You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
buzz: a hosted harness loses every channel's conversation and sandbox on each deploy (buzz-acp session/new after restart) #774
A hosted Buzz agent's harness (Fountain.Buzz.Harness, buzz-acp in the pod) restarts on every deploy — Horde moves it to the new pod. buzz-acp keeps its channel→ACP-session map in memory only (crates/buzz-acp/src/pool.rs, SessionState.sessions: HashMap<Uuid, String>) and never calls session/load/session/resume; after a restart the next mention in a channel gets a fresh session/new. Through fountain acp that is a new conversation and a new sandbox. Observed 2026-08-16 with FizzTheShark: after the harness restart (#773), a DM continued in a new conversation f313b4e4 instead of 1d239529; the old sandbox is suspended with the workspace on it, unreachable from Buzz.
With ~24 deploys/day, a hosted agent's per-channel context and workspace last hours. Buzz's own base prompt assumes the opposite ("after a session restart, resume silently — rebuild state from your todos, memory…" — a persistent cwd).
Why the gateway cannot fix it alone today
session/new from buzz-acp carries cwd, mcpServers, _meta.sessionTitle (agent + channel name, DMs get no name), and _meta.systemPrompt. No channel id — the only channel-id carrier is BUZZ_GIT_ORIGIN_CHANNEL_ID, and only for stream channels with MCP servers configured. The id does appear in the first prompt's [Context] Channel: DM (#…) text, which is not a key to build on.
Proposed
Upstream (block/buzz), the real fix — drafted alongside this issue: (a) _meta.channelId / _meta.channelType on session/new; (b) buzz-acp persists channel→session per identity and calls session/load on restart when the agent advertises loadSession (Fountain does), falling back to session/new.
Gateway sticky resume (once 1a lands, or as _meta becomes available): fountain acp maps (identity, channelId) → conversation and treats session/new for a known channel as a resume of that conversation, so continuity does not depend on buzz-acp's memory.
Fountain-side, independent of upstream: a Buzz identity's new conversations reuse the identity's previous sandbox/workspace (or a per-identity persistent volume), so memory/todos survive even when the ACP session does not. Matches Buzz's "rebuild from your workspace" contract and is worth doing regardless of 1/2.
Gateway half shipped in #775 (channel-bound conversations, hosted buzz-acp built from the fork). Remaining: #776 (repin to upstream when block/buzz#6088 ships), #778/#779 (24 h-ceiling cliff, deferred — not worried about 24 h agents for now).
Closing — the reported break is fixed on the build we ship; the one remaining idea is better tracked elsewhere.
Conversation continuity across a deploy works today.#775 added the gateway sticky resume: session/new carrying _meta.channelId maps (identity, channel) to a conversation and resumes it rather than creating one (cli/internal/acp/session.go:126,186 → Conversations.start_or_resume_conversation/2). A hosted harness runs our pinned buzz-acp, which sends that member, so a Horde move on deploy no longer strands the channel. The suspended sandbox comes back with the conversation (ADR 0017), so the workspace returns too.
3 (a Buzz identity's new conversations reuse its previous sandbox/workspace) — never built, and it is not really this bug. It is a per-identity persistent machine, which is exactly the reframe Design note: the sandbox as a first-class Machine — what Fountain looks like with agent and sandbox decoupled (companion to ADR 0023) #805 already holds as its north star ("the sandbox is the fifth primitive; a conversation is a binding of an agent to a machine"). Tracking it as a fix for a deploy-restart symptom that no longer occurs would misfile it. It belongs to that design or to an ADR extending it.
Nothing here is lost: #776 for the pin, #805 for the persistent workspace, #936 for the ceiling that still costs an agent its memory.
Activity
feat(conversations): channel-bound conversations — resume by channel_…
jhgaylor commentedon Aug 17, 2026
Gateway half shipped in #775 (channel-bound conversations, hosted buzz-acp built from the fork). Remaining: #776 (repin to upstream when block/buzz#6088 ships), #778/#779 (24 h-ceiling cliff, deferred — not worried about 24 h agents for now).
feat(acp): honour _meta.freshSession — !rotate opens a new channel co…
feat(acp): honour _meta.freshSession — !rotate opens a new channel co…
feat(agui): Fountain speaks AG-UI, so an OpenBot coworker can be a Fo…
feat(agui): Fountain speaks AG-UI, so an OpenBot coworker can be a Fo…
jhgaylor commentedon Aug 22, 2026
Closing — the reported break is fixed on the build we ship; the one remaining idea is better tracked elsewhere.
Conversation continuity across a deploy works today. #775 added the gateway sticky resume:
session/newcarrying_meta.channelIdmaps (identity, channel) to a conversation and resumes it rather than creating one (cli/internal/acp/session.go:126,186→Conversations.start_or_resume_conversation/2). A hosted harness runs our pinned buzz-acp, which sends that member, so a Horde move on deploy no longer strands the channel. The suspended sandbox comes back with the conversation (ADR 0017), so the workspace returns too.Where each proposal landed:
_meta.channelId+ a session store) — buzz-acp: resume channel sessions across a restart, and name the channel on session/new block/buzz#6088, still open. We carry it as a fork pin; buzz-acp: retire fork after all required behavior ships upstream #776 owns dropping that pin, and now covers all three forked PRs rather than just this one. Desktop agents, which run the released buzz-acp rather than ours, stay affected until it ships — stated on buzz-acp: retire fork after all required behavior ships upstream #776.Nothing here is lost: #776 for the pin, #805 for the persistent workspace, #936 for the ceiling that still costs an agent its memory.