Dlab CLI

decision-lab (dlab)

dlab runs autonomous coding agents in frozen Docker environments with domain-specific skills and parallel subagents. You package the environment, prompts, and skills into a decision-pack, point it at data, and get back reports and recommendations that hold up to scrutiny.

When to use this skill

• Creating a new decision-pack (interactive or programmatic)
• Designing agent system prompts for data science workflows
• Understanding how parallel agents, consolidators, and retry protocols work
• Analyzing a completed session's logs, outputs, and artifacts
• Running or configuring dlab CLI commands

Workflow overview

1. Create a decision-pack — scaffold with dlab create-dpack wizard or generate_dpack() programmatically
2. Design agents — write orchestrator, subagent, and parallel agent configs
3. Run a session — dlab --dpack <path> --data <data> --prompt "..."
4. Monitor — dlab connect <work-dir> (live TUI) or dlab timeline <work-dir> (Gantt chart)
5. Analyze results — browse session directory, logs, parallel instance outputs

Key concepts

decision-pack: A directory containing config.yaml, docker/, and opencode/ (agents, skills, tools, permissions). Everything an agent needs to run.

Orchestrator (mode: primary): Coordinates the workflow, spawns parallel agents, evaluates results, writes reports. One per decision-pack.

Subagents (mode: subagent): Execute focused tasks. Each runs ONE strategy per run. If it fails, it writes diagnosis and stops — the orchestrator coordinates retries.

Consolidator: Auto-generated read-only agent that compares parallel instance results. Never picks a winner.

Parallel exploration: Fan out multiple agents with structurally diverse approaches (different priors, models, data prep). Check if they agree before recommending.

Critical methodology rules

These are non-negotiable for any data science agent system:

1. Never fabricate — no mocking data, no silently swallowing errors, no try/except: value = 0
2. Understanding over fitting — a model that doesn't converge is evidence, not failure
3. Know when to stop — hard round limits, conflict detection, degenerate problem reports
4. Templates, not implementations — no concrete numbers in prompts, use <PLACEHOLDER> syntax
5. Uncertainty, not point estimates — always report intervals, distinguish model vs structural uncertainty
6. Recommendations must be computed — no napkin math, multiple scenarios, realistic actions
7. Document everything — including failures, two reports (business + technical)

Load references/agent-design.md for the full methodology guide.

References

Load these as needed — don't read all upfront:

• references/agent-design.md — Full methodology: anti-fabrication, retry protocol, epistemic humility, conflict detection, prompt design, parallel exploration, degenerate problems, agent prompt structure, runtime directory layout, YAML config
• references/create-dpack.md — Programmatic decision-pack creation: generate_dpack() API, config keys, package managers, permissions, Modal integration
• references/create-dpack-interactive.md — Interactive wizard guide: how to interview a user and call generate_dpack() with the right config
• references/run-analyzer.md — Session analysis: directory layout, log format (NDJSON events), how to navigate parallel runs, what to look for