Module Deep Dives
Follow module objects, interfaces, lifecycles and boundaries across six layers. Only module-focused essays belong in L1–L6; cross-module topics and engineering retrospectives remain related reading and do not count as module articles.
Local implementation and offline checks do not establish production readiness, real-model performance or financial research quality. Each article retains its specific limits and outstanding checks. Links marked “Chinese original” lead to articles without a published English edition.
Counts include articles in this category only; related reading is listed separately.
L1 · Task Input and Triggers
Scheduled tasks and other authorized triggers start work; input does not replace research decisions.
L2 · Identity, Sessions and Context
Identity, capability admission, context and durable interactions follow their primary Owners across layers.
- Keep Chatting Is Not Keep Executing: Agent Conversation Admission and Task Continuation
- How Does an Agent Keep Each User’s Research Private?
- Models Understand; Code Authorizes: Run-Scoped Authorization Through Recovery
- Extension Host: From Feature Phone to Smartphone
- Main Context Compaction Without Rewriting Facts
- Cross-Run Compaction: Why an Agent Cannot Simply Summarize History
- Why Agents Need a Workspace: From One-Off Conversations to Long-Term Research
- Memory: From Saving and Recall to Safe Forgetting
- Prompt Injection Is More Than Malicious Text: An Agent Content Trust Boundary
- Why an Agent Cannot Execute a Stop Command Immediately
- After an Agent Asks: Durable Human Response Adoption
- A Custom Expert Should Be More Than a Prompt Field
- Agent Usage: Requests, Tokens and Shadow Credits
- How Is a Skill Different from a Prompt?
- Action Reminders: From Candidate Advice to the Model Request
L3 · Agent Decisions and Execution
Entry, Main, Child, Provider, tool presentation and budgets support bounded execution.
- Entry: Why the Entry Point Became Simpler
- When Should an Agent Stop Researching and Start Answering?
- Provider: Why a Unified API Still Needs Shared Admission
- Why the Next Model Request Still Queues After HTTP 200
- Tool Surface: Why the Model Cannot See an Available Tool
- Multi-Ecosystem Hooks: Claude, Pi and Codex
- Checkpoint System Design: What Does an Agent Save, and What Lets It Continue?
- What a Model Attempt Leaves Behind: Capturing, Retaining, and Reading Agent Research Evidence
L4 · Tools and Domain Capabilities
Controlled reading, isolated computation, MCP, market time and investment facts; receipts are not conclusions.
- Web Research: From Search to Deep Reading
- PDF: From Text to Evidence
- When an Agent Says Today: User Time and Market Sessions
- Python: From Tool to Computational Workbench
- Exact Numbers Are Not Correct Investments: A Numeric Kernel for Units, Rounding, and Residuals
- Deterministic Fund Allocation: From Holdings to Reproducible Amount Advice
- Why Can’t a Portfolio Report Be Written While It Is Still Being Calculated?
- Connecting Portfolio Data Through a Read-Only MCP
- The RAG Lifecycle: From Document Ingestion to Cited Answers
- MCP Host: From Discovery to Usable Tools
- Why a Research Report Must Become an Asset Before an Agent Can Reuse It
- Externalized Results and Input Completeness
L5 · Result Publication and Delivery
Separate accepted content, committed revisions, conversation delivery and feedback facts.
L6 · User Interface
Visible answers, progress and correct Run attribution; reading placement does not imply exclusive frontend ownership.