How DormWay Builds With AI Agents: Inside Our Dev Workflow

DormWay treats AI as a real teammate — not a chatbot you occasionally copy-paste errors into. We built an engineering workspace where agents can track context, run tests, search our internal knowledge base, and hand off work cleanly across sessions. It’s a big part of how two people shipped an 80k-line, production-ready platform in six months.

1. Session Commands: The Guardrails

Every dev session starts with a slash command. Think of them as rituals that stop the AI from guessing and force clean boundaries.

/session-start

• Reads progress files and git status
• Scans Obsidian worklogs
• Checks active feature checklists
• Outputs a structured session briefing

/session-end

• Runs verification (tests/build)
• Blocks session end if anything fails
• Updates feature checklists
• Appends a daily worklog summary
• Creates a clean handoff for the next session

Rule: if tests aren’t green, the session can’t end.

2. Context Loading: Preventing AI Guesswork

/load-context [topic]

Topics: temporal, auth, api, ios, database, realtime, playbook

• Searches the Obsidian vault
• Extracts constraints, patterns, and past decisions
• Responds with DormWay-specific architectural guidance

/search-docs [query]

BM25 → rerank → MMR diversity retrieval:

• High‑signal matches
• Proximity‑boosted results
• Non‑redundant selections

3. Self-Correcting Test Harness

scripts/smart-run.sh automatically:

1. Captures + filters logs
2. Extracts error signatures
3. Searches for similar past failures
4. Truncates logs (first 100 + last 1000 lines)
5. Generates a context file for /analyze-logs

The agent sees not just what failed — but whether we solved it before.

4. Log Capture With PII Redaction

scripts/capture-logs.sh server --clean --redact

• Redacts emails, student IDs, JWTs, API keys
• Produces AI‑safe logs
• scan-risks.sh checks prompts for PII

Temporal Workflow Debugging

• Captures workflow history
• Compresses ~90%
• Extracts failure chain, stack traces, retry state

5. Parallel Development With Git Worktrees

./scripts/claude-start.sh dormway-platform feat/my-feature cursor

Each worktree includes:

• Independent Claude context
• Auto‑generated TASK.md
• Isolated git state
• VS Code workspace config

This keeps multiple AI‑assisted features running in parallel without context bleed.

6. JSON Feature Tracking

Features live in .workspace/features/*.json.

Why JSON? Models corrupt Markdown checklists but respect schema.

Includes:

• Feature list w/ completion flags
• Verification command + status
• Session logs

/session-end only completes when verification passes.

7. Obsidian Vault as Long‑Term Memory

Stores:

• Daily worklogs
• Architecture docs
• Error history
• Playbooks w/ acceptance criteria

Search via retrieve.js:

• BM25 + cosine scoring
• Phrase/proximity boosts
• Path/date/tag filters

8. Core Principles

1. Context first — agents never guess.
2. Verification interlocks — no shipping broken code.
3. PII protection — FERPA‑safe by default.
4. Parallel development — worktrees prevent context loss.
5. Error memory — reuse old fixes.
6. Structured handoffs — no lost context.

Why This Matters

DormWay didn’t get built with heroic all‑nighters. It got built with a system that makes two people feel like ten — a development loop where AI is a teammate with its own memory, rituals, and responsibilities. Want to work on this with us?