qingye-lab Project guidance
A user's explicit request to create, refresh, or refine guidance authorizes that work within the stated file scope; the user need not name this Skill. Review/validation requests are read-only. Ordinary coding tasks and natural-language repository guidance do not authorize persistent edits. The SessionStart Hook is always read-only.
The existing repository-authored exception is the exact standalone marker:
<!-- rootloom:refine-once version=1 -->
It permits one refinement of that marked file only. Remove it with the successful edit; validation does not consume it. Preserve any explicit independent approval requirement.
Existing guidance
Resolve each target's real path and symlink chain before editing. Read its applicable global and root-to-target guidance and only the evidence needed for the requested change. An entry pointing outside the authorized repository is external content: propose its diff unless the user has separately authorized that real target. Preserve backups and unrelated edits; do not reset or overwrite user work to simplify refinement.
For wording-only refinement, read semantic-refinement.md and edit authorized unmarked content directly. No manifest/CI/module probe is needed. Keep model judgment outside generated managed markers. Add only durable, supported rules that change a real decision; remove redundant wording without dropping unique constraints.
Generate or refresh project facts
Resolve this Skill directory and use its deterministic probe and writer:
python3 <skill-dir>/scripts/seed_project_guidance.py probe --cwd "$PWD"
python3 <skill-dir>/scripts/seed_project_guidance.py seed --cwd "$PWD"
The writer owns only its generated project block. It must not overwrite unmarked guidance,
AGENTS.override.md, symlinked or external targets, disabled projects, or vendor/cache trees.
Respect its skip reasons. Only after an exact untrusted_project response and the user's
explicit persistence request for this repository may it retry once with --allow-untrusted;
this does not change host trust or authorize other writes.
Either .rootloom/disable-project-guidance or legacy
.codex/disable-project-guidance-seeding disables session context and persistent seeding.
Use --target path/to/module only for a genuine manifest-backed boundary with distinct
commands, ownership, or contracts; create at most three nested files per pass within three
levels of the Git root. Do not mirror the directory tree.
Verify
For edited unmarked guidance or blocks generated by this project seeder:
python3 <skill-dir>/scripts/seed_project_guidance.py validate --file path/to/AGENTS.md
The global working-agreement block belongs to Setup, even though it shares the marker prefix. Compare that block with the Setup-owned template and use Setup's plan/status for installation state; do not treat it as a generated project block or refresh it with this writer.
Check changed rules against the inherited chain and a concrete misuse scenario. Compare the final diff with the authorized paths and preserve unrelated content. Report actual checks and remaining gaps. Use no-cache verification where practical; remove only temporary artifacts created by this task. Continue any independent authorized work after guidance verification.