Summary
Changes since v0.12.0 (14 commits / PRs #529–#533).
The terminal UI (TUI) is removed: seiso tui, its install path, harnesses, and tests are gone — the Tauri desktop app (start) and web UI (seiso forge) are the only interfaces. The agentic chain-of-thought path was reworked so thinking stays off the TTFT (time-to-first-token) critical path, and the swarm now escalates sensibly (verify, add subagents, prompt continuation). A chat hang and slow token decode were fixed, and tool/subagent defaults were corrected.
What's new
TUI removed (#530)
- Deleted the entire terminal UI:
seiso/tui/package,seiso tuiCLI command,install_tui.py, TUI harnesses, and all TUI tests - Install script and docs updated to drop the TUI;
startremains Tauri-desktop-only
Agentic chain-of-thought off the TTFT path (#533)
- Thinking no longer runs on the TTFT path: default thinking cap 128 → 64 tokens (thinking decodes before the first visible token, so this halves the worst-case TTFT tax); reasoning models on general chat get a brief-thinking fraction of the content budget
- Swarm escalates sensibly: verify, add subagents, prompt continuation
- Kept multi-subagent concurrency (reverted a one-subagent experiment) while retaining the thinking TTFT fixes; CI lint format + mypy narrowing in
_build_feedback
Chat hang, slow decode, tools/subagents defaults (#532)
- API
request()and SSE streams now have timeouts (60s request / 30s connect) so a wedged backend can't leave chat stuck in "loading" forever streamDisplayreveals text ~2.4x faster; streaming caret classes cleared on stream endallow_toolsdefaults on (web_search + write_artifact, safe on localhost); ChatPage Tools toggle defaults on when available- Swarm
register_subagentenforcescan_spawn(agent cap + available RAM ≥1GiB + CPU <90%); cap configurable viaSEISO_MAX_CONCURRENT_AGENTS;get_resource_inforeports live available/total memory
CI & docs
Notes
- TUI users should use the Tauri desktop app (
start) or the web UI (seiso forge)