ax

AX — Project Knowledge Sedimentation

Extract valuable knowledge from AI coding sessions and save it in repo-native paths.

Usage

/ax:ax
/ax:ax <prompt>
/ax:ax architecture
/ax:ax skill <name>

Workflow

Follow these steps in order.

Step 0: Pick Exactly One History Adapter

Use the adapter that matches the current agent. Do not read another agent's local history in the same run.

Agent	Adapter
Claude Code	Current conversation + recent git changes + local Claude Code session history under ~/.claude/projects
Codex	Current conversation + recent git changes + files read or written in this session

For Codex: do not guess private transcript paths under ~/.codex, ~/.config, or similar locations. If the current conversation plus git evidence is insufficient, ask the user for direction instead of scraping local history.

Step 1: Collect Data

1a. Current session context

Gather from the current conversation:

• files read or written
• bugs found
• root causes identified
• solutions applied
• key design decisions

1b. Recent git changes

git log --oneline -10
git diff --stat

1c. Claude Code adapter only: local session history

If the current agent is Claude Code, do a two-level scan of recent local session history.

Level 1: quick scan

PROJECT_DIR=$(pwd)
SESSION_BASE="${HOME}/.claude/projects"
PROJECT_HASH=$(echo -n "${PROJECT_DIR}" | sed 's|/|-|g' | sed 's|^-||')
SESSION_DIR="${SESSION_BASE}/${PROJECT_HASH}"

if [ ! -d "$SESSION_DIR" ]; then
  SESSION_DIR=$(ls -d ${SESSION_BASE}/*${PROJECT_DIR##*/}* 2>/dev/null | head -1)
fi

if [ -d "$SESSION_DIR" ]; then
  find "$SESSION_DIR" -name "*.jsonl" -mtime -7 -type f | xargs ls -t 2>/dev/null | head -10
fi

For each candidate session file, extract only the first 10 user messages:

python3 -c "
import json, sys
count = 0
with open(sys.argv[1]) as f:
    for line in f:
        line = line.strip()
        if not line:
            continue
        d = json.loads(line)
        if d.get('type') != 'user':
            continue
        msg = d.get('message', {})
        content = msg.get('content', [])
        if isinstance(content, str):
            texts = [content]
        elif isinstance(content, list):
            texts = [c.get('text', '') for c in content if isinstance(c, dict) and c.get('type') == 'text']
        else:
            continue
        for text in texts:
            text = text.strip()
            if not text:
                continue
            print(text[:200])
            count += 1
            if count >= 10:
                sys.exit()
" <session_file>

Write a one-line summary for each session and decide which 1-3 sessions are worth deeper extraction.

Level 2: deep extract

For each selected session, extract only user and assistant text:

python3 -c "
import json, sys
with open(sys.argv[1]) as f:
    for line in f:
        line = line.strip()
        if not line:
            continue
        d = json.loads(line)
        msg_type = d.get('type')
        if msg_type not in ('user', 'assistant'):
            continue
        msg = d.get('message', {})
        content = msg.get('content', [])
        if isinstance(content, str):
            texts = [content]
        elif isinstance(content, list):
            texts = [c.get('text', '') for c in content if isinstance(c, dict) and c.get('type') == 'text']
        else:
            continue
        for text in texts:
            text = text.strip()
            if not text:
                continue
            if len(text) > 500:
                text = text[:500] + '...[truncated]'
            print(f'[{msg_type}] {text}')
" <session_file>

Ignore tool calls, tool results, file snapshots, and other noise.

Step 2: Analyze and Extract

Identify knowledge worth saving.

Type	Target Path	When
Architecture knowledge	docs/ai-context/{topic}.md	Core design, data flow, system principles
Project skill	.agents/skills/{name}/SKILL.md	Repeatable workflows such as debugging, deployment, review
Module documentation	{module}/AGENTS.md	Module-specific context, APIs, gotchas

Rules:

• If the user provided a prompt, focus extraction around that topic.
• Reuse or update existing files when the content overlaps.
• Keep each written file under 200 lines.

Step 3: Preview and Confirm

For each file you want to write, show the user:

1. target path
2. content preview
3. why this knowledge is worth saving

Then ask for explicit confirmation in the conversation. Do not rely on tool-specific UI helpers. Do not write any file without explicit user confirmation.

Step 4: Validate

Run these checks before writing.

4a. Line count gate

echo "<draft content>" | wc -l

If the draft exceeds 200 lines, split it and use @ references.

4b. Duplication scan

grep -rl "<2-3 core keywords from draft>" docs/ai-context/ .agents/skills/ 2>/dev/null || echo "no duplicates"

If an overlapping file exists, update it instead of creating a duplicate.

4c. @ reference validity

for ref in $(python3 -c "
import re, sys
draft = sys.stdin.read()
for ref in re.findall(r'@([\w/.,-]+\.\w+)', draft):
    print(ref)
" <<'EOF'
<draft content>
EOF
); do
  test -f "$ref" || echo "BROKEN: $ref"
done

Fix any broken reference before writing.

4d. Actionable content

Reject vague advice. Saved knowledge must include concrete commands, paths, patterns, or steps.

Step 5: Write Files

For confirmed files:

1. Write the file content.
2. Update the nearest parent AGENTS.md with an @ reference when needed.
3. If AGENTS.md does not exist at the project root, create it with sedimentation rules.
4. If CLAUDE.md does not exist, create it as a symlink or pointer to AGENTS.md.

Step 6: Summary

Report:

• files written or updated
• files skipped
• any follow-up review the user should do before committing

Never auto-commit.

Output Conventions

File Constraints

• Each markdown file must stay under 200 lines.
• Use @ references to link related docs.
• Follow the existing AGENTS hierarchy when it exists.

Cross-Agent Compatibility

沉淀产物使用通用格式，确保所有 coding agent 都能消费：

产物	路径	消费方式
项目入口	AGENTS.md	Codex 原生读取；Claude Code 通过 CLAUDE.md 软链接读取
架构知识	docs/ai-context/*.md	所有 agent 通过 @ 引用读取
项目技能	.agents/skills/*/SKILL.md	通用 skill 格式，各 agent 按自身机制发现
模块上下文	{module}/AGENTS.md	所有 agent 通过目录遍历发现

不要把知识写入 agent 专属路径（如 .claude/）。所有沉淀只写入上述通用路径。

AGENTS.md Template

# {Project Name}

## Instructions

### AGENTS.md 读取规则
修改或设计涉及某目录的代码前，必须先读该目录及父目录的 AGENTS.md。

### 知识沉淀
项目知识只写入 `.agents/skills/`、`docs/ai-context/` 和模块 `AGENTS.md`。
所有沉淀与知识修订都必须通过 `/ax:ax` 流程完成。

## Quick Start
Key commands and setup.

## Architecture
High-level architecture summary.

## Module Index
- @{module}/AGENTS.md — Module description

## Deep Dive Docs
- @docs/ai-context/{topic}.md — Topic description

Key Principles

• Human in the loop: every write requires explicit confirmation.
• Repo-native knowledge: project docs and project skills live outside any agent-specific directory.
• Cross-agent compatible: all output uses AGENTS.md + .agents/skills/ + docs/ai-context/ — formats every coding agent can read.
• One adapter per run: only read the history source that matches the current agent.