lore-digest

/digest — Knowledge Buffer Flush

When to Trigger

• Stop hook detects buffer has reached flush threshold (5+ entries)
• Post-commit hook after git commit
• User explicitly invokes this skill
• Session end with buffered knowledge

Purpose

Convert raw conversation knowledge (stored in .ai-context/buffer.md) into a structured, versioned increment file in .ai-context/increments/. Each increment includes author, timestamp, confidence, affected files, and evidence for full source tracing.

Process

Step 1: Read Buffer

Read .ai-context/buffer.md. If empty or contains only lock markers, exit silently — no knowledge to flush.

Step 2: Analyze and Structure

Parse the buffer entries and reorganize them into a proper increment format with full metadata.

Buffer format (raw):

[round: 1] [DECISION] 支付超时改为 5s 因为上游 SLA 从 3s 升级
  - Detail: 上游接口 SLA 从 3s 升级为 5s，支付超时需要同步调整...

Transform to increment format:

# YYYY-MM-DD payment-timeout-adjust

## Meta
- author: {{current agent name}}
- timestamp: YYYY-MM-DD HH:MM:SS
- trigger: auto-dialogue
- confidence: high

## Affected Files
- `PaymentService.java` → DECISION (payment timeout threshold)
- `payment-config.yaml` → CONSTRAINT (timeout rule)

## Changes
- `PaymentService.java`: timeout threshold changed from 3000ms to 5000ms

## Decisions
- Payment timeout adjusted to 5s
  - 原因: upstream SLA changed from 3s to 5s, old timeout would trigger false rollback
  - 讨论中提及: rule of thumb — timeout ≤ upstream SLA × 1.2

## Constraints
- Payment timeout must be ≤ upstream SLA × 1.2
  - 来源: user explicitly stated in dialogue

## Evidence
- 来源: 利普, 2026-04-23, development dialogue
- 验证状态: unverified
- 验证人: —

Step 3: Determine Confidence

Assign confidence level based on source:

Source	Confidence
User corrected agent (direct feedback)	high
User explained business context	high
Agent inferred from code analysis	medium
Agent guessed from usage patterns	low