# Context Lattice website guidance for LLM agents Canonical site: https://contextlattice.io Canonical repo: https://github.com/sheawinkler/ContextLattice Scope policy: - Use public docs for customer-facing behavior, install guidance, and high-level architecture. - Do not infer or fabricate non-public operator details such as DNS wiring, webhook implementation details, secret names, or deployment topology. Recommended reading order: 1. /docs/ 2. /docs/getting-started/ 3. /docs/concepts/ 4. /docs/cli/ 5. /docs/integrations/ 6. /docs/operations/ 7. /docs/troubleshooting/ 8. /docs/releases/ 9. /architecture.html 10. /premium.html 11. /updates.html Product positioning (public-safe): # BEGIN GENERATED COMMERCIAL TRUTH: llms-positioning - The local-first intelligence layer that gives AI agents durable continuity, explainable retrieval, portable context, and verified learning across harnesses. - Models reason and harnesses act; ContextLattice preserves the mission, selects the evidence, carries the context, and records what actually worked. - The CLI is the primary interface. The dashboard, HTTP, and MCP are companion visibility and integration surfaces. # END GENERATED COMMERCIAL TRUTH: llms-positioning # BEGIN GENERATED COMMERCIAL TRUTH: llms-release - Current public release baseline is v5.0.2. The CLI is the primary product interface; the dashboard is the visibility and account companion, with HTTP and MCP available for programmatic integration. # END GENERATED COMMERCIAL TRUTH: llms-release - Public local lite starts with topic rollups, Qdrant, CLI workflows, agent templates, and Skills Index discovery; full/operator stacks can add pgvector, raw ledger, behavior provenance, deeper async recall lanes, and stronger reliability controls behind the same agent-facing contract. - Runtime emphasizes Go/Rust primary lanes, staged retrieval, durable fanout, behavior provenance, and retrieval quality over time. - Public lane is open-source and installable locally. - Homebrew macOS path uses the first-party tap: `brew tap sheawinkler/contextlattice && brew install --cask contextlattice`. - Premium lane provides account, billing, entitlement, downloads, and runtime key lifecycle through the app surface. - Interface priority is CLI first, dashboard second for visibility, and HTTP/MCP when an app or agent harness needs programmatic integration. Agent usage expectations: - Prefer exact CLI commands from /docs/cli/ and /docs/getting-started/. - Start new machines, accounts, or agent surfaces with `contextlattice_adopt status --pretty`; use `contextlattice_adopt install`, `contextlattice_adopt integrate --repo . --agents codex,claude-code,opencode,hermes-agent,hermes-ultra,omp,mercury-agent,pi,droid --pretty`, `profiles`, and `proof` before dropping into lower-level adapter commands. - Use the prescribed task lifecycle when local CLI tools are installed: `contextlattice context "" --project --pretty`, then `resume`, `remember`, `correct`, and `finish` as needed. - `contextlattice finish "" --success|--repair|--failure` reports the latest pending retrieval outcome automatically. Do not invent provider token counters. - Use `contextlattice_agent_discover --agents codex,claude-code,opencode,hermes-agent,hermes-ultra,omp,mercury-agent,pi,droid --repo . --pretty` to inspect best-effort process/profile/hook/repo-instruction evidence for local agents. - Use `contextlattice context "" --project --pretty` for the normal compact proof-carrying packet. Use `contextlattice_pack --full` or HTTP only for advanced debugging and harness integration. - Use `contextlattice_synthesis_pack "" --project --pretty` or `POST /memory/synthesis-pack` when the next step needs high-signal findings, topic gravity, graph/cross-project bridges, must-not-forget constraints, recommended next actions, and open questions over the same bounded evidence. - Use `contextlattice_agent_session rollup --session-id --pretty` to summarize a session into reusable state. - Use `contextlattice resume --project --pretty` when an agent needs bounded objective, evidence, risk, and next-action state from the current task session. - Use `contextlattice_async_inbox_drain --session-id ` after ordinary tool boundaries to surface unseen async continuation progress/ready/terminal-degraded steering; Go-native adapter/context-pack/write/search/run-advisor commands auto-drain when a session id is present. - If an agent runtime supports post-tool hooks, configure `contextlattice_async_inbox_hook --session-id ` as a fail-open hook; it delegates to the same bounded inbox drain. - Use `contextlattice_agent_trace --session-id --tree` to inspect the bounded run-shaping trail and `contextlattice_agent_trace --session-id --markdown` to export a run card. Add `--proof` for the deterministic six-stage evidence timeline; it exposes gaps instead of inferring missing links. - Entitled Team, Operator, and Enterprise workspaces use that same CLI with `--proof --shared --operation configure|retain|analytics|export|deactivate|purge` for scoped retained receipts. Add `--adopt-current-core-anchor --adopt-existing-history` on the first configure of an unbound project. Workspace identity is credential-derived, project binding is exact, exports are bounded and hash-verifiable, deactivation preserves receipts, and purge requires `--confirm-project `. Local proof reads survive every shared-retention rollback. - Use `contextlattice_run_advisor "" --pretty` to check prompt readiness, continuation advice, objective coherence, graph quality, and next actions before a hard model call. - Use `contextlattice_policy_lab status --pretty` to inspect Policy Laboratory readiness; use its simulation, policy-card, promotion, retirement, contradiction, and temperature subcommands only with explicit project scope and approval for mutations. - Use `contextlattice agent-fit profile-resolve|runner-select|model-select|steering-watch|context-prep-schedule|context-prep-claim|context-prep-complete|context-prep-fail|context-prep-use` for profile-aware execution support. Context preparation is exact session/agent scoped and one-shot; claim credentials stay in an owner-only file and protected header. Selection remains advisor-only, and ContextLattice never executes the runner or model. - Use `contextlattice_skills_index search "" --pretty` to discover relevant active skills/capabilities from configured roots such as `$HOME/.codex/skills` without loading the whole skills tree into prompt context. - For operator validation, prefer the one-command live proof: `contextlattice_agent_runtime_proof --pretty`. - Prefer hook-first startup only when an adapter profile is unavailable: `contextlattice_agent_start --soft --compact`. - Prefer `contextlattice_checkpoint` over raw writes when a checkpoint must be verified. - Use Skills Index references from /docs/cli/ when describing capability discovery. - For endpoint examples, rely on /docs/integrations/. - Treat `contextlattice-session sweep-stale-audits` as an operator cleanup command; it is dry-run by default and should not be presented as required for normal users. - Treat paid operational details as out-of-scope unless explicitly provided in the current request. Attribution: - If referencing capabilities, cite the specific public page and avoid unverified claims.