No Organization at all. Given a person who has never created an Organization, when they open the conversation sidebar section, then there is no organization channel row, an explanatory line reads "Create an Organization to broadcast to every Agent at once", and one-to-one and group Conversations work normally.

S-27 — An Agent tries to talk to an Agent it may not talk to. Given two Agents where the first does not list the second as an allowed collaborator, when the first attempts to open a Conversation with the second, then the attempt is refused, no Conversation row is created, and the refusal is recorded against the initiating Agent's execution rather than surfaced as a person-facing error.

S-28 — An Agent pair that will not stop. Given an Agent-to-Agent Conversation that has reached 20 consecutive Agent-authored messages with no person's message between them, when either Agent attempts a twenty-first, then it is refused, the Conversation shows a Paused — waiting for you state with the reason, and it resumes only when the person posts into it.

S-29 — Message too large. Given a person pasting 40 KB of log output into the composer, when they attempt to send, then the send is blocked before it leaves the browser with "This message is too long (40 KB of 16 KB). Attach it as a file instead", and the Attach control is offered inline.

S-30 — A secret in the composer. Given a person pasting text that contains what looks like an API key, when they send, then the message is rejected before it is stored, with "This message looks like it contains a credential. Store it in the connection's settings instead" — matching how Agent instruction files and Task comments already behave.

S-31 — Empty states. Given an Agent with no Conversations, a group section with no groups, an organization channel with no posts, and an Agent conversations list with no pairs, when each is opened, then each shows its own explanatory empty state with the one action that resolves it, and none of them renders an empty list frame with no explanation.

S-32 — The reply that never comes. Given a person who posts into a group of four Agents without mentioning anyone, when none of them has anything to add, then the Conversation shows a quiet, non-alarming line — "No one had anything to add. Mention someone by name to ask directly." — and no error, no spinner and no notification are produced.

S-33 — Reading someone else's Conversation. Given a person who obtains the address of a Conversation they do not participate in, when they open it, then they get the same not-found response as for a Conversation that does not exist.

4. Functional requirements

Every number below is normative.

4.1 Conversations, addressing and naming

• FR-1 A Conversation MUST have exactly one kind: direct (one person, one Agent), group (one person, 2–8 Agents), organization channel (one per Organization), or agent pair (two Agents, with the person as a joinable observer).
• FR-2 A direct Conversation MUST be addressed at exactly one Agent, and that address MUST NOT change for the life of the Conversation.
• FR-3 Sending a message from an Agent's composer while no Conversation is open MUST create a new Conversation rather than appending to the most recent one.
• FR-4 Every Conversation MUST be identified in lists by its name when it has one, rendered above its first-message preview; otherwise by the first-message preview alone. A list row MUST NEVER render as "Untitled".
• FR-5 A person MUST be able to set, change and clear a Conversation's name. The name is capped at 200 characters.
• FR-6 Setting a name MUST permanently disable automatic titling for that Conversation. Clearing the name MUST re-enable it.
• FR-7 Clearing the name of a group Conversation MUST fall back to a name derived from its participants (e.g. "Nova, Orion & you"), never to blank.
• FR-8 The organization channel and agent pair Conversations MUST keep their own names and MUST NOT be renamed or participant-derived.
• FR-9 A Conversation MAY carry one attached context object — a Mission, Task, Work, Idea or Agent — set at creation and immutable thereafter.
• FR-10 When a Conversation carries attached context, the addressed Agent MUST receive that object's identity and current summary in its execution context, and the Conversation header MUST show the object as a chip linking to it.
• FR-11 The existing global assistant Conversation — no Agent address, no participants — MUST keep working unchanged and MUST be treated as a direct Conversation with no Agent.
• FR-12 Conversation lists MUST be ordered by most recent activity, MUST default to 50 rows, and MUST NOT accept a page size above 200.

4.2 The docked conversation panel

• FR-13 The panel MUST persist across route changes within the dashboard and MUST NOT be closed by navigation, by opening a dialog, or by any action other than its own close control.
• FR-14 The panel MUST provide three views — a Conversation, an Agent's Conversation list, and a participant switcher — navigable in that order by a single back control.
• FR-15 The participant switcher MUST list the person's Agents, their group Conversations, the organization channel (when one exists) and an entry into the Agent conversations list.
• FR-16 The panel MUST be resizable by dragging its left edge, bounded to ≥ 350 px and ≤ 50 % of viewport width.
• FR-17 Double-clicking the panel's left edge MUST reset its width to 420 px.
• FR-18 Panel width MUST persist per browser and MUST be restored without a visible reflow on first paint.
• FR-19 Below 768 px viewport width the panel MUST render as a full-screen overlay with the same three views and the same close semantics; resize affordances MUST be hidden, not disabled.
• FR-20 The panel MUST restore the last open Conversation on reload, and MUST fall back to the Agent's Conversation list if that Conversation is gone.
• FR-21 Opening a Conversation from any list, card action or link elsewhere in the product MUST load it into the docked panel rather than navigating away from the current screen.
• FR-22 New messages arriving in the open Conversation MUST appear without a manual refresh within 5 seconds.
• FR-23 When live delivery is unavailable, the panel MUST fall back to polling at 30 seconds and MUST NOT show an error for the downgrade.
• FR-24 The panel MUST show an unread marker on Conversations in its lists that have messages newer than the person's last read position, and MUST clear it when the Conversation is opened and scrolled to the newest message.

