Plan execute

Plan-Execute Skill v2 (KIK-609)

The Orchestrator leads a 7-agent team to carry out workflow design → execution → autonomous loop → review.

Trigger

Phrases such as "in plan mode," "plan it," "make a plan," "execute in plan mode," etc.

Agent Configuration (7 agents)

Agent	Type	Role	Participation
Orchestrator	Parent + Facilitator	Plan formulation, execution instructions, result evaluation, autonomous loop judgment, review integration	Always
Context Analyst	Execution support	Pre-analyzes historical context of financial markets + macro environment as a unit (leverages LLM knowledge)	On investment decisions
Strategist	Plan	Workflow design. Designs options from 4 perspectives: Growth / Value / Macro / Contrarian	Always
Lesson Checker	Plan	Checks whether past lessons (constraints) are reflected in the workflow	Always
Devil's Advocate	Plan	Points out blind spots, biases, oversights, and contrarian perspectives	Always
Quantitative Reviewer	Review	Quantitative checks + Gate Keeper function (numerical consistency, tax costs, currency allocation, lot size, constraint satisfaction, confirmation that all steps were executed)	Always
Qualitative + Risk Reviewer	Review	Qualitative checks (thesis alignment, respecting conviction, catalysts) + risk (geopolitical, macro, market overheating, PF structure, stock-specific)	Always

Orchestrator Rules

• Quantitative basis requirement: When rejecting an agent's input, always provide quantitative justification. "Feels small" or "impact is limited" are prohibited. Show PF ratios and comparisons against thresholds as numbers.
• Autonomous progression: Phase transitions proceed without user confirmation. Confirmation is only required for action execution approval in Phase 6.
• Progress display: Display a one-line progress update to the user at the start of each Phase:

  [1/6] Retrieving context...
  [2/6] Designing workflow...
  [3/6] Executing analysis...
  [4/6] Evaluating results (autonomous loop)...
  [5/6] Reviewing...
  [6/6] Generating final report...
  

Execution Flow

Phase 1: Pre-Plan

1. Run python3 scripts/get_context.py "<user input>" to retrieve graph context
2. Run python3 scripts/extract_constraints.py "<user input>" to get lesson constraints (when an investment decision is possible)
3. Reference user assumptions from config/user_profile.yaml (use defaults if file does not exist)
4. If an investment decision is involved → have Context Analyst retrieve historical context of financial markets
5. Recent event scan: Pre-check earnings dates and catalysts for held stocks via WebSearch or yfinance. If any stock has earnings within 7 days, instruct Phase 2 Strategist to "design workflow considering earnings timing."
6. Individual context retrieval for target stocks: In addition to the full PF query, run get_context.py for each stock likely to be an action target to retrieve past memos, theses, and lessons.

Context Analyst Perspectives

Category	Examples
Market cycle	"Fed rate hikes in 2022 caused growth stocks -30%"; "15 months after hike pause: +15%"
Theme history	"AI theme started in 2023, bubble concerns in 2024, re-evaluated on real demand in 2025"
Bubble patterns	"Dot-com bubble (2000) P/E >100 vs. current AI semiconductor P/E >100"
Geopolitical precedents	"2018 US-China trade war: AMZN -20%"; "2022 Russia-Ukraine: energy surge"
F&G history	"Average correction after F&G >80 sustained for 1 month: -8%, recovery 45 days"
Rate cycle	"Between rate hike pause and first cut, equities historically perform well"

Phase 2: Plan (3 agents in parallel)

Launch Strategist + Lesson Checker + Devil's Advocate in parallel.

Strategist's 4 Perspectives

The Strategist designs the workflow from 4 perspectives (not separate agents — one agent with 4 perspectives):

• Growth perspective: EPS growth rate, theme early-stage signals, Forward PER
• Value perspective: P/E, P/B, undervaluation score, dividend yield
• Macro perspective: Rate cycle, F&G, sector rotation
• Contrarian perspective: Is the analysis being swept along by consensus? Any oversights?

Strategist's Additional Required Perspectives

• Cash-equivalent inventory: For MMFs and short-term bond ETFs (SHV, etc.) in the PF, consider "continue holding (secure yield) vs. liquidate (improve flexibility)." Especially important in cash-preservation strategies.
• ETF action candidates: ETFs like GLDM/JEPI/SHV are also considered as action targets (even if health check shows "no issues," they can be sell candidates if not needed strategically).
• Role classification of all PF holdings: Classify as risk assets / safe assets / cash equivalents and explicitly state "reason to hold" for stocks not targeted for action.

