tw-market-data

What YOUR agents have actually done, from the durable audit trail. Every resident agent records what it did and ON WHAT BASIS — the rule and the two closes behind an alert, the coverage and limits behind a risk finding, the run_id and declared trial count behind a factor verdict. This is that trail, and it survives deploys. Reports which agents have recorded NOTHING (`coverage.missing`), because "the monitor has been quiet" and "the monitor is not running" look identical from the records alone and only one of them means your alerts work. Args: limit: how many records to return, newest first (max 200).

Record a HUMAN's approval of a proposed action. This writes an audit record naming who approved what, and when. It does NOT execute the action: TWMD has no order or funds path, by design. Execution, if any, happens elsewhere and is performed by a person. Args: action_id: from `list_pending_actions` or a report's `proposed_actions`. approver: who is approving. Required — an unattributed approval is not an approval.

Answer a plain-language question in Taiwanese-market vocabulary, sentence by sourced sentence. ★ FOR BEGINNERS WHO DO NOT KNOW WHICH DATASET THEY WANT. Ask "PBR 是什麼" or "什麼叫漲跌停" in ordinary words; routing happens on our side. `describe_dataset` explains a table you already named, `search_filings` digs through company disclosures, and `query_dataset` returns rows — this one turns a beginner's wording into a sourced explanation instead. ★ EVERY SENTENCE CARRIES A CITATION OR IS NOT RETURNED. Vocabulary sentences cite `glossary:<id>`. Figures cite the dataset row they came from. A sentence whose number cannot be traced to a retrieved row is DROPPED — it appears in `dropped[]` with status `unverifiable`, and never in `claims[]`. There is no path by which this tool composes a number from its own memory. ★ TERMS OUTSIDE THE GLOSSARY RETURN `unsupported_term`, NOT A GUESS. The corpus is 99 curated Taiwan-market terms. CAPM, options greeks and general finance vocabulary are not in it, and the honest answer is that we do not cover them. ⚠️ Explanations only. It states what a term means and what a figure was; it does not tell you what to do about either. Args: question, optional as_of (YYYY-MM-DD).

Sort corporate dates into what is still ahead and what has already passed. ★ TWO DATES, NOT ONE. What is "upcoming" is decided by the date the event HAPPENS; `as_of` filters on the date it was ANNOUNCED. An ex-dividend declared on 2026-08-01 for 2026-09-15 is both already known and still ahead on 2026-08-10. Collapsing the two fields either hides every future date or reports last month's ex-dividend as though it were coming. ★ ELAPSED DATES ARE SEPARATED, NOT DISCARDED. They come back under `past` — the previous ex-dividend is useful context for a question about the next one — but they can never appear under `upcoming`. ★ NOT `search_filings` AND NOT `query_dataset`. Those retrieve disclosures and rows; this one only arranges dated corporate events on a timeline relative to now. ⚠️ A scheduled date is a schedule, not a promise; companies move them. Args: rows (dated events), today (YYYY-MM-DD), optional as_of.

Turn rows you already fetched into a Vega-Lite drawing your chat window can render. ★ IT DRAWS; IT DOES NOT FETCH. Hand it the output of `query_dataset` — this tool never touches the database, so it cannot bypass the `as_of` filter those rows were selected under. A plotting tool that fetched its own numbers would be a second data path, and a second path eventually disagrees with the first about what was knowable when. ★ GAPS BREAK THE LINE INSTEAD OF BEING BRIDGED. A missing value is emitted as null, so the rendered line stops rather than sloping smoothly across a period where nothing was published. The absent positions are also listed in `data_gaps`, because a break is easy to misread as a flat stretch. Zero is never substituted — zero draws a real low point for something that never happened. ★ THE DRAWING POINTS BACK AT THE PROOF. `citation` and `inclusion_pointer` travel with the spec, so a picture and the rows behind it name the same published checkpoint. ⚠️ Rendering, not analysis. A shape you notice in a picture is not a forecast. Args: rows, x_field, y_field, optional title and mark (line/point/bar/area).

