aitp-load-profile-resolver

Determine whether a topic should use the full or light runtime load profile, using semantic reasoning instead of regex matching.

AITP Load Profile Resolver

Environment gate

  • This skill runs inside an AITP session, typically at session-start or resume.
  • Research mode should already be classified before this skill is invoked.

Classification task

Choose between two load profiles:

ProfileWhen to choose
fullThe topic requires deep runtime infrastructure: full action queue materialization, detailed trust audits, promotion candidates, multi-step validation loops, or subtopic decomposition. Typical for formal_derivation and first_principles topics with bounded research questions.
lightThe topic is exploratory, early-stage, or does not yet have a well-bounded question. The runtime stays minimal: topic state, operator console, research question contract, control note. No full queue materialization or trust audit scaffolding.

Reasoning priority

  1. Explicit request: If the user said "full profile", "deep dive", "full runtime", or explicitly requested comprehensive infrastructure, choose full.
  2. Research mode signal: formal_derivation and first_principles topics with bounded questions usually need full. exploratory_general and early toy_model topics often work with light.
  3. Question boundedness: Is the research question well-scoped with clear target claims and a validation route? Yes → full. No or still being scoped → light.
  4. Topic maturity: A topic that has already produced multiple action rounds or has L2 promotion candidates should use full.
  5. Default: Choose light. Upgrading later is cheap; running full prematurely wastes resources.

Recording the classification

After reasoning, call the MCP tool:

aitp_record_classification(
    topic_slug=<current topic>,
    classification_type="load_profile",
    value=<"full" or "light">,
    rationale=<1-2 sentence explanation>,
    signals_used=<list of signals>,
    source="ai_reasoning"
)

Hard rules

  • Do not use regex matching or keyword scanning on the human request text.
  • The profile can be upgraded from light to full at any point; record a new classification when this happens.
  • Never downgrade from full to light within the same session without explicit user direction.