| name | post-dev |
| description | Post-development cleanup: update AGENT.md and knowledge files to reflect code changes. MUST activate before git commit (or when user asks to commit/push). Also activate after any code modification that adds/removes files, changes architecture, or modifies core behavior. |
Knowledge Management
Maintain a living knowledge base so future sessions (with zero memory) can work effectively.
Iron Rules
- Every file referenced in AGENT.md MUST exist on disk. Before adding a reference, create the file. Before removing a file, remove its reference. Broken references are worse than no references.
- Knowledge files are the primary deliverable, not AGENT.md. AGENT.md is just an index. When you learn something non-obvious, write it into the appropriate knowledge file. Only update AGENT.md's index entry if the file list changed.
- Read before write. Before updating any knowledge file, read it first. Before creating AGENT.md references, verify the target file exists.
- Do NOT copy the structure from this skill into AGENT.md. Every project is different. Observe the actual project structure and document what exists, not what a template says should exist.
Two-Layer Architecture
AGENT.md (auto-injected into prompt — the single source of truth)
→ project summary, build commands, architecture overview
→ GOTCHAS: critical pitfalls written directly here (no separate file)
→ Knowledge Files index: tells you WHERE to look for details
→ should make you want to Read specific files, not answer questions directly
Knowledge files (the actual knowledge, on disk)
→ agent reads them with Read tool when needed
→ each file is self-contained on one topic
→ AGENT.md references them; agent uses them
AGENT.md
Auto-loaded into system prompt (up to 10000 chars). Keep it concise but information-dense.
Purpose: tell your future self where to look AND what will kill you.
What belongs:
- One-line project summary
- Architecture overview (2-3 sentences, link to detail file for more)
- Build/test/lint commands
- GOTCHAS section: critical pitfalls that MUST be visible without opening extra files. These are non-negotiable — if a pitfall can waste hours or crash the process, it goes here. Keep it to bullet points, no prose.
- Knowledge Files section: list of existing files with one-line descriptions
- Key conventions that don't fit elsewhere (max 5 bullets)
What does NOT belong:
- Anything that belongs in a knowledge file (except critical gotchas)
- Specific line numbers, function signatures, or code snippets
- Information already in README
- Minor/niche gotchas that are package-specific → put those in the relevant knowledge file
Knowledge Files
These are where the real knowledge lives. Create them freely — one file per topic, no need to consolidate. More small files is better than fewer large ones.
Directory Structure
Mirror the repository's directory structure under the knowledge root (e.g. docs/agent/). This makes it trivial to find the right file:
docs/agent/ ← knowledge root
architecture.md ← cross-cutting: message flow, pipeline, conventions
conventions.md ← cross-cutting: coding style, error handling
agent.md ← agent/ package: loop, engine, middleware
channel.md ← channel/ package: CLI, Feishu, Web, QQ
llm.md ← llm/ package: OpenAI, Anthropic, retry, streaming
tools.md ← tools/ package: built-in tools, sandbox, hooks
memory.md ← memory/ package: letta, flat, providers
session.md ← session/ package: multi-tenant sessions
storage.md ← storage/ package: SQLite, vector DB
config.md ← config/ package: JSON config, env overrides
prompt.md ← prompt/ package: templates, embed, rendering
When you explore a new subsystem or package, create its knowledge file. Don't worry about having too many — future sessions will use AGENT.md's index to find exactly the right file.
When to create a new knowledge file
- You explored a subsystem deeply enough to document it
- A knowledge file would save future sessions from re-exploring the same code
- A topic is growing too large in an existing file — split it
When to update an existing knowledge file
- You discovered something that changes or contradicts what's documented
- The project structure changed (files moved, APIs renamed)
- You found a bug, workaround, or gotcha worth recording
Decision Flow
After completing a task, update knowledge:
- Did I encounter or fix a gotcha/pitfall? → Write it directly into AGENT.md's GOTCHAS section. Critical gotchas (crashes, silent data loss, hours-wasting traps) go into AGENT.md so they're always visible without opening extra files. Package-specific/niche gotchas can go into the relevant knowledge file instead.
- Did I learn something about the codebase (APIs, dependencies, conventions, architecture)? → Write it into the relevant knowledge file
- Did the file/knowledge list change? → Update AGENT.md's Knowledge Files index
- Did any existing documentation become stale due to my changes? → Update it in place
- Did I add gotchas to AGENT.md? → Remove
docs/agent/gotchas.md if it exists and all its content has been migrated to AGENT.md. AGENT.md should be the single source of truth for critical pitfalls.
Default to updating. Every session should leave the knowledge base more accurate than it found it. The only exception is truly trivial changes (typo fixes, comment-only edits).
Accuracy Maintenance
- Before writing: Read the existing file first
- After writing: Verify AGENT.md references match actual files on disk
- When deleting/renaming files: Update all references in AGENT.md and other knowledge files
- Do NOT just append — revise outdated content
- Keep each file focused on one topic; split when a file grows beyond ~200 lines
- After significant updates: spot-check that AGENT.md index entries match actual files on disk
