buildTestArgs generates test arguments, using tempFileName for file path params.

buildTestArgsDefault generates default test arguments.

classifyTaskType determines the cognitive type of the incoming task. This drives effort allocation, model selection, and confidence calibration.

cleanCodeResponse strips unnecessary content from code responses.

cleanReasoningResponse cleans up reasoning responses.

cleanSummaryResponse ensures summaries are concise.

Get common failure patterns for a specific tool.

Get recall results as a formatted context string. Ready to inject into an LLM prompt. Returns: A string like:

Count memories in a namespace.

Run decay pass — remove memories that have decayed below threshold. Returns number of memories removed.

detectCodeLanguageFromContent guesses language from code content.

detectLanguage extracts a programming language name from a natural language query.

Phase 4: Execute + Test

explainChoice generates a human-readable reason for the routing decision.

extractOwner pulls the org/owner from a repo item for diversity checks.

findAlternative finds the best alternative model. Must be called with r.mu held.

Forget all memories in a namespace.

Ingest from all available Runtime files at once. Call this once at startup to bootstrap experience from existing data.

Get performance stats for a model across tasks.

Semantic recall — find memories most relevant to the query. Multi-signal ranking: - Semantic similarity (cosine) — 50% wei

Recall across ALL namespaces for this agent. Single query, no per-namespace loop.

stripNoiseWords removes common filler words from a query, keeping only meaningful domain keywords. GitHub search works best with focused terms.

toolNames returns sorted list of registered tool names (for error messages).

trimToolOutput reduces raw tool output to essential content to save tokens. JSON arrays get truncated to maxItems. Long strings get cut. This alone ca