4.3 The composer — mentions, references, attachments

• FR-25 Typing @ in any conversation composer MUST open a mention picker.
• FR-26 The picker MUST debounce input by 150 ms, return at most 8 candidates, and rank Agents addressed most recently by this person first, then name prefix, then name substring.
• FR-27 The picker MUST offer only Agents and Organization members the person can actually address in this Conversation's kind, and MUST exclude archived Agents.
• FR-28 A mention inserted from the picker MUST render as a highlighted chip carrying the target's display name; two-word names MUST highlight as one unit.
• FR-29 An @ word that does not resolve MUST remain plain, unstyled text. A highlight MUST therefore always mean the mention will land.
• FR-30 Mention matching MUST be case-insensitive and MUST require the full display name or the full slug — never a prefix.
• FR-31 A message MUST carry at most 10 resolved mentions; @ tokens beyond the tenth MUST stay plain text and the composer MUST say so before sending.
• FR-32 Unresolved mention tokens MUST be stripped from what the Agent receives, so no Agent ever sees a reference to a participant that does not exist.
• FR-33 Typing # MUST open the Knowledge Base document picker and insert a reference, per AW-06 FR-49–FR-60. Until that epic ships, the existing document-reference syntax MUST continue to work unchanged in this composer.
• FR-34 A resolved document reference MUST render to the person as the document title as a link, and MUST cause the referenced document's content to be supplied to the addressed Agent's execution.
• FR-35 The composer MUST support attaching files by picker, drag-and-drop and paste, at most 10 per message, each at most 25 MB.
• FR-36 The send control MUST be disabled while any attachment upload is in flight, and MUST show which attachment is still uploading.
• FR-37 A message body MUST NOT exceed 16 KB. The composer MUST block a longer send locally with the actual size in the message and offer attachment as the alternative.
• FR-38 A message whose body matches the platform's credential patterns MUST be rejected before storage, with the same wording used for Agent instruction files.
• FR-39 The composer MUST preserve its text when a send is refused for any reason.
• FR-40 The composer MUST keep its existing affordances — voice dictation, per-message model pin, provider selection — in every Conversation kind where they apply today.

4.4 Sending, failure and retry

• FR-41 Every send MUST carry a client-generated identifier so that a retry of the same message can never create a second message.
• FR-42 A message that fails to send MUST be shown in the Conversation in a failed state with the reason in plain language, not an error code.
• FR-43 A failed message MUST survive a page reload and MUST remain retryable.
• FR-44 A failed message MUST offer Retry and Discard. Retry MUST reuse the original identifier; Discard MUST remove it locally and permanently.
• FR-45 The system MUST NOT retry a failed send automatically.
• FR-46 Rate-limit failures MUST be distinguished from provider failures and from network failures in the person-facing text.
• FR-47 Sends MUST be limited to 30 messages per minute per person, and posts to the organization channel additionally to 10 per hour per person.
• FR-48 Exceeding either limit MUST produce a failed message with a retry, never a silent drop and never a lost composer.

4.5 Group Conversations

• FR-49 A group Conversation MUST be creatable in two ways: explicitly, from a New group action; and implicitly, by mentioning an Agent inside a direct Conversation.
• FR-50 Implicit creation MUST carry the last 20 messages or the last 7 days of the originating Conversation, whichever is fewer.
• FR-51 Implicit creation MUST record a link on both Conversations pointing at each other, and both MUST display it.
• FR-52 Implicit creation MUST NOT modify, move or truncate the originating Conversation.
• FR-53 A group MUST contain the creating person plus 2 to 8 Agents.
• FR-54 Attempting to exceed 8 Agents MUST leave the mention unresolved, send the message as prose, and explain the limit inline.
• FR-55 A person MUST be able to add and remove Agents in an existing group. Removing an Agent MUST preserve every message it already wrote.
• FR-56 Concurrent implicit creations from the same originating Conversation and the same mentioned Agent MUST produce exactly one group.
• FR-57 A group MUST be archivable from its list row, and archiving MUST preserve the transcript in full.
• FR-58 Archived groups MUST be listed under an Archived entry showing at most 100 entries, most recently archived first.
• FR-59 Restoring an archived group MUST return it to the position its last activity earns it, not to the top of the list.
• FR-60 An archived group MUST NOT accept new messages and MUST NOT dispatch Agent replies.
• FR-61 A group MUST be renamable from inside the Conversation; archiving MUST be offered on its list row.
• FR-62 v1 groups MUST contain exactly one person. Multi-person group Conversations are out of scope (§7).

