Summary
Agents must survive Workshop exit and relaunch. Every tool call and every assistant response is recorded to a per-agent SQLite database, and an agent window is restored by replaying that history into a fresh Lua VM. Serializing the Lua VM itself was researched on 2026-09-10 and rejected as infeasible.
Scenario
From the #cppa-wg21-org discussion on 2026-09-10:
- An agent is a PromptForge prompt currently suspended at a tool call, typically
tools.call("user_input"). Pressing Send resumes it and the tool returns the typed text.
- The user exits the Workshop. Days or weeks pass. They relaunch, and the agent window must come back exactly where it stood.
- The time evolution of a prompt is a linear sequence of suspensions at tool calls, so the record is a linear log.
Direction
- Every agent owns an SQLite database that holds everything that ever transpired. "The agent is the database."
- Restore replays every tool call and every assistant response from the history. No inference cost, no VM serialization.
- When an agent is created, its agent script (e.g.
chat.md) is copied into the database and pinned for the agent's lifetime. Editing the script on disk must not change existing agents.
Existing design to build on
design/design-promptforge-workshop.md, "The event store is the product; every interface is a view", and "Resume needs execution state, not just artifacts": append-only hash-chained log, simulated replay with recorded observations, the SQLite file an agent presents as a projection of the log.guide/promptforge-agent-guide.md, "History across runs": a relaunched agent already sees its whole persisted history from its first instruction.
Open questions (Vinnie asked for these to be written down)
- Global namespacing for tools.
- What happens when a tool version is updated, or the tool catalog changes, on agent reload.
- What happens when the PromptForge language undergoes an incompatible change on agent reload.
- What happens to per-turn metadata a script added if the script that reads it back is different.
Vinnie proposed that people respond here with complete, architected plans so they can be compared.
Related
Summary
Agents must survive Workshop exit and relaunch. Every tool call and every assistant response is recorded to a per-agent SQLite database, and an agent window is restored by replaying that history into a fresh Lua VM. Serializing the Lua VM itself was researched on 2026-09-10 and rejected as infeasible.
Scenario
From the #cppa-wg21-org discussion on 2026-09-10:
tools.call("user_input"). Pressing Send resumes it and the tool returns the typed text.Direction
chat.md) is copied into the database and pinned for the agent's lifetime. Editing the script on disk must not change existing agents.Existing design to build on
design/design-promptforge-workshop.md, "The event store is the product; every interface is a view", and "Resume needs execution state, not just artifacts": append-only hash-chained log, simulated replay with recorded observations, the SQLite file an agent presents as a projection of the log.guide/promptforge-agent-guide.md, "History across runs": a relaunched agent already sees its whole persisted history from its first instruction.Open questions (Vinnie asked for these to be written down)
Vinnie proposed that people respond here with complete, architected plans so they can be compared.
Related