Produce a bibliographic citation for TWMD data — APA, BibTeX, and a re-verifiable token. ★ FOR PAPERS, REGULATORY FILINGS AND ANYTHING A REVIEWER WILL RE-CHECK LATER. A dataset is corrected, backfilled and re-run. Three years from now a reviewer opening our API sees different numbers than the paper, and nobody — author, reviewer, or us — can tell whether the data changed or the author mis-transcribed. So the citation carries `as_of`, the Merkle `checkpoint_root`, and a `verify_url`, plus a signed `reproducible_token` that binds those fields TO EACH OTHER. ★ PASS `as_of` IF YOUR WORK IS POINT-IN-TIME. Without it a citation is still produced, but it is marked `point_in_time: false` and the APA line reads "Retrieved <date>" instead of naming a knowledge horizon — because re-running the same query later can legitimately return different figures, and nothing in the data would mark the difference. ★ WHAT IT PROVES: which data was used, and that those fields were signed by us. **NOT that the figures are correct** — a faithfully committed wrong figure cites and verifies exactly like a right one. `limitations` says so in the response; keep it when you quote the citation. Args: dataset: the dataset id you queried, e.g. 'valuation'. as_of: the knowledge cutoff your work used (YYYY-MM-DD). Omit only for present-day lookups. row_key: pin the citation to one specific row, if you are citing one row.

一檔股票的誠實體檢:成長、獲利品質、估值、籌碼、紅旗,每個數字可驗證。 ★ 這是**事實整理**,不是投資建議、不是預測、不是目標價。輸出帶 `does_not_prove`,轉述時請一併保留。 ★ POINT-IN-TIME:`as_of` 逐資料集尊重揭露時差 —— 月營收用法定截止日推得的 估計時點,財報用出表日,估值是當日。不給 `as_of` 就用「現在可知的最新」, **不是今天**。 ★ 沒看到的東西會列在 `cannot_see`,並說明是「查了沒有」「查詢失敗」還是 「散佈權未取得」—— 三者不同。⚠️ 不要替 `cannot_see` 裡的項目估算或補值。 ★ `checked_and_clear` 是「查過而且確認沒有」(例如近一年無裁罰), 和 `cannot_see` 的「不知道」是兩回事。 ★ 代號解析不到會回 `needs_disambiguation` 附候選 —— 請回問使用者,不要挑一個。 ★ PRECONDITIONS:這支工具**自己**去拿九個資料集,所以 rows 不需要先用 `query_dataset` 取好。它 requires 的是身分:呼叫端 must be 一個已識別的 session(X-API-Key 或已驗證的 OAuth 登入),因為體檢會寫進呼叫者的工作區 脈絡。方案沒有涵蓋某個資料集時,那一項會以 `not_licensed` 出現在 `cannot_see`,而不是讓整支工具失敗 —— 所以不需要 entitlement 先檢查。 ★ SIDE EFFECTS:read-only。does not 寫入任何資料表,不建立警示、不下單。

Lay two to five named companies side by side on the same measures, gaps marked as gaps. ★ AN ABSENT FIGURE STAYS ABSENT, AND THE COMPANY STAYS ON THE TABLE. A blank cell is reported as `available: false`, never filled with a zero, a previous period, or by quietly dropping the column. Dropping is the worst of the three: it converts "we do not hold this measure for that company" into "you did not ask about that company". ★ EACH CELL NAMES ITS OWN SOURCE. Margins and institutional flows come from different datasets, so one citation for the whole table would imply every figure came from one query. ★ NOT `screen` AND NOT `find_related`. `screen` finds symbols from a description when you have none; `find_related` walks supply-chain links. This one needs you to already know which two to five companies you mean. ⚠️ Placing figures next to each other is not ranking them. A measure one company publishes and another does not is a coverage difference, not evidence about either. Args: tickers (2-5), metrics, optional pre-fetched rows_by_ticker, optional as_of.

Cancel one armed price trigger permanently, by its rule id. Disarms a single watch so it will not fire again — the opposite of `set_price_alert`, and unlike `list_alerts` it changes state rather than reporting it. Cancellation is irreversible: re-arming means creating a fresh trigger. Naming somebody else's rule id cancels nothing at all. Args: rule_id.

FULL semantics of one dataset: grain (what a row is), field meanings+units, ★TIME-CORRECTNESS rules (knowledge_time_field / point_in_time_safe — read before backtesting), relations for cross-table reasoning, agent_hints (when to use), quant_use (which factors). Args: dataset_id.

