name	agent-operational-governance
description	Keep AI coding agent work tree-clean and enforce evidence only for high-impact operations such as non-scratch deletion, database writes, production/runtime changes, cross-machine sync, backups, models, services, cron, hook changes, or deployment evidence.

Agent Operational Governance

Use this skill when agent work needs clean Git handoff or high-impact evidence.

Three Working Rules

Temporary outputs only go under scratch/<session-id>/.
Prefer updating existing docs over creating new docs; if a new doc is needed, first identify why no existing slot fits.
Final response order: user task result, actual actions, file tree impact, validation, then governance evidence.

Zone Model

Classify work by reversibility and blast radius, not by action words.

Green: read-only queries, any operation inside scratch/, ordinary edits to Git-tracked source/docs, and local Git status/diff/add/commit/log work. Green work needs no extra journal or manifest because Git is the evidence chain.
Yellow: new files outside scratch/, new Markdown outside the approved documentation slots, and remote-affecting but Git-backed actions such as git push. Yellow work should be visible in the final response but should not block ordinary handoff.
Red: deleting non-scratch files, database writes, production/runtime config, service/cron/systemd/docker changes, cross-machine scp/rsync, protected artifacts, model/data artifacts, hook changes, and deployment evidence. Red work must close with a Git commit or a manifest.

Path/resource overrides beat the default zone. A Git-tracked production config, deployment selector, hook, secret, model, dump, or service definition is not green just because Git can revert it.

Evidence Policy

Git is the primary evidence. Do not create a second proof file for changes that are already represented clearly by a commit or a readable diff.

Use a manifest only for:

large or runtime artifacts kept out of Git
deletion of non-scratch files
cross-machine sync, backup/offload, or restore work
production/runtime work that cannot be fully represented in Git

Daily worklogs are for human continuity, not a universal stop-hook requirement. Keep entries short and append-only. Do not create per-turn reports unless the user asks for a report or the work produces an external artifact that needs a manifest.

Standard Closeout

Before final response:

Run git status --short for the touched repo.
Summarize changed source/docs/config.
Run relevant checks, or name skipped checks and why.
Keep large data, logs, models, dumps, and CSVs out of Git.
Commit code/docs when appropriate, or explain residual dirty state.
For red-zone work, commit or record a manifest that includes the session marker shown by the hook, plus path, size, file or row count, checksum, source machine, target machine, and whether external sync is required.

Use bundled scripts when possible:

python3 scripts/agk_journal_update.py --domain ops --item "what changed"
python3 scripts/agk_closeout_check.py --repo .
python3 scripts/agk_closeout_check.py --repo . --allow-dirty

Domain Mapping

ops: governance repo, inventory, audit, hook, backup, machine boundary
infra: SSH, tunnels, tmux, cron, systemd, machine setup
prod: production services and deployment selectors
research: experiments, evaluation, feature engineering, backfills, training outputs

Protected Artifact Rule

Do not commit runtime artifacts:

.env, .ssh, secrets, auth files
models/, data/, logs/, tmp/, exports/, research_eval/
large .csv, .gz, .jsonl, .pt, .pkl, .bin, .npz, .parquet, .dump, .sqlite, .db, .log

If such files matter, write a manifest and sync them through your approved backup path.

Stop Hook Interaction

The Agent Governance Kit Stop hook blocks for missing red-zone evidence or protected artifacts. Material but non-red-zone work defaults to warning mode unless AGK_MATERIAL_CLOSEOUT_MODE=enforce is set. If the hook continues the turn, complete the needed commit/manifest or protected-artifact cleanup, then answer.

name	agent-operational-governance
description	Keep AI coding agent work tree-clean and enforce evidence only for high-impact operations such as non-scratch deletion, database writes, production/runtime changes, cross-machine sync, backups, models, services, cron, hook changes, or deployment evidence.