types
diapason.agents.hybrid.skillorchestra.types
¶
Core data types for SkillOrchestra.
- Skill
- AgentProfile
- BetaCompetence
- ModeMetadata
- RoutingInsight
- CostStats
AgentProfile
dataclass
¶
Agent profile for skill-aware orchestration.
overall_success_rate
property
¶
Overall success rate (trajectory-level when available, else skill-level).
category_competence(category_prefix)
¶
Aggregate competence on all skills under a category (skill_id prefix).
E.g. category_competence('entertainment_knowledge') = avg of get_competence(s) for all s where s.startswith('entertainment_knowledge.').
category_competence_for_skills(active_skill_ids)
¶
Category-level competence for hierarchical tie-breaking.
Extracts parent categories from active_skill_ids (e.g. 'entertainment_knowledge' from 'entertainment_knowledge.episodic_competition_outcome'), computes category_competence for each, returns average.
get_competence(skill_id)
¶
Get empirical success rate for a skill. Returns 0 if unseen.
get_competence_dist(skill_id)
¶
Get full Beta distribution for a skill, creating with prior if unseen.
update_competence(skill_id, success)
¶
Update competence estimate for a skill.
weighted_competence(skill_weights)
¶
Compute weighted competence: sum w_{t,sigma} * alpha/(alpha+beta).
BetaCompetence
dataclass
¶
Bayesian competence estimate for an agent on a specific skill.
skill_scores / get_competence use empirical_rate (successes/attempts)
CostStats
dataclass
¶
Execution cost statistics for an agent under a specific mode.
Tracks both total cost (prompt + completion) and completion-only cost separately, since completion cost is the variable component that differs most between models (prompt cost is roughly constant for the same query).
update(prompt_tokens, completion_tokens, latency_s, cost_usd, completion_cost_usd=0.0, prompt_cost_usd=0.0)
¶
Incremental running-average update.
ModeMetadata
dataclass
¶
Mode-level routing metadata.
RoutingInsight
dataclass
¶
A single routing insight learned from execution traces
Skill
dataclass
¶
SkillProvenance
dataclass
¶
Tracks how and why a skill was discovered.