dteam

A Pi extension: lets Pi gather facts, explore evidence-backed candidate paths, and do mechanical work on cheap small models while your strong model judges and closes — less token, less time.

What it is

dteam is an extension for Pi. Once installed, when the main model faces a non-trivial coding task (reading code, gathering facts, exploring technical paths, making changes, running checks), it can dispatch fact-finding, evidence-backed candidate exploration, and mechanical work to cheap, fast small models (T3) and get reports back, escalating to a standard model (T2) or flagship (T1) when stronger capability is needed. The strong model you're using stays in charge of understanding the task, making decisions, and doing the final review — it just no longer reads every file or runs every small check itself.

In one line: small models do the work, the strong model makes the calls.

What you'll feel after installing

• The same task, cheaper and faster: reading files, gathering facts, exploring evidence-backed candidate paths, and mechanical changes go to cheap models; the strong model's call count and token use drop noticeably.
• The strong model focuses on what actually needs it: problem understanding, design decisions, conflict resolution, critical review.
• You don't pick models by hand: dteam routes tiers by task need; use /dteam to inspect or take over.

Relationship to dgoal

dteam and pi-dgoal are two independent Pi extensions, orthogonal and composable:

• dteam: multi-model tier routing — cheap models do the work, the strong model judges.
• dgoal: single-model build-check loop — freeze an acceptance contract for a goal and run independent audits.

Use dteam to save token / time; use dgoal when you need a frozen acceptance contract and independent audit; combine when both apply. See the trigger protocol for how to choose.

Quick start

dteam requires a personal config file; it never guesses tiers from a model name and never silently falls back to the current model.

# 1. Configure tier models at ~/.pi/agent/pi-dteam.json:
# {
#   "tiers": {
#     "T1": ["openai-codex/gpt-5.6-terra:high"],
#     "T2": ["openai-codex/gpt-5.6-luna:medium"],
#     "T3": ["openai-codex/gpt-5.3-codex-spark:low"]
#   }
# }

# 2. Install into Pi
pi install "$(pwd)"

# 3. /reload in Pi

# 4. Give the main model a non-trivial coding task and watch it dispatch T3 to gather facts

If the config is missing or incomplete, dteam warns at startup and refuses to dispatch. Each tier is an ordered candidate array of provider/model[:thinking]; later entries are fallbacks.

Tiers

Tier	Model	Thinking	Use
T1 flagship	flagship model	high	thinking, decisions, critical review, fallback redo
T2 standard	standard model	medium	regular implementation
T3 fast	fast / local model	low	mechanical small tasks, fact-finding probes

• Read-only by default; any write must be explicitly authorized via addTools at dispatch with a project-relative writeScope.
• Each worker has an initial work-tool budget (T3 60 / T2 120 / T1 180); dteam_signal and the final dteam_report don't count against it.

The four tools (used by the main model; you usually don't touch them directly)

dteam_dispatch (dispatch), dteam_respond (respond to worker requests, including cancel), dteam_recover (timeout recovery), dteam_wait (wait for a worker you depend on). You use /dteam to inspect and manage background workers, and press m in its modal to edit the global T1/T2/T3 candidate chains; saving affects future dispatches only.

Going deeper

• Documentation overview — entry to all docs
• Glossary
• Tool API reference
• When to use dteam vs dgoal
• Project roadmap
• CHANGELOG

Chinese version: README-zh.md.