Design goal: Agents that replace humans at the frontier of knowledge — freely exploring and delivering high-level research. They are power users of open-source tools (Cursor, Trae, Aider, MCP, etc.), not rebuilders of them, and they are built for deep thinking: framing questions, debating approaches, judging evidence. Current runnable focus is the scheme phase; see scheme phase blueprint.
finisher is CIO-led (see docs/policies/cio.md): one orchestrating thread turns intent into blueprints and acceptance evidence—aligned with single-operator quant work (full stack, no fake departmental walls). Implementation routes through MD policy skills registered in docs/skills/manifest.json (e.g. data-scientist, quant-researcher, quant-dev, quant-soul) as contracts, not as separate “department” agents. Policies, docs/reference/quant_tech_stack.md, and knowledge/ supply methodology and technique; optional Layer 1 / Layer 2 / Reviewer labels in docs/agent/AGENTS.md are shorthand only. Execution leans on open-source tools and external agents (e.g. Cursor/Trae). Core loop: mandate → evidence → gates → sign-off.
- Literature Review: Searches Arxiv to support or refute hypotheses.
- Deterministic Experiments: Runs reproducible experiments with explicit assumptions.
- Evidence Gates: Time-safety and validation requirements are enforced as acceptance criteria.
- Skill-First Extensibility: Deterministic
BaseSkilltools plus MD policy skills indocs/skills/manifest.json(CIO, quant-soul, data-scientist, quant-researcher, quant-dev)—contracts for stages, not separate “department” agents.
git clone https://github.com/isomorphicor/finisher.git
cd finisher
pip install -r requirements.txtconfig/settings.yaml— API keys / backend preferences.config/agents.yaml— scheme phase:scheme_phase.default_agent_model(main tool-loop LLM),reviewers,translator. Override per run withpython scripts/run_scheme_agent.py --model ...or envINVERST_SCHEME_AGENT_MODEL.
Session layout: <runs_root>/<project>/<session>/ — e.g. ~/Project/projects_generated/algo_alpha/main. In-repo default: out/<project>/<session>/. Artifacts under artifacts/, code under project/src/, run outputs under project/outputs/. --project and --session are the two path segments (default session name main unless INVERST_DEFAULT_SCHEME_SESSION is set or timestamp). More examples: docs/guides/command_examples.md.
One-shot — scheme, execution prep, and IDE coding in one process (recommended):
python scripts/run_research_session.py --project algo_alpha --ide-max-rounds 200 \
"Your research task for the scheme phase"Equivalent legacy name: python scripts/run_scheme_then_ide.py (same pipeline via core/research_session_pipeline.py). Optional: --ide-task "…" (extra instruction for the IDE step only), --skip-execution-prep, --abort-ide-on-scheme-partial. See --help on either script.
Deprecated: scripts/run_autonomy_loop.py — use run_research_session.py or run_scheme_then_ide.py instead; see docs/skills/ARCHITECTURE.md.
Skill layering (policy / ops / runtime): docs/skills/ARCHITECTURE.md.
Smoke test (plan + code in one run): docs/experiments/smoke_plan_and_code.md — ./scripts/smoke_research_session.sh (requires LLM backend).
If the scheme phase stops before writing all four artifacts/*.md, the combined run exits with a JSON field scheme_artifacts_missing — raise --scheme-max-rounds or re-run run_scheme_agent on the same session.
Step by step — same pipeline, three commands:
python scripts/run_scheme_agent.py --project algo_alpha "Your research task"
python scripts/run_execution_prep.py --project algo_alpha --resume-latest --workspace-root .
python scripts/run_ide_execution_agent.py "out/algo_alpha/$(cat out/algo_alpha/LATEST)" \
--workspace-root . --max-rounds 200(Use straight quotes; put a space between --max-rounds and 200.)
Utilities (skills / tools):
python scripts/run_routed_tool.py --intent workspace.write --args-json '{"path":"demo.txt","content":"hello"}'python scripts/install_skill.py <path_to_skill_package>python scripts/list_skills.py --capability workspace.writepython scripts/create_skill_template.py my_skill --output-dir /tmp
Suggested skill flow: create_skill_template → edit runtime.py → install_skill → run_routed_tool.
- Docs Index: Start page for contracts and standards.
- Project Status: Current roadmap and active tasks.
- Worklog: Minimal change log (only key behavior changes).
- Skill Catalog: List of