Skip to content

Roadmap: Hermes/Animus Integration & Mobile Builder Surface

Status: Canonical roadmap · Created 2026-06-14 · Owner: ARETE
Methodology: METHOD_AGENT_WORKFLOW.md — 5-phase multi-agent build loop.

Strategic Goal

Turn the Animus kernel from an extractable engine into a daily-use autonomous builder that runs on local GPU, is controlled from a phone, and costs $0 per month.

This means: 1. Hermes model integration — tool-calling prompts, role-based routing, schema validation. 2. Terminal builder — iterative read-plan-edit-test-retry loop for actual code changes. 3. Web UI — minimal chat interface accessible from phone browser on local network. 4. Discord — status checks and approval gates without opening a terminal. 5. Ollama-first — works with zero cloud API subscriptions.

Milestones

Milestone Priority Status Depends On Deliverable
M1: Hermes Model Integration 1 Spec ready — Hermes prompts, role router, schema validator
M2: TerminalAgent 2 Spec ready M1 Iterative builder loop (read→plan→edit→test→retry)
M3: Minimal Web UI 3 Spec ready M1, M2 FastAPI endpoint + vanilla-JS chat page
M4: Discord Bot 4 Spec ready M1 /build slash commands for status/approve/queue
M5: Ollama-Default Wiring 5 Spec ready M1 Auto-detect offline, default to local models

Dependency Graph

M1 (Hermes) ──┬── M2 (TerminalAgent) ── M3 (Web UI)
              ├── M4 (Discord)
              └── M5 (Ollama-Default)

M1 unlocks everything. Sequential within a milestone, parallel across milestones where dependency-free.

Task Inventory

Task Name Milestone ET Dependencies
TASK-001 Hermes Prompt Templates M1 800 —
TASK-002 ProviderRouter Role-Based Tuning M1 600 TASK-001
TASK-003 Tool-Call Schema Validation M1 700 TASK-002
TASK-004 CommandRunner M2 600 —
TASK-005 TerminalAgent Loop M2 1,200 TASK-003, TASK-004
TASK-006 FastAPI Chat Endpoint M3 800 TASK-005
TASK-007 Mobile HTML UI M3 600 TASK-006
TASK-008 Discord Slash Commands M4 900 TASK-002
TASK-009 Ollama Default Detection M5 500 TASK-002
TASK-010 Kernel Integration Tests ALL 1,500 TASK-001–009

Total Estimated Budget: ~8,200 ET (well within the 15,000 ET pool reserved for this roadmap).

Execution Phases

Phase A: SPECIFICATION (Complete ✓)

  • Status: Done 2026-06-14
  • Deliverables:
  • docs/METHOD_AGENT_WORKFLOW.md (reusable methodology)
  • docs/ROADMAP_HERMES_2026-06.md (this file)
  • TASK_SPECS/001_hermes_prompts.md through 010_integration_tests.md

Phase B: DELEGATE (Pending)

  • Goal: Spawn Builder agents for each task in dependency order.
  • Order:
  • Wave 1 (no deps): TASK-001, TASK-004
  • Wave 2 (deps satisfied): TASK-002, TASK-003, TASK-009
  • Wave 3: TASK-005, TASK-008
  • Wave 4: TASK-006, TASK-007
  • Wave 5: TASK-010
  • Agents: 10× Builder, operating in parallel within waves.
  • Budget: ~8,200 ET.

Phase C: EVALUATE (Pending)

  • Goal: Score each output against its rubric.
  • Agents: 10× Reviewer.
  • Budget: ~1,640 ET (20% of delegate).

Phase D: REPAIR (Pending)

  • Goal: Re-run rejected tasks.
  • Expected rejections: 30% (3 tasks).
  • Budget: ~1,640 ET.

Phase E: INTEGRATE (Pending)

  • Goal: Merge, run tests, commit.
  • Budget: ~410 ET.

Risks & Mitigations

Risk Likelihood Impact Mitigation
Hermes prompts don't match model expectations Medium High Include dual-format (JSON) fallback; test early with actual model
CommandRunner false positives on safe commands Low Medium Whitelist tests; safety patterns reviewed by human
Discord bot rate limits Medium Low Batch embeds; exponential backoff
Ollama model not loaded when UI starts High Medium UI shows "loading model..." with ollama pull hint
Browser CORS issues in local network Low Low Allow * origin in dev mode; tighten in prod
Web UI > 50KB with budget bar + queue Low Low Use inline SVG for icons, no external libs

Open Decisions (Need Human Input)

  1. M3 vs M4 priority: Should Web UI ship before Discord, or vice versa? Discord is lower-friction for approvals. Web UI is richer for exploration.
  2. TerminalAgent iteration limit: Currently capped at 10. Too restrictive for complex refactors?
  3. Hermes vs Qwen for Builder: The spec assumes Hermes for Builder. If Hermes underperforms on your GPU, should we dynamically fall back to Qwen?
  4. Auto-commit: TerminalAgent rolls back on failure but does not auto-commit on success. Should successful builds auto-commit or queue for human approval?

Success Criteria (End of Roadmap)

  • All 10 task specs marked COMPLETE in TASK_SPECS/.
  • pytest packages/kernel/tests/ passes.
  • python packages/kernel/scripts/verify_imports.py passes.
  • curl http://localhost:8000/chat streams a response from local Ollama.
  • /build status in Discord shows active builds and budget.
  • Zero cloud API keys required for daily use.

References

  • [[Methodology]] docs/METHOD_AGENT_WORKFLOW.md
  • [[Task Specs]] TASK_SPECS/*.md
  • [[Kernel Map]] packages/kernel/KERNEL_FILE_MAP.md
  • [[Interface Vision]] docs/INTERFACE_BOOTSTRAP_VISION.md
  • [[Engine Assessment]] docs/ENGINE_VS_SHELL_ASSESSMENT.md

Part of the Animus system. Built with the multi-agent workflow methodology.