What

SDK-authoritative probe of 28 Tushare endpoints, persisted per project conventions: a reviewable audit doc, a token-free machine-readable capability registry, and a gitignored raw-probe artifact. No new factor, no backtest-logic change, no token in the repo.

MCP single-shot probing only proves permission, not batch viability (rate limit), so the Python SDK result is the source of truth.

Findings (2026-06-13)

• Core P0–P4-2 endpoints (daily, adj_factor, index_weight, suspend_d, namechange, stock_basic, stk_limit, daily_basic, fina_indicator, index_member_all, income): all authorized + proven batch-viable in real runs.
• EXPLORATORY candidates (stk_factor, stk_factor_pro, cyq_perf, cyq_chips, moneyflow, moneyflow_dc/ths/hsgt, limit_list_d, top_list, block_trade, margin_detail, stk_surv): authorized, but per-batch rate limit NOT stress-tested.
• stk_mins: first probe hit 频率超限(1次/小时); a same-day re-probe was 3/3 in ~12s — the 1/hour cap is lifted, so minute data is now feasible at the data layer (exact per-minute ceiling still unstress-tested). Registry/doc updated accordingly.
• hm_detail: authorized but throttled to 1 call/hour (not re-checked this round) → callable but not batch-viable.
• stk_auction_o / stk_auction_c: 40203 no_permission (call-auction family needs a separate tier).

Files

• docs/data/tushare_permissions.md — audit table: permission vs rate-limit vs batch viability vs PIT timing, plus explicit post-close look-ahead warnings for cyq_/moneyflow/limit_list_d candidates.
• data/feed/tushare_capabilities.py — token-free, importable capability registry (metadata + usage intent + dated observed_status); explicitly NOT a runtime permission gate.
• artifacts/permissions/tushare_probe_20260613.json — raw probe output (gitignored, token-free, errors scrubbed).

PIT discipline

Candidate sources cyq_/moneyflow/limit_list_d/top_list/hm_detail/block_trade/margin_detail are post-close data — T-day values must never feed T-day close selection (look-ahead); usable only at T+1. Any candidate entering the P3 matrix still follows the EXPLORATORY + independent-sample discipline.

Test plan

• py_compile data/feed/tushare_capabilities.py OK; module imports, 28/28 endpoints, no token.
• pytest tests -q — all pass (428); docs/registry are additive, no code/test behavior change.
• ruff check . clean.
• token-leak scan across tracked files + artifact = 0; artifacts/permissions/ confirmed gitignored.
• Probe used the SDK + external token file with minimal single queries; rate-sensitive endpoints not brute-forced (stk_mins re-probe limited to 3 small calls).