v0.1.13 — The grounding repair, a finance-math layer, and institutional research
pip install -U vibe-trading-ai
408 commits and 162 merged pull requests since v0.1.12 (2026-07-22) — the largest release to date, from 35 contributors. Full detail in CHANGELOG.md.
🛡 The headline is a fix, not a feature
The single most user-visible defect in the 0.1.12 line: a well-formed question would spend minutes on real tool calls and then refuse to answer, returning
当前无法安全确认标的身份或价格证据，因此没有生成交易结论。请确认候选证券代码和交易所后再继续。
(cannot safely confirm instrument identity or price evidence)
The identity/grounding gate was rejecting answers it already had the evidence for. Seven independent root causes, each now fixed and covered by two-sided guard tests (healthy samples must pass; bad samples must still be blocked) plus mutation tests:
| # | Root cause | Effect |
|---|---|---|
| 1 | .SS and .SH treated as different instruments |
Every Shanghai ticker was permanently ambiguous |
| 2 | A+H dual listings and bare HK codes collapsed to ambiguous |
Dead end instead of a shortlist |
| 3 | Yahoo returns HTTP 400 for every CJK query; recorded as a source failure | Escalated to blocking invalidated instead of "not listed here" |
| 4 | A failed side query could demote an already-locked identity | Sticky aggregate status with no recovery |
| 5 | A hardcoded per-tool whitelist decided which bare tickers could match | 11 of the 17 documented argument spellings were blocked (3/14 → 14/14 after the fix) |
| 6 | Source and currency had to be written in ASCII | Chinese answers rejected for writing 雅虎/腾讯, or 元 instead of 人民币 |
| 7 | A thousands separator split the clause mid-number | ¥1,309.22 compared as 1 against the observed range → false price conflict |
Conceptual questions with no instrument at all, and comparison reports, no longer dead-end either.
A quote outside recorded OHLC evidence is still refused. The gate got more precise, not weaker — that is the point of the two-sided tests.
Verified end to end against a real LLM, real tools and real network: a 600519.SH query that previously took 6m14s and ended in the canned refusal now answers in 7–9s with zero rejections; 茅台 in 14s; a knowledge question in 14s.
This also closes the narrower cases from the 0.1.12 line — numbers that were never prices (confidence scores, indicator readings, moving-average windows, year-less dates like 8/5, percentage ranges, and a trading plan's own trigger levels, where close ≥ 6.45 is a condition rather than a quote) — and makes a many-candidate shortlist count as an answer rather than a stalled resolution (#1001, #983, #955).
🧮 src/quantlib — a tested finance-math layer
265 tested functions across 19 modules — every module the tool allowlists now exports. Skills now import these instead of carrying formulas inside markdown code blocks — if you find a pricing formula living in a SKILL.md, that is a bug, not a pattern.
| Module | Coverage |
|---|---|
options |
Black-Scholes price + greeks, implied-volatility inversion |
fixedincome |
Bond math, Nelson-Siegel / Svensson curve fitting |
credit |
Altman Z-score, Merton / KMV distance-to-default |
timeseries |
Stationarity, cointegration, GARCH, bootstrap |
risk · var_backtest |
VaR / CVaR / EVT and their backtests |
attribution |
Brinson-Fachler decomposition |
performance · fundmath |
TWR / MWR / Modified Dietz; XIRR / MOIC / DPI / TVPI |
factormodel · eventstudy |
Factor regressions, event studies |
multipletesting · crossvalidation |
Deflated significance, purged CV |
impact |
Market-impact models |
The read-only quantlib_call tool reaches all of it through one contract, so the finance math works on the CLI, the Web UI, the REST API and MCP — including deployments where bash is gated off. It is structurally not a shell — module allowlist, __all__-only dispatch, export_* refused. Econometrics needs the stats extra; those functions lazy-import and name the missing one.
💰 Valuation engine
A valuation engine that refuses to invent its own inputs. The one rule in contracts.py: a missing input makes a model NOT RUNNABLE and is never silently defaulted — every default in a valuation model is an opinion wearing a constant's clothes.
run_dcf— FCFF bridge, WACC build, mid-year discounting, net-debt bridge, WACC×g sensitivity grid. Dual terminal value: each method is cross-checked against the other's implied multiple and implied g.run_comps— EV bridge, LTM + calendar-year calendarisation, multiple matrix. A peer with a non-positive denominator is excluded and reported, never averaged in as a negative multiple.threestatement— linked projection with a hard balance assertion, an explicit revolver plug, and an iterated interest↔debt circularity that must converge or raise.
