wyckoff-agent-skill（威科夫特工技能）

Wyckoff Trading Agent

This skill operates in two modes: Setup (Step 0) and Analysis (Steps 1–9). Step 0 runs only when prerequisites are missing; once everything is configured, jump directly to analysis.

Step 0: Environment Detection & Guided Setup

Run these checks in order. Stop at the first failure and guide the user to fix it before continuing.

0.1 CLI Installation Check

Run wyckoff --version.

• Success → proceed to 0.2.
• Failure → guide installation:

  pip install youngcan-wyckoff-analysis

Or one-line install:

  curl -fsSL https://raw.githubusercontent.com/YoungCan-Wang/Wyckoff-Analysis/main/install.sh | bash

After install, verify with wyckoff --version again.

0.2 Auth Check

Run wyckoff auth status.

• Logged in → proceed to 0.3. (Credentials are auto-saved; token expiry triggers automatic re-login.)
• Not logged in → guide registration and login:

1. Tell user to open https://wyckoff-analysis-youngcanphoenix.streamlit.app/ to register an account. 2. After registration, run: wyckoff auth login <email> <password> (credentials will be persisted, no need to login again). 3. Verify with wyckoff auth status.

0.3 Data Source Check

Run wyckoff config show.

• tushare_token present → OK.
• tushare_token missing → guide:

- Tushare token is free, get it from https://tushare.pro/ after registration. - Then run: wyckoff config tushare <token>

• tickflow_api_key present → OK.
• tickflow_api_key missing → guide:

- TickFlow provides real-time and historical A-share data. - Purchase at: https://tickflow.org/auth/register?ref=5N4NKTCPL4 - Then run: wyckoff config tickflow <api_key>

Note: At least one data source (tushare or tickflow) must be configured. Both are recommended for better coverage.

0.4 Model Check

Run wyckoff model list.

• Has models → proceed to Step 1.
• No models → guide: wyckoff model add (interactive) or wyckoff model set <name> <provider> <api_key> --model <model_name>.

When all checks pass, print a brief summary and proceed to analysis.

CLI Operational Commands

When the user's intent is operational (not analysis), route directly to the appropriate CLI command instead of running the analysis pipeline:

For the full CLI reference, see rules/cli-setup-guide.md.

Input Protocol

Accept any combination of:

• Stock symbol(s), single or multiple.
• holdings: [symbol+cost+qty, ...].
• cash: available cash amount.
• candidate: optional non-holding symbol.
• CSV file(s), image(s), text constraints/goals.

Infer scenario automatically:

• holdings non-empty + candidate: rotation comparison + per-holding actions.
• holdings non-empty + no candidate: per-holding add/reduce/hold/exit.
• holdings empty + cash: empty-position cash deployment.
• No portfolio fields: symbol analysis flow.

Do not require users to explicitly say "switch/add/reduce/empty-position."

Full Capability Orchestration (Steps 1–9, Required Order)

1. Parse and normalize inputs.

- Normalize symbols to exchange-qualified format when possible. - Parse holdings into {symbol, cost, qty}. - Preserve raw user inputs for output audit.

1. Acquire current time via system/tool first.

- Fetch current timestamp from tool/system. - Convert to Asia/Shanghai. - Print 当前北京时间：YYYY-MM-DD HH:MM（UTC+8）.

1. Decide trading availability with authoritative calendar checks.

- Judge weekday and continuous-auction windows: 09:30-11:30, 13:00-15:00 (Beijing time). - Query authoritative trading calendar when holiday/adjusted-workday uncertainty exists. - If not tradable, downgrade to post-market review + next-session plan + T+1-safe order strategy.

1. Fetch online data with source fallback.

- Follow rules/source-fallbacks.md strictly for each symbol. - Perform schema and row-count validation before accepting a source. - Log fallback attempts and final source per symbol.

1. Integrate CSV/image modalities when provided.

- Use CSV as supplemental historical structure input and reconcile with fetched data. - Treat chart images as micro-structure evidence; explicitly acknowledge image reception. - Continue analysis if one modality fails and state exact failure cause.

1. Run Wyckoff structural analysis first.

- Analyze latest available window (target 500 trading days) with MA50/MA200. - Identify only evidenced phases/events (SC/ST/Spring/LPS/SOS/UTAD). - Use event-date news search only for validation context, never as primary trade logic.

1. Produce portfolio decisions after structure analysis.

- For each holding, output one explicit action: add / reduce / hold / exit. - If candidate exists, compare against structurally weakest current holding and decide switch / partial switch / hold. - If holdings are empty and cash exists, output staged cash deployment suggestion.

1. Render plots only when session rules allow.

- Skip plotting during tradable intraday windows. - When plotting is allowed, enforce all constraints in rules/alpha-system-prompt.md.

1. Apply capability degrade policy.

- Never fabricate OHLCV rows, event timestamps, or trading status. - If all sources fail for a symbol, mark data_unavailable and continue others. - If valid rows < 30, report insufficient structure depth and avoid hard phase labels. - If image parsing fails, explain reason and continue CSV/text/online path.

For detailed capability-routing policy, read rules/system-capability-playbook.md.

Fixed Output Contract

Always output in this order:

1. 当前北京时间：YYYY-MM-DD HH:MM（UTC+8）
2. Trading verdict: 当前是否可盘中交易：是/否 (with reason if no).
3. Data audit table per symbol: symbol, source_used, rows_kept, window_end_date, fallback_count.
4. Wyckoff analysis: current cycle background and phase (only evidenced), key events with rationale, action boundaries respecting T+1.
5. Portfolio action section (portfolio flow only): holdings snapshot, per-holding actions, candidate comparison, cash suggestion, final summary in Wyckoff tone.
6. Plotting section (only when allowed by session rules).

Hard Constraints

• Do not change the fixed prompt wording unless explicitly requested.
• Do not fabricate missing OHLCV rows.
• Do not ignore image input if image is parseable.
• Do not use opaque white text boxes in chart annotations.
• If fetching data requires running Python scripts, run them only in a sandboxed environment.
• Prefer direct web/API fetch first; use Python scripts only when needed for fallback, parsing, or normalization.

Resources

• rules/alpha-system-prompt.md: fixed role and hard rules.
• rules/source-fallbacks.md: online source switching policy.
• rules/system-capability-playbook.md: full system capability routing and degrade policy.
• rules/cli-setup-guide.md: CLI installation, registration, and command reference.