4.6 The organization channel

• FR-63 Each Organization MUST have at most one organization channel, created on first use.
• FR-64 The channel MUST be pinned to the top of the Conversations section and MUST NOT be reordered by activity.
• FR-65 The channel MUST NOT be archivable, renamable or deletable.
• FR-66 Every member of the Organization MUST be able to read and post to the channel.
• FR-67 A post with no mentions MUST be addressed to every Agent in the Organization whose status is active.
• FR-68 A post with mentions MUST be addressed only to the mentioned Agents, and Agents not mentioned MUST receive nothing.
• FR-69 Every channel post MUST record a reach receipt listing, per Agent, one of: delivered, queued (with the queue reason), skipped (with the status that caused it — paused, draft, error or archived), or refused (with the guardrail that refused it).
• FR-70 The reach receipt MUST be visible from the message itself and MUST show a summary count ("Reached 7 of 9") without being opened.
• FR-71 A post MUST be refused when the Organization has more than 200 active Agents, with an explanation and no partial delivery.
• FR-72 A post to an Organization with zero addressable Agents MUST still be stored and displayed, with a receipt saying so.
• FR-73 Delivery to Agents MUST happen through the platform's background job runtime, never inline in the request that posts the message.
• FR-74 The post request MUST return as soon as the message is stored — it MUST NOT block on delivery.
• FR-75 The channel MUST show each message's author, whether a person or an Agent.
• FR-76 Losing Organization membership MUST remove the channel from the person's surfaces and MUST refuse further posts with an explanation, not a generic error.

4.7 Agent-to-Agent Conversations

• FR-77 An Agent MAY open a Conversation with another Agent only when the second Agent is on the first Agent's collaborator allow-list.
• FR-78 There MUST be at most one Agent-pair Conversation per unordered pair of Agents.
• FR-79 Agent-pair Conversations MUST be listed, one row per pair, with the latest message and its time.
• FR-80 A person MUST be able to open any Agent-pair Conversation in their Organization and read it in full.
• FR-81 A person MUST be able to post into an Agent-pair Conversation; their message MUST be attributed to them by name and MUST be delivered to both Agents.
• FR-82 An Agent-pair Conversation MUST pause after 20 consecutive Agent-authored messages with no person's message between them, and MUST display a Paused — waiting for you state naming the reason.
• FR-83 A paused Agent-pair Conversation MUST resume when a person posts into it, resetting the counter.
• FR-84 Agent-authored messages across all Agent-pair Conversations MUST be limited to 200 per day per Organization; beyond the limit, further Agent messages MUST be refused and recorded.
• FR-85 Agent-pair Conversations MUST NOT be archivable or deletable by a person in v1; they are a record of what happened.
• FR-86 A person MUST be able to disable Agent-to-Agent Conversations for the whole Organization with a single setting, defaulting to enabled.

4.8 The reply contract

• FR-87 In a direct Conversation the addressed Agent MUST reply to every person's message.
• FR-88 In a group Conversation or the organization channel, a message with mentions MUST start a reply for each mentioned Agent and MUST NOT start one for any other participant.
• FR-89 In a group Conversation or the organization channel, a message without mentions MUST be delivered to every addressable Agent, each of which independently decides whether to reply. Not replying is a correct outcome and MUST NOT be reported as a failure anywhere.
• FR-90 At most 8 reply executions MUST be started from any single message; beyond that the remainder MUST be recorded as queued in the receipt.
• FR-91 Mentioning the same Agent more than once in one message MUST start exactly one reply.
• FR-92 A mention of an Agent that already has a live execution started from this Conversation MUST be delivered into that execution rather than starting a second one, and the Conversation MUST show that it was delivered to a run in progress.
• FR-93 When a message addressed without mentions produces no replies at all, the Conversation MUST render a neutral line stating that no one had anything to add and how to ask directly.

4.9 Permissions, scope and privacy

• FR-94 A Conversation MUST be readable only by its participants, and additionally — for the organization channel and Agent-pair Conversations — by members of the owning Organization.
• FR-95 A request for a Conversation the caller may not read MUST be indistinguishable from a request for one that does not exist.
• FR-96 Mention candidates MUST be scoped to what the caller can already see; the picker MUST NOT reveal the existence of any Agent, person or document otherwise hidden from them.
• FR-97 Conversations, participants and messages MUST carry the platform's tenant and Organization scope, and every list MUST filter on the active scope.
• FR-98 Switching the active Organization MUST switch which organization channel and which Agent-pair Conversations are visible, without leaking rows across scopes.
• FR-99 Attachments MUST inherit the visibility of the Conversation they were posted in.
• FR-100 Deleting a Conversation MUST delete its messages, its participants and its reach receipts. The organization channel MUST NOT be deletable while the Organization exists.
• FR-101 Deleting an Agent MUST NOT delete Conversations it participated in; its messages remain and its participation is marked as departed.
• FR-102 The existing "delete all my conversations" behaviour MUST continue to delete only the caller's own direct and group Conversations, never the organization channel and never Agent-pair Conversations.
• FR-103 No message body, mention or attachment name may be written to application logs.

4.10 Attention, cost and receipts

• FR-104 A person MUST be notified when they are mentioned by name in a group Conversation or the organization channel.
• FR-105 An Agent's ordinary reply in a Conversation the person is currently viewing MUST NOT produce a notification.
• FR-106 Notification routing, batching and quiet hours are owned by AW-13; this epic MUST emit the events and MUST NOT implement a second routing mechanism.
• FR-107 Every Agent-authored message MUST link to the execution that produced it, and that execution's receipt MUST show what it cost.
• FR-108 Every organization-channel post MUST show the total cost of the executions it caused, updating as they complete.
• FR-109 A message whose reply execution was refused for budget reasons MUST say so in the Conversation, naming the cap that refused it.
• FR-110 Conversation activity MUST be recorded in the platform activity record at the granularity of Conversation created, named, archived, restored, and message posted — never the message body.

5. Key entities

Concept	Status	Notes
Conversation	Exists, extended	Today: one user, an optional machine title, a provider and a model. This epic adds a kind, an optional addressed Agent, a name that a person owns, an optional attached context object, an archived state, a last-activity time and a link to a related Conversation. No rename, no replacement.
Conversation message	Exists, extended	Today: role, content, parts, model, token usage. This epic adds an author (a person or an Agent), resolved mentions, attachments, a send status, a client identifier for safe retry, and — on channel posts — a reach receipt.
Conversation participant	NEW	The one genuinely new entity. Today a Conversation has exactly one user and no way to express a second party. A participant is a person or an Agent in a Conversation, with a role (owner, member, observer), a joined and optionally departed time, a last-read position and a mute flag. Everything in this epic that is not one-to-one needs it: group membership, channel membership, Agent pairs, unread state and read receipts. It is added to the program vocabulary table in the same change.
Agent	Exists, unchanged	Addressed by Conversations and by mentions. Its status lifecycle (draft, active, running, paused, error, archived) is what the reach receipt reports against. Its collaborator allow-list is what authorises an Agent pair.
Run	Exists, extended	An Agent reply is a Run like any other. This epic adds a link from a Run back to the Conversation message that triggered it, and a trigger kind for it.
Organization	Exists, unchanged	Owns exactly one organization channel and scopes Agent-pair visibility.
Mention	Exists as a field, generalised	Task comments already carry resolved mentions of people, Agents and documents. The same shape is reused; nothing new is invented.
Reach receipt	New field, not an entity	Per-Agent delivery outcome recorded on a channel post. Deliberately not a table: it is written once, read with its message, and bounded by the Agent count.
Knowledge Base document	Exists, referenced	The # reference target. Owned by AW-06.

5.1 Conversation states and transitions

                    (person sends first message)
                               │
                               ▼
   ┌──────────┐  name set  ┌──────────┐  name cleared   ┌──────────┐
   │ UNNAMED  │───────────►│  NAMED   │────────────────►│ UNNAMED  │
   │ (preview)│            │ (bold)   │                 │ (preview)│
   └──────────┘            └──────────┘                 └──────────┘
        │                        │
        │  @mention of a second Agent (direct kind only)
        ▼                        ▼
   ┌────────────────────────────────────────────┐
   │  GROUP created — carries last 20 messages, │
   │  both Conversations cross-linked           │
   └────────────────────────────────────────────┘
        │                          ▲
        │ archive (group only)     │ restore (to activity position)
        ▼                          │
   ┌──────────┐                    │
   │ ARCHIVED │────────────────────┘
   │ read-only│
   └──────────┘

  ORGANIZATION CHANNEL:  created-on-first-use ──► active (terminal; never archived)
  AGENT PAIR:            active ⇄ paused-waiting-for-you (20 consecutive Agent messages)

5.2 Message states

  composing ──send──► SENDING ──ok────► SENT ──(agent replies land as new messages)
                         │
                         └──error────► FAILED ──retry──► SENDING
                                          │
                                          └──discard──► (gone, locally)

SENT is terminal for a person's message. An Agent's message is SENT on write.

5.3 Reach outcomes for one channel post

  post ──► for each addressable Agent:
             ├── delivered   (a reply execution started)
             ├── queued      (admitted later; reason recorded, e.g. concurrency ceiling)
             ├── skipped     (Agent paused / draft / error / archived)
             └── refused     (a guardrail or a budget cap said no; the rule is named)

6. UX

Copy in this section is the exact user-visible text.

