Carry the investigation across people and AI assistants

A mail discussion belongs to a particular canonical message and its selected context. A topic provides a separate conversation for an issue that spans several messages or accounts. Attaching mail to a topic links the sources without merging originals, moving their comments or overwriting their private tracking. Removing that link preserves the mail and its own discussion, while the topic keeps appropriate membership history.

Topics are useful when a business exchange continues through another company address, or when a decision depends on separate requests. Topic participation and source access remain separate. A person may participate in a topic without reading every linked original. The assistant must retrieve each source under its own current rights rather than treating membership as permission to open everything.

A useful comment explains the claim, evidence examined and uncertainty that remains. It can quote a permitted passage or point to a visible earlier finding, but an attribution such as checked the contract is still a participant's claim. The discussion should make the checked and unchecked scope understandable rather than presenting an ordinary comment as a certified execution receipt.

Comments keep their author and original content. Corrections and replies add a new attributed step, so another reader can see how a claim changed. A correction is not automatically the truth, and a label such as question or decision does not create a separate task engine. The discussion records reasoning; external sending, account changes and other effects retain their own authority and approval rules.

Each enrolled assistant has its own work state in a discussion context. It may accept work, plan it for later, work on it, pause, wait, complete its part or decide participation is unnecessary. Those choices describe that assistant's judgment. There is no global done flag that makes all assistants finished because one participant closes its work.

Reading and work are also independent. An assistant can acknowledge a summary without starting analysis, or continue unfinished work without being forced to reread the whole original. Pausing does not mean a comment silently resumes execution, and waiting does not mean an incoming message automatically satisfies the awaited condition. A completed assistant can retain interest in discussion, while its completed disposition remains intact until it chooses otherwise.

Mail information and comments have distinct reading channels. A summary acknowledgment covers that summary; it does not claim the assistant read the original body, inspected every attachment or understood the content. Acknowledging a comment page covers the visible comments on that page, not the originals mentioned in them. Merely listing, searching or opening a history view does not acknowledge everything returned.

This separation gives rechecks a precise meaning. A new finding may require reading a comment, a specific original passage or an attachment. The assistant can decide what evidence is needed, use eligible prior body knowledge where available and report the remaining gap. Its private reading state does not become a public attendance register. A human opening the page likewise does not fabricate an agent's reading or change the provider's Seen flag.

A visible comment directed to a known assistant can bring new discussion to its attention even after that assistant completed its part. This creates a reason to inspect the comment; it does not reopen the work, prove original rereading or compel execution. The receiving assistant evaluates the evidence and chooses whether to resume, ask for more information, correct its conclusion or explain why it still stands.

General comments follow the participant's discussion interest and prior participation. They are not a broadcast command to every registered agent. This keeps renewed attention useful while preserving private work decisions. A configured client may use supported subscriptions or checks to discover that attention, but the discussion itself does not guarantee that an offline assistant wakes up or answers immediately.

The discussion provides a common record for assistants used through different AI services. One may be good at finding historical mail, another at checking a numerical argument, and a person may supply the business constraint that neither can infer. Verified participant identity and delegated rights matter more than a model name written in a comment. Each client uses the connection and operations it actually supports.

People participate directly by asking questions, adding context and recording decisions. Posting a discussion comment does not send email. Addressing a participant also differs from selecting the audience: a comment can be directed to one assistant yet readable to the permitted group, or restricted to a smaller set. An author cannot give a reader source access merely by naming that reader in a message.

A useful final decision states what was chosen, why, what evidence supported it and what remains conditional. It should distinguish the source's facts from the participants' interpretation. When the issue returns months later, an assistant can retrieve the decision and inspect the relevant original instead of treating the last confident sentence as an unquestionable answer.

Private conversations and explicitly shared discussions stay separate. Hidden comments must remain hidden, including their activity, rather than appearing as tantalizing placeholders. Source deletion also needs an explicit retention choice for the discussion and work records; keeping a conversation is not permission to keep every original forever. These boundaries make the record useful to later participants while preserving who may inspect its evidence and reasoning.