architect
Ask when consequential design choices must be settled before or during implementation — investigates context, compares…
Sessions are stored per project data directory, not directly in the genome root. The data directory contains session metadata, event logs, project memory files, and per-session subprocess logs.
How it fires
How this agent gets triggered: by you, by Claude, or both.
Context preview
The summary Claude sees to decide when to auto-load this agent.
Sessions are stored per project data directory, not directly in the genome root. The data directory contains session metadata, event logs, project memory files, and per-session subprocess logs.
Sessions are stored per project data directory, not directly in the genome root. The data directory contains session metadata, event logs, project memory files, and per-session subprocess logs.
Common paths:
Source of truth:
Metadata stores session id, root agent, status, turn/context counts, model selection, resolved model, timestamps, and cached root memory surfacing data. If a previous run crashed while status was `running`, resume marks it `interrupted` before starting new work.
Source of truth:
Agents emit structured events such as session start/end, perceive, recall, plan start/end, act start/end, verify, warnings, errors, and learn events. The current event taxonomy lives in source, not this document.
Source of truth:
Resume reconstructs message history from the JSONL event log, loads metadata for model selection and cached memory surface, extracts child handles, and loads session plus child event logs for the UI.
Source of truth:
There are two compaction concepts:
git history contains their audit trail.
Source of truth:
After a completed root run, session events can be summarized into a segment and extracted memories. This uses configured memory summary/extraction models and records project activity.
Source of truth:
Session JSONL can contain large prompts, tool calls, and full LLM messages. Prefer metadata-first filtering and command-runner extraction over loading raw logs directly into an agent context.
A self-improving, multi-agent AI system that decomposes tasks and delegates to specialist agents. Sprout is an autonomous coding assistant built on a recursive agent architecture — where every action is a goal-directed delegation to a subagent.
Ask when consequential design choices must be settled before or during implementation — investigates context, compares…
Memory investigation, synthesis, and authorized curation when surfaced memories are insufficient
Systematically diagnose and fix bugs: find root cause before attempting any fix
Observe Sprout's live session behavior and send concise guidance when it is drifting, stuck, or missing an important instruction
Explore and understand unfamiliar codebases, map architecture, and synthesize structural insights
Ask when you need to know what Sprout agents, tools, or MCP servers are available, need a reusable capability plan, or need a new…