6.1 The docked panel — Conversation view

┌────────────────────────────────────────────────────┐
│ ‹  Nova                                    ⌄   ✕   │  ‹ = up to Nova's conversations
│    Q4 pricing page                         ✎       │  ⌄ = switch participant
│    ◈ Mission: Refresh Q4 pricing                   │  ✎ = name this conversation
├────────────────────────────────────────────────────┤
│                                                    │
│  You · 09:14                                       │
│  Can you check whether the pricing page still      │
│  claims the old trial length? #Pricing policy      │
│                                       ▲            │
│                                       └ document link
│                                                    │
│  Nova · 09:14                                      │
│  It does — three places. I have drafted the fix.   │
│  ┌──────────────────────────────────────────────┐  │
│  │ 📄 Pricing page — trial length corrections   │  │
│  │    Knowledge Base · written 09:14            │  │
│  └──────────────────────────────────────────────┘  │
│  Run receipt · 41s · $0.06                         │
│                                                    │
├────────────────────────────────────────────────────┤
│ ┌────────────────────────────────────────────────┐ │
│ │ Message Nova…                                  │ │
│ └────────────────────────────────────────────────┘ │
│ 📎  🎙  Auto model ⌄                    Send  ↵    │
└────────────────────────────────────────────────────┘
     ▲
     └ drag to resize · double-click to reset

6.2 The docked panel — an Agent's Conversation list

┌────────────────────────────────────────────────────┐
│ ‹  Nova                                    ⌄   ✕   │
├────────────────────────────────────────────────────┤
│  + New conversation                                │
├────────────────────────────────────────────────────┤
│  Q4 pricing page                          ● 09:14  │  ● = unread
│  Can you check whether the pricing page…           │
├────────────────────────────────────────────────────┤
│  Can you summarise yesterday's inbound…    Mon     │  (no name → preview only)
├────────────────────────────────────────────────────┤
│  Launch checklist                          12 Aug  │
│  Here is where we are on the launch…               │
└────────────────────────────────────────────────────┘

EMPTY
│  No conversations with Nova yet.                   │
│  Send a message to start one.                      │
│  [ Message Nova ]                                  │

LOADING
│  ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒                                │
│  ▒▒▒▒▒▒▒▒▒▒▒▒                                      │
│  ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒                                  │

ERROR
│  Could not load conversations.                     │
│  [ Try again ]                                     │

6.3 The docked panel — participant switcher

┌────────────────────────────────────────────────────┐
│ ‹  Switch                                      ✕   │
├────────────────────────────────────────────────────┤
│  Search agents and conversations…                  │
├────────────────────────────────────────────────────┤
│  ◆ Organization channel                    2h      │
│    Everyone — launch moved to the 14th             │
├────────────────────────────────────────────────────┤
│  AGENTS                                            │
│  ● Nova            Research               09:14    │
│  ● Orion           Content                Mon      │
│  ○ Piper           Support        Paused  12 Aug   │
├────────────────────────────────────────────────────┤
│  GROUPS                                            │
│  ● Launch week                            11:02    │
│  ● Nova, Orion & you                      Fri      │
├────────────────────────────────────────────────────┤
│  Agent conversations                         ›     │
├────────────────────────────────────────────────────┤
│  Archived (3)                                ›     │
└────────────────────────────────────────────────────┘

● = active Agent, ○ = not currently addressable, ◆ = the pinned organization channel.

6.4 The Conversations section in the sidebar

▾ Conversations                                 + ⌄
  ◆ Organization channel                        2h
  ● Launch week                              11:02
  ● Nova, Orion & you                          Fri
  ● Nova                                     09:14
    Archived (3)                               ›

NO ORGANIZATION
  Create an Organization to broadcast to every Agent
  at once.                        [ Create Organization ]

EMPTY
  No conversations yet.                [ Start one ]

6.5 The name control

┌────────────────────────────────────────────────────┐
│  Name this conversation                            │
│  ┌──────────────────────────────────────────────┐  │
│  │ Q4 pricing page                              │  │
│  └──────────────────────────────────────────────┘  │
│  Clearing the name shows the first message         │
│  instead.                                          │
│                        [ Cancel ]  [ Save ]        │
└────────────────────────────────────────────────────┘

OVER LIMIT
│  Names are at most 200 characters. This one is 214.│

6.6 Composer — mention picker and live highlighting

┌────────────────────────────────────────────────────┐
│  Nova, can you get @or| to check the copy?         │
│                     ┌─────────────────────────────┐│
│                     │ ● Orion         Content     ││
│                     │ ○ Ordering desk  Support    ││
│                     ├─────────────────────────────┤│
│                     │ ↑↓ move · ↵ insert · esc ✕  ││
│                     └─────────────────────────────┘│
└────────────────────────────────────────────────────┘

AFTER INSERT — the chip is highlighted, the plain word is not
┌────────────────────────────────────────────────────┐
│  Nova, can you get (@Orion) to check the copy?     │
│  I asked @someone-else already.                    │
│            ╰── stays plain: this will not be a     │
│                mention                             │
└────────────────────────────────────────────────────┘

PICKER — NO MATCH
│  No agent or teammate matches "zz".                │
│  Keep typing, or press esc to leave it as text.    │

PRE-SEND HINTS (under the composer, non-blocking)
ⓘ Mentioning Orion will start a group conversation with Nova and Orion.
ⓘ Only 10 mentions land in one message. The rest stay as plain text.
ⓘ A group holds up to 8 agents. This one is full — start a new group?