Artifacts are input-hashed and versioned, with xlsx / pptx export.
🏛 Institutional research
Six slash commands — /comps /dcf /attrib /memo /earnings /screen — each carrying a step skeleton and an arithmetic-consistent worked example: the Brinson decomposition sums exactly to active return, and the earnings bridge sums exactly to the EPS delta.
Investor lenses become a standalone skill: named-investor reasoning frameworks as stackable analysis overlays, decoupled from the data layer. Each lens is an operating procedure — priority signals, disqualifying conditions, typical misuse — not a biography, and names no tool.
Five ready-to-schedule research playbooks (premarket brief, earnings-season tracker, portfolio checkup, A-share money flow, institutional-holdings diff), reachable three ways: auth-gated REST routes, a vibe-trading playbook CLI subcommand, and a /playbook slash command. Templates state their data needs in natural language rather than naming tools, so coverage can grow without editing them, and every one mandates naming a missing input instead of filling it from memory.
📊 Four new read-only data tools — all on free public sources
get_institutional_holdings— SEC 13F-HR books in manager / ticker-holders / top-managers modes, with quarter-over-quarter position diffs for factor use. Cover-page totals carryvalue_units+ basis, because pre-2023 filings report in thousands.etf_holdings— cross-market look-through. SEC N-PORT for US; for A-shares the semi-annual/annual reports carry the full book, not the quarterly top ten:510300returns 342 rows covering 98.66% of net assets versus 10 rows / 22.74%.coverageseparatesfull_portfoliofromtop_n_disclosed, and every response stamps the report period — a full book is always the older disclosure.prediction_market— event-contract search / event / market / history, with prices converted to implied probability and the unit labelled, so downstream never reads 0.63 as dollars. Read-only by construction: no order path.research_papers— arXiv + OpenAlex search/read with source-anchored claim extraction. Anything not literally present in the source is left empty and markednot stated in source, and a paper's claimed performance is never presented as our backtest result.
🧾 Governance wired into every run
- A run manifest hashes the prompt, the skill contents, the tool registry and the package versions — so "what methodology produced that number?" is answerable a month later.
- The audit ledger chains each record to its predecessor's hash and fsyncs. Editing or deleting a record is detectable, and an edit that recomputes its own hash is still caught one record later via
prev_hash_mismatch. Timestamps are always caller-supplied; no module here callsdatetime.now(). - Trace redaction is sink-aware:
contentis released only in the tool-RESULT sink and stays redacted in the fail-closed ARGUMENTS sink used by tool-call arguments and the live audit ledger.envis never released. Result strings are pattern-scrubbed, since shell output arrives as a JSON envelope. - All 30 swarm presets were re-audited — a deliverable no granted tool can compute is now declared as such instead of invented.
🧱 Platform
- Desktop shell — a source-first Electron host owning the backend lifecycle: random loopback port, per-launch secret, five-locale startup recovery, owned-process cleanup (#923). Windows packaging assembles a checksum-pinned embedded Python 3.12 runtime with x64 NSIS review/signing paths, plus Electron
safeStoragefor an allowlisted credential set — the renderer can set or clear secrets but never read them, plaintext config migrates once, and both unsigned-review and signed builds fail closed on the wrong signature state (#1015). No installer artifact was published from that PR. - eToro joins as the 13th broker connector with path-separated demo/real profiles; demo keys structurally reach only
/demopaths (#989). - Korea (KRX: KOSPI/KOSDAQ) becomes the 9th backtest engine — execution-time ±30% band on the unified tick grid, structurally long-only, config-driven 2026 0.20% securities transaction tax (#693).
