GUI agent-overview: cheap-model session recaps + NEEDS_INPUT/COMPLETED status chips

[justrach]

From the prime-agent analysis (daemon-session-summarizer.ts). Their agents dashboard runs a cheap model over each session's last ~8 messages (600 chars each, tool calls by name only) on a slow sweep with a settle debounce, emitting a one-line recap plus NEEDS_INPUT|COMPLETED — silently disabled when no auth exists for the cheap model.

graff fit: maps directly onto the agent-overview GUI panel (gui/src/components/agent-overview/). Per the TUI-first rule, define the event in the protocol first: a session_recap event in the --json stream (schema.zig), produced harness-side, consumed by GUI and TUI alike.

Notes: strict input bounds (message count, chars, tool calls by name) keep it cheap; debounce so streaming turns don't thrash it; degrade to a heuristic (last event kind) when no cheap model is available.

[justrach]

Landed on feat/agent-inbox (commits 18e3f31 + sdk regen f5d3ea3):

Protocol — session_recap event in the --json stream, schema_version 0.10 → 0.11: {text, status: needs_input|completed|failed, source: heuristic|model}. SDKs regenerated (ts + remote).

Harness — src/recap.zig (bounded digest: last 8 messages, 600 chars each, tool calls by name only; STATUS:/RECAP: reply contract with a strict parser) + src/mainloop_recap.zig (detached jobs modeled on the AI-title jobs; a per-turn generation counter is the settle debounce, so streaming turns never thrash it). Cheap model = root provider's tier-ladder small/mid rung; silently skipped when the provider has no ladder rung or no key — the heuristic (last-turn text; a closing ? reads as needs_input) is the always-on floor and is emitted before the terminal turn event so stream-until-turn clients always see it.

GUI — stream_turn (and the control-ack wait loop, which can surface a between-turns model recap) parse it into per-conversation state; it rides ConversationViewSnapshot.recap into the agent-overview projection, which prefers live signals (running request, pending followup, error) and otherwise shows the recap's chip + one-liner. Settled-but-not-current orchestrators with a recap now appear in recent activity instead of vanishing.

Verified — tier 1 green (incl. sdk drift), tier 2 green (45 cases; 3 new: recap-precedes-turn, needs-input heuristic, model-silent-without-cheap-rung), 1135 unit tests, 203 bun tests, 31 cargo tests, tsc + eslint clean.

Known limits, by design:

• In --json mode the harness blocks on stdin between turns, so a model recap that completes mid-idle is emitted at the next loop poll (just before the next prompt is read). The GUI keeps it then; the heuristic recap is always immediate. Fixing this properly means an idle-wakeup in the json read path — a Epic: separate the REPL from the engine — typed event contract, sink inversion, TUI as first client #422-slice candidate.
• Error-path turns (interrupt/stall/drop/api) emit no recap; the GUI's own last-error signal still marks those sessions failed.
• The TUI stores the recap (root.session_recap) but has no overview surface to show it — deferred to the Epic: separate the REPL from the engine — typed event contract, sink inversion, TUI as first client #422 TUI-client work.

[justrach]

Shipped in v0.0.244 (18e3f31): session_recap on the event stream and GUI agent-overview chips (needs_input / completed). Remaining TUI overview surface is tracked under #422, not this issue.