Where did this number come from, and when could anyone have known it? (WP-6) ★ THIS IS THE GROUNDING TOOL. Every other tool's answer is supposed to be expandable through this one: given a ticker, a field and a date, it names the dataset that serves that field, the knowledge column, and WHICH as_of rule applies on that date. Use it whenever you are about to state a number as fact. ★ IT DOES NOT OVERLAP WITH THE OTHER TWO PROVENANCE TOOLS, and they are not substitutes: `get_inclusion_proof(dataset, row_key)` proves a ROW was in a published snapshot (Merkle). `cite_this(dataset, as_of)` produces a citation for a DATASET. `explain_provenance(ticker, field, date)` answers WHERE A CELL came from and WHEN it was knowable. Nothing else answered that. ★ FOUR VERDICTS, and `partial` is not `resolved`: resolved dataset, knowledge axis, as_of rule AND the official endpoint. partial ⚠️ the first three, but NOT the official endpoint — no dataset declares a per-cell endpoint today. "When was it knowable" and "which official endpoint produced it" are different questions; conflating them is exactly the vagueness this tool exists to remove. ambiguous_field ⚠️ the field lives in several datasets. Every candidate is listed and NONE is picked for you: picking one would hand you a confident wrong answer. Re-ask with `dataset=`. unknown_field no served dataset declares that field. This is "we do not have that column", NOT "no data found" — a different next step for you. ★ READ `limitations`. It says when the knowability date is DERIVED (statutory deadline, a conservative upper bound) rather than observed, and whether the rule came from declarations or from the live schema. A declared capability is not the same as a working one. Args: ticker: the security, e.g. '2330'. field: the column you are about to quote, e.g. 'close', 'revenue'. date: the date your claim is about (YYYY-MM-DD). dataset: pin the answer to one dataset when the field is ambiguous.

Traverse the knowledge graph for cross-table / supply-chain reasoning. - dataset_id: returns join-able datasets (+why) to plan multi-table analysis. - ticker: returns its industry value-chain node + peers in the same node (supply-chain reasoning). Args: dataset_id (e.g. 'equity_daily_prices') and/or ticker (e.g. '2330').

Retrieve a previous backtest by run_id — the full record, including why it was rejected. Only runs in YOUR namespace are visible; a run_id belonging to someone else is simply not found. Args: run_id (the `twmd_bt_…` handle returned by run_backtest).

Emit a copy-pasteable HTTP snippet wired to the real endpoint, header and parameter names. ★ FOR WRITING YOUR OWN CLIENT, NOT FOR GETTING DATA. Every other tool here answers a question; this one hands you source code so your program can ask it directly over HTTPS. Nothing is fetched and no rows come back. ★ THE CONSTANTS ARE READ OUT OF THE SERVER, NOT REMEMBERED. Base URL, the `X-API-Key` header spelling, and the route path all come from the code that serves them. The worst kind of broken example is one that looks right: a mistyped path 404s and the reader blames their own key. ★ AN INTENT IT HAS NO ROUTE FOR IS REFUSED. It will not point at a plausible-looking path it has not confirmed. Python output is passed through Python's own compiler before it is returned; the JavaScript variant is structure-checked only, and says so. ⚠️ The credential in the snippet is an obvious placeholder, never a realistic-looking string — a convincing fake gets pasted, sent, and then fails somewhere nobody can trace. Args: intent, language (python/javascript), ticker, optional as_of.

Prove a row was in the snapshot TWMD published — and check it yourself. Returns the Merkle sibling path, the signed root, and the checkpoint it belongs to. It returns the PATH rather than a yes/no on purpose: a service that answers "yes, it is included, trust me" is the opposite of verifiable. Recompute the root from the leaf and the path; the verifier is ~30 lines and is written out in docs/VERIFIABLE_DATA.md. ★ WHAT IT PROVES: integrity (the row was not altered after publication) and origin (the root was signed by TWMD). **It does NOT prove the numbers are correct** — if the exchange published a wrong figure, TWMD faithfully committed to the wrong figure. Do not present a passing proof as a correctness guarantee. ★ THREE STATUSES, and they must not be collapsed: ok proof enclosed; verify it. not_in_snapshot that row_key was NOT a leaf of the snapshot. A true answer, NOT a failure and NOT evidence of tampering. no_checkpoint no snapshot was ever built for that dataset/version. 🔴 VERIFY ON THE CLIENT SIDE. There is deliberately NO `verify` tool here, and that is the point: this server is the ISSUER of the proof, so a "we checked our own proof and it is fine" tool would carry almost no evidential weight while looking exactly like independent verification. Run the check where you are: * Python SDK `twmd.Client.flight(..., proof_api=...).verify(dataset, row_key)` -> {"verified": bool, "verification": {status, proves, checkpoint, unverified_reason}} with FOUR values (verified / failed / unproven / absent) — `unproven` is NOT `failed`. * any language `scripts/twmd_verify_proof.py` — copy it; stdlib only, imports nothing from our package (a verifier that needs our package asks you to trust our package). * the core `twmd.proof_core` is the SAME SOURCE TEXT as that script, guarded per symbol by tests/test_proof_core_is_one_source.py, so the two cannot drift. Args: dataset: e.g. 'daily_price'. row_key: the dataset's LOGICAL key joined by '|', e.g. '2330|2026-08-14'. snapshot_version: omit for the most recent checkpoint.

