Agent teams

hipmmcode's core superpower: the lead agent can spawn teammates — full agents with their own context, tools, and even their own provider/model — and coordinate them through shared task lists and mailboxes.

Spawning teammates

Just ask, or be explicit:

› Spawn three named background Agents; they automatically join this session's implicit team:
  - name=gpt     provider=openai    model=gpt-4o-mini        prompt="find recent X"
  - name=claude  provider=anthropic model=claude-haiku-4-5   prompt="find recent Y"
  - name=gemini  provider=gemini    model=gemini-2.5-flash   prompt="find recent Z"
  Poll each with TaskOutput and aggregate.

Three teammates run in parallel on three different vendors. Key Agent parameters:

Param	Effect
background: true	Returns immediately with an agent_id; poll via TaskOutput, stop via TaskStop
provider / model	Any configured channel — mix vendors freely
agent	Use a named profile (team-designer, Explore, …)
allowed_tools	Restrict the teammate's tool surface
isolation: "worktree"	Run in a private git worktree (parallel file edits without conflicts)
effort	Per-teammate reasoning effort

Foreground agents block until done and return their result; background agents stream progress into the team UI.

Watching the team

Surface	What you see
Live task tree	Teammates appear in the in-turn tree with per-agent colors
/teammates (/team)	Panel: every background teammate, status, transcript view, stop
/panes (/tiles)	Auto-refreshing tiled view, teammates side-by-side
/foreground <name> (/fg)	Pin one teammate's live output above the prompt; --full for fullscreen
/preview	Cycle the persistent preview row: off / pills / tree
Mid-turn keys	Shift+↑/↓ select · f foreground · k stop · Enter expand

/teammates in action:

hipmmcode

› /teammates

Teammates  team: audit-pod · 2 running · 1 done
❯ security  anthropic/claude-opus-4-8 · running · 4.2k tok · Read src/auth.rs
  perf      deepseek-anthropic/deepseek-v4-flash · running · 2.9k tok · Grep "clone("
  deps      gemini/gemini-2.5-flash · completed · 1.8k tok

Enter view transcript · f foreground · k stop · Esc close

Teams persist under ~/.hipmmcode/teams/<name>/ (config, inboxes, tasks). hipmmcode team list / hipmmcode team rm <name> manage them from the CLI; /teams and /tasks from the REPL.

Agent profiles

A profile is a markdown file with YAML frontmatter — name, description, tool allowlist, model, color, effort — plus the system prompt as the body.

hipmmcode agents                          # list all profiles (bundled + yours)
hipmmcode agent new reviewer \
  -d "strict code reviewer" \
  -t Read,Grep,Glob \
  -m anthropic/claude-sonnet-4-6 \
  -p "You are a meticulous reviewer..."   # or pipe the prompt via stdin
hipmmcode agent new x --project           # write to <cwd>/.hipmmcode/agents/ instead of ~/.hipmmcode/agents/

In the REPL, /agents opens the manager (create / edit / delete). --agents '<json>' injects session-only profiles from the CLI.

Bundled profiles (zero setup): team-designer (designs agent teams for a goal, interviews you, writes the files), judge (scores results against a rubric), optimizer (improves prompts/agents), Explore (read-only codebase search), Plan (read-only planning), code-reviewer, general-purpose.

Adopt a profile as the lead persona: hipmmcode chat --as team-designer.

Legions

A legion is a named bundle of agents + skills + collaboration patterns — a reusable team:

~/.hipmmcode/legions/startup-cognitive-team/
├── legion.md
├── agents/{blacksmith,assassin,sage,prism}.md
└── skills/delivery-rhythm.md

hipmmcode legions                          # list
hipmmcode chat --legion startup-cognitive-team   # activate for one session
hipmmcode config set activeLegion=startup-cognitive-team   # persist

In the REPL: /legion <name> activates, /legion off deactivates, /legions lists. When active, the legion's agents and skills take precedence over your global ~/.hipmmcode/{agents,skills}/, which in turn override the bundled defaults.

Let the bundled designer build one for you:

› Agent name='designer' agent='team-designer'
  prompt='Design a 4-person startup cognition team: planner / competitor-scout / user-sage / coordinator'

It interviews you via AskUserQuestion, writes the legion files, and iterates until you approve.

Background agents in practice

▸ Agent name='runner' provider=openai model=gpt-4o-mini background=true
  prompt='research Rust 2024 edition'
[runner spawned, agent_id=agent_a3f9…]

▸ TaskOutput name='runner' block_ms=10000
[runner · openai/gpt-4o-mini · running · 1430 tok]
[tool WebSearch] {"query":"Rust 2024 edition release"}

▸ TaskOutput name='runner' block_ms=10000 offset=2
[runner · completed] Rust 2024 Edition stabilized on 2025-02-20 with Rust 1.85 …

Teammate defaults: teamMateDefaultModel config picks the model spawned teammates use (leader = inherit the lead's); autoJudge=true auto-scores teammate results; agent_color in a profile pins its color in the UI.