Curiosity Backlog

Prioritized frontier threads from SWAP and the current AI/ notes.

1. Agent Workflow Observability for Personal Runtimes

Core question: What is the smallest useful observer system for improving personal agent workflows?
Why interesting now: OpenClaw Observer VM for Agent Observability already frames an async reviewer, evaluator loops, event schemas, and scorecards. Pathfinder Orchestration Architecture adds the orchestrator/worker split that would generate meaningful traces.
Source links: OpenClaw Observer VM for Agent Observability, Pathfinder Orchestration Architecture
Related concepts: LLM observability, evaluator loops, process mining, agent trajectory analysis, Tailscale, OpenClaw
Suggested next exploration: Draft an MVP event schema and scorecard for completed Codex/OpenClaw sessions, then test it on one historical session.
Confidence: high
Status: active

Core question: Can you build a self-hosted personal spending pipeline that captures transactions early, reconciles them later, and tracks reimbursements/reward limits?
Why interesting now: SWAP contains an unusually concrete project seed with constraints, source links, ingestion paths, MVP question, and good-to-have features.
Source links: SWAP
Related concepts: email parsing, bank statement ingestion, Apple Shortcuts, reconciliation algorithms, SGFinDex, credit card rewards, personal analytics
Suggested next exploration: Build a source-of-truth model first: transaction, evidence, account/card, merchant, reimbursement, category, reward rule, reconciliation status.
Confidence: high
Status: active

Core question: What does a trace-native MCP layer need beyond raw ClickHouse SQL access?
Why interesting now: ClickHouse MCP for Dataservice Trace Analysis argues that the winning pattern is MCP over curated trace abstractions, not unrestricted SQL over raw span tables.
Source links: ClickHouse MCP for Dataservice Trace Analysis, OpenClaw Observer VM for Agent Observability
Related concepts: ClickHouse, OpenTelemetry, MCP, semantic layers, incident copilots, span trees, service graphs
Suggested next exploration: Define 6-8 trace-specific MCP tools such as find_slow_services, get_trace_tree, compare_latency_before_after, and find_error_clusters.
Confidence: high
Status: active

Core question: Can Cursor act as the primary runtime behind OpenClaw well enough that Codex is no longer required in normal usage?
Why interesting now: Bastion - Cursor as Primary Runtime Plan lays out a concrete experimental deployment plan, while Pathfinder Orchestration Architecture explains why worker/runtime boundaries matter.
Source links: Bastion - Cursor as Primary Runtime Plan, Pathfinder Orchestration Architecture
Related concepts: ACP, Cursor, OpenClaw, Bastion, runtime providers, session persistence, Telegram gateway
Suggested next exploration: Convert the Bastion phases into a small checklist with explicit exit criteria and unknowns for each phase.
Confidence: high
Status: active

Core question: What state should a personal R&D companion maintain so it feels continuous rather than stateless?
Why interesting now: Pathfinder - Proposed Curiosity Agent Stack and Pathfinder - Curiosity Stack Implementation already define Notes Distiller, Frontier Scout, Curiosity Ranker, Topic Composer, and state artifacts.
Source links: Pathfinder - Proposed Curiosity Agent Stack, Pathfinder - Curiosity Stack Implementation
Related concepts: interest maps, open loops, backlog freshness, anti-slop ranking, recently explored memory
Suggested next exploration: Define a simple scoring rubric for backlog candidates and apply it to this curiosity-backlog.md.
Confidence: high
Status: active

Core question: Which backend internals would most improve your practical system-design depth?
Why interesting now: SWAP clusters several topics: Kafka, Elasticsearch, OLAP vs OLTP, Snowflake, PostgreSQL/serverless coupling, SSO, Spring, JVM/GC, VPCs, Kubernetes, sidecars, SMS vs APN.
Source links: SWAP
Related concepts: distributed systems, storage engines, backend frameworks, cloud networking, messaging systems
Suggested next exploration: Pick one spine topic, probably “OLAP vs OLTP and storage engines”, then link Kafka, ClickHouse, Snowflake, and Elasticsearch around it.
Confidence: medium
Status: unexplored

Core question: How far can Obsidian plus agent-maintained notes go as a personal operating system for research, projects, and memory?
Why interesting now: This R&D frontier folder is itself a test of the pattern. Pathfinder - Proposed Curiosity Agent Stack expects durable artifacts; SWAP is a backlog of unstructured intentions.
Source links: Pathfinder - Proposed Curiosity Agent Stack, SWAP
Related concepts: Obsidian wikilinks, durable Markdown, agent-maintained indexes, source-linked synthesis
Suggested next exploration: Compare which tasks belong in handwritten source notes versus AI-maintained synthesis notes.
Confidence: medium
Status: active

Core question: When does being informed create better decisions rather than just more consumption?
Why interesting now: SWAP asks “does it pay to be informed?”, “Best ways to be kept updated with the news”, and several company/market-history questions. This could become a meta-framework for choosing what to read.
Source links: SWAP
Related concepts: information value, technical taste, company history, markets, decision quality
Suggested next exploration: Write a note separating information that changes action from information that only feels useful.
Confidence: medium
Status: unexplored

Core question: Can an Obsidian plugin make local graph exploration more useful by highlighting a node and its neighbors?
Why interesting now: SWAP has a concrete plugin idea that connects directly to the vault-as-system direction.
Source links: SWAP
Related concepts: Obsidian plugin API, graph view, local neighborhoods, note navigation
Suggested next exploration: Check whether Obsidian exposes enough graph-view internals; if not, prototype a separate note-neighborhood view.
Confidence: medium
Status: unexplored

Core question: Can chat history become a useful LLM context source without turning private, messy conversation logs into an unsafe memory dump?
Why interesting now: SWAP includes “use the chat history to create a LLM bot”, and the AI notes repeatedly explore personal agents, memory, and orchestration.
Source links: SWAP, Pathfinder - Proposed Curiosity Agent Stack
Related concepts: memory extraction, privacy boundaries, retrieval, summarization, personal assistants
Suggested next exploration: Define what should be extracted from chat history: preferences, project state, recurring questions, or conversational style.
Confidence: medium
Status: unexplored