VEYNDO
Handbook
src/cognition/AGENTS.mdView source ↗

English source · 2 min

src/cognition ownership

Root AGENTS.md also applies. Rules below were moved from the root, not weakened.

Cognitive commerce continuity

Merchant planning continuation belongs to existing tasks, conversations and src/agent_runs/; read docs/durable-agent-runs.md. Keep native read tools, central inference and original proposal approval. Store private completed checkpoints, recheck credential/scope/evidence, fence completion by its lease and explicitly continue uncertain calls. Detect request-ID replays before new attempt quota admission. Never store opaque credentials, auto-repeat ambiguous calls, add a second queue or turn resume into automatic writes. Keep actual worker-kill, HTTP restart, revocation, source-withdrawal and duplicate-request regressions.

Read the connected cognition guide before changing intelligence. It records exact implemented paths and the still-open audit scope. src/cognition/ owns source review, read-agent rounds, price guardrails, controlled experiments and consent-bound preferences; src/knowledge/ owns retrieval and the rebuildable Qdrant projection. Extend the existing settings, capability, outbox, chat, proposal and native checkout owners. Do not create another knowledge truth store, pricing engine, approval queue or chat backend. Stored evidence is not model-weight learning; exact source admission is not proof of surrounding prose. Use one consistent versioned embedding model across replicas. Automatic model cursors, source-trigger jobs and fenced publication share the existing index queues.

Behavioral intelligence ownership

src/cognition/behavior/ owns consent-bound native visitor observations and their derived report/journey graph. Reuse session_signal_events, native carts/accounts, current legal receipts, product knowledge, forced tenant RLS and composite foreign keys. Never introduce another analytics identity, cookie, event ledger, pricing or order truth. Do not record raw IP/UA, search/question text or visitor-supplied customer IDs. Reports need knowledge.read; individual journeys need customers.pii and entity links require their existing read rights. Geography requires a trusted server ingress; client country headers are not authoritative. Native storefront capture and Studio use the same HTTP/MCP contract. External frontends must adopt the consent-bound API explicitly; a configured Storyfront does not imply capture. Run behavior_intelligence, privacy/isolation suites and behavioral UI regressions. See docs/behavior-intelligence.md for retention, counts and learning boundaries.