mirror of
https://github.com/PR0M3TH3AN/TORCH.git
synced 2026-10-27 07:51:52 +00:00
docs: add generic agent onboarding and handoff guidance
This commit is contained in:
1 parent
1452e3f964
commit
8d5e7e0220
8 files changed
+211
-3
No files matched your search
@@ -0,0 +1,69 @@
|
||||
# AGENTS.md
|
||||
|
||||
Repository-wide instructions for any autonomous or semi-autonomous coding agent.
|
||||
|
||||
## Purpose
|
||||
|
||||
This repository is intended to be reusable across many projects. Keep changes and documentation generic unless you are actively working inside a specific downstream repo that requires project-specific customization.
|
||||
|
||||
## Startup checklist (every run)
|
||||
|
||||
1. Read this file (`AGENTS.md`).
|
||||
2. Read `CLAUDE.md` for operating conventions.
|
||||
3. Read `CONTRIBUTING.md` for commit/PR quality expectations.
|
||||
4. Check `KNOWN_ISSUES.md` before making changes.
|
||||
5. If present, check `docs/agent-handoffs/` for recent agent learnings.
|
||||
|
||||
## Generic-first policy
|
||||
|
||||
- Do not hardcode organization, product, customer, or brand-specific details.
|
||||
- Do not assume a single deployment environment unless explicitly configured.
|
||||
- Prefer neutral names such as `project`, `service`, `integration`, `consumer`, `provider`.
|
||||
- Add project-specific notes only when working in that project's repo and label them clearly.
|
||||
|
||||
## Agent learning + handoff protocol
|
||||
|
||||
Agents are expected to leave concise, reusable notes for later agents.
|
||||
|
||||
### Where to write
|
||||
|
||||
- `docs/agent-handoffs/README.md` for folder usage.
|
||||
- `docs/agent-handoffs/learnings/` for new practical learnings.
|
||||
- `docs/agent-handoffs/incidents/` for issue/incident writeups and mitigations.
|
||||
|
||||
### Naming convention
|
||||
|
||||
Use generic, descriptive filenames:
|
||||
|
||||
- `YYYY-MM-DD-short-topic.md` (example: `2026-02-14-lock-timeout-tuning.md`)
|
||||
|
||||
Avoid brand, team, or person-specific prefixes in filenames.
|
||||
|
||||
### Required note structure
|
||||
|
||||
Each note should contain:
|
||||
|
||||
- Context
|
||||
- Observation
|
||||
- Action taken
|
||||
- Validation performed
|
||||
- Recommendation for next agents
|
||||
|
||||
Keep each note short and actionable.
|
||||
|
||||
## Safety and quality baseline
|
||||
|
||||
- Make smallest safe change first.
|
||||
- Prefer explicit commands over assumptions.
|
||||
- Validate with available tests/checks.
|
||||
- If blocked by environment limits, document the blocker and fallback.
|
||||
|
||||
## Priority order
|
||||
|
||||
When instructions conflict, follow this order:
|
||||
|
||||
1. System/developer/user instructions in-session
|
||||
2. `AGENTS.md`
|
||||
3. `CLAUDE.md`
|
||||
4. `CONTRIBUTING.md`
|
||||
5. Inline comments and prompt files
|
||||
Reference in new issue
Block a user