Retrieve one of YOUR previous research reports. Others' runs are simply not found. Args: research_id — the identifier `run_research` returned when it started that run.

Show the price-trigger rules you have armed, and whether each is still armed. A read-only inventory of thresholds you asked to be watched — it arms nothing and cancels nothing (`set_price_alert` arms, `delete_alert` cancels). Another customer's triggers are simply not visible here.

Browse an INDEX of your past backtest runs — ids and headline metrics only, no re-execution. Use when you want to find a run whose id you have forgotten. It never re-computes anything: `run_backtest` executes a new one, `get_backtest` opens a single record in full, and `replay_backtest` re-derives one to check reproducibility. This is the catalogue, not any of those three. Args: optional strategy_id filter, limit.

List available Taiwan-market datasets (discovery entry point). Returns id / 中文名 / category / tier / one-line description for each. Use this first to find the right data. Args: category: optional, e.g. 'chip'(籌碼) 'fundamental'(基本面) 'price'(行情) 'macro'(總經) 'relation'(關聯/產業鏈) 'derivatives'(期權) 'event'(事件) 'rag_text'(文本). tier: optional minimum plan: 'free' 'starter' 'pro' 'max' 'developer' 'enterprise'.

Verdicts from the overnight factor search on YOUR namespace — including the rejections. The rejections are returned deliberately. A research log that keeps only the winners is the highlight reel overfitting lives in, and the acceptance RATE is the number that tells you whether the anti-overfitting gate is doing its job: a search that accepts most of what it tries has a broken gate, not a talent for finding alpha. Every verdict carries the trial count it was judged against, so it can be re-checked. `coverage.missing` names hypotheses that were proposed but never judged because a cost ceiling was reached — those are UNTESTED, not rejected. An accepted factor is a FINDING with a run_id, not an allocation. Nothing here trades. Args: limit: how many verdicts to return, newest first (max 100).

Financial actions proposed by your research runs that are waiting for a human decision. Nothing here has been executed or ever will be by this system. These are proposals.

總經 Regime:景氣信號、利率與殖利率曲線、匯率、資金流。 ★ 這支講的是**整個市場**,不是任何一檔個股 —— 所以它不收 ticker。 ★ 事實整理,不是投資建議、不是預測。輸出帶 `does_not_prove`。 ★ ⚠️ `macro_worldbank` 類的來源會**回溯修訂**歷史年度:今天讀到的舊年度數字 和當時看到的不是同一個,`pit_notes` 會標出來。 ★ Args:只有 `as_of`(選填)—— 它**不收 ticker**,因為主體是整個市場。 ★ Returns:同一個誠實信封;`subject` 固定是 "TW_MARKET"。 ★ 例如問「現在景氣循環在哪個位置」就是這一支,而不是問某一檔。 ★ PRECONDITIONS:自己取數;requires 一個已識別的 session。 ★ SIDE EFFECTS:read-only,does not 寫入。

Read the tickers on one named watchlist, as it stands right now. A curated roster you maintain — distinct from `memory_search`, which digs through everything you ever recorded, and from `list_alerts`, which is about price triggers rather than symbols you are following. Args: key — the roster's name, default 'default'.

Re-run a remembered query by its `twmd_q_…` id, through the read API's own replay store. This is what makes a recalled finding checkable: the memory says where it came from, and this fetches that same data again. It never re-executes the query by another route — two implementations of "replay" would be two answers to a question whose whole value is having one. `status` is one of: found the bytes are here, with `result_hash` to check them against too_large it WAS served, but exceeded the size ceiling: `result` is absent and `result_hash` is authoritative — you can still verify a copy you hold not_found never recorded (or pruned) — the citation cannot be resolved Args: query_id — the `twmd_q_…` reference carried on a remembered finding's citation.

Remember something, with its sources and its knowledge time. Nothing is ever overwritten: saving a `factor_def` or `watchlist` under an existing key SUPERSEDES the previous version (both rows survive, so "what did I believe in June?" stays answerable), and saving identical content twice is a no-op rather than a duplicate. Args: kind: 'query' | 'factor_def' | 'watchlist' | 'finding' | 'note'. content: the thing to remember, as an object. key: the stable name — REQUIRED for 'factor_def' and 'watchlist' (that is what makes a definition reusable next session instead of re-invented). as_of: the knowledge time this memory is about. Recall can bound on it, which is what keeps a memory from leaking the future into a point-in-time question. source_query_ids: the `twmd_q_…` ids behind this. REQUIRED for 'finding' — a conclusion that cannot point at its data is not evidence, and will be refused. agent_id: optional label for which of your agents wrote this.

