caasi-dong3

Delegate tasks to external chat endpoints (OpenAI-compatible and LM Studio native API) as a subagent

Describe multi-step agent workflows using an Arrow-style DSL (>>>, ***, &&&, |||, ?, loop, \, let...in, (), ;) with epistemic operator conventions and structural validation

Socratic dialogue for reflecting on human-AI stewardship

Platform-specific fetch strategies for content that resists simple WebFetch

Assisted review loop — local reviewers answer blind and in parallel, each returns its own convergence verdict (loop-until-dry, facilitator owns K + escalation); roster enrolled per host via /review-loop:init; never merges autonomously; ships an always-on UserPromptSubmit hook that writes nothing until you opt in via /review-loop:init

Git-native preparation & convergence — four routines: init/prepare/converge/prune. prepare is local-first: work stays on local prepare/* by default; remote push is a cross-machine opt-in (push-prepare-branches: yes); prepare freshness-rebases in-progress branches onto default (schema 7; opt-out rebase-prepare-onto-default), and converge offers the refresh. Human takeover recognized by containment; redundant prepare/<slug> surfaced for cleanup, never auto-deleted. converge accept defaults to a handoff-only rename (with a human-marked maintenance exception); prune sweeps unused local + remote branches (human-present, read-only until approved). Committed .tsugu/ is a WIP-knowledge layer (policy.md + context.md + knowledge/) plus a one-line .gitattributes (context.md merge=union); state derived from refs (single-layer, no intake notes); never auto-merges; invokes no user-installed skill by default.