6.7 The organization channel

┌──────────────────────────────────────────────────────────────┐
│ ‹  ◆ Organization channel                            ⌄   ✕   │
│    One message to every active agent                         │
├──────────────────────────────────────────────────────────────┤
│                                                              │
│  You · 08:02                                                 │
│  Launch moves to the 14th. If you are mid-mission on          │
│  anything that assumes the 7th, say so here before you        │
│  change course.                                              │
│                                                              │
│  ✓ Reached 7 of 9                                       ⌄    │
│  ┌────────────────────────────────────────────────────────┐  │
│  │ Delivered                                              │  │
│  │   Nova · Orion · Iris · Kai · Devon                     │  │
│  │ Queued — organization is at its run ceiling            │  │
│  │   Sasha · Quinn                                        │  │
│  │ Skipped — paused                                       │  │
│  │   Piper                                                │  │
│  │ Skipped — archived                                     │  │
│  │   Halcyon                                              │  │
│  ├────────────────────────────────────────────────────────┤  │
│  │ 5 replies · $0.31 so far                               │  │
│  └────────────────────────────────────────────────────────┘  │
│                                                              │
│  Nova · 08:03                                                │
│  Mission "Launch email" assumes the 7th. Rescheduling.       │
│  Run receipt · 12s · $0.04                                   │
│                                                              │
├──────────────────────────────────────────────────────────────┤
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ Message everyone… mention someone to narrow it           │ │
│ └──────────────────────────────────────────────────────────┘ │
│ 📎  🎙                                          Send  ↵      │
└──────────────────────────────────────────────────────────────┘

NARROWED POST
│  ✓ Reached 2 of 9 — you mentioned Nova and Orion        ⌄    │

NO ADDRESSABLE AGENTS
│  ⚠ Reached 0 of 4 — every agent is paused.                   │
│    [ Open agents ]                                           │

OVER LIMIT
│  This organization has 214 active agents. Broadcasts reach   │
│  at most 200. Mention the agents you need instead.           │

NO LONGER A MEMBER
│  You no longer have access to this Organization's channel.   │
│  Your message is still in the box.                           │

6.8 Group Conversation

┌──────────────────────────────────────────────────────────────┐
│ ‹  Launch week                                       ⌄   ✕   │
│    Nova · Orion · Iris · you                         ✎  ⋯    │
│    ↩ Continued from your conversation with Nova              │
├──────────────────────────────────────────────────────────────┤
│  ⓘ Carried the last 20 messages from that conversation.      │
│                                                              │
│  You · 10:58                                                 │
│  (@Orion) can you take the copy pass on this?                │
│                                                              │
│  Orion · 10:58                                               │
│  On it. First draft in about ten minutes.                    │
│  Run receipt · 8s · $0.01                                    │
│                                                              │
│  You · 11:02                                                 │
│  Anything anyone wants to flag before we lock this?          │
│                                                              │
│  ⌁ No one had anything to add. Mention someone by name to    │
│    ask directly.                                             │
├──────────────────────────────────────────────────────────────┤
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ Message the group…                                       │ │
│ └──────────────────────────────────────────────────────────┘ │
└──────────────────────────────────────────────────────────────┘

ROW MENU (⋯)
┌──────────────────────────┐
│  Rename                  │
│  Add an agent…           │
│  Remove an agent…        │
├──────────────────────────┤
│  Archive                 │
└──────────────────────────┘

ARCHIVED VIEW
│  This conversation is archived. Restore it to reply.         │
│                                           [ Restore ]        │

6.9 Agent conversations (Agent-to-Agent)

┌──────────────────────────────────────────────────────────────┐
│ ‹  Agent conversations                               ⌄   ✕   │
│    What your agents say to each other                        │
├──────────────────────────────────────────────────────────────┤
│  Nova ⇄ Orion                                        11:40   │
│  Orion: sending you the corrected copy now                   │
├──────────────────────────────────────────────────────────────┤
│  Nova ⇄ Iris                                         Mon     │
│  Nova: the source list is stale, can you refresh it?         │
├──────────────────────────────────────────────────────────────┤
│  Kai ⇄ Devon                              ⏸ waiting for you  │
│  Devon: I still think we should hold. Kai?                   │
└──────────────────────────────────────────────────────────────┘

EMPTY
│  Your agents have not needed to talk to each other yet.      │
│  They will appear here when they do.                         │