Recall your own memories — hybrid (semantic + exact-term), with provenance attached. Every result carries where it came from (`source_query_ids`, replayable), when it was believed (`valid_from`/`valid_to`) and what knowledge time it is about (`as_of`), plus a `recall` block stating which model and which filters produced the answer. Args: query: what you are looking for, in words. kinds: restrict to some of 'query' 'factor_def' 'watchlist' 'finding' 'note'. as_of: knowledge-time bound — pass a backtest's as_of and nothing recorded later can come back. Use this for anything point-in-time. key: the stable name, when you know it (e.g. a factor name). agent_id: narrow to one of your agents. include_superseded: also return old versions (the audit view). Default is current only.

籌碼結構:法人分項、融資融券、借券使用率、大戶集中度、董監質押。 ★ 事實整理,不是投資建議、不是預測、不是目標價。輸出帶 `does_not_prove`。 ★ POINT-IN-TIME:`as_of` 逐資料集尊重揭露時差;不給就用「現在可知的最新」。 ★ 看不到的在 `cannot_see`,並分「查了沒有 / 查詢失敗 / 散佈權未取得」。 ⚠️ 逐券商分點進出**永遠**在 cannot_see —— 散佈權未取得,不要替它估算。 ★ PRECONDITIONS:自己取數,rows 不需要先用 `query_dataset` 備好;requires 一個 已識別的 session。方案沒涵蓋的資料集以 not_licensed 進 cannot_see,不整支失敗。 ★ Args:`ticker`(代號,例如 "2330")、`as_of`(選填的知識時間界線)。 ★ Returns:一個信封,`concerns`/`observations` 各帶觸發它的數字與驗證連結, 缺的東西 listed in `cannot_see`。 ★ SIDE EFFECTS:read-only,does not 寫入任何資料表。

Query rows with built-in look-ahead protection. ★ POINT-IN-TIME: pass `as_of` (YYYY-MM-DD) for backtesting/agent-learning. For non-point-in-time-safe datasets (fundamentals, monthly_revenue, dividend_policy…) rows are filtered by DISCLOSURE date <= as_of, so the agent only sees what was public at that moment. Omit as_of only for present-day lookups (warned). ★ IF A VALUE IS IN `coverage.missing`, IT IS NOT AVAILABLE. Say it is not available. **Never estimate it, interpolate it, infer it from a neighbouring period, or carry the last known value forward.** `coverage.missing` lists exactly what was requested and not returned, with a reason (e.g. "9999 在 as_of 當日未上市"). An empty `data` array alongside a populated `missing` list is a complete and correct answer to "what do you have" — not an invitation to fill the gap. ★ EVERY VALUE IS ATTRIBUTABLE. `provenance` carries {source, source_role, ingested_at, revision, provenance_uri}; `meta.query_id` names this exact question. Quote the query_id when reporting a number — `replay_query(query_id)` returns the bytes that were served, so the claim can be checked later. `freshness.is_stale` is computed server-side against the dataset's own cadence budget; `null` means it could not be determined, which is NOT the same as fresh. Args: dataset_id: see list_datasets. tickers: e.g. ['2330','2317']. start/end: 'YYYY-MM-DD' range. as_of: knowledge-time cutoff 'YYYY-MM-DD' (use for backtests). limit: <=5000. Returns: {meta:{table,coverage,row_count,as_of_applied,point_in_time_safe,warnings,query_id}, data:[...], provenance:{...}, coverage:{requested,returned,missing,reason}, freshness:{status,latest_available,expected_lag,is_stale}} Example: query_dataset('income_statement', tickers=['2330'], as_of='2023-06-30')

The Taiwan business-cycle light (NDC monitoring indicator) as a monthly series. ★ THIS IS A REVISED FIGURE, NOT A POINT-IN-TIME ONE. Every response carries `revision_basis: as_revised`. Our source holds exactly one row per month — the CURRENT value, not the value as first published — and it records no publication date. ★ `as_of` IS REFUSED, AND THE REFUSAL IS THE POINT. There is no honest point-in-time answer here yet. Do NOT work around it by asking for a date range that ends at your as_of: the NDC publishes a month's light about 27 days AFTER that month ends, so a range ending 2026-03-05 still contains the February light that was not public until late March. That is look-ahead, it raises no error, and it makes a backtest look better than it was. ★ USE IT FOR CONTEXT, NOT AS A BACKTEST INPUT. Describing what regime the market is in today, or labelling historical periods for narrative, is fine. Feeding it into a simulated decision that claims to have been made at the time is not. Args: optional start/end (YYYY-MM-DD), limit. `as_of` returns a refusal explaining the above.

