docs: add repository instructions and the agent-loop skill

.github/copilot-instructions.md

@@ -0,0 +1,7 @@
+# Copilot instructions
+
+This repository is control-plane practice for the GH-600 path. The stack is bash, GitHub Actions, and Markdown. There is no application, no package manager, and no `pytest` or `npm test`.
+
+Read `AGENTS.md` first. Follow `.github/skills/agent-loop/SKILL.md` for the claim, branch, pull request, and stop procedure. Stay inside the issue's Allowed scope.
+
+Do not add an application, a test framework, or a `package.json`. Do not add a Plan Gate check or a hook pattern unless the issue you are on asks for one. Plan Gate stays a traditional workflow with `contents: read`.

.github/instructions/scripts.instructions.md

@@ -0,0 +1,10 @@
+---
+applyTo: ".github/scripts/**"
+---
+
+# Script rules
+
+- Start each script with `set -euo pipefail`.
+- Keep the self-test in the script. Run it with `--self-test`.
+- A hook deny prints the reason and exits 0. A non-zero `preToolUse` exit would deny every later shell command.
+- Do not add a Plan Gate check or a hook pattern unless the issue you are on asks for one.

.github/instructions/workflows.instructions.md

@@ -0,0 +1,10 @@
+---
+applyTo: ".github/workflows/**"
+---
+
+# Workflow rules
+
+- Plan Gate (`.github/workflows/plan_gate.yml`) stays a traditional GitHub Actions workflow and stays the merge check.
+- Plan Gate keeps `permissions: contents: read`.
+- A new workflow does not get write permissions. `contents: write` or `pull-requests: write` would let that workflow change the branch or the pull request and bypass Plan Gate.
+- A later agentic workflow may comment. It does not replace `require-plan`.

.github/skills/agent-loop/SKILL.md

@@ -0,0 +1,45 @@
+---
+name: agent-loop
+description: Claim a ready-for-agent issue, branch from main, open a pull request inside Allowed scope, and stop after one failed check. Use when starting work from a GitHub issue in this repo.
+---
+
+# Agent loop
+
+The contract record is `docs/adr/agent-loop-contract.md`. Follow this procedure. Do not invent a second process.
+
+## Claim
+
+1. Start only when the issue body has **Goal**, **Allowed scope**, and **Success criteria**, and a human has applied `ready-for-agent`.
+2. Assign yourself. Remove `ready-for-agent`. Add `in-progress`.
+3. Do not apply `ready-for-agent` to any issue.
+
+## Branch
+
+Branch from `main` only. Name it `feat/#<n>-…`, `fix/#<n>-…`, or `chore/#<n>-…`. Do not commit or push on `main` or `master`.
+
+## Pull request
+
+1. Open the pull request ready for review.
+2. Include `Fixes #<n>` or `Closes #<n>`.
+3. Fill **Goal**, **Scope**, **Steps**, and **Success criteria**. **Scope (paths/files):** is one line of comma-separated repo paths. A changed file must equal a path or sit under it. Stay inside the issue's Allowed scope. The rule is `docs/02-scope-match.md`.
+4. For a code change, fill **Risks** and **Rollback**. For docs-only work those fields are optional.
+5. When a check or scan ran, fill **Evidence**. Otherwise write `n/a`. The heading is exactly `## Evidence`.
+6. Leave the review checklist for the human reviewer.
+
+## Stop
+
+Revise the branch once when a check fails. If that same check fails again, stop. Do not push a third time for the same failure. On the pull request, write what failed, what you tried, the evidence, and the next step for a human. The rule is `docs/02-fail-twice.md`.
+
+If you cannot finish inside the Allowed scope, remove `in-progress`, add `needs-info` or `ready-for-human`, unassign yourself, and comment why.
+
+## Bans
+
+1. Do not push or commit on `main` or `master`.
+2. Do not merge your own pull request.
+3. Do not change rulesets or branch protection. Do not weaken or remove `CODEOWNERS`, Plan Gate, or the hook.
+4. Do not read or print secrets. Do not commit tokens.
+5. Do not edit a path outside the issue's Allowed scope unless a human comment widens it. Update the Scope line in the same push.
+6. Do not apply `ready-for-agent`.
+7. Do not force-push.
+
+A human merges. Merge to `main` finishes the loop. The closing keyword closes the issue. Remove `in-progress`.

AGENTS.md

@@ -2,7 +2,9 @@
 
 This repository is control-plane practice for the GH-600 path. The stack is bash, GitHub Actions, and Markdown. There is no application, no package manager, and no `pytest` or `npm test`.
 
-Follow `docs/adr/agent-loop-contract.md` for how work enters, is claimed, branched, reviewed, and finished, and for the seven bans. Label meanings are in `docs/agents/triage-labels.md`. Do not invent a second process, and do not add a check unless the issue you are on asks for one.
+This file is the index. The procedure is `.github/skills/agent-loop/SKILL.md`. The contract record is `docs/adr/agent-loop-contract.md`. Label meanings are in `docs/agents/triage-labels.md`. The shell hook is `docs/02-hooks.md`. Do not invent a second process, and do not add a check unless the issue you are on asks for one.
+
+Branch from `main` only. Name it `feat/#<n>-…`, `fix/#<n>-…`, or `chore/#<n>-…`.
 
 ## Commands