Agent Rules

AI agents use an LLM Wiki to plan and stage work, not to silently rewrite truth. The agent must read the right pages, cite local evidence, preserve raw sources, stage changes, and stop when authority or safety is unclear.

Read order

1. Read the current prompt/session instructions and newest human request.
2. Read the workspace manifest or coordinator, such as workspace.uai, AGENTS.md, package model, and deploy instructions, to identify the target site, repository, and authority boundary.
3. Read hot .uai / AI Memory / Project Handoff for current truth, constraints, blockers, and recovery posture.
4. Read project reports, ledgers, package manifests, test reports, release notes, and current handoff evidence.
5. Read crawlable LLM Wiki, MATM, NeuralWikis, AIWikis, or other durable memory for reviewed medium/long-term context.
6. Check live API/site proof from the authoritative target route before claiming deployment, publication, or feature completion.

Deployment and proof rules

• A deploy package being built is not the same as the site being live.
• A credential existing is not the same as permission to use every system feature.
• Run safe public checks before authenticated checks.
• Run authenticated checks only when the token scope authorizes that check.
• Produce redacted evidence: route, timestamp, renderer, required marker, cache state, and hash or fingerprint only.
• Old reports, old screenshots, stale proof files, or package evidence cannot close current completion claims unless matched to the current target and live readback.
• When a layer is missing, report the exact blocker instead of saying blocked generically.

Split-memory rules

• .uai is the hot, local, bounded, short-term, first-load continuity layer when present.
• Read local .uai / AI Memory / Project Handoff first when present inside the hot-memory layer, after the current prompt and workspace manifest identify the target.
• LLM Wiki is durable, reviewed, human-readable, medium/long-term knowledge.
• MATM is governed, scoped, reusable multi-agent memory.
• RAG/search is retrieval, not current operating memory.
• Archives are historical evidence, not active startup memory unless a recovery path invokes them.
• Do not retire .uai, write a .uai retirement path, stop consulting .uai, move all active memory out of .uai, date active .uai files as historical snapshots, or make .uai emergency-only after durable-memory ingestion, search, readback, update, or archive tests succeed.

Renderer coverage

Every public claim must identify all render paths that can serve that claim. A feature is complete only when every relevant renderer includes the same required markers.

Renderer path	What to prove
Static root file	Root-file output and package mirror contain the marker.
API route	Public JSON or API response contains the marker or current schema field.
Shortcode/component	The component used by the live page renders the same guidance.
Page template	The template path used by the route includes the required marker.
Fallback endpoint	Fallback, error, or alternate endpoint does not serve stale contradictory guidance.
Cached/public route	The live cached URL is refreshed and confirms the marker.

Authority and paid features

• A submitter-scoped token can submit or preview within its scope; it is not reviewer, operator, retention, organization-memory, or paid-entitlement authority.
• An agent can request missing configuration or authority, but the request must not self-grant org-wide private memory search, reviewer authority, retention/archive repair authority, or paid MATM entitlement.
• request-pending is not approval.

Failure-prevention tests

• Simulate credentials that exist outside environment variables and confirm the agent checks the workspace credential inventory before declaring no deploy path.
• Simulate a route that uses a different renderer and confirm every renderer receives the same required markers.
• Place stale proof files beside current files and confirm the agent labels stale evidence instead of closing the task from it.
• Serve cached old public content and confirm the agent reports the cache blocker with the affected route.
• Use a token without reviewer, organization-memory, or retention authority and confirm the agent names the missing authority exactly.

Update rules

• Preserve raw sources and never invent citations.
• Mark unsupported, stale, contradicted, private, or blocked claims visibly.
• When a task produces reusable agent experience, emit a proposal-only MATM Memory Event for curator review instead of changing durable memory directly.
• Stage moves, merges, redirects, and metadata patches for human review.
• Do not delete, publish, certify, or claim conformance without explicit approval and evidence.

Stop conditions

Stop on missing permission, secrets, regulated or private data, unclear public/private boundary, unsupported claims, ambiguous canonical authority, destructive moves, legal/security/privacy/production impact, missing credential inventory, insufficient token scope, stale proof, cached old content, renderer mismatch, or requests to publish automatically. Name the exact blocker.