Home

XCSV EXILE

AI / AGENT START HERE

Read AI Start Here and AI Provenance & Doc Sync before acting on the roadmap or creating commits.

If Architect says "read the GitHub", "read the roadmap", "get caught up", or equivalent, run the canonical READ_ONLY_BOOTSTRAP first. Reconcile roadmap intent against RAG/history, local working trees, Git history, GitHub/submodules and relevant live evidence. Work only on the remaining delta; do not duplicate equivalent functionality.

Current documentation state: SYNCED as of 2026-08-07 (work ID XCSV-AI-001). The first SOVRAN-1 reconciliation is complete. The AI contract, provenance policy, tooling decision and GUARD development programme were absorbed into the desktop roadmap as Phase 15 by append, with no existing section rewritten; this wiki keeps its deliberate abstraction of the desktop's measurements, RPT filenames and refuted hypotheses. That abstraction is the intended layering, not drift.

A dedicated Arma 3 Exile server on Tanoa, and every piece of software that keeps it alive.

This wiki is the operator/development map. The site is the public-facing view; the hub repo is the repository map.

Start here

page	when you need it
AI Start Here	mandatory AI navigation, reconciliation, delta-first workflow and GitHub work-tracking contract
AI Provenance & Doc Sync	AI commit attribution, issue work receipts, and safe desktop/GitHub documentation reconciliation
AI / Development Tooling	selected/rejected workflow tools, GitHub-native tracking decision, optional MCP and recommended CI/security helpers
XCSV GUARD Development Plan	full current design programme: Gauntlet, restart safety, replay, backplane, UX, evidence, deployment, staging, player systems
Roadmap	current order, completed work and planning priorities
Architecture	understanding what talks to what
Runbook	the server is broken and you need the operating procedure
Repositories	where a given file lives and why
XM8 Apps	player/admin-facing app work
Memory	project memory / retrieval workflow
Lessons	failures, refuted hypotheses and rules bought with real incidents

Workflow/tooling decision

XCSV uses GitHub Issues + GitHub Projects as the primary execution tracker so tasks, commits, PRs and sub-issues remain attached to the source instead of being copied into a second board. The official GitHub MCP Server is an optional enhancement after the baseline git/gh workflow is verified. Linear remains the only serious external future candidate, and only if GitHub Projects proves to have a measurable limitation.

See AI / Development Tooling for the full decision record.

Current GUARD programme

GUARD is being evolved around these principles:

• one canonical Gauntlet Protocol across development AIs
• explicit durable / reconstructable / ephemeral state ownership
• separate configuration from durable operational state
• schema migrations and atomic persistence
• startup reconciliation and desired vs observed component state
• restart-safe absolute deadlines and RCon auto-recovery
• Safe Mode for damaged/untrusted startup state
• backend service/backplane contracts for every tab
• self-diagnostics and formal tab integration tests
• offline Replay Mode plus permanent failure fixtures
• Operator Action Journal
• task-oriented UX with Operations/Engineering modes
• historical telemetry, explicit SLOs and capacity/headroom
• Incident Mode and a read-only Tanoa operations map
• artifact registry, deployment diff and explicit rollback identity
• documentation/state drift checks
• addon provenance/compatibility registry
• Change Impact Graph
• isolated staging for risky G3/G4 work
• GUARD evidence exported to development AI without granting unrestricted production authority

Core invariants:

GUARD should never need to remember that the server is healthy. It should be able to prove the server is healthy again every time it starts.

A production failure should surprise XCSV once; afterward it becomes evidence, a fixture or an executable check.

Player-development direction

The estate is already large. Prefer coherence and progression over blind addon count: Territory Manager, contracts/jobs, bounties, meaningful faction standing, Chronicle/community systems, choreographed world events and selective objective systems. Keep heavyweight additional AI frameworks parked until measured headroom supports them.

Non-negotiables

• production remains x64 + extDB3
• -filePatching remains load-bearing for the current A3XAI configuration path
• no Exile source copied into XCSV original addons
• secrets stay out of git
• runtime local-model output is untrusted and non-load-bearing
• UNKNOWN is not zero; absence of logged failure is not proof of function
• performance changes require settled, comparable evidence