Keprix Agent Parity Report¶

Agent parity reviewed against the Hermes Agent reference tree. Inventory completed 2026-07-12. Prompts 327-334 were archived before this verification pass. Re-proofed 2026-07-27 as part of private OSS ship-ready prompts 365-370.

Executive Summary¶

Keprix core agent runtime is broadly aligned with Hermes Agent in the areas listed below. The current repo has a local parity gate and focused tests for architecture, compatibility, TUI, package entry points, and deterministic agent smoke behavior.

2026-07-27 gate evidence (this workspace):

• bash scripts/check-agent-parity.sh -> 10/10 passed
• bash scripts/check-tui-parity.sh -> TUI parity contracts 100/100 passed
• bash scripts/check-tui-surpass-hermes.sh -> surpass contracts passed
• bash scripts/check-private-ship-gate.sh -> private ship gate passed
• Frontend npx tsc --noEmit -> exit 0
• .venv/bin/python -m pytest tests/api -q -> 86 passed, 2 skipped

Focused auth, billing, architecture, TUI, agent, and API suites are green in this pass. Prefer scripts/check-private-ship-gate.sh before a private tag.

Keprix deliberately differs in surface UI/UX, product architecture, visual identity, and branding. These are intentional product decisions, not gaps.

Parity Areas Reviewed¶

Keprix-Better Areas¶

Keprix extends the Hermes baseline in these areas:

• Layered prompt assembly: Ordered prompt layers (identity, budget, safety, tools, tone, execution, memory continuity, domain, persona, product) with opt-out per layer.
• Agent loop modularization: 5,467-line monolith decomposed into conversation_loop, turn_context, turn_retry_state, tool_executor, errors, and prompt_builder modules.
• Memory edit gate: Blocks false confirmation when the memory tool wasn't actually called (Prompt 295).
• Connector router: Prefers MCP connectors over browser scraping for intent-matched queries (Prompt 296).
• Product hook system: Before-tool, after-tool, and after-turn hooks via registries.product_hooks with error isolation per hook.
• Product prompt layers: Product modules register prompt layers at runtime via register_product_prompt_layer() with enable/disable support.
• Memory continuity layer: Stable-tier memory context injected via the layered prompt system.
• Domain layers: Context-specific prompt injections for code, legal, medical, and property domains.
• Ponytail ladder: Seven-rung minimal-code generation enforced via agent/ladder.py and agent/ladder_mode.py.
• Persona engineering: 10 typed personas (NEXUS, FORGE, WARDEN, SAGE, BEACON, COMPASS, PRISM, EMBER, ECHO, CODEX) with routing guide enforcement.
• Provider normaliser: Multi-provider tool schema conversion (Anthropic, OpenAI, Google formats).
• Tool ACL per product: Tool access control lists scoped by product context, enforced via security/tool_acl.py.
• Network egress policy: Per-product network restrictions enforced via security/egress_policy.py.
• Resource quotas: Fairness scheduler for compute and token budgets.

Deliberate Differences¶

These are intentional, not gaps:

• TUI: Keprix uses Python Textual. Hermes uses TypeScript Ink.
• Gateway: Channel Shield owns gateway session management in Keprix. Hermes gateway is a monolith.
• Visual identity: Keprix has its own branding, navigation, color system, and surface UI. Hermes identity is preserved only in attribution and upstream references.
• Desktop app: Hermes ships a desktop Tauri app. Keprix has a separate app under apps/.
• Product architecture: Keprix has a core/product boundary enforced by import tests. Hermes has no such boundary.

Compatibility Notes¶

• State paths: .keprix is the primary state directory. .hermes is readable for migration compatibility. KEPRIX_* env vars preferred; HERMES_* accepted as fallback.
• Binary names: keprix is the entry point. Legacy hermes references in upstream tracking modules preserved for monitoring.
• Config: Old config at .hermes/config.yaml is readable. New writes go to .keprix/config.yaml.
• Checkpoints: Ref path migrated from refs/hermes/ to refs/keprix/.

Product Extensions Preserved¶

Keprix product modules are cleanly isolated from core agent runtime:

• Channel Shield: safe content, memory guard, channel config
• Agent OS: run ledger, glass panel, mutation hooks
• Scout: governance, telemetry, policy signals, evidence packs
• Billing: wallet enforcer, rate limiting, usage tracking
• Agent Apps: workspace runtime, built app shell
• Brain: graph viz, session replay, health dashboard
• Voice: Twilio handler, VAD pipeline, TTS client
• Playbooks: canvas compiler, variable context, NL builder
• Extensions: product isolation, lifecycle hooks, config merging

Keprix UI/UX Identity Preserved¶

Keprix has its own: - Navigation system with sidebar and top bar - Color theme and CSS aliases - Launcher with skill cards - Admin dashboard - Marketing site and docs - Onboarding flow - Authentication pages

Hermes visual identity (logos, branding, desktop app shells) is NOT copied. Upstream attribution is explicit in docs and license.

Test Evidence¶

bash scripts/check-agent-parity.sh
python -m pytest tests/architecture/test_core_product_boundaries.py \
  tests/config/test_hermes_compatibility.py \
  tests/config/test_doctor_hermes_compatibility.py \
  tests/cli/test_package_entrypoint.py \
  tests/tui tests/parity -q

# Focused verification observed in this audit cycle:
#   145 passed

TUI comparison and surpass¶

Keprix TUI targets Hermes behavior parity and then surpasses Hermes in Command Center, runtime timeline, tool cards, and operator proof harnesses. Visual identity stays Keprix (Textual theme, copy, layout). See:

• docs/architecture/tui-hermes-behavior-parity-contract.md
• docs/architecture/tui-surpass-hermes-contract.md
• docs/features/tui.md (parity matrix)

Remaining Gaps¶

No specific Hermes-better core behavior is currently identified as an untracked blocker in the inventory. Remaining risk is test depth and release confidence:

• The parity eval suite is deterministic smoke coverage and should be strengthened against real agent loop, tool dispatch, checkpoint, memory, and product hook code paths.
• Prefer scripts/check-private-ship-gate.sh for private soft-ship checks.
• Public GTM still needs legal entity fill, live domain proof, and a clean tagged release cut separate from a dirty working tree.