vpp-ai-platform/CLAUDE.md

# CLAUDE.md — agent operating manual

Design docs live in `docs/` (00–13, Chinese). They are the specification; this
file distills what is **binding** when writing code. When code and docs conflict,
docs win — or write an ADR changing the doc first.

## Current phase

Docs complete; M1–M5 implemented: schemas + contracts pipeline; deterministic
services (ledger, stores, policy engine incl. dispatch rules, envelope with
deviation-streak suspension, authority/permits, file-export + shadow + simulation
gateways, events, lineage assemblers, revenue + power-balance simulation,
review service, Case Desk, kill-switch hierarchy L0–L4 with loss breaker and
abnormal-day protocol, shadow clearing/scoring, KPI computation); Python skill
service (forecasts, bid MILP, report, potential assessment, dispatch
optimization); L2 eval harness with committed baseline (synthetic data); Mastra
runtime with the proposal-lifecycle safety chain (durable suspend/resume),
situation / bid / award-decomposition / review / envelope-review / shadow-close
workflows, trigger service (incl. shadow schedule and live-data ingestion), LLM
port with LLM-down mode, Case Desk API with insight cards, KPI dashboard and
breaker endpoints, shadow replay CLI (`npm run shadow -w @vpp/runtime`).
Storage is file-backed reference semantics (ledger, time series, counters and
repositories all persist under the data dir) — Postgres/Timescale adapters
later. Phase 1 is feature-complete; what remains is running the shadow period on
real Hubei data (`ROADMAP.md` M5 acceptance: 20+ consecutive days) and the
governance steps that need business answers (`docs/open-questions.md`). Build
order is ROADMAP.md (M1→M5). A skill change that moves an L2 metric
must update the baseline in the same change (`npm run eval -w @vpp/evals -- --write-baseline`).
Workflows are per-runtime factories (`createProposalLifecycle(ctx)` etc.) — never
module singletons, or a second Mastra instance steals their context.
Do not build phase-2 items (edge control links, federation,
interaction/load-control agents, programmatic bid submission, acting on
envelope-widening recommendations without the docs/03 envelope-review approval)
unless explicitly asked.

## Hard rules (from docs/01 invariants — treat as review criteria)

1. **LLM never computes numbers.** Any numeric field in a `Proposal` payload must
   be a `{toolCallId, path}` reference into recorded tool-call lineage,
   dereferenced by the assembler. If you find yourself parsing a number out of
   LLM text into a payload, stop — that's the architecture's one forbidden move.
2. **No LLM in any control path.** Nothing under `services/` or the execution
   path may import or await an LLM call. Periodic workflows must run with the
   LLM backend down (invariant I6 — there is a test for this).
3. **External effects only via `(Proposal, ExecutionPermit)`.** Gateways/adapters
   never accept a bare plan. Permits are short-lived and revocable.
4. **AI cannot approve.** No code path may transition a Proposal to
   APPROVED/AUTHORIZED without either a matched envelope + passing checks, or a
   human `resume()` with an approver identity distinct from the origination chain.
5. **Everything auditable.** State transitions, tool calls, approvals, permits,
   and gateway receipts are event-sourced. Snapshots referenced by lineage are
   immutable.
6. **Schemas live in `packages/domain` only.** Python models are generated from
   `contracts/*.schema.json` — never hand-edit generated files, never define a
   business object schema anywhere else.

## Conventions

- **Naming**: use the canonical code names in GLOSSARY.md. Do not invent new
  English names for domain terms that already have one.
- **Language**: architecture docs Chinese; code, identifiers, comments, commit
  messages, ADRs English.
- **Data representation** (docs/11 §3.3): money/energy/prices as fixed-point
  decimal **strings**; units in field names (`power_mw`, `price_yuan_per_mwh`);
  timestamps ISO8601 UTC; market intervals as `{date, interval_index}` with
  `interval_minutes` explicit; all IDs strings; enums UPPER_SNAKE string literals.
- **Mastra**: do not trust memorized APIs — check `node_modules/@mastra/*/dist/docs/`
  (or the mastra skill) against the installed version before writing framework code.
- **Testing floor**: every invariant above has at least one automated test;
  schema changes run golden-fixture validation on both TS and Python sides;
  policy pack changes require their rule tests green.

## Do not

- Do not invent business parameters (market deadlines, envelope bounds, loss
  budgets, buffer coefficients). Check `docs/open-questions.md`; if a needed
  parameter is listed there, wire it as named config with a placeholder value
  and a `// OPEN-QUESTION:` comment, and say so in your summary.
- Do not relitigate decisions recorded in `docs/adr/`. If a decision must
  change, propose a superseding ADR.
- Do not create root-level summary docs that duplicate `docs/` content
  (no architecture.md / tech-stack.md — README points to the sources of truth).
- Do not add a sixth agent, merge the safety chain into business workflows, or
  bypass the proposal lifecycle for "internal" effects — these were considered
  and rejected (see ADRs 0005–0007).

## Reading order for common tasks

| Task | Read first |
|---|---|
| Domain schemas / contracts | docs/11, docs/00 §4, GLOSSARY.md |
| Runtime, workflows, agents | docs/09, docs/02 |
| Safety chain, approvals, permits | docs/03, docs/01 invariants |
| Skill services (Python) | docs/05 §1–2, docs/11 §3 |
| Anything touching money or bids | docs/07 (scenario), docs/13 §1 |
| Eval harness | docs/12 |