- Canadian equities end to end —
.TO/.Vclassified in CAD, Yahoo → yfinance → local, Canada-specific GlobalEquity rules,XIC.TObenchmark, mixed-currency aggregation refused (#1024, #1019, #1037, closes #952). - OpenBB Workspace bridge (#817) and a read-only Taiwan snapshot tool (#848).
src/entities— a typed entity + irregular dated cash-flow substrate for NAVs, capital calls and coupons, deliberately parallel to the bar engines so anavcolumn can never reach one and get priced as a close. Surfaced bycashflow_performance.orderbook_depth— crypto L2 ladder via ccxt: spread bps, depth imbalance, and impact cost of a stated notional walked through the real book both ways.- ModelScope joins the built-in providers (#1011);
vibe-trading updateself-upgrades and distinguishes wheel installs from editable checkouts (#1020); Settings gains live model discovery and records the immutable provider/model/reasoning identity that actually served each reply (#924). - The MCP surface grows to 70 tools.
alpha_zoo+ boundedalpha_bench(#979) and QVeris discovery/inspect/execute (#976, cost quote read from the marketplace rather than trusted from the caller) join, and so do six read-only analytics tools that had reached the agent but never MCP —quantlib_call,cashflow_performance,orderbook_depth,sentiment,technical_indicatorsandget_fundamentals. Order-placing tools remain structurally un-exposed: the mirrored registration path refuses any class whoseis_readonlyis notTrue. - Memory Tier 2 — hierarchical routing, BM25 semantic linking, TF-IDF compression and an FTS5 index with CJK bigram tokenization, behind a one-line
VT_MEMORY=off|on|fullpreset. All off by default (#815, #733). - New read-only tools:
sentiment(#939),technical_indicators(#921), analyticoptions_payoff(#946).
🔒 Sandbox
A closed gap: generated strategy code could import the broker layer, and could reach socket / subprocess / os.system / ctypes through a renamed binding. Both were accepted before; both are now refused. src.quantlib still imports.
✅ Correctness — the expensive ones
- SEC reporting periods are keyed on their
(start, end)span. A 10-Q files the true quarter and the year-to-date frame under the same end date and fiscal period, soperiod="annual"had been returning a single quarter for AAPL FY2018–2020 — a 4.2× understatement — and every fiscal-Q4 slot in a quarterly series carried the full-year figure. - Tushare A-share prices are corporate-action adjusted in both the factor bench and backtests. A raw close-to-close return across an ex-date was off by up to 47 percentage points (300750.SZ, 2023-04-26). The CSI300 bench also masks each date to its point-in-time index membership.
bar_returnsno longer erases the move across a trading halt longer than the forward-fill window — the resumption move was silently recorded as 0, understating volatility and inflating Sharpe.- Annualisation now covers all 24 data sources at every interval, with a coverage test that fails CI when a loader lands without entries.
- Cross-market composite backtests refuse a mixed-currency code set instead of summing CNY, USD and KRW into one equity curve.
- Option legs are marked at the volatility they were opened at, removing a fabricated day-zero P&L of up to +93% of premium.
- A 19-PR interval-normalization sweep — lowercase
1h/4h/1d/1waccepted everywhere, unsupported intervals fail fast instead of silently returning daily bars. - Two quantlib modules were allowlisted but unreachable.
attribution(Brinson-Fachler) andimpact(market-impact models) carried no__all__, andquantlib_calldispatches on__all__alone — so the tool listed zero functions for both while the package docstring advertised them. Both now export, and a guard test fails when any allowlisted module exposes nothing; the existing test only asserted that each module imports, which could never catch this. - The published MCP manifest under-reported the server.
SKILL.md's tool count was derived by counting@mcp.tooldecorators, which ignores every tool registered through the mirrored path — so four institutional-research tools were live over MCP and absent from the manifest, and the contract test asserted that absence was correct. Both tests now measuremcp.list_tools(). - Resource leaks closed: HTTP throttle sweeps stale buckets interval-aware, the rate limiter no longer grows unboundedly with unique client IPs, the event bus notifies and removes subscribers on clear, and
_json_loadsis guarded against corrupted JSON in database columns.
🙏 Contributors
@santhreal · @shadowinlife · @Robin1987China · @he-yufeng · @QCYTSN · @Shizoqua · @honginp · @cgycorey · @wiliao · @ngoanpv · @x-lambda · @ofeksh-tr · @00EVA · @zwrong · @yrk111222 · @su322 · @hhj123123 · @dineeshd · @sambazhu · @ddy4633 · @tyj147454413-cmd · @y85998607 · @JungHoonGhae · @shugaoye · @TSENGCHIENFENG · @darkknight4563 · @MuggleJinx · @klmtseng · @ebujinovch · @g0rdonL · @AmirF194 · @Echoandelementwebsites · @yagnikpipaliya · @dvirarad · @1anter
Full changelog: v0.1.12...v0.1.13
