Decision support

Decision Support

Generate position-aware action guidance for: $ARGUMENTS

Overview

• Implementation status: code-backed
• Local entry script: <bundle-root>/decision-support/run.py
• Primary purpose: convert current market state plus account constraints into an explicit action and sizing frame with systematic risk disclosure
• Research layer: decision support, not execution planning or portfolio instruction
• Workflow stage: stage 6 Risk Management & Position Sizing
• Local executor guarantee: produce a baseline action, quantity, stop, take-profit, and risk-budget frame from the current snapshot and account inputs

Use When

• The user asks whether to buy, watch, hold, reduce, or avoid.
• The user wants quantity sizing under capital and risk constraints.
• The user wants stop-loss and take-profit anchors without the full narrative brief.
• The user needs a conditional decision frame that can be evaluated against explicit account assumptions.

Do Not Use When

• The user only wants market interpretation. Use market-analyze.
• The user wants a full report with data, analysis, and strategy sections. Use market-brief.
• The user wants execution style and zone design rather than action sizing. Use strategy-design.
• The user wants a deeper thesis memo or explicit variant view discussion. Use analysis.
• The user wants executable trading instructions. This skill produces conditional guidance, not orders.

Inputs

• Normal case: one stock symbol.
• Optional --csv PATH: use a local CSV instead of the default market source.
• Optional --capital, --cash, --position, --risk, --max-position: control sizing and risk limits.
• Optional --start, --end, --source: constrain the data-loading path.
• If symbol is omitted, the skill may reuse last_symbol from the same execution context.
• Important assumption boundary: if account inputs are omitted, the local defaults still produce an answer. The caller must mark that answer as conditional on defaults rather than as portfolio-specific advice.

Execution

Step 1: Confirm the portfolio context

Use decision-support when the user wants a decision on whether and how much to own. If the user mainly wants how to express an already accepted view in the tape, route to strategy-design.

Step 2: Run the local executor

python3 <bundle-root>/decision-support/run.py <symbol> [--cash N] [--position N] [--capital N] [--risk N]

Step 3: Deliver the result as three-part conditional guidance

Every decision-support output must be structured in three mandatory parts:

Part 1: Conditional Action Frame

• Action: buy, watch, hold, reduce, or avoid
• Heuristic conviction score (labeled explicitly as non-probabilistic)
• Target position size
• Quantity (may be zero if action is watch/avoid or if constraints bind)
• Reference price
• Stop loss level
• Take profit level
• Risk budget allocation
• Reasoning bullets (3-5 specific factors)

Part 2: Explicit Assumptions State which inputs were user-supplied vs. defaulted:

• Capital: [user-supplied: X / defaulted: Y]
• Cash available: [user-supplied: X / defaulted: Y]
• Current position: [user-supplied: X / defaulted: Y]
• Risk tolerance: [user-supplied: X / defaulted: Y]
• Max position limit: [user-supplied: X / defaulted: Y]
• Data window: [start date] to [end date]
• Symbol source: [explicit / session reuse]

Part 3: Non-Modeled Risks Systematically disclose material execution and market risks not captured in the decision frame:

• Liquidity risk: if average daily volume suggests position size may impact market, state that explicitly
• Gap risk: opening gaps can bypass stop levels; state whether symbol has history of gap behavior
• Limit behavior: A-share price limits (typically ±10% or ±20%) can prevent stop execution; acknowledge this constraint
• Stop slippage: actual stop fills may differ from reference levels, especially in volatile conditions
• Benchmark/style mismatch: if the decision frame does not consider benchmark tracking or style constraints, state that explicitly
• Catalyst uncertainty: if the action depends on an expected catalyst, acknowledge timing and outcome uncertainty
• Regime sensitivity: if the decision is regime-dependent, state what regime change would invalidate the frame

Step 4: Frame as conditional guidance, never as instruction

The output must be presented as:

• "Conditional decision frame based on [stated assumptions]"
• "This is advisory guidance requiring human judgment, not an executable order"
• "Actual position sizing should incorporate portfolio-level constraints not modeled here"

Output Contract

• Minimum local executor output: human-readable text beginning with 决策支持.
• Fields: action, confidence, target position, quantity, reference price, stop loss, take profit, risk budget, and reasoning bullets.
• Side effects: updates session memory for the current execution context.
• Caller-facing delivery standard:
  • Three-part structure mandatory: every delivery must include (1) conditional action frame, (2) explicit assumptions, (3) non-modeled risks
  • Assumption transparency: distinguish user-provided capital, cash, position, risk, and max-position inputs from local defaults
  • Confidence labeling: treat confidence as "heuristic conviction score" and label it explicitly as such in every delivery; never present as calibrated probability or expected hit rate
  • Conditional framing: label the output as conditional on the current snapshot and stated account assumptions
  • Risk disclosure: surface non-modeled risks systematically, not optionally
  • Zero-quantity handling: if quantity is zero because the action is watch, hold, or avoid, or because a constraint binds, state the binding constraint explicitly
  • No order language: never use language that implies executable instructions, automated trading, or portfolio mandates

Failure Handling

• Parse and argument errors: non-zero exit with a readable 命令错误 message.
• Data-loading or normalization errors: readable failure text beginning with 执行失败:.
• Missing symbol with no reusable session symbol: readable guidance instead of a traceback.
• If material account context is missing, continue with conditional guidance rather than pretending exact portfolio fit.
• If risk disclosure cannot be completed due to missing data (e.g., volume history unavailable), state that gap explicitly.

Key Rules

• Respect lot-size and max-position constraints and state when they bind.
• Treat the output as a decision frame, not as an automated execution order.
• Keep the question of whether to own the name separate from the question of how to execute it.
• Route to strategy-design when the user wants execution mechanics after the decision is accepted.
• Systematic risk disclosure is mandatory, not optional. Every delivery must include Part 3 (Non-Modeled Risks).
• Default assumptions must be surfaced explicitly. Never let defaulted inputs masquerade as user-approved constraints.
• Confidence scores are heuristic conviction, never probabilities. This must be stated in every delivery.

Composition

• Builds on the same upstream data and snapshot logic as market-data and market-analyze.
• Often pairs with strategy-design or appears inside market-brief and analysis.
• Should feed into backtest-evaluator when retrospective evaluation is needed.