Imported from bniceley50/GravenSpire (AGENTS.md). Install upstream with npx skills add bniceley50/GravenSpire. Copyright stays with the author.

AGENTS.md — Gravenspire Behavioral Contract

Source of truth for universal workflow rules: docs/brian-system-prompt-v4-6.md. This file contains project-specific extensions only. If AGENTS.md and the system prompt conflict, the system prompt wins.

This file is read by Claude Code at session start alongside the system prompt. Every agent, skill, and hook in this project inherits these rules via CLAUDE.md's Read First block. The 49 agent files in .claude/agents/ do not re-state these rules — they inherit them.

0. Canonical Repo Path

The canonical working directory for Gravenspire on this machine is:

N:\GravenSpire

(Windows host. From a POSIX shell like Git Bash, also addressable as /n/GravenSpire. The camelCase folder name is identifier-form; the in-world/prose name is "Gravenspire".)

Legitimate alternate paths: parallel agents (e.g. Codex) may operate from a dedicated worktree — canonical worktree path for Codex is N:\GravenSpire-codex (see "Codex onboarding" in .agents/ if present). Other worktrees follow the same pattern: outside the primary checkout, one per feature branch.

If you find yourself operating on a path that is NOT the primary checkout or a registered worktree (~/Projects/Gravenspire, /Users/brian/Desktop/..., a stale clone anywhere): stop and verify. The stale-checkout risk is real — the clinic-notes-ai 2026-03-22 lesson (inherited via the governance migration, DECISIONS.md D005) cost a full day chasing a bug in a stale checkout. That lesson predates Gravenspire but the failure mode is universal.

1. Session Start Ritual

Read, in order, at session start:

1. docs/brian-system-prompt-v4-6.md — universal rules
2. AGENTS.md — this file
3. DECISIONS.md — locked architecture decisions (D-numbered)
4. production/session-state/active.md — current work state (may not exist yet)
5. design/gdd/game-concept.md — Gravenspire concept / pillars
6. tasks/lessons.md — accumulated lessons

Then read files relevant to the current task. CCGS's session-start.sh hook previews active.md automatically; the rest is on you.

2. EDIT_OK Protocol

Before writing or editing any file, request approval with the phrase:

May I write this to [filepath]?

The user replies with EDIT_OK, EDIT_OK [filepath], or a denial. Batch approvals (execute on a pre-listed set of files) cover the whole listed batch. Do not extend a batch approval to files not in the original list.

This matches CCGS's existing Collaboration Protocol in CLAUDE.md — AGENTS.md just names the protocol token.

3. Evidence Rule

Every claim in a PR, audit, or status report must cite file:line + verification method.

• ❌ "Save integrity is handled."
• ✅ "Save integrity handled in src/core/save/SaveSerializer.cs:42 via HMAC signature; verified by tests/unit/save/SaveSerializer_Tamper_test.cs:18."

"Configured" never means "I think I configured it." Open the file, paste the line.

4. Source-of-Truth Rule

When two sources disagree, the canonical source wins. Canonical sources by domain:

If a GDD and a DECISION disagree, update one of them; don't paper over it. Usually the DECISION wins (it's the lock), and the GDD gets amended.

5. Output Contract

Every substantial response should end with:

• What changed (files touched, with paths)
• What's next (one sentence, concrete)
• Blockers or open questions (if any)

Don't narrate internal deliberation. Don't restate what the diff already shows. See CCGS's CLAUDE.md tone-and-style guidance — this just reinforces it.

6. Tier-Gated Work Policy

Gravenspire has four scope tiers. Work is gated by tier. Do not implement Tier N+1 features during a Tier N sprint without an explicit tier-transition decision appended to DECISIONS.md.

Current tier: T1 (per DECISIONS.md D003).

7. Pre-PR Adversarial Review (4 Questions)

Before opening or requesting review on any PR, answer these in the PR description:

1. What does this PR claim to do? (one sentence)
2. What evidence proves it works? (test file:line, playtest note path, screenshot path)
3. What's the blast radius if this is wrong? (which systems break, which data corrupts)
4. What did I not test? (be honest — name the untested edge case)

Tier-gated RED_TEAM signoff (see §6) is a separate block in the PR template.

