MiroFish Market Validation Integration

MiroFish provides swarm intelligence market validation via simulated user/market

reception. Results are advisory only and never gate the RARV cycle or

Completion Council decisions.

When This Module Applies

Load this module when ANY of the following are true:

• Your prompt contains a MIROFISH MARKET VALIDATION section
• The project has .loki/mirofish-context.json
• The session was started with --mirofish flag
• Tasks in .loki/queue/pending.json have source: "mirofish"

If none of the above are true, do not load this module.

Understanding MiroFish Advisory Data

When present, MiroFish context includes:

• overall_sentiment: positive, negative, or mixed
• sentiment_score: 0.0-1.0 (higher = more positive)
• confidence: low, medium, or high (based on simulation depth)
• key_concerns: Risk items identified by simulated personas
• feature_rankings: Features ranked by simulated reception score
• recommendation: proceed, review_concerns, or reconsider

Risk-Driven Prioritization

When MiroFish identifies key concerns, address them proactively:

Sentiment-Aware Feature Ordering

Use feature reception scores to influence implementation order:

• Score >= 0.8: High demand -- implement first (validated demand)
• Score 0.6-0.79: Moderate interest -- standard priority
• Score 0.4-0.59: Lukewarm -- consider redesign before building
• Score < 0.4: Low reception -- deprioritize or cut from MVP

Confidence Score Interpretation

Pipeline Status

If your prompt contains MIROFISH_STATUS instead of full validation data, the

pipeline is still running. This is normal -- MiroFish simulations take 15-45 minutes.

• Continue with implementation using available context
• Results will appear in subsequent iterations when the pipeline completes
• Do not wait or block for MiroFish results

What MiroFish Does NOT Do

• Does NOT override Completion Council decisions
• Does NOT block the RARV cycle
• Does NOT evaluate code quality (that is the quality gates system)
• Does NOT make architecture decisions (those are yours)
• Results are directional signals, not precise measurements

Known Limitations

• LLM convergence bias: Simulated agents may converge on similar opinions
• Garbage-in-garbage-out: Poor PRD = poor simulation seed = unreliable results
• No real user data: Synthetic personas, not actual market research
• Cultural bias: Agent behavior reflects LLM training data distributions
• Timing: Results arrive 15-45 min after launch, may miss early iterations

Configuration

MiroFish requires these environment variables for its Docker container:

• LLM_API_KEY (required) -- API key for MiroFish's LLM provider
• ZEP_API_KEY (required) -- Zep Cloud key for graph memory

CLI flags:

• --mirofish [URL] -- Enable with optional custom URL (default: localhost:5001)
• --mirofish-docker IMAGE -- Auto-start MiroFish Docker container
• --mirofish-rounds N -- Simulation rounds (default: 100)
• --mirofish-timeout S -- Max pipeline wait (default: 3600s)
• --mirofish-bg -- Run in background mode
• --no-mirofish -- Disable even if LOKIMIROFISHURL is set

MiroFish Task Queue

When MiroFish completes, actionable items from the report are added to the task

queue with source: "mirofish". These tasks represent risk mitigations and

recommendations. They are advisory -- prioritize alongside normal development tasks

based on the concern severity and your judgment.