80835138e9
- README: orientation, doc map, target package layout (pointers only, no duplicated architecture content) - CLAUDE.md: agent operating manual — invariants as code-review rules, conventions, do-not list, task reading order - ROADMAP: M1-M5 with verifiable acceptance criteria, phase-2 fence - GLOSSARY: canonical Chinese-term → code-name mapping - docs/adr/: eight ADRs recording settled decisions and rejected alternatives - docs/open-questions.md: consolidated TODO(业务) tracker by owner and blocking milestone - .gitignore; untrack .DS_Store Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019u5SLNweVio6ozJX7yfxQr 🔮 View transcript: https://logs.lojong.info/s/e8u90k3t33w590r7b5y7yzqh
17 lines
794 B
Markdown
17 lines
794 B
Markdown
# ADR-0003: Provider-abstracted LLM
|
|
|
|
**Status**: accepted · 2026-09
|
|
|
|
**Context**: Production mandates 光明电力大模型 (on-prem, spec unknown at design
|
|
time); development can't wait for access.
|
|
|
|
**Decision**: All LLM calls go through a capability-oriented abstraction with a
|
|
model router: commercial API in dev, local open model in preprod, GuangMing
|
|
adapter in prod. Conservative assumptions until spec arrives: OpenAI-compatible,
|
|
8K context, no native tool calling (degrade to prompt+JSON+retry). Backend swap
|
|
is gated by the L1 eval suite plus L4 shadow comparison (docs/12 §3), not judgment.
|
|
|
|
**Consequences**: No vendor-specific features in agent code; prompts and tool
|
|
definitions maintained provider-portable; phase-1 acceptance must not depend on
|
|
GuangMing access (docs/13 §7).
|