xete-mcp: Instructions file for Claude Code

Xete Agent Operating Rules

These rules bind every agent session (Claude Code, Remote Control, owl-alpha, or any other model) working in Xete repositories. They are not suggestions. The pre-commit hook enforces the review gate mechanically; the rest are enforced by you, the agent, refusing to skip them.

Hard rules

1. No code before spec for (a) any work touching protected paths and (b) any work handed between people or agent sessions. Use skills/spec-and-plan/SKILL.md. Exploratory work on a feature branch that stays in one head and off protected paths needs no spec — but the moment it's handed off or aimed at a protected path, write the spec before continuing.

2. No contract-path commit without a DDR. Any change under a protected path (see .githooks/protected-paths) requires a completed doubt-driven review (skills/doubt-driven-review/SKILL.md) with a staged reviews/DDR-*.md carrying Verdict: SHIP. The hook will block you; do not use --no-verify unless the human explicitly authorizes it in the current session, and record that authorization in the DDR.

3. Security gates run inside the DDR for any crypto, key-handling, or on-chain change. Use skills/solana-security-hardening/SKILL.md. Mark any unresolved item OPEN-GATE — the hook blocks commits containing open gates. Do not delete an OPEN-GATE marker; resolve it or escalate to the human.

4. Every commit follows git discipline (skills/git-discipline/SKILL.md): atomic, green, ~100 lines, spec-referenced. Run the session start ritual before your first change.

5. Fresh-context doubt is real, not simulated. When the DDR calls for adversarial review, spawn a subagent or instruct the human to open a fresh session. Reviewing your own reasoning inside the same context does not count and must be recorded as self-review-only in the DDR, which downgrades the verdict to BLOCK for contract paths.

6. The agent runs the gates; the human reads verdicts. Writing the spec, running the DDR loop, spawning the doubt subagent, filling the security checklist, and authoring reviews/DDR-*.md are all agent work, done without asking permission. The human's involvement is reading the finished DDR's doubts and verdict — not producing paperwork. If a gate turns a 10-minute task into an hour of ceremony for a non-contract change, the protected-paths patterns are miscalibrated: flag it to the human with a proposed pattern fix rather than suffering silently.