Epic: Runtime Migration — LangChain Removal + SOTA Patterns (8 weeks)

[meiiie]

Goal

Remove langchain* dependencies entirely from Wiii runtime. Adopt SOTA patterns from Anthropic Claude Code + OpenAI Codex internal architecture + Unsloth lane-first design. Zero new framework deps (no Pydantic AI, no OpenAI Agents SDK, no Strands).

Aligned with Anthropic Engineering: "avoid frameworks that obscure prompts. Start simple. Look to human engineers for inspiration."

Decisions (committed)

• D1: Xóa triệt để langchain — including langchain-core (39 messages + 30 tools + 12 model wrappers imports = ~150 LOC custom replacement)
• D2: Stack final = pure raw SDK (anthropic, openai, google-generativeai, mcp) + DIY messages/tools/registry + existing UnifiedLLMClient
• D3: Execution model = Hybrid (architect Claude Code session + parallel worker sessions per phase + user as approval gate)
• D4: 8 weeks, 7 phases, eval harness included (Phase 6 — THE moat per Anthropic)

Phase checklist

• Week 0 — Stabilize: PR fix(runtime): fast-path host UI routing #205 merged (8d468dac) + PR fix(runtime): bound sync supervisor structured-route latency #208 merged (0a42c060, closes runtime: bound sync conversation supervisor latency #206)
• Phase 0 (Week 1) — Drift fix + scaffold — PR chore(runtime): Phase 0 — drop drift langchain deps + scaffold runtime/ #209 merged (48a6000b). Removed 4 langchain drift deps + scaffold app/engine/runtime/ + 2 settings fields.
• Phase 1 (Week 2) — Messages layer DIY — 39 imports, RỦI RO CAO NHẤT — kickoff prompt ready: .Codex/reports/runtime-migration-briefs/phase-1-kickoff-prompt.md
• Phase 2 (Week 3) — Tool registry DIY — 30 imports
• Phase 3 (Week 4) — Provider wrappers — flip enable_unified_providers=True
• Phase 4 (Week 5) — Lane-first architecture (Unsloth pattern) — Pointy V1 official test
• Phase 5 (Week 6) — Harness/Session/Sandbox split (Anthropic pattern) — p50 TTFT giảm ≥30%
• Phase 6 (Week 7) — Eval harness as production tooling — THE moat
• Phase 7 (Week 8) — Cleanup + canary + Pointy LMS GA

Tracking issues

• Test drift: 4 pre-existing unit test failures on main (config validators + embedding fingerprint + provider default) #210 — 4 pre-existing local-env test failures (not blocking, separate cleanup)

Source documents (local, gitignored — paste content into worker session prompts)

• Master plan: .Codex/reports/RUNTIME-MIGRATION-EPIC-2026-05-02.md
• Per-phase briefs: .Codex/reports/runtime-migration-briefs/phase-{0..7}-*.md
• Phase 1 kickoff prompt (paste into worker session): .Codex/reports/runtime-migration-briefs/phase-1-kickoff-prompt.md

Verification gates per phase (testing strategy update 2026-05-02)

• Targeted per commit: pytest tests/unit/{relevant_module}/ tests/unit/test_feature_tiers.py tests/unit/test_*config*.py -q (under 30s)
• Collection smoke before push: pytest tests/unit/ --collect-only -q (catches import errors)
• CI = full unit gate (10-15 min, server-side)

Owner / Coordination

• Architect = primary Claude Code session
• Workers = parallel Claude Code sessions, one per phase. Sonnet 4.6 for mechanical phases (0/1/2/3/7); Opus 4.7 for architectural phases (4/5/6).
• Reviewer/Merger = repo maintainer

When a worker session ships a PR, tag the architect session for review.

🤖 Generated with Claude Code