Claude Code is the execution surface for this repo: it plans and performs authorized work under AGENTS.md, under single-writer-per-branch discipline. ASK is the source-of-intent and authorization apex, the relay across surfaces, and the final adjudicator. The standard ASK setup configures a non-writing GPT advisor surface. When configured, it supplies external challenge, reconstruction, and verification from outside this thread and holds no repo-write authority.

Direct execution — ASK driving this thread without an advisor pass — is the bounded variant for work where a separate pass would not materially reduce uncertainty. It is a proportional path within the same model, not a separate model.

Compactly: multi-surface in reasoning, single-writer in mutation, single-apex in authority.

The workflow rules live in repo-local AGENTS.md and are written agent-agnostically — they would apply to any executor.

Before any meaningful repo work:

1. Verify repo attachment.
2. Stop on mismatch.
3. Report the exact working directory, git toplevel, origin remote URL, current branch, and working-tree state.
4. Read repo-local entry points before responding to the substantive request.

Run:

pwd
git rev-parse --show-toplevel
git remote get-url origin
git branch --show-current
git status --short

After verification:

1. README.md
2. AGENTS.md, plus CLAUDE.md — the required adapter for a Claude-operated repo, whose @AGENTS.md import is what actually delivers the resolved carrier into context (if it is missing or does not import, say so rather than proceeding on prose alone)
3. docs/architecture.md if present
4. docs/index.md if present
5. The latest milestone or finding artifact relevant to the task
6. Repo-external grounding note (if the task involves intent, audience, philosophy, foundational premises, or durable loose threads)

The full rule set lives in repo-local AGENTS.md. The essentials:

• Plan-Before-Execute. Before executing a meaningful repo change, state the plan: files in scope, scope in vs out, non-actions, expected terminal state. The plan is an inspectable reasoning object before mutation: it lets ASK, the executor, and any configured advisor challenge scope, assumptions, non-actions, and the intended terminal state before work becomes harder to unwind.
• Exact scoped diff gate. Stop at exact scoped diff before commit, push, or PR creation, unless ASK has already approved commit / push / PR.
• Structured Change Summary. Meaningful changes require why-this-exists / what-changed / what-did-not-change / what-remains-out-of-scope, captured in the PR description or the approval record.
• Default hold or carry-through. After ASK approves the scoped diff, continue through commit → push → PR. When the adversarial-collaboration preconditions are met — hardened backbone, active architectural uncertainty, configured advisor surface — hold at PR created for advisor relay; otherwise continue through merge → branch cleanup. Do not ask ASK to re-authorize predetermined transitions inside either path.
• Single-writer per branch. Treat repo and remote as the audit trail when sessions disagree. Stop on suspected concurrent mutation.
• Scope discipline. Smallest honest unit for implementation; largest tractable structural question for architecture work. Don't bundle unrelated work.
• Repo + grounding note are the durable sources. Per-conversation state (this thread, task lists, in-flight context) is ephemeral. Do not promote it into the durable sources.
• Claude Code auto memory (MEMORY.md and its topic files) is persistent, not ephemeral — a non-authoritative, read-mostly operator cache, never a durable owner. Route every reusable learning to its visible owner first, and treat any memory write as a separate approval unit per AGENTS.md §Learning Disposition and §Private-Memory Write Gate.

Stop and ask before:

• widening scope mid-task
• committing or pushing without explicit scoped-diff approval
• destructive operations (force push, branch deletion, file deletion at scale)
• mutations to external systems where the target is ambiguous

When in doubt, prefer asking over assuming authorization.