context-retrospective（背景回顾）

Context Network Retrospective

Purpose

Analyze agent-user interaction transcripts to identify context network maintenance needs and guidance improvements. Extract actionable insights for enhancing both network structure and agent instructions.

Core Principle

Learn from every interaction. Each transcript reveals gaps in context, navigation issues, and guidance problems that can be systematically fixed.

Analysis Dimensions

1. Knowledge Gap Identification

Look For:

• Questions agent couldn't answer from existing context
• Information discovered during task that should be pre-documented
• Repeated lookups of same information
• Agent confusion about structure, relationships, dependencies

Questions:

• What foundational knowledge was missing?
• Which relationships weren't documented?
• What context about history, decisions, or constraints was absent?
• Which domain boundaries were unclear?

Output: Missing information nodes and relationship gaps

2. Context Boundary Violations

Look For:

• Planning documents created outside context network
• Implementation files placed in context areas
• Agent uncertainty about where to place information
• Mixed concerns within single documents

Questions:

• Did agent distinguish context from project artifacts?
• Were "planning stays in context network" rules followed?
• What guidance would prevent future confusion?

Output: Boundary violations and guidance improvements needed

3. Navigation and Discovery Patterns

Look For:

• How agent found (or failed to find) information
• Sequences of information access
• Dead ends or inefficient paths
• Information that should have been connected

Questions:

• What navigation paths did agent follow?
• Which information should be more discoverable?
• What logical connections were missing?
• What hub documents would improve efficiency?

Output: Navigation improvements and missing connections

4. Task-Context Alignment

Look For:

• Mismatches between task needs and available context
• Information at wrong abstraction levels
• Context too detailed or too high-level
• Task patterns revealing organizational weaknesses

Questions:

• Was information at appropriate abstraction for the task?
• Did context support decision-making needs?
• Were there cognitive load issues from organization?
• What restructuring would support this task type?

Output: Abstraction adjustments and reorganization needs

5. Relationship Mapping Deficiencies

Look For:

• Agent difficulty understanding dependencies
• Missing context about how changes affect other areas
• Lack of clear interface definitions
• Implicit relationships that should be explicit

Questions:

• What relationships were implied but not documented?
• Which dependencies were discovered during task?
• What impact relationships were unclear?
• Where would explicit documentation have helped?

Output: Missing relationships and documentation needs

6. Guidance Effectiveness

Look For:

• Agent behavior suggesting unclear guidance
• Task approaches deviating from optimal patterns
• Mode switching decisions and appropriateness
• Tool usage relative to restrictions

Questions:

• Did agent follow mode-appropriate patterns?
• Were mode transitions handled effectively?
• What guidance was missing or unclear?
• Did restrictions support the purpose?

Output: Guidance refinements and rule clarifications

Retrospective Process

Phase 1: Preparation

1. Context Gathering

- Load current context network state - Identify agent mode(s) used - Note task type and complexity - Review applicable rules

1. Baseline

- Map context available at task start - Identify active guidance - Note recent network changes - Document expected vs. actual behavior

Phase 2: Transcript Review

1. Chronological Analysis

- Track information seeking patterns - Note decision points where context influenced choices - Identify struggle points - Map actual navigation paths

1. Critical Incidents

- Flag confusion or inefficiency - Identify boundary violations - Note "rediscovery" of information - Mark where better context would have helped

1. Pattern Recognition

- Recurring information needs - Systematic gaps in knowledge areas - Consistent navigation difficulties - Successful context utilization

Phase 3: Gap Analysis

1. Information Architecture

- Map knowledge coverage gaps - Evaluate abstraction appropriateness - Assess relationship completeness - Review navigation effectiveness

1. Guidance System

- Analyze mode-specific guidance - Review boundary rule clarity - Evaluate instruction completeness - Assess prompt override needs

1. Prioritization

- Critical: Caused task failure or significant inefficiency - High: Required real-time discovery - Medium: Would enhance efficiency - Low: Nice-to-have improvements

