Skip to content

Session card: what a session did while you were away, from facts flue already has (no model) #134

Description

@karngyan

Status

No model, no AI features, for now. Decided 2026-08-21. Everything below is deterministic: facts flue already has from the transcript and the session state. Model-backed summaries are parked; see the last section for why and what might reopen them.

Why

The phone moment is "what happened while I was away". A transcript of three hundred messages is not an answer. Most of the answer is already sitting in the transcript and the session state, and nothing renders it as one card today.

The card

One card per session, on the session list hover, in the session switcher, and at the top of the chat view (#133):

  • what it was asked: the first user prompt, and the last
  • what it did: files edited, commands run, tool call counts, from the transcript
  • where it is: running, waiting on a permission or a question, finished, exited, from the agent hooks issue (Agent hooks: let Claude Code, Codex and Pi tell flue when a session needs you, finished, or wants a permission #132)
  • the agent's own summary when it has one. Claude Code writes summary lines and an ai-title into its transcript for its resume picker. internal/agentstore/claude.go parses ai-title and custom-title today and skips summary; rendering it costs nothing.

All of it comes from internal/agentstore and the hooks state. No network call, no token, works on every machine whether or not any agent CLI is installed there.

Later, maybe

  • A fleet digest: one line per active session, same facts.
  • "Where did I fix the nginx config": agentstore search already answers this with hits; a card per hit is the deterministic version.

Parked: model-backed summaries

Not now. Not everyone has Claude, Codex or Pi installed, and the routes that looked free are not:

  • claude -p and the Agent SDK on subscription auth are metered separately from interactive use at API rates (Anthropic, June 2026; rollout reportedly paused, line drawn).
  • Reading a CLI's credential store, proxying its HTTP, driving a hidden TUI, or exposing a completion endpoint are all programmatic use in a costume. They put users' accounts at risk, are easy to fingerprint on a public project, and turn the vendor hostile to the rest of flue.

If this reopens, the honest options are: bring your own key, a locally hosted model, or a flue subscription that pays for the model. Any of those is opt-in, on demand, and never on a timer.

Where to look

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requestepicLarge feature. Decide and break down before building.

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions