core — shared agent infrastructure

Language: 中文 | English · Back to module index

Agent-loop building blocks shared by the CLI and GUI: prompts, persona assembly, capability tiers, runtime state.

File	Description
`agent_tiers.py`	`TierProfile` and `TIER_ORDER = (low, medium, high, xhigh, max, ultra)`; `get_tier_profile()` gives each tier's max tool turns, subagents, reflection, verification, and compaction threshold.
`agent_runtime.py`	Ephemeral per-run state: `AgentRecord`, `AgentRunState`, `AgentLoopResult`.
`prompts.py`	Shared system prompts (single source for CLI/GUI), including scripted MCP/skill install workflows.
`persona.py`	Assembles the `[CHARACTER CARD] / [EXAMPLE DIALOGUES] / [USER CARD] / [EMOTION STATE]` prefix block.
`strategy_injector.py`	Synthesises the `## Learned Strategies` block.
`experience_injector.py`	Builds the `## Experiences` index block.
`context_compaction.py`	Structure-aware context compaction for long tasks.
`event_bus.py`	In-process event bus (`EventType` enum); the scheduler emits through it.
`formula_engine.py`	Safe arithmetic evaluator for character-card emotion formulas (only `+ - * /`, numbers, variables, parens; clamped).
`time_context.py`	Renders real-world time as prompt context: the date, how long the user was away, and the age of recalled memories. Pure functions — `now` is always passed in, never read from the clock.
`birthday.py`	Birthday trigger decisions and the letter the app writes in its own voice. Pure functions and static text; `server/birthday_runner.py` performs the effects.

Source docs-public/modules/core.mdView on GitHub ↗