Analysis Templates

Knowledge Gap

## Gap: [Name]

**Discovery Context:** [When/how revealed]
**Task Impact:** [How it affected completion]
**Information Type:** [Domain/Process/Relationship/Decision criteria]
**Recommended Action:** [Specific node or relationship to add]
**Priority:** [Critical/High/Medium/Low]
**Related Gaps:** [Connected gaps]

Navigation Issue

## Issue: [Name]

**Problem Pattern:** [What difficulty occurred]
**Information Sought:** [What agent wanted]
**Current Path:** [How agent actually found it]
**Optimal Path:** [How it should be discoverable]
**Recommended Improvement:** [Specific changes]
**Affected Tasks:** [What else would benefit]

Guidance Assessment

## Guidance: [Mode/Rule Area]

**Expected Behavior:** [What guidance should produce]
**Actual Behavior:** [What agent did]
**Deviation Analysis:** [Why different]
**Guidance Clarity:** [Current clarity level]
**Recommended Changes:** [Specific modifications]
**Test Scenarios:** [How to validate]

Quality Metrics

Completeness

• Information Coverage: % of questions answerable from context
• Relationship Completeness: Documented vs. discovered relationships
• Navigation Efficiency: Steps vs. optimal paths
• Boundary Compliance: % correct domain placements

Effectiveness

• Task Completion Quality: Success rate with available context
• Agent Confidence: Frequency of uncertainty expressions
• Context Utilization: % of relevant context actually used
• Discovery vs. Lookup: New discoveries vs. existing use

Evolution

• Context Network Growth: New nodes/relationships rate
• Guidance Refinement: Rule update frequency
• Pattern Recognition: Recurring improvement themes
• System Maturity: Decreasing structural changes

Common Patterns & Solutions

Implementation Priority

Phase 1: Critical Infrastructure

• Fix boundary violations
• Add missing foundational knowledge
• Repair broken relationships

Phase 2: Navigation Enhancement

• Improve discoverability
• Create hub documents
• Strengthen cross-domain connections

Phase 3: Guidance Refinement

• Update mode-specific instructions
• Clarify ambiguous rules
• Enhance prompts for common tasks

Phase 4: Optimization

• Fine-tune abstraction levels
• Optimize for discovered workflows
• Enhance metadata systems

Anti-Patterns

1. The Blame Game

Pattern: Attributing interaction failures to agent capability rather than context gaps. "The agent should have known..." Why it fails: Agents operate from context. If context is incomplete, even capable agents fail. Blaming agents prevents systemic improvement. Fix: Assume context gaps first. Ask "what information would have prevented this?" before "why didn't the agent figure it out?"

2. The Completeness Illusion

Pattern: Believing context networks can capture everything. Adding more and more information hoping to prevent all failures. Why it fails: Context networks grow without bound. Navigation becomes impossible. Signal-to-noise ratio degrades. Maintenance becomes unsustainable. Fix: Focus on high-impact gaps. Prioritize what actually caused failures. Remove outdated information as aggressively as you add new.

3. The Surface Fix

Pattern: Fixing the specific issue without identifying the pattern. Adding a fact that was missing without asking why it was missing. Why it fails: Treats symptoms, not causes. The same class of gap will appear elsewhere. Whack-a-mole maintenance. Fix: Classify gaps by type. If the gap is "missing relationship documentation," the fix is improving relationship capture, not adding one relationship.

4. The Retrospective-Only

Pattern: Running retrospectives but never implementing changes. Analysis paralysis or action avoidance. Why it fails: Insight without action produces no improvement. Accumulating analysis without implementation wastes the analysis effort. Fix: Every retrospective must produce at least one actionable change. Schedule implementation before finishing retrospective.

5. The Guidance Overdose

Pattern: Adding more rules and restrictions after every failure. Context networks become constraint lists. Why it fails: Excessive guidance produces paralysis. Agents become afraid to act. Guidance conflicts emerge. Nobody reads the rules. Fix: Before adding guidance, consider removing it. Simplify before complexifying. Test if clearer boundaries achieve more than more rules.

Integration Points

Inbound:

• After any significant agent interaction
• After task failures or inefficiencies
• During context network maintenance

Outbound:

• To context network updates
• To guidance/instruction improvements

Complementary:

• Context Networks framework
• Agent mode configurations