Gormes

Phase 4 — The Brain Transplant

Phase 4 — The Brain Transplant (Powertrain)

Status: ⏳ planned

Deliverable: Native Go agent orchestrator + prompt builder.

Phase 4 is when Hermes becomes optional. Each sub-phase is a separable spec.

Phase 4 Sub-phase Outline

SubphaseStatusDeliverable
4.A — Provider Adapters⏳ plannedNative Go adapters for Anthropic, Bedrock, Gemini, OpenRouter, Google Code Assist, Codex (mirrors agent/{anthropic,bedrock,gemini_cloudcode,openrouter_client,google_code_assist}_adapter.py)
4.B — Context Engine + Compression⏳ plannedPort agent/{context_engine,context_compressor,context_references}.py; manage long sessions without blowing the model context window
4.C — Native Prompt Builder⏳ plannedPort agent/prompt_builder.py; assemble system + memory + tool + history into a model-ready prompt
4.D — Smart Model Routing⏳ plannedPort agent/smart_model_routing.py + agent/model_metadata.py + agent/models_dev.py; pick the right model per turn
4.E — Trajectory + Insights⏳ plannedPort agent/trajectory.py + agent/insights.py; self-monitoring telemetry surface
4.F — Title Generation⏳ plannedPort agent/title_generator.py; auto-name new sessions
4.G — Credentials + OAuth⏳ plannedPort agent/google_oauth.py, agent/credential_pool.py, tools/credential_files.py; token vault + multi-account auth
4.H — Rate / Retry / Caching⏳ plannedPort agent/{rate_limit_tracker,retry_utils,nous_rate_guard,prompt_caching}.py; provider-side resilience

Once 4.A–4.D are shipped Gormes can call LLMs directly. The :8642 health check becomes optional.

Build Priority Context

Phase 4 is optimization, not differentiation. The Python bridge works. Replace it only after Skills (Phase 6 / Phase 2.G) and Subagents (Phase 2.E) prove the architecture is correct. Phase 4 is the last step in the priority sequence:

Skills System → Subagent System → Multi-Platform Gateway → Native Agent Loop

The rule: Build Skills before Subagents, Subagents before wider Gateway, Gateway before Phase 4. Each layer validates the architecture for the next.