Glossary — canonical reyn terms

Authoritative names for reyn concepts. Use these terms verbatim in skill DSL files, phase instructions, and documentation. Translations are for prose only — code, frontmatter keys, and CLI flags stay in English.

Core layers

English	日本語	Definition
Agent	エージェント	A long-lived ChatSession with its own profile, history, memory layer, and inbox. The interpreter of user intent. Persisted under .reyn/agents/<name>/.
Agent Profile	エージェントプロファイル	.reyn/agents/<name>/profile.yaml declaring name, role, created_at, optional allowed_skills.
Agent Registry	エージェントレジストリ	Process-scoped owner of all loaded ChatSession instances. Routes attach/detach and inter-agent messaging.
Skill	スキル	A directory defining a phase graph and final output schema.
Phase	フェーズ	A reusable processing unit declaring only its input and instructions.
OS	OS	The runtime executor; sole owner of control flow.
Workspace	ワークスペース	The shared store for files and artifacts.
Artifact	アーティファクト	Structured data passed between phases.
Event	イベント	A recorded state change.

Multi-agent

English	日本語	Definition
Topology	トポロジー	A declared communication structure (network / team / pipeline) listing members and edge rules. Persisted at .reyn/topologies/<name>.yaml.
_default topology	デフォルトトポロジー	Auto-managed network containing every agent that does NOT belong to any user-declared topology. In-memory, recomputed on demand.
Chain	チェイン	One logical request path from a top-level user submission, possibly spanning multiple agents and hops. Identified by chain_id.
chain_id	チェイン ID	uuid4 hex minted by submit_user_text; propagated through every inbox payload, history meta, and event in the same chain.
Pending Chain	ペンディングチェイン	State held in a delegating agent while it waits for delegate responses (deferred reply). Cleared when waiting_on becomes empty.
allowed_skills	許可スキル一覧	Optional list[str] \| None in profile.yaml. None = unrestricted, [] = router-only, [a, b] = allowlist. stdlib router/compactor are not subject. (FP-0011 removed skill_narrator; the router LLM narrates inline.)
Hop Depth	ホップ深度	Number of agent-to-agent forwards from the original user request. Bounded by safety.loop.max_agent_hops.

Execution

English	日本語	Definition
Context Frame	コンテキストフレーム	The read-only payload built per phase visit and given to the LLM.
Control IR	Control IR	List of side-effect ops the LLM may emit (file, ask_user, run_skill, lint, shell).
Preprocessor	プリプロセッサ	Deterministic pre-LLM steps a phase may declare (run_skill, iterate, validate, python).
Decision	決定	OS-level value: continue / finish / abort. Never skill-specific.
Transition	遷移	A move from one phase to another, validated against the skill graph.
Final Output	最終出力	The artifact produced when a skill finishes; validated against the skill's final_output declaration.
Visit Count	訪問回数	Number of times a single phase has been entered in the current run.

DSL files

File	Purpose
skill.md	Skill declaration: entry, graph, final_output, permissions.
phases/<name>.md	Phase declaration: input, optional preprocessor, instructions in body.
artifacts/<name>.yaml	Artifact schema (JSON Schema fragment).
eval.md	Optional eval spec: cases + per-phase quality criteria.
<name>.py	Optional Python module for python preprocessor steps.

Permission verbs

Verb	Meaning
allow	Always permit, no prompt.
ask	Prompt the user the first time; persist the choice.
deny	Reject without prompt.

Modes

Mode	Where	Meaning
safe	Python preprocessor	AST-validated, sandboxed builtins, allowed-modules-only imports. No extra flag needed.
unsafe	Python preprocessor	Free Python; requires --allow-unsafe-python at the CLI plus a permissions.python entry with mode: unsafe in skill.md. Stdlib skills in mode: unsafe are auto-trusted — the flag is still required but no approval prompt fires (the skill is vendored with reyn, not user-supplied).

DO NOT confuse

• continue (decision) vs "next phase = X" (transition) — continue is OS-level; the actual phase name is in next_phase.
• finish (decision) vs end (graph token) — end appears in graph: adjacency lists to mark terminal transitions; finish is the LLM's decision value when terminating.
• Skill vs "skill" — capitalized when referring to the architectural concept; lowercase in CLI commands and file names.