Summary
There is no way to regenerate from an edited assistant message. Editing an assistant response works — the edited copy becomes the session leaf and the next user message continues from it — but the model cannot produce its own continuation from that edited text without a new prompt.
Why it cannot be done today
AgentSession.prompt(text: string, options?) requires text. There is no promptless continuation anywhere in the stack:
- the agent's prompt API takes a required string;
- a downstream desktop adapter therefore cannot advertise a
promptlessTurnContinuation capability, and its provider layer refuses a promptless turn without one;
- only a different provider's adapter declares that capability today.
So "edit an assistant response, then let the model continue from the edited version" is unreachable for this agent, while it is reachable for a provider that supports promptless continuation.
Impact
A message-edit feature shipped with three of its four intended cells: users can edit their own message with or without re-running it, and can edit an assistant response. The fourth — edit an assistant response and have the model regenerate from it — was documented as knowingly reduced, because closing it requires this capability rather than any UI change.
The shipped workaround is honest rather than silent: after saving an assistant edit the UI states that the next message continues from the edited response, and the disabled retry control explains that retrying would re-run the original prompt and discard the edit.
What would close this
A way to start a turn with no new prompt, continuing from the current leaf — whatever shape fits the session model. Once it exists, the consumer side is a small change: advertise the capability in the adapter and re-enable the action. No new addressing or id scheme is needed; navigate_tree already gained an explicit intent: "resume" for setting the leaf to an exact entry.
Acceptance
- A turn can be started from a leaf with no prompt text, and the resulting assistant entry is appended under that leaf.
- Editing an assistant message and then continuing produces a new response derived from the edited text, with the original preserved on an abandoned branch as today.
- Existing prompt-carrying turns are unchanged.
Summary
There is no way to regenerate from an edited assistant message. Editing an assistant response works — the edited copy becomes the session leaf and the next user message continues from it — but the model cannot produce its own continuation from that edited text without a new prompt.
Why it cannot be done today
AgentSession.prompt(text: string, options?)requires text. There is no promptless continuation anywhere in the stack:promptlessTurnContinuationcapability, and its provider layer refuses a promptless turn without one;So "edit an assistant response, then let the model continue from the edited version" is unreachable for this agent, while it is reachable for a provider that supports promptless continuation.
Impact
A message-edit feature shipped with three of its four intended cells: users can edit their own message with or without re-running it, and can edit an assistant response. The fourth — edit an assistant response and have the model regenerate from it — was documented as knowingly reduced, because closing it requires this capability rather than any UI change.
The shipped workaround is honest rather than silent: after saving an assistant edit the UI states that the next message continues from the edited response, and the disabled retry control explains that retrying would re-run the original prompt and discard the edit.
What would close this
A way to start a turn with no new prompt, continuing from the current leaf — whatever shape fits the session model. Once it exists, the consumer side is a small change: advertise the capability in the adapter and re-enable the action. No new addressing or id scheme is needed;
navigate_treealready gained an explicitintent: "resume"for setting the leaf to an exact entry.Acceptance