Skip to content

feat: ambient mode detection tweaks — refine lean classification approach #178

@dean0x

Description

@dean0x

Overview

After working with the new lean ambient mode architecture (SessionStart classification rules + UserPromptSubmit preamble + router SKILL), there are remaining tweaks to refine. The core approach is solid but needs tuning based on real usage.

Current Architecture

  • SessionStart hook injects ~30-line classification rules as additionalContext (once per session)
  • UserPromptSubmit hook injects one-sentence preamble per message triggering classification + router loading
  • Router SKILL is a pure lookup table (~50 lines) mapping intent×depth to domain and orchestration skills
  • Three-tier depth: QUICK (direct response), GUIDED (focused skills), ORCHESTRATED (full agent pipelines)

Areas for Potential Tweaks

Classification accuracy, depth calibration, edge cases, and router mappings may need adjustment after real-world usage. Specific areas to evaluate:

  • Intent signal accuracy — are prompts being classified correctly?
  • Depth boundaries — is the QUICK/GUIDED/ORCHESTRATED split calibrated right?
  • Edge cases — ambiguous prompts, multi-intent messages, follow-ups
  • Router skill mappings — are the right skills loading for each intent×depth?
  • Preamble wording — is the one-sentence injection effective?

Tasks

  • Collect observations from real ambient usage sessions
  • Identify misclassification patterns (wrong intent or wrong depth)
  • Propose and apply classification rule adjustments
  • Propose and apply router mapping adjustments
  • Verify changes don't regress existing correct classifications

References

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions