Agent-Native Architecture Re-Audit (v2) — Templiqx
Date: 2026-07-13 Scope: current production-readiness branch; application service, CLI, MCP, filesystem composition, and conformance harness Baseline: v1 audit (2026-07-12) — 66% overall Current score: 94% (strong agent-native contract surface; remaining ceiling is mostly intentional host/UI scope)
Score summary
| Core principle | Score | % | Evidence-backed status |
|---|---|---|---|
| Action parity | 26/26 | 100% | Every canonical operation has Rust, CLI, and MCP behavior coverage |
| Tools as primitives | 25/26 | 96% | test_package intentionally remains a workflow convenience over eval primitives |
| Context injection | 6/6 | 100% | Catalog, packages, workspace resource, roots, empty state, and prompts |
| Shared workspace | 2/2 | 100% | CLI and MCP use the same confined filesystem composition |
| CRUD completeness | 16/16 | 100% | Contract, package/manifest, and workspace-artifact lifecycle operations |
| UI integration | 4/5 | 80% | Structured envelopes and filesystem visibility; no live push/watch surface |
| Capability discovery | 7/7 | 100% | Catalog/resources/instructions plus bootstrap and run-eval prompts |
| Prompt-native features | 8/12 | 67% | Contract behavior is declarative; deterministic compiler and host policy remain code/host concerns |
Weighted total: 94/100. This is a current architecture score, not a claim that CRM3 production integration is complete.
What changed since the earlier 86% snapshot
26-operation parity
templiqx_application::CAPABILITY_CATALOG is the authority. CLI and MCP expose
the same 26 names, including package identity/sign/verify, package update and
delete, and workspace-artifact delete. The catalog-derived conformance test in
crates/templiqx-conformance/tests/crm3.rs fails when an operation lacks a
Rust/CLI/MCP behavior case. This replaces name-only parity with behavior parity.
test_package is the only workflow-lite operation. Agents can instead compose
the primitive list_evals and run_eval operations, so retaining the wrapper
does not create a separate semantic path.
Complete repository-owned lifecycle
| Entity | Create | Read | Update | Delete |
|---|---|---|---|---|
| Contract | put_contract |
inspect_contract |
put_contract with CAS |
delete_contract |
| Package / manifest | create_package |
discover_packages, identity export |
update_package with CAS |
delete_package |
| Workspace artifact | render/execute writes | list/read | confined overwrite | delete_workspace_artifact with CAS |
Mutations preserve path/symlink confinement, compare-and-swap behavior,
signature invalidation, and dependency/untracked-content guards. Package-local
sha256-keyed trust is a development/CI tamper-evidence mechanism; it is not
the production OCI trust root.
Agent context and discovery
The MCP surface provides templiqx://catalog, templiqx://packages, and
templiqx://workspace, dynamic root/workspace instructions, empty-state
guidance, and bootstrap / run-eval prompts. The MCP binary accepts explicit
package and workspace roots, so agents and humans operate on the same files
without a hidden agent sandbox.
Remaining architectural ceiling
The remaining six points are not missing catalog operations:
test_packageis deliberately a convenience workflow rather than a single primitive;- stdio MCP is stateless and has no push/subscription, recent-activity, or live watch UI;
- preferences, approval, authorization, retrieval, tenant context, and trace composition are host-owned;
- validation/compilation, DOCX safety, and deterministic rendering correctly remain code, while authorable interaction/eval behavior remains in contracts.
Adding a web UI, session store, or CRM3 policy to chase a literal 100% would weaken the current boundary. A future host may add live resource subscriptions or activity context without duplicating Templiqx semantics.
Production-readiness relationship
Agent-native readiness and CRM3 production readiness are separate claims. Templiqx can own and release the standalone 26-operation compiler surfaces, package trust contract, deterministic DOCX compatibility corpus, and 8/8 synthetic scenario/conformance artifacts. A real CRM3 ModelGateway, tenant/auth/retrieval/approval/audit policy, sanitized customer data, and host deployment acceptance remain blocked on the Basenet host.
Verdict
| Question | Answer |
|---|---|
| Current agent-native score | 94% (up from 66% v1 and the earlier 86% v2 snapshot) |
| Rust/CLI/MCP action parity | 26/26 with catalog-derived behavior coverage |
| Repository-owned CRUD | Complete for contracts, packages/manifests, and workspace artifacts |
| Literal 100% desirable? | No — remaining scope is workflow convenience or intentional host/UI ownership |
| CRM3 production-ready? | Not claimed — standalone release/conformance readiness is distinct from host integration |