9.8 KiB
9.8 KiB
| artifact_type | name | author | created | status | source_artifacts | next_action | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| handoff | viewpoint-runtime-agent-refactor-handoff-2026-06-07 | CCPE System | 2026-06-07 | active-handoff |
|
refactor viewpoint runtime from chunk-first discovery to gestalt-first topic graph workflow |
Handoff: Viewpoint Runtime Agent Refactor
1. Current Situation
The user is testing Viewpoint Discussion Distillation Runtime from knowledge-vault against this source:
C:\Users\wangq\Documents\Codex\knowledge-vault\diary\2026\2026-06\2026-06-05.md
The Knowledge run created a local capability layer:
C:\Users\wangq\Documents\Codex\knowledge-vault\skills
It then ran real topic-discovery worker participants and produced a second 01-topic-map.md draft under:
C:\Users\wangq\Documents\Codex\knowledge-vault\discussions\viewpoint-development\2026-06-05-失望性情感隔离\
The second draft improved over the first flat list. It now proposes:
ROOT:
从失望性情感隔离到认知生成与关系治理
A 入口与边界层
B 机制解释层
C 认知主权层
D 架构升级层
E 局部深化层：青少年与家庭系统
But the user identified that the structure is still not stable enough for formal material routing.
2. User's Current Judgment
The user accepts:
- ROOT topic.
- A/B/C/D/E as the broad frame.
- A2 should be a subsection of A1, not an independent document.
- E2 家庭三角认知场 should remain as a strong local deepening model candidate.
The user does not accept the current C/D layer logic.
Key correction:
巨人方舟 is the new name for the upgraded multi-layer architecture.
QPI is not a parallel first-level topic. It is a tool inside the metacognitive layer of 巨人方舟.
方法论调度与信任分级 should be understood inside the governance / coordination mechanism of 巨人方舟, not as a flat peer topic.
Reality Lab / 行动实验层 is a key new layer that turns 巨人认知 from an explanatory model into 巨人方舟: explanation + intervention.
家庭三角认知场 is a local deepening model extracted through the youth / family-system case.
The user also specified the upgrade content for 巨人认知 2.0:
1. 强化 GL0：加入身体调节、情绪耐受、神经系统状态。
2. 增加 GL-R：关系/环境层，防止社会系统废热被内化为个人问题。
3. 增加行动实验层：认知生成必须靠现实反馈。
4. 限制 GL4 独裁：元认知是协调者，不是情绪和身体的统治者。
After these changes, the model should be treated as:
巨人认知 2.0 -> 巨人方舟
解释模型 -> 解释 + 干预模型
3. Do Not Proceed Yet
Do not enter:
02-material-routing-log.md
topics/*.md generation
coverage audit
downstream-task-routing
Reason:
The topic graph has major parent-child and layer-placement uncertainty.
Routing source blocks now would place evidence into the wrong hierarchy and make topic docs harder to repair.
The next Knowledge step should be:
Hierarchy Repair Pass
not formal material routing.
4. Major Runtime Design Defect Found
The current Runtime is too chunk-first / worker-first.
Current flow:
source registration
segmentation
source excerpt packs
worker topic discovery
topic synthesis
human confirmation
material routing
topic docs
audit
downstream routing
Problem:
For coherent long discussions under about 200k Chinese characters, worker-first discovery can become blind-men-touching-elephant.
Each worker sees only part of the elephant. The synthesis then tries to reconstruct the whole from lossy local fragments.
This can recover the broad frame but often misplaces tools, submodels, action layers, examples, and model-upgrade relationships as flat peer topics.
The user proposed a better pattern:
For initial source drafts under about 200k characters and high topic coherence:
first pass the whole source to understand the elephant,
then segment and route with that global picture in hand.
This should be accepted as a Runtime design improvement.
5. Required Runtime Refactor
Upgrade the Runtime from 8 steps to 9 steps by adding a new Stage 0:
0. Whole-Source Gestalt Pass
Recommended new flow:
0. Whole-Source Gestalt Pass
A high-context real participant reads the full source when feasible and outputs global topic picture, core structure, model evolution line, key tensions, and likely hierarchy.
1. Source Registration
Register source path, length, boundary rules, and no-modification policy.
2. Conversation Segmentation
Segment into stable source blocks.
3. Source Excerpt Context Packs
Build excerpt packs, but now under the global picture.
