code-HFD / AGENTS.md
Tinghong-Ye22's picture
Add files using upload-large-folder tool
aa02c28 verified
|
Raw History Blame Contribute Delete
2.59 kB
# Repository workflow
Do not create standalone documentation unless the user explicitly requests it.
Routine work updates belong in the existing CHANGELOG/DIALOG (or consolidated
doc/工作留痕.md), not additional dated documents.
This repository is the standalone code backup for Humanoid-Full-Duplex (HFD).
It must not push to, inherit new work from, or otherwise mutate the old
Psi0-yth repository. Follow these rules for every change made in this working
tree.
1. Use one branch per HFD change or project version. Do not reuse old Psi0-yth
version prefixes such as `v26-`. Every newly created branch must include a
`YYMMDDHHMM` timestamp in its name.
2. After changing source code, configuration, scripts, tests, or documentation,
add a concise but complete Chinese entry to `CHANGELOG.md`.
3. Run checks appropriate to the change, then create a focused commit with a
concise message that says what changed.
4. Push the commit to the same branch on `origin`, then merge that tested branch
into the repository's configured default branch and push the default branch
as well. Preserve the feature commit and its authorship in the merge history.
If either push or merge is blocked, report it explicitly; do not silently
leave unpushed or default-branch-only work.
5. Never commit training runs, checkpoints, datasets, runtime indexes, caches,
credentials, `.env` files, or repository SSH keys.
6. Keep the repository-specific SSH key and Git identity local to `.git/` and
`.git/config`; do not modify global Git or SSH configuration for this repo.
7. Record each substantive user request and its outcome in `DIALOG.md` as well
as the Chinese change summary in `CHANGELOG.md`. Preserve the user's original
wording in DIALOG and summarize the assistant's work. Experiments must also
be recorded in both logs, with evidence limits stated explicitly.
8. In this formal working tree, `experiments/` is for Markdown reports only:
one consolidated report per experiment, named with its start date. Update
that report instead of creating dated progress, plan, audit, and conclusion
fragments. Keep formal implementation and usage/check details in the normal
source locations and DIALOG/CHANGELOG, not in an experiment report.
9. Keep temporary experiment code, raw metrics, checkpoints, caches, and audit
traces outside this formal tree in an isolated experiment workspace. Before
removing redundant artifacts, verify the exact scope and preserve unique
useful evidence; never clean unrelated user files or active training output.