Read the FULL TEXT of filings and announcements — with proof links and a knowledge cutoff. ★ NOT `search_filings`. That one ranks passages by similarity and hands you fragments; this hands you whole documents so you can read what was actually said and where it sat in the filing. Similarity is not importance, and a fragment cannot show you its own context. ★ POINT-IN-TIME: pass `as_of` (YYYY-MM-DD). The cutoff is applied in SQL on the source's declared knowledge-time column BEFORE the row limit, so a bounded read is a true prefix of what was knowable, not a random subset of it. **Without `as_of` the read is NOT point-in-time** and says so in `warnings`. ★ NO SENTIMENT, NO SCORES — deliberately. Judging the text is your job. A stored score is one model's output on one day; after that model changes, the stale number still sits in the table looking exactly like a fresh one. ★ READ `corpus_reality` BEFORE CONCLUDING ANYTHING. The full-text corpus is SMALL and the response says how small. One source carries ~1M rows of TITLES ONLY — a large row count there is breadth, not depth, and "what did they say about it" is not answerable from titles. ★ A ticker that returns nothing appears in `coverage.missing`. That means nothing is held for it under those filters — NOT that the company disclosed nothing. Do not fill the gap. Args: source: which corpus, e.g. 'announcements_fulltext' or 'mops_major_event'. tickers: restrict to these codes, e.g. ['2330']. as_of: knowledge cutoff (YYYY-MM-DD). since: optional lower bound on the same knowledge-time column. limit: max documents (these are whole documents; keep it small).

Re-run a stored backtest and report whether it still produces the same numbers. Same spec, same `as_of`, same data questions. If the numbers moved, either the engine version changed or the underlying data was restated — both are reported, neither is smoothed over. Args: run_id.

Measure a portfolio you state against limits you state, on official point-in-time prices. Reports concentration and peak-to-trough drawdown, and NAMES every position it could not price rather than quietly assessing the rest — an assessment covering 60% of a portfolio without saying so is worse than none. Any breach produces a PROPOSAL (e.g. "reduce 2330") that requires a human decision. Approving a proposal records that decision; it executes nothing. TWMD has no order path. Args: positions: `[{"ticker": "2330", "quantity": 100}, ...]`. YOUR stated holdings — nothing is read from a brokerage account, because no such connection exists. as_of: knowledge cutoff, `YYYY-MM-DD`. Defaults to the latest available data. max_position_weight: single-name limit as a fraction (0.35 = 35%). max_drawdown: peak-to-trough limit as a fraction (0.25 = 25%).

事件與市場結構風險:隱含波動率、注意處置、裁罰訴訟、放空限制、因子。 ★ ⚠️ **部位集中度與回撤不在這裡** —— 那是 `risk_assess` 的職責。重算一份會產生 第二個答案,而兩個都帶著我們的名字。 ★ `checked_and_clear` 是「查過而且確認沒有」(例如近一年無裁罰),和 `cannot_see` 的「不知道」是兩回事。 ★ 事實整理,不是投資建議、不是預測。輸出帶 `does_not_prove`。 ★ Args:`ticker`(例如 "2330")、`as_of`(選填)。 ★ Returns:誠實信封;事件風險 listed in `concerns`,查過確認沒有的 listed in `checked_and_clear`。 ★ PRECONDITIONS:自己取數;requires 一個已識別的 session。 ★ SIDE EFFECTS:read-only,does not 寫入。

