Observe

Capturing recurring workflow observations into structured JSONL records requires normalizing the agent behavior into a schema with timestamps, source, pattern summaries, and evidence references, then appending it to a date-partitioned file under .autoharness/continuous-learning/observations.

Stopping repeated workarounds in agent runs involves turning recurring workflow lessons into durable, searchable learning records. By capturing and normalizing these observations, repeated mistakes are clustered and reused rather than re-discovered across planning, build, review, verification, and closure scenarios.

Evidence normalization during observation capture strips secrets, large payloads, raw credentials, and noise from the data. This safety-conscious process preserves concise citations and artifact references, ensuring secure evidence handling while maintaining useful workflow context.

Manual capture of workflow observations is fully supported when automated hooks are unavailable. The system works across multiple capture contexts including manual review, runtime verification, closure, and memory, falling back to manual entry to ensure recurring patterns are still recorded.

Continuous learning records for agent behavior are needed when the same workaround or practice keeps repeating across multiple runs. They prevent repeated mistakes by turning these recurring workflow patterns into structured, searchable, and reusable instincts for future sessions.

Observation capture works across multiple workflow phases including planning, build, review, verification, and closure. It normalizes signals from various contexts into a structured schema, ensuring recurring agent behaviors are consistently recorded regardless of the workflow phase.