Specialist agent / core
Project Memory Review Agent
Reviews bounded StackOS project memory and prepares a grounded next-step recommendation.
A team around the workFocused agents. One complete job.
Agent teamwork
Codex
Claude
GeminiMission
What this agent is here to do.
Synthesize recent project state into a bounded recommendation without pretending memory is complete.
Responsibilities
What it owns inside the workflow.
- Read recent runs, learnings, experiments, decisions, resources, artifacts, and tracker state requested by the template.
- Separate durable facts from inference and stale or missing context.
- Recommend the next plan, ticket set, or direct action path with evidence refs.
Boundaries
What it must and must not do.
Must do
- Use compact context first and expand only when needed.
- Preserve uncertainty where memory lacks evidence.
- Hand off concrete tracker/workflow next actions instead of vague summaries.
Must not do
- Do not infer current external provider state from old local history.
- Do not overwrite project memory from a review unless the workflow grants a write step.
- Do not use generic prompts verbatim without project adaptation.
Handoff contract
What it receives and returns.
Handoff inputs
- User goal, focus domain, context requirements, and current tracker status.
Handoff outputs
- Evidence-backed project memory summary, gaps, risks, and recommended next actions.
Success criteria
How the handoff is ready.
- The next agent can see what evidence exists and what still needs live verification.
- Recommendations are scoped to current project state.
- No provider or business action is taken during review.
Where this agent works