INSIDE A PAIR
┌──────────────────────────────────────────────────────────────┐
│ ‹  Nova ⇄ Orion                                          ✕   │
├──────────────────────────────────────────────────────────────┤
│  Nova · 11:38                                                │
│  The pricing page copy is wrong in three places.             │
│                                                              │
│  Orion · 11:40                                               │
│  Sending you the corrected copy now.                         │
│                                                              │
│  You · 11:41                                                 │
│  Hold on — check with me before publishing.                  │
│  ╰── your message, attributed to you, delivered to both      │
├──────────────────────────────────────────────────────────────┤
│ │ Step in…                                                 │ │
└──────────────────────────────────────────────────────────────┘

PAUSED
│  ⏸ Paused — waiting for you.                                 │
│    Kai and Devon exchanged 20 messages without you. Post to   │
│    continue.                                                 │

6.10 Failed send and retry

┌──────────────────────────────────────────────────────────────┐
│  You · 09:31                                          ⚠      │
│  Can you pull the last quarter's numbers?                    │
│  ⚠ Not sent — you are sending faster than the rate limit     │
│    allows. Nothing was lost.                                 │
│                              [ Retry ]   [ Discard ]         │
└──────────────────────────────────────────────────────────────┘

OTHER REASONS (same shape, different line)
⚠ Not sent — the model provider is unavailable right now.
⚠ Not sent — you are offline.
⚠ Not sent — this message looks like it contains a credential.
  Store it in the connection's settings instead.
⚠ Not sent — this message is too long (40 KB of 16 KB).
  [ Attach as a file ]

RETRY IN FLIGHT
│  ↻ Sending…                                                  │

6.11 Budget refusal inside a Conversation

│  ⚠ Nova did not reply — the agent's spend cap for this        │
│    period was reached. [ Open caps ]                          │

6.12 Keyboard affordances

Key	Where	Effect
Enter	Composer	Send
Shift+Enter	Composer	New line
Esc	Composer with picker open	Dismiss the picker, keep the text
Esc	Composer, no picker	Move focus out of the composer; never closes the panel
@	Composer	Open the mention picker
#	Composer	Open the document picker
↑ ↓	Picker open	Move the selection
Enter / Tab	Picker open	Insert the highlighted candidate
↑	Empty composer	Focus the last message for keyboard reading
Alt+←	Panel	Back one panel view (Conversation → list → switcher)
Alt+↑ / Alt+↓	Panel list	Previous / next Conversation
Ctrl/Cmd+Enter	Failed message focused	Retry
Tab	Panel edge handle	Focus the resize handle; ←/→ resize by 16 px; Home resets

Every control named in §6 is reachable by keyboard, has an accessible name, and the panel traps nothing — focus can always leave it with Tab.

7. Out of scope

• Multi-person group Conversations. A group holds one person and up to 8 Agents. Two people in one Conversation is a teammate-access question owned by AW-18.
• Message editing, deletion, reactions and quoting. Existing unused edit strings are left exactly where they are; this epic neither uses nor removes them.
• Nested replies inside a group or the channel. Group and channel Conversations are flat.
• Full-text search across Conversations. AW-01 reaches Conversation names; searching message bodies is explicitly deferred.
• Conversation export. Owned by AW-22.
• The # document picker itself. Owned by AW-06; this epic consumes it and keeps today's reference syntax working until it lands.
• Notification routing, batching, quiet hours and digests. Owned by AW-13; this epic only emits.
• Bridging Conversations to external messaging surfaces. The existing external chat bridge keeps working exactly as it does today and is not extended, rewired or given participants.
• Changing the assistant's tool-calling behaviour, its tool gating, its canvas or its provider and model resolution. Untouched.
• Voice or video calls, presence and typing indicators. Not in v1.
• Retention or truncation policies for Conversations. Nothing is auto-deleted by this epic.
• Per-Conversation model or provider overrides beyond what ships today.

8. Acceptance criteria

Conversations and naming

• A message sent from an Agent's composer with nothing open creates a new Conversation addressed at that Agent.
• A Conversation list row shows a bold name over a preview when named, and a preview alone when not; no row ever reads "Untitled".
• Setting a name stops automatic titling permanently; clearing it restarts it.
• A 201-character name is refused with the count shown.
• A group with no custom name displays a participant-derived name.
• The organization channel and Agent-pair Conversations cannot be renamed.
• A Conversation opened from a Mission card shows the Mission chip and the Agent receives the Mission as context.

The panel

• Navigating across five screens does not close the panel and does not lose the scroll position.
• The back control walks Conversation → Agent's list → switcher, and the switcher lists Agents, groups, the channel and Agent conversations.
• Dragging the edge resizes between 350 px and 50 % of the viewport; double-clicking resets to 420 px; the width survives reload with no visible reflow.
• Below 768 px the panel is a full-screen overlay with the same three views.
• A new message in the open Conversation appears within 5 seconds without a refresh, and the surface degrades to a 30-second poll with no error shown.