The Orchestrator integrates results from the 3 agents and revises the workflow if Lesson Checker returns FAIL (up to 2 times).

Phase 3: Execute

Execute skills/scripts sequentially according to the plan.

Phase 4: Result Evaluation + Autonomous Loop

The Orchestrator evaluates execution results and autonomously performs additional execution and plan revisions as needed.

Evaluation Result	Action
No issues	Proceed to Phase 5 (Review)
Information gap detected	Run additional scripts pinpoint → return to Phase 4
New fact revealed (earnings date, etc.)	Revise plan → return to Phase 3
Action candidate detected	Auto-run what-if and generate numerical proposal → return to Phase 4

Autonomous Loop Examples

Example 1: Earnings date discovered Phase 3 runs health → Phase 4 reveals "NFLX earnings are today" → Orchestrator: "Remove NFLX take-profit from plan" → return to Phase 3 and re-execute with revised plan

Example 2: Unrealized gain concentration detected Phase 3 runs health → Phase 4 detects "AMZN unrealized gains 68% concentrated" → Orchestrator: auto-runs what-if --remove "AMZN:5" and "AMZN:7" → generates comparison table → proceed to Phase 5

Example 3: Theme gap detected Phase 3 runs health → Phase 4 detects "AI theme only, all other themes at 0%" → Orchestrator: adds theme-specific candidate screening → proceed to Phase 5

Autonomous Loop Limits

• Additional execution / plan revision: up to 2 times
• On the 3rd attempt, cut off and proceed to Phase 5

Autonomous Research Rules (strictly observed)

• When a new fact is detected (earnings result, surge/plunge, news) → immediately research via WebSearch. Do not ask the user "Should I check this?"
• Do not ask the user whether to collect information or conduct additional analysis
• Only ask the user for confirmation in Phase 6 (final summary) to obtain approval for action execution
• Phase 1→2→3→4→5→6 proceeds autonomously without user confirmation
• Users retain the right to interrupt (say "stop" at any time)

Information Collection Priority

Priority	Method	Time	Use Case
1st	WebSearch	Seconds	Immediate confirmation of earnings results, news, breaking news
2nd	yfinance	Seconds	Stock prices, financial data, earnings date retrieval
3rd	run_research.py	30-60s	Deep research via Grok API (only when detail is needed)

Phase 5: Review (2 agents in parallel)

Launch Quantitative Reviewer + Qualitative/Risk Reviewer in parallel.

Quantitative Reviewer Checklist (including Gate Keeper function)

Check	Verdict
Were all Orchestrator steps executed?	PASS/FAIL
Is there an action proposal when an issue is detected?	PASS/FAIL
Does the proposal include share count, amount, and tax cost?	PASS/FAIL
Are lot sizes correct? (Japan: 100 shares, SGX: 100 shares, etc.)	PASS/FAIL
Does currency allocation stay within 60% limit?	PASS/FAIL
Are user_profile assumptions referenced?	PASS/FAIL
Numerical consistency (what-if fund balance, HHI changes, etc.)?	PASS/FAIL
Is tax cost calculation accurate (including purchase FX rate)?	PASS/FAIL

Qualitative + Risk Reviewer Perspectives

Qualitative checks:

• Thesis alignment (is the take-profit reason based on thesis breakdown, or purely technical?)
• Respecting lesson/conviction (is the analysis rejecting a stock the user bought with conviction based on numbers alone?)
• Catalyst verification (have earnings dates, catalysts, and theme trends been confirmed?)
• Theme validity (does the recommended theme align with the market environment?)

Risk checks:

• Geopolitical risk (US-China tensions, Taiwan conflict, Middle East situation, sanctions → supply chain impact on PF stocks)
• Macro risk (interest rates, FX, inflation, recession probability)
• Market risk (F&G overheating, VIX spike, earnings season)
• PF structural risk (currency concentration, sector concentration, theme concentration, unrealized gain concentration)
• Stock-specific risk (liquidity, regulation, country risk)

Handling Review FAIL

• FAIL → Orchestrator identifies deficiencies and re-executes only the deficient parts (not a full redo)
• Example: Quantitative FAIL (tax cost not reflected), Qualitative PASS → add only tax cost calculation → re-review Quantitative only

Phase 6: Final Summary

All PASS → Orchestrator presents the final report in the following 8-section structure.