Triggering skill: /pre-pr-review.

8. RED_TEAM Reference

Adversarial review protocol lives in RED_TEAM.md (9-section template) and RED_TEAM_RUBRIC.md (scoring guide). Run it when the tier table in §6 requires it. Save outputs to docs/audits/RED_TEAM_[YYYY-MM-DD]_[scope].md.

9. LESSONS Promotion Ritual

New lesson lives in tasks/lessons.md. Tag taxonomy:

• [GLOBAL] — applies across all Brian projects (also worth pushing up to the universal system prompt)
• [CI] — build/test pipeline
• [TEST] — testing discipline
• [TYPES] — C# type / generics / nullability
• [SCOPE] — tier discipline, scope creep
• [DATA] — data-driven config, ScriptableObjects, Addressables
• [DETERMINISM] — faction sim, physics, replay integrity
• [SAVE] — save format, versioning, migration
• [LLM] — prompt hardening, output moderation, latency
• [NETCODE] — FishNet, replication, authority
• [BALANCE] — combat formulas, economy

Promotion path: lesson repeats 2–3 times → promoted to CLAUDE-patterns.md (cross-cutting) or a specific .claude/rules/*.md file (path-scoped). Mark the original lesson entry with [PROMOTED to <target> YYYY-MM-DD].

Triggering skill: /update-memory-bank.

10. Session Self-Assessment

At the end of any session that touched more than one file, include:

• Files modified (full paths)
• Decisions made (D-number if one was added)
• Lessons worth logging (even if not formally added yet)
• What I got wrong or corrected mid-session

This is how lessons actually make it to tasks/lessons.md.

11. Secrets and PII (Always On)

Gravenspire has a much smaller PII surface than a HIPAA system, but:

• Never commit Steam API keys, signing certs, server auth tokens
• Never log player email / username / IP at severity INFO or above — use DEBUG and scrub in production builds
• Any LLM call that includes player-authored chat must sanitize (strip control chars, cap length) before dispatch
• Save files may contain player-chosen names — treat as user-controllable strings (length-bound, validate on load)

12. Pre-Build Verification (Inherited Lesson from clinic-notes 2026-04-10)

Before any Steam upload (or for T1, any playtest handoff), validate:

• ProjectSettings: build target, scripting backend, IL2CPP/Mono, color space, graphics API list
• Addressables groups: all referenced groups exist and build without errors
• Version stamp: Application.version matches the intended git tag
• (T3+) Server config: required env vars present in the target deployment

Checklist lives in production/pre-build-checklist.md. This is the Gravenspire analogue of clinic-notes's missing pre-deploy env-vars check. Every item produces evidence (file:line, screenshot, console output), not self-report.

13. Tool / File Writes

Honor the CCGS Collaboration Protocol (CLAUDE.md §Collaboration Protocol). This file does not override it — just names the approval token (EDIT_OK).

14. Commit & Push Cadence

CLAUDE.md's "No commits without user instruction" prevents speculative auto-commits. This rule prevents the inverse failure: approved work accumulating uncommitted.

At every approval checkpoint, the agent must prompt for commit + push:

• /design-review APPROVED verdict on a GDD
• EDIT_OK + post-edit verification on a substantive batch
• Test-passing implementation milestone (T1 = local gate per §6)
• Session about to end with approved work uncommitted

Default cadence: per-approval. One approval = one commit + push, scoped to that approval's files (GDD + review log + row updates + registry deltas). Batch commits only when the user explicitly says "batch these N items" or when the work is a coordinated cross-doc change.

The agent never commits without explicit user go. This rule shifts only the agent's prompting obligation, not commit authority.

Forbidden patterns:

• Approved work crossing a /clear or session boundary uncommitted
• Silent deferral of the commit question after approval
• Force push to main (already Bash-tool-guarded)

Failure mode this prevents:

2026-04-25 catch-up — ~5,000 lines of approved design work accumulated across multiple sessions because no protocol forced the commit question at approval moments. Recovered via single catch-up commit f1df1c5.

15. When AGENTS.md and System Prompt Conflict

System prompt wins. Always. File a lesson tagged [GLOBAL] so the conflict propagates up to the universal prompt if the system-prompt rule is wrong for all projects.