The Skills Registry: Hook Points, Auto-Activation, and 38 Skills

The pipeline in agent.py doesn't import skill implementations directly — it asks skills.py which skills are active for a given task, then calls them at the appropriate stage. This indirection means adding a new skill is a two-step operation: register it in the REGISTRY dict, then implement it. The pipeline never needs to change.

Six hook points

The full registry

Auto-activation in practice

Five skills activate without a slash command. Each has an auto lambda that receives the task string and the planner's Plan object:

# /deep: force max search rounds when task implies thoroughness
"auto": lambda task, plan: bool(re.search(
    r"\bcomprehensive\b|\bthorough\b|\bexhaustive\b|\bin.depth\b|\bdeep.dive\b",
    task, re.IGNORECASE
)),

# /panel: run 3-persona evaluation when planner sees a complex task
"auto": lambda task, plan: plan.complexity == "high",

# /contextualize: inject self-knowledge when task is self-referential
"auto": lambda task, plan: bool(re.search(
    r"\byour(?:self)?\b|\bwhat (?:can|are) you\b|\bthe harness\b",
    task, re.IGNORECASE,
)),

The panel's auto-activation is the most interesting case: it depends on the planner's output, which means the pipeline runs two LLM calls (planner) before it knows whether a third post-synthesis call (panel) will be needed. This is by design — the planner's complexity classification is more reliable than keyword matching for detecting tasks that warrant multi-perspective evaluation.

Lazy loading

Importing skills.py does not import any skill implementation. The REGISTRY is pure data — strings, lambdas, and None. When agent.py needs to actually run a skill, it calls skills.run_post_synthesis(), skills.run_annotate_standalone(), etc., which import the relevant module at call time. This keeps startup fast and avoids circular imports between agent.py and the skill modules that call back into inference.py.