Composer

• @ opens a picker within 150 ms of the second character, showing at most 8 candidates.
• An inserted mention is highlighted; a hand-typed non-matching @ word is not.
• A two-word name highlights as a single unit.
• The eleventh mention in one message stays plain and the composer says why.
• # opens the document picker and the referenced document reaches the Agent's execution.
• Eleven attachments are refused; a 26 MB attachment is refused; send stays disabled while an upload is in flight.
• A 40 KB body is blocked in the browser with the real size in the message.
• A body containing a credential pattern is rejected before storage.
• Every refusal leaves the composer text intact.

Failure and retry

• A failed send shows a failed message with a plain-language reason, a Retry and a Discard.
• The failed message survives reload.
• Retrying twice in quick succession produces exactly one delivered message.
• The 31st message in a minute fails with a rate-limit reason and a working Retry.

Groups

• Mentioning a second Agent in a one-to-one Conversation creates a group carrying 20 messages, cross-linked both ways, leaving the original intact.
• Two simultaneous promotions from the same Conversation create exactly one group.
• The ninth Agent cannot be added and the limit is explained inline.
• Archiving preserves the transcript, removes the group from the active list, and lists it under Archived.
• Restoring returns it to its activity position, not the top.
• An archived group refuses new messages and starts no executions.

The organization channel

• Exactly one channel exists per Organization and it is pinned first.
• It cannot be archived, renamed or deleted.
• An unmentioned post reaches every active Agent; a mentioned post reaches only those named, and the receipt distinguishes the two cases in words.
• The receipt separates delivered, queued, skipped and refused, and names the reason for each non-delivery.
• Posting with the Organization at its run ceiling still succeeds and shows queued Agents.
• Posting with zero addressable Agents shows "Reached 0 of N" and a link to the Agents list.
• An Organization with 201 active Agents refuses the broadcast with an explanation and delivers nothing.
• The post request returns before delivery completes.
• A person removed from the Organization is refused with the stated message and keeps their composer text.

Agent conversations

• An Agent cannot open a Conversation with an Agent that is not on its collaborator allow-list.
• Exactly one Conversation exists per pair.
• A person can read any pair in their Organization and post into it; their message is attributed to them and reaches both Agents.
• After 20 consecutive Agent messages the pair pauses with the stated reason and resumes on a person's post.
• Disabling Agent-to-Agent Conversations for the Organization stops new pairs being created.

The reply contract

• A mentioned Agent replies; an unmentioned Agent in the same group is not dispatched.
• An unmentioned post reaches every addressable Agent and produces zero, one or several replies without any of those being reported as an error.
• Mentioning one Agent twice starts one reply.
• Mentioning an Agent with a live execution from this Conversation delivers into it and the Conversation says so.
• A message that produces no reply renders the neutral "no one had anything to add" line.

Permissions and cost

• A Conversation the caller does not participate in is indistinguishable from one that does not exist.
• Switching Organizations switches the channel and the Agent pairs with no cross-scope leakage.
• Every Agent message links to a Run receipt showing its cost.
• A channel post shows the running total cost of the executions it caused.
• A budget-refused reply says which cap refused it.
• No message body appears in any log line.

9. Open questions

• [NEEDS CLARIFICATION: should Agents be able to post to the organization channel unprompted?] This spec allows an Agent to reply in the channel but does not let one start a channel post. Allowing it would make the channel a genuine team feed; refusing it keeps the channel a person-initiated broadcast. The reach receipt is defined only for person-authored posts today.
• [NEEDS CLARIFICATION: what exactly is an "active" Agent for broadcast purposes?] This spec reads it as status active or running. An Agent in error is treated as skipped. Product should confirm that an Agent that auto-paused after repeated failures should be skipped rather than queued for later delivery.
• [NEEDS CLARIFICATION: does a broadcast reach an Agent that becomes active later?] Today's rule is point-in-time: the receipt is computed at post time and never re-evaluated. A "catch-up on resume" rule would need a durable pending-delivery record.
• [NEEDS CLARIFICATION: should groups be per-person or per-Organization?] v1 makes a group belong to its creator. If two people in an Organization each pull the same two Agents into a group, they get two groups. AW-18 may want one shared group instead.
• [NEEDS CLARIFICATION: how much history should a promoted group carry?] 20 messages / 7 days is a guess calibrated on the 16 KB body cap and typical context budgets. It should be validated against real transcript lengths once the surface has usage.
• [NEEDS CLARIFICATION: unread semantics for the organization channel.] With every Organization member reading the same Conversation, an unread count is per-person, but a very active channel could produce a permanently non-zero badge. A "mark all read" is specified; whether the channel should badge at all is a Design call, coordinated with AW-13.
• [NEEDS CLARIFICATION: should the 200-per-day Agent-message ceiling be per Organization or per pair?] Per Organization is specified because that is where the spend lives, but it means one chatty pair can consume another pair's allowance.
• [NEEDS CLARIFICATION: what happens to a group when its last Agent is archived?] The group currently persists as a read-only record. Product may prefer auto-archiving it.