Submit a point-in-time backtest. NON-BLOCKING: returns a run_id at once; poll get_backtest. `job_status` tells you where it is: `cached` (this exact backtest was already computed — the stored result is returned immediately), `done` (it finished within `wait_seconds`), or `queued` / `running` (call `get_backtest(run_id)` until it is `done` or `failed`). The same spec always has the same run_id, so re-submitting is free and never computes twice. The run may only see data stamped on or before `as_of` — that is enforced structurally, not by convention. Results arrive with the data `query_ids` behind them and an anti-overfitting verdict (out-of-sample, deflated Sharpe, multiple-comparison, crash stress); a run that fails the gate is returned REJECTED with reasons rather than hidden. Args: strategy_id: a registered strategy, e.g. 'buy_and_hold' or 'cross_sectional_momentum'. start / end: the測試期間 (YYYY-MM-DD). `end` must not be after `as_of`. as_of: the knowledge cutoff. REQUIRED — there is no "today" default. tickers: required when universe_kind='explicit'; ignored for 'point_in_time'. universe_kind: 'point_in_time' (survivorship-safe, resolved from listing/delisting dates at each rebalance) or 'explicit' (a list you supplied). market: optional market filter for a point-in-time universe. rebalance: 'daily' | 'weekly' | 'monthly' | 'quarterly'. cost_bps: one-way transaction cost in basis points. params: strategy parameters, e.g. {'lookback_days': 60, 'top_k': 5}. wait_seconds: how long this call may wait for a FAST run to finish before answering `queued`/`running` (0 = answer at once; capped at 10). It never waits for a slow run. costs: the trading costs to simulate, e.g. {"model": "tw_default_model"} (standard: 0.1425% commission both sides + 0.3% securities tax on sells + slippage), {"model": "tw_default_model", "fee_discount": 0.28} (a 2.8-fold broker discount on the commission; the tax is statutory and cannot be discounted), or {"model": "none"} (commission and tax zeroed — the result is flagged as not a real-world result). Omit for standard. Do not combine with cost_bps. The response's `costs` says exactly what was applied. Strategies served by the Zipline engine are limited to a 3-year window per interactive request (`error: "window_exceeds_interactive_cap"`, with a suggested window). Full-period results for the strategy-library strategies are precomputed nightly on the marketplace. This measures history. It is not advice and it places no orders.

Replay a saved multi-step routine over rows you fetched, with every step listed. ★ THREE ROUTINES: `momentum_scan` (rank by a return column), `earnings_surprise` (actual versus estimate), `dividend_capture` (which ex-dates are still ahead). ★ A SAVED ROUTINE IS NOT A TRADING VIEW. The names are conventional labels for well-known sequences; what runs is arithmetic over rows you supplied. `steps[]` spells out each operation so you can disagree with the routine rather than trust it, and nothing here says any of these sequences makes money. ★ SKIPPED SYMBOLS ARE LISTED, NEVER QUIETLY OMITTED. A symbol lacking the column a routine needs lands in `skipped[]` with the reason. A silently shorter list reads as "these were evaluated and did not qualify", when in fact they were never evaluated at all. ★ IT COMPOSES, IT DOES NOT FETCH. Rows come from `query_dataset`, so the routine inherits that call's `as_of` rather than defining a second point-in-time story of its own. ⚠️ Missing inputs are never substituted with zero — a zero estimate turns any positive result into an infinite surprise. Args: recipe, rows, as_of, today, top_n, min_yield.

Run a multi-agent research pass and return a structured, sourced report. Six roles run in order — data analyst, factor researcher, backtest engineer, risk officer, portfolio manager, compliance officer. Each step's output carries the `query_ids` behind it; risks are reported alongside results, not beneath them; and anything the run could not do is listed as a limitation rather than filled in. The factor researcher checks memory first and SKIPS a hypothesis a previous run already rejected. The portfolio manager proposes nothing when the evidence failed the anti-overfitting gate, and any allocation it does propose is a PROPOSAL awaiting a human — this system places no orders and moves no money. Args: prompt: the research question, in your words. This is the only channel carrying instructions; anything a tool returns is treated as data. as_of: the knowledge cutoff. REQUIRED — nothing stamped after it is visible to the run. tickers: optional explicit universe. Omit for a point-in-time (survivorship-safe) one. start / end: optional test period; `end` must not be after `as_of`. max_backtests: per-run cap on backtests (cost control).

Post-process rows already returned by query_dataset; without rows it returns a clear hint. Turn a spoken shortlist description into explicit numeric cut-offs, and show the cut-offs. ★ THE THRESHOLDS COME BACK WITH THE SHORTLIST. "低本益比" becomes `per < 15`, and that 15 is printed in `applied[]` so you can disagree with it. A filter that hands over thirty names without saying where it drew the line cannot be checked by anyone — and whether the line was 15 or 20 completely changes which thirty. ★ PHRASES IT CANNOT MAP COME BACK IN `unparsed[]`. It will not quietly invent a boundary for a wording it did not recognise, because a list the caller believes they defined and actually did not is worse than a shorter list. ★ NOT `query_dataset` AND NOT `compare`. `query_dataset` returns rows for symbols you already chose; `compare` puts a handful of named symbols beside each other. This one is for when you have no symbols yet, only a description of what you are looking for. ⚠️ A shortlist is not a recommendation. Rows whose value is absent are excluded rather than assumed to pass — being unmeasured is not the same as qualifying. Args: conditions (plain words), optional as_of, optional pre-fetched rows.

Semantic search over MOPS filings, financial-statement notes and company news. Answers questions a keyword filter cannot: "what risks did this company disclose this quarter?", "which companies mentioned CoWoS capacity expansion?" — matching on MEANING, so a paragraph that never uses your exact words still ranks. ★ POINT-IN-TIME: pass `as_of` (YYYY-MM-DD). Chunks are filtered `published_at <= as_of` in SQL BEFORE ranking, so a backtest cannot retrieve a filing that did not exist yet. `meta.as_of_applied` echoes the cutoff that actually ran — check it. **Without `as_of` the results include the most recent filings and are look-ahead UNSAFE for backtesting**; the response says so in `meta.warnings`. ★ REFERENCE CONTEXT, NOT AUTHORITY. Every hit carries `source`, `source_tier` ("official" = MOPS/exchange, "third_party" = press) and a `url`. Read the chunk and judge it yourself; the ranking is similarity, not importance, and similarity is not evidence. **Nothing here is investment advice** (`not_investment_advice: true`). ★ A ticker you asked about that returns nothing appears in `coverage.missing` with a reason. That means NOTHING IS INDEXED for it under those filters — it does NOT mean the company disclosed nothing. Do not fill the gap. Args: query: what to look for, in Chinese or English (e.g. '匯率風險', 'CoWoS capacity'). tickers: restrict to these codes, e.g. ['2330','2317']. doc_type: e.g. 'mops_major_event', 'financial_note', 'company_news'. as_of: knowledge-time cutoff 'YYYY-MM-DD' — use it for anything backtest-shaped. source_tier: 'official' to exclude third-party press. limit: <= 100. Returns: {data:[{ticker,doc_type,published_at,source,source_tier,url,chunk_text,similarity}], meta:{as_of_applied,point_in_time_safe,embedding_model,warnings,query_id}, provenance:{...}, coverage:{requested,returned,missing,reason}, freshness:{...}}

Leave a standing instruction: tell me when this symbol crosses this price. The alert OUTLIVES this conversation. It is evaluated against official daily closes by a resident agent and delivered to your realtime stream and to any webhook endpoints you have registered — signed, retried, and de-duplicated so one crossing is one notification. This is a NOTIFICATION, not an order. Nothing in TWMD can place a trade. Args: rule_id: your name for this alert. Re-using one UPDATES it rather than adding a second. symbol: the Taiwan ticker, e.g. "2330". direction: "below" or "above". threshold: the price level, in TWD. label: optional human-readable note carried on the alert. edge_triggered: True (default) fires on the CROSSING only — one long slump does not notify you every day. False fires on every bar that is past the level.

供應鏈與同業:產業鏈位置、同業分組、轉投資、外銷訂單 vs 營收。 ★ 事實整理,不是投資建議、不是預測。輸出帶 `does_not_prove`。 ★ ⚠️ `company_peer_groups` 無日期欄:指定過去的 `as_of` 時,回的是**現況**分組, `pit_notes` 會標「非當時分類」。 ★ ⚠️ 質化的客戶/供應商關係與分部營收在年報附註,那條來源**尚未建置** —— 固定列在 `cannot_see`(status=no_loader),不是這次查失敗。 ★ Args:`ticker`(例如 "2330")、`as_of`(選填)。 ★ Returns:誠實信封;外銷訂單與營收的背離會 listed in `concerns`。 ★ PRECONDITIONS:自己取數;requires 一個已識別的 session。 ★ SIDE EFFECTS:read-only,does not 寫入。

Hand an unregistered caller a short taste of an open dataset, plus where to unlock the rest. ★ WHAT AN ACCOUNTLESS CALLER GETS INSTEAD OF A BARE REFUSAL. Somebody arriving through a chat connector with no plan would otherwise meet a flat rejection, which the host model relays as "this service turned you down". ★ WHAT IS ACTUALLY FREE, STATED CONCRETELY. The reference resources read with no key at all, and the five sample tickers (2330, 2317, 2454, 0050, 2603) answer through `query_dataset` with no plan. Querying beyond those over MCP starts at the Pro plan. ★ OPEN DATASETS YIELD A FEW MARKED ROWS; PAID ONES YIELD NONE. For a paid dataset the row list is not even consulted, so no figure can escape through this path regardless of what the caller passes in. `rows_withheld` states how many were held back, because a taste that does not say it is a taste reads as the whole thing. ★ NOT `query_dataset`. That one serves entitled callers in full. This exists only for the moment somebody hits the edge of what they are entitled to. ⚠️ Withdrawn datasets stay refused here too — a licence ruling is not a tier, so no amount of signing up unlocks them. Args: dataset, keyless_eligible, rows.