4. Worker Topic Extraction
Workers do not discover from zero. They validate, enrich, challenge, and localize the gestalt topic graph.
5. Topic Graph Synthesis
Produce hierarchy, parent-child relations, cross-links, merge/split candidates, model/tool/action/example distinctions.
6. Human Structure Confirmation
User confirms hierarchy before formal routing.
7. Material Routing
Route source blocks only after hierarchy is stable.
8. Topic Docs + Coverage Audit
Generate topic docs, then audit coverage and distortion.
9. Downstream Task Routing
Route to writing, model update, scale extraction, engineering/system work, todos, continued research, or archive-only.
6. Required Agent / Skill Changes
6.1 New Agent Or Skill Needed
Create one of the following:
Option A:
agents/agent-specs/whole-source-gestalt-reader.agent.md
Option B:
skills/workflow/whole-source-gestalt-pass.skill.md
Recommended:
Create an Agent first.
Reason:
This is not only mechanical summarization. It requires judging global structure, model evolution, tool/layer/case distinctions, and source coherence. It is an Agent judgment role, not merely a formatting Skill.
6.2 Topic Discovery Router Must Change
topic-discovery-router.agent.md should be revised so it does not only discover topics from source blocks.
It should accept:
- whole-source gestalt report
- source-map
- source excerpt context packs
- worker returned outputs
- user corrections
It should output:
- topic graph, not only topic list
- parent-child candidates
- layer membership
- model/tool/action/example distinctions
- merge/split candidates
- cross-topic links
- downstream task candidates
- human confirmation questions
6.3 Runtime Must Add Hierarchy Repair Pass
Add an explicit stage before material routing:
Hierarchy Repair Pass
Trigger:
Human rejects or amends parent-child relations, model/tool/layer distinctions, or global topic graph.
Output:
01-topic-map.md updated as topic graph draft
handoff or run-notes entry recording accepted / rejected / repaired relationships
Hard rule:
Do not generate 02-material-routing-log.md until Hierarchy Repair Pass is accepted.
7. Invocation Boundary Still Applies
Do not weaken the no-simulation rules.
Relevant protocol:
runtimes/hybrid/agent-invocation-packet.protocol.md
Current rule:
No formal participant output is accepted without:
- agent-invocation-packet.md plus returned real output
- prompt-to-send.md plus returned external output
- skill-execution-record.md for local Skill execution
If a Whole-Source Gestalt Reader is added, it must also use the same invocation protocol.
If the whole source is too long for one participant, use a quasi-global strategy:
- split into 2-3 large source packs,
- run gestalt workers,
- synthesize a global gestalt report,
- then run ordinary worker extraction under that global picture.
8. Files To Modify Next
Primary files:
runtimes/hybrid/viewpoint-discussion-distillation.runtime.md
agents/agent-specs/topic-discovery-router.agent.md
workbench/analysis/knowledge-vault-viewpoint-runtime-handoff.md
workbench/analysis/viewpoint-discussion-distillation-creation-brief.md
Likely new file:
agents/agent-specs/whole-source-gestalt-reader.agent.md
Optional later file:
skills/workflow/hierarchy-repair-pass.skill.md
Do not start by editing Knowledge Vault. Update CCPE definitions first, then send a concise update back to Knowledge.
9. Suggested Start Prompt For The Next CCPE Session
请继续改造 Viewpoint Discussion Distillation Runtime。
先阅读：
C:\Users\wangq\Documents\Codex\ccpe-system\workbench\analysis\viewpoint-runtime-agent-refactor-handoff-2026-06-07.md
目标：
把当前 chunk-first / worker-first 的流程，升级为 gestalt-first topic graph workflow。
重点：
1. 新增 Whole-Source Gestalt Pass。
2. 增加 Whole-Source Gestalt Reader Agent。
3. 修订 Topic Discovery Router，让它输出 topic graph / hierarchy，而不是 flat topics。
4. 增加 Hierarchy Repair Pass，且在它通过前禁止 material routing。
5. 保留 Agent Invocation Packet / no-simulation 硬规则。
不要处理 knowledge-vault 的源文件；先改 ccpe-system 中的 Runtime / Agent / handoff 规格。
10. Status
Current status:
handoff prepared for next CCPE refactor session.
Do next:
read this file, then refactor CCPE specs.
Do not do next:
do not route material, do not generate topic docs, do not claim Knowledge processing is complete.