Required 8 Sections

1. Executive Summary (1-2 line conclusion)
2. PF Current Scorecard (one table: total value / number of stocks / P/L ratio / F&G / VIX / USD ratio)
3. Required Actions (in priority order; each action includes share count, amount, tax cost, and exit strategy)
4. Candidate Stock List (presented even if "not buying now." Includes minimum investment, theme, region, P/E)
5. Comparison with "Do Nothing" (expected value of each action vs. expected value of hold)
6. Risk Map (3 levels: high/medium/low. Geopolitical / macro / market / PF structure / stock-specific)
7. Unresolved Items (items flagged with WARN in Review)
8. Next Checkpoint (what to check and when)

Inter-Phase Output Schema

Each Phase passes the following required fields to the next Phase:

Phase	Required Output
Phase 1	context, constraints[], user_profile, upcoming_events[]
Phase 2	workflow_steps[], lesson_check: PASS/FAIL, devils_advocate_concerns[]
Phase 3	health_results, screening_candidates[], what_if_results[], market_data
Phase 4	revised_plan?, additional_findings[], action_proposals[], autonomous_research[]
Phase 5	quantitative: PASS/FAIL + reasons[], qualitative: PASS/FAIL + reasons[]
Phase 6	Required 8-section report

Retry Rules

Phase	Max Count	On Exceeding
Phase 2 Lesson Checker FAIL	2 times	Continue with WARN
Phase 4 Autonomous Loop	2 times	Proceed to Phase 5
Phase 5 Review FAIL	2 times	Output with WARN

3rd FAIL is cut off: "⚠️ The following items are unresolved but results are presented"

Issue Detection → Auto-Proposal Triggers (applied in Phase 4)

Detection	Auto-Proposal
Unrealized gains concentrated in 1 stock at >50% of PF unrealized gains	Specific partial take-profit plan (share count, sale proceeds, estimated tax cost)
RSI >70 + dead cross occurring simultaneously	Specific take-profit review plan (how many shares to sell, net after-tax proceeds)
Shareholder return rate declining for 3+ consecutive years	Specific sell plan (full sell or replacement candidate screening)
EXIT judgment in health check	Sell + screen 3 replacement candidates in same sector/theme
Theme gap	Present top 3 candidates by theme (with minimum investment amount)
F&G >80 + new buy-add proposal	Attach "market overheating" warning. Present comparison with cash-preservation strategy

Proposal Constraints

• Lot size validation: Japan stocks in 100-share lots, SGX in 100-share lots
• Currency allocation check: USD 60% cap. Warn if replacement is USD-denominated
• user_profile.yaml: Auto-calculate fees and tax costs (use defaults if file does not exist)
• F&G >80: Attach "market overheating" warning to new buy-add proposals

Required Elements for Each Proposal

All action proposals must include the following together:

• Exit strategy: Stop-loss line / take-profit target / review conditions
• Re-entry criteria: Conditions for buying back after take-profit (if applicable)
• Time limit: "Review if no change within N weeks"
• Candidate stock list: Reinvestment candidates after sale (present even when "not buying now")

Candidate stock list must include:

• Top 3–5 candidates derived by reverse-engineering PF gaps (region, theme, currency)
• Each candidate's minimum investment, theme, region, P/E, dividend yield
• Improvement effect on PF (currency allocation change, regional diversification improvement)
• When cash preservation is recommended: re-entry trigger conditions (F&G, RSI, VIX, etc.)

Escalation Criteria

Convene Context Analyst and run the full 3-agent parallel Plan Phase when any of the following apply:

• User's intent involves buying, selling, replacement, rebalancing, or adjustment
• extract_constraints.py returns action_type: swap_proposal / new_buy / sell / rebalance / adjust
• The plan contains what-if / adjust / rebalance commands
• An action proposal is generated during Phase 4's autonomous loop

For information queries only (snapshot, analyze, health, etc.), Context Analyst is not needed and Plan Phase can run in lightweight mode (Strategist only).

Available Skills / Scripts

Skill	Script	Purpose
screen-stocks	run_screen.py	Screening
stock-report	generate_report.py	Individual stock report
stock-portfolio	run_portfolio.py	PF management (snapshot/analyze/health/forecast/what-if/adjust/rebalance/simulate/review)
stress-test	run_stress_test.py	Stress test
market-research	run_research.py	Market / industry / stock research
watchlist	manage_watchlist.py	Watchlist
investment-note	manage_note.py	Investment memos
graph-query	run_graph_query.py	Knowledge graph search
—	market_dashboard.py	Market conditions dashboard
—	get_context.py	Graph context retrieval
—	extract_constraints.py	Lesson constraint extraction