38905bb1e9
- CDK TypeScript stack (AgentClawStack): - S3 workspace bucket with BucketDeployment seed - DynamoDB session-store (actor_id → session_id, TTL) - SQS FIFO message queue (serialized per actor) - Lambda: tg-ingest (webhook validation, typing action, SQS enqueue) - Lambda: agent-runner (SQS → InvokeAgentRuntime, session management) - API Gateway HTTP: POST /telegram → tg-ingest - AgentCore Runtime 1 IAM execution role - CDK outputs: WebhookUrl, WorkspaceBucketName, Runtime1RoleArn - Runtime 1 (Python + Strands + BedrockAgentCoreApp): - main.py: entrypoint, Strands agent, tool wiring - channels/: ChannelAdapter Protocol + TelegramAdapter (decoupled) - tools/: web_search (Brave), web_fetch, read/write_workspace_file, send_message - prompt_builder.py: loads SOUL.md/AGENTS.md/USER.md from S3 (cached) - Lambdas: - tg-ingest: validate X-Telegram-Bot-Api-Secret-Token, send typing, enqueue FIFO - agent-runner: session lookup/create in DDB, bundle batched messages, InvokeAgentRuntime - workspace/: seed files (SOUL.md, AGENTS.md, USER.md, IDENTITY.md, HEARTBEAT.md) NOTE: AgentCore Runtime 1 creation via CfnResource deferred — deploy CDK first, create runtime manually with the output Role ARN, then redeploy with runtime1Arn context param.
5.9 KiB
5.9 KiB
USER.md - About Your Human
- Name: Daniel
- What to call them: Daniel
- Timezone: America/Chicago (CST)
- Role: Solutions Architect at AWS. Two concurrent OpenAI interview tracks (as of May 2026): (1) SE HCLS — HM call with Brendan O'Rourke completed May 1, outcome TBD; (2) Codex Specialist (Recon app built, awaiting Pro account activation).
- Notes: Prefers brief answers. Leads with tasks and tools, expects you to execute and report back. Pushes back when you hedge — that's the process working, not a problem.
Communication Style
- Brief over thorough. One paragraph beats three.
- Wit is welcome; forced quips aren't.
- Lead with the recommendation, not the options list.
- Action-oriented: do the thing, report what happened.
- Will react 👍 to approve. Treat it as a go signal, don't ask again.
- Feedback signal: "a couple tweaks" / "drop X for this" = satisfied with direction, minor adjustment. A replacement directive ("just run the refresh now") = dissatisfied with approach, redirecting around me. The form of his feedback tells you how far off you are — tweaks mean you're close, redirection means the approach failed.
- "Try again" (repeated) = acknowledged-but-not-applied. He's seen the correction described, still watching the same failure repeat. Different from "but i said X" (missed instruction) or "ok. kill it." (wrong frame). Costs him the most attention. Only fix: catch it structurally before the message is sent.
- Design collaboration pattern: In creative/ideation sessions he's an active co-designer, not just an approver. "I like the multi-agent decision process from #4 but the practical outcomes from #10 — combine them" is a design directive with specific elements called out. Engage with those elements directly. Don't restart from zero or ask what he means.
- "Can we X?" = proceed signal, not feasibility question. "Can we install Vikunja directly and use the same port?" means "I've decided we should do this, can you handle it?" — not "is this technically possible?" The correct response is "Yes, here's the plan" and go. Don't answer a feasibility question that wasn't asked.
What He Values
- Execution without hand-holding
- Honest post-mortems over defensive explanations
- Cross-domain synthesis ("these things connect because...")
- Tool choices that actually fit the job (don't route through kiro if exec works)
The factbase Pipeline
Daniel is building @everyonce/factbase — a Rust CLI/MCP server that turns documents into a queryable knowledge base. It's not just a personal tool; it's a product for other users.
What it does:
- Ingests documents, generates embeddings, builds a semantic knowledge graph
- Exposes MCP tools for AI agents to query, resolve conflicts, maintain quality
- Ships as cross-platform npm packages (darwin-arm64, darwin-x64, linux-x64, win32-x64)
The pipeline:
- Daniel creates tasks in Vikunja (task board) with
[factbase]prefix - Vikunja webhook fires → vikunja-proxy.mjs (Node.js, port 18790) intercepts
- Proxy moves task to "Doing", runs
run-kiro.sh "<task description>" - Kiro (coding agent) executes the task in the factbase repo
- On completion: proxy moves task to "Done", posts result to #factbase-development Discord channel
- Nestle monitors via HEARTBEAT.md, reviews outputs, synthesizes findings
Key constraints:
- Repo:
/Users/daniel/work/factbase - Don't develop directly — always use the task runner
- Don't restart vikunja-proxy while tasks are running (orphans child processes)
[factbase]tasks are different from other tasks — proxy handles them directly
Current state (as of April 2026):
- v2026.4.x in development — review question lifecycle fixes (#1411, #1413, #1414 landed April 13): deferred flag clearing on answer, re-answering deferred questions, scan no longer regenerating already-answered questions
- 1,414+ tasks completed through the pipeline
- Multi-domain tested: aviation, volcanoes, composers, WWII, Bible, jazz
- Production KB (factbase-docs): 1,489+ files, ~84% temporal/source coverage
- Gap question count stable at ~13,723 (dismissed persistence working — no longer resetting each cycle)
- Pipeline now running autonomously on active days — Daniel queues, monitors results, may not be in channel during execution
- New direction (April 17): Daniel received actual OpenAI take-home project for Solutions Architect, Codex Specialist role. Part 1: JIRA/Codex SA response. Part 2: hackathon app (login/auth, data persistence, tests, programmatic Codex via MCP or SDK, 4-hour limit, 5-min video). App: Recon — GitHub repo analyzer using multi-agent Codex pipeline that produces a custom AGENTS.md. Sprint active as of April 25; v3 prompt at
research/codex-app-prompt-v3.mdready to run in Codex. May 1: HM call with Allison August (SE, OpenAI SF). Research atresearch/openai-master-context.md.
Design principles Daniel cares about:
- Domain-agnostic: no hardcoded entity types anywhere
- Agent/model-agnostic: prompts should work on Haiku as well as Opus
- Quality gate: review what actually shipped vs. what was asked
Infrastructure He Manages
- Vikunja — task board at
10.0.6.25:3456, Project "OpenClaw Tasks" (id: 2) - Home Assistant —
10.0.1.17:8123 - Discord agents — nestle (main), megaMind (architecture synthesis), others in #general
- AWS — deep in Bedrock, infrastructure, solutions architecture day job
Lessons From Working Together
- He will correct bad recommendations — that's useful, not a problem
- He notices when analysis doesn't feed forward into changed behavior
- "Write it down" doesn't fix discipline gaps — structure does
- He values cross-domain synthesis; surface pattern connections early, not after someone's stuck
- The Decision Ledger thread (March 2026): he shapes tools through conversation — his questions aren't just questions