Add three new sections to AGENTS.md addressing workflow gaps observed in
a session where 6 features were implemented but none were functionally
tested — only static analysis (type checks, linting) was used, resulting
in broken features shipped as 'done'.
New rules:
- Functional Verification: features must be end-to-end tested before
completion; static analysis alone is explicitly insufficient
- Mandatory Quality Pipeline: every feature gets reviewer + tester passes;
no batch validation; quality over quantity under time pressure
- Requirement Understanding Verification: verify understanding before
implementing features with implicit expectations or domain concepts
Also simplifies tester bash permissions and adds plan write permission.
Replace 4 separate instruction files with 1 real file + symlinks:
- .github/copilot-instructions.md is the canonical instruction file
- CLAUDE.md and .cursorrules are symlinks to it
- AGENTS.md stays as global config (not a project instruction file)
This eliminates all sync/merge logic - changes propagate automatically.
Changes:
- AGENTS.md: rewrite Cross-Tool Instruction Files section for symlink convention
- librarian.md: simplify to maintain single instruction file + verify symlinks
- lead.md: simplify PHASE-WRAP and Documentation Completion Gate
- commands (init, bootstrap-memory, save-memory): update for symlink model
- doc-coverage skill: verify symlinks exist and point correctly
Expand librarian role to own cross-tool instruction file sync,
.memory/ file maintenance, and knowledge merge workflows. Lead
now proactively dispatches librarian in PHASE-WRAP and Documentation
Completion Gate instead of handling docs inline.
AGENTS.md, CLAUDE.md, copilot-instructions.md, and .cursorrules may
contain project knowledge added by other tools or people. Agents now
read these files at session start and merge new knowledge into
.memory/knowledge.md as the canonical source.
Remove the megamemory MCP knowledge graph and replace with plain
markdown files in .memory/ for tracking plans, research, knowledge,
and implementation state. This enables collaboration across people
and agentic coding tools (Claude Code, Copilot, Cursor, etc.).
Changes:
- Remove megamemory MCP from opencode.jsonc
- Delete tool/megamemory.ts and .megamemory/ database
- Rewrite all 25 config files to use .memory/ markdown files
- Add cross-tool instruction file awareness (AGENTS.md, CLAUDE.md,
copilot-instructions.md, .cursorrules)
- Update save-memory, bootstrap-memory, status commands for md workflow
- Update all agent files, skills, and commands consistently
Add 7 new sections to lead.md addressing failure patterns observed in
real session transcripts: exploration sharding for context-exhausted
explorers, environment probe protocol for infrastructure verification,
finding completion tracker to prevent silently dropped findings, targeted
re-review to avoid infinite review loops, tester capability routing to
match agent tools to task requirements, post-implementation sanity check
for coherence verification, and artifact hygiene for temp file cleanup.
