Merge pull request #1215 from yinwm/refactor/agent

docs: add agent refactor working notes
2026-06-12 18:08:54 +00:00 · 2026-03-08 00:25:47 +08:00
parent 826f92cf53 726a87b70f
commit 4768edc67b
1 changed files with 145 additions and 0 deletions
@@ -0,0 +1,145 @@
+# Agent Refactor
+
+## What this directory is for
+
+This directory is the working area for the current Agent refactor.
+
+The purpose of this refactor is simple:
+
+the project needs a smaller, clearer, and more stable Agent model before more Agent-related behavior is added.
+
+The codebase already contains meaningful Agent behavior. What it still lacks is a sufficiently explicit and stable semantic boundary around that behavior.
+
+This refactor exists to fix that first.
+
+---
+
+## Refactor stance
+
+This is a maintenance-led consolidation effort.
+
+It is not a general invitation to expand Agent behavior in parallel.
+
+During this refactor window, Agent-related work should converge on the current refactor track instead of branching into new semantics.
+
+That means:
+
+- concept clarification before feature expansion
+- boundary tightening before abstraction growth
+- semantic consolidation before new behavior
+
+---
+
+## Core rule: minimum concepts only
+
+This refactor follows one hard rule:
+
+**do not introduce a new concept unless it is strictly necessary**
+
+More explicitly:
+
+- if an existing concept can be clarified, reuse it
+- if an existing boundary can be made explicit, do that first
+- if a behavior can be expressed without a new abstraction, do not add one
+- "future flexibility" is not enough justification on its own
+
+The goal of this refactor is not to grow the model.
+
+The goal is to reduce ambiguity.
+
+---
+
+## What is being clarified
+
+This refactor is currently concerned with the following questions:
+
+1. what an `Agent` is
+2. what an `AgentLoop` is
+3. what the lifecycle of `AgentLoop` is
+4. what the event surface around `AgentLoop` is
+5. how persona / identity is assembled
+6. how capabilities are represented
+7. how context boundaries and compression work
+8. how subagent coordination works
+
+These are the current working boundaries.
+
+If they need to be adjusted, they should be adjusted explicitly rather than drift implicitly in code.
+
+---
+
+## Status of this directory
+
+The documents here are working materials.
+
+They are not final or immutable.
+
+If current notes are incomplete, incorrectly split, or too broad, they should be revised. This directory should evolve with the refactor rather than pretending the first draft is complete.
+
+---
+
+## Suggested document split
+
+This directory may eventually contain notes such as:
+
+- `agent-overview.md`
+  - what an Agent is
+- `agent-loop.md`
+  - AgentLoop contract, lifecycle, event surface
+- `persona.md`
+  - persona and identity assembly
+- `capability.md`
+  - tools / skills / MCP capability semantics
+- `context.md`
+  - context scope, history, summary, compression
+- `subagent.md`
+  - subagent coordination rules
+
+These files should be added only when they help clarify the current refactor work.
+
+This directory should not turn into a generic architecture dump.
+
+---
+
+## What this directory is not for
+
+This directory is not intended for:
+
+- broad speculative architecture
+- future multi-node protocol design not required by the current refactor
+- parallel feature planning unrelated to Agent consolidation
+- adding new concepts before current ones are made clear
+
+If a topic does not directly help reduce ambiguity in the current Agent model, it probably does not belong here yet.
+
+---
+
+## Relationship to implementation
+
+Implementation changes should not keep redefining Agent semantics implicitly.
+
+If a PR changes or depends on Agent semantics, those semantics should either already exist here or be clarified in a linked issue first.
+
+This directory is here to make implementation narrower and more disciplined.
+
+---
+
+## Relationship to GitHub tracking
+
+The umbrella issue for this refactor should point here.
+
+The issue is the coordination surface.
+
+This directory is the repository-local working surface.
+
+---
+
+## Summary
+
+The main question of this refactor is not:
+
+- what more can Agent do
+
+The main question is:
+
+- what is the smallest stable model that current Agent behavior can be organized around