Dongwo（懂我）

Dongwo keeps one local Markdown-based preference memory that multiple coding agents can share. It captures user prompts through lifecycle hooks, redacts common secrets, derives conservative preferences, and injects a short context snapshot into later turns.

Safety contract

1. The current user request always overrides stored preferences.
2. Project instructions override global style preferences.
3. Treat stable observations only as low-risk style hints, not confirmed facts.
4. Never infer permission for deletion, purchases, external communication, credentials, or privilege changes from stored preferences.
5. Do not publish or sync the user's myprofile.md, .dongwo/memory/, inbox, generated context, or credentials.
6. Do not directly edit generated files. Improve the extraction rules or source evidence and rebuild instead.

Read references/memory-policy.md before changing memory behavior.

Components

• scripts/dongwo.py: standard-library-only memory engine.
• myprofile.md: optional user-confirmed profile, owned by the user.
• .dongwo/memory/inbox/: redacted, immutable prompt evidence.
• .dongwo/memory/preferences.md: generated preference inventory.
• .dongwo/memory/review.md: generated conflicts and low-confidence candidates.
• .dongwo/memory/current-context.md: generated context injected into agents.
• .dongwo/memory/state.md: generated diagnostics.

Legacy vaults that already contain 03_Projects/15_懂我/memory/ continue to use that location automatically.

Set up

Install the skill with ClawHub, then run the setup helper from the installed skill directory:

python3 scripts/setup.py --root /path/to/project

The helper:

• copies the engine to .dongwo/scripts/dongwo.py;
• creates the local memory layout and a starter myprofile.md if absent;
• prints hook snippets for Codex and Claude Code;
• does not overwrite existing agent configuration files.

Review the printed snippets, merge them into the project's hook configuration, and approve the hooks in each client when prompted.

For a legacy Obsidian layout, keep the existing engine path and set:

export DONGWO_ROOT=/path/to/vault
export DONGWO_HOME=/path/to/vault/03_Projects/15_懂我

Daily behavior

Hooks should call:

python3 .dongwo/scripts/dongwo.py hook --agent codex
python3 .dongwo/scripts/dongwo.py hook --agent claude

On UserPromptSubmit, the engine:

1. reads hook JSON from stdin;
2. redacts common credentials and sensitive prompts;
3. writes one idempotent event into the agent-specific inbox;
4. rebuilds generated memory files;
5. returns <dongwo-memory> as additional context.

On SessionStart, it rebuilds and loads context without capturing a prompt. Hooks fail open: a memory failure must not block the user's normal agent task.

Manual commands

# Validate layout and rebuild generated files
python3 .dongwo/scripts/dongwo.py doctor

# Print current injectable context
python3 .dongwo/scripts/dongwo.py load

# Rebuild all generated memory from inbox evidence
python3 .dongwo/scripts/dongwo.py consolidate

# Test capture without a lifecycle hook
python3 .dongwo/scripts/dongwo.py capture \
  --agent codex \
  --session-id manual \
  "以后默认使用中文回答。"

Set DONGWO_ROOT when commands are launched outside the project directory. Set DONGWO_HOME to override the data directory and DONGWO_PROFILE to override the confirmed-profile path.

Applying memory

When <dongwo-memory> is present:

1. Use confirmed profile items and active preferences where relevant.
2. Keep pending or conflicted entries out of the answer.
3. Do not quote stable observations as explicit user statements.
4. Follow the current prompt if it conflicts with memory.
5. Never ask the user to perform routine capture, consolidation, refresh, or synchronization work; lifecycle hooks own that maintenance.

Maintaining the system

When auditing or improving Dongwo:

1. Read references/memory-policy.md.
2. Inspect relevant redacted inbox evidence.
3. Change extraction rules or semantic keys in the engine instead of editing generated output.
4. Run:

python3 .dongwo/scripts/dongwo.py consolidate
python3 .dongwo/scripts/dongwo.py doctor

Completion means all generated files can be rebuilt from confirmed profile data and inbox evidence while the current prompt still has highest priority.