cc-quant-team

CC_quant_team

Multi-agent team orchestration for quantitative research and trading projects in Claude Code.

Quant-specialized fork of CCteam-creator by jessepwj. Keeps the core harness (file-based planning, team-lead control plane, 3-Strike escalation, team-snapshot fast-resume), but replaces web-development roles and checks with quant-specific ones.

Current version: 0.1.4

What it does

Turns a single Claude Code session into a coordinated team of quant agents. You talk to the team-lead (the main conversation); team-lead dispatches specialized subagents working in parallel. Progress, findings, and decisions persist in .plans/<project>/ so nothing is lost across context compressions.

Team Roster (11 roles)

Role	Model	Multi-instance	When to include
backend-dev	sonnet	no	Live trading infra (exchange API, order DB)
frontend-dev	sonnet	no	Dashboard / viz (lazy-loaded)
algorithm-engineer	sonnet	yes (volume)	C++ / Python perf, pybind11, memory optimization
hft-researcher	sonnet	yes (direction)	Tick-level / L2 feature engineering
lft-researcher	sonnet	yes (direction)	Daily / minute feature engineering
data-engineer	sonnet	no	Historical pipeline, parquet schemas, data quality
model-researcher	opus	no	DL / ML signal / alpha modeling
strategy-researcher	opus	no	Portfolio construction, execution, simulator
code-validator	sonnet	no	Code review + corner-case tests
research-validator	opus	no	Lookahead / survivorship / leakage / overfitting checks
risk-manager	opus	no	Fund verification, strategy risk, live monitoring

Team-lead = the main conversation (you). Not spawned as an agent.

Quant-specific mechanisms

• 21 pre-filled invariants covering temporal correctness, data integrity, reproducibility, statistical discipline, execution realism, and code hygiene
• 5 review dimensions: Research Correctness (CRITICAL) / Reproducibility / Performance / Capacity & Risk / Observability
• 9 automated checks (4 universal + 5 quant): timestamp monotonicity across parquet + pkl.gz data, unseeded randomness detection, type-annotation import verification, stdlib name shadowing, pd.read_csv without dtype
• Two mandatory gates: research-validator gate before any research output is consumed downstream; risk-manager gate before any strategy goes live

Progress Tracking Discipline

Both team-lead and spawned agents have mandatory progress-tracking protocols.

Agents: break tasks into 5-15 min sub-steps and append to progress.md immediately after each sub-step — not at task end. Anti-patterns (empty progress.md, single vague entry, late batched update) are called out explicitly in the onboarding.

Team-lead owns 5 root files at .plans/<project>/:

File	When team-lead updates
task_plan.md	After every dispatch / completion / verdict / phase change
progress.md	After every significant event (dispatch, completion, verdict, decision)
findings.md	At milestones — consolidates scattered agent findings into [MILESTONE] entries
decisions.md	New D<N> entry for every non-trivial architecture / pipeline / tool-choice decision
team-snapshot.md	Regenerated on roster change; staleness-checked at phase boundaries

Milestone Consolidation (at phase boundaries / significant completions):

1. Read agent task-folder findings → synthesize into a [MILESTONE] entry in root findings.md with key metrics, conclusions, and links
2. Extract durable knowledge (schemas, interfaces, capacity numbers) → update docs/data-schemas.md / docs/pipeline-flow.md / docs/strategy-contracts.md / docs/invariants.md
3. Update docs/index.md if any docs/ file changed

This means that when Phase 3 starts (say, model-researcher consuming Phase 2 features), everything they need is in one place — root findings.md + docs/ — instead of scattered across 5 agent directories.

Prerequisites

This skill depends on Claude Code's experimental agent teams feature — the primary dispatch tools (TeamCreate, TaskCreate, TaskUpdate, TaskList, SendMessage) only exist when this feature is enabled. Without it, the skill's instructions reference tools Claude Code cannot call.

Requirements:

• Claude Code v2.1.32 or later (check: claude --version)
• Environment variable CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Enable (recommended) via ~/.claude/settings.json:

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

Then restart Claude Code. Team tools load only at session start.

Verify: after restart, ask Claude "Is TeamCreate available?" — it should say yes.