Consolidate plan documents: close vision-implementation gap

[Coldaine]

Problem

There are 13 plan documents in docs/plans/ totaling thousands of lines of spec. Most describe systems that don't exist yet and may never be built as specified. The plans have outpaced the foundations:

Plan Document	Lines	Implementation Status
recovery_agent_architecture.md	1062	Zero lines implemented
durable_agent_architecture_design.md	~500	Zero lines implemented
recovery_agent_implementation_plan.md	~300	Zero lines implemented
llm_driven_gameplay_architecture.md	~300	Zero lines implemented
tool_node_design_review.md	251	Zero lines implemented
durable_agent_system_plan.md	~200	Zero lines implemented
interactive_state_viz_plan.md	~200	Zero lines implemented
local_vlm_setup.md	~200	Zero lines implemented
scheduler_status_jsonl_plan.md	~100	Partially done (JSONL exists)
task_tag_logging_plan.md	~100	Not implemented
operationalize_llm_piloting.md	~100	Partially done
phase_0_login_tool_spec.md	~100	Done (login tool exists)
phase_2_gemini_cli_plan.md	~100	Not implemented

Meanwhile, the actual system has 7 MCP tools and can't tell if ADB is connected.

The Gap

The plans describe:

• SQLite/Redis checkpoint storage backends
• LangGraph durable execution with sub-graphs
• Multi-agent supervisor/follower patterns
• Vision diagnosis agents with Gemini Flash
• Circuit breaker patterns
• Chaos engineering test suites

The system needs:

• An adb_health tool
• Non-blocking MCP calls
• CLI debugging mode

Proposed Actions

1. Archive pure-spec plans

Move plans with zero implementation to docs/archive/plans/. They're not wrong, but they're not actionable right now:

• recovery_agent_architecture.md
• durable_agent_architecture_design.md
• recovery_agent_implementation_plan.md
• durable_agent_system_plan.md
• phase_2_gemini_cli_plan.md

2. Keep actionable plans

Plans that describe near-term work stay:

• operationalize_llm_piloting.md (active work)
• local_vlm_setup.md (March 2026 milestone)
• interactive_state_viz_plan.md (March 2026 milestone)
• task_tag_logging_plan.md (Phase 1 observability)

3. Close completed plans

• phase_0_login_tool_spec.md — login tool exists, mark done
• scheduler_status_jsonl_plan.md — JSONL exists, mark done

4. Update ROADMAP.md

Align milestones with reality. Current March 2026 milestones are:

• Local VLM serving (L1-L2) — not started
• Interactive state viz V1 — not started
• Dashboard/state tools — not started

These should be re-sequenced behind the MCP v2 foundation (Issue #35).

5. Close Issue #29

The replay harness (Issue #29) has been implemented via PRs #31 and #32. The issue should be closed.

Related

• Issue MCP server v2: non-blocking primitives, health probe, CLI mode #35: MCP server v2 (the actual foundation work)
• Issue Reconcile llm_pilot.py with MCP-only control mandate #36: llm_pilot.py reconciliation
• Issue arch: Tool Node complexity and feedback loop design #21: Tool Node design (valuable but depends on MCP server v2: non-blocking primitives, health probe, CLI mode #35)
• Issue Feature: Deterministic Replay Testing Harness #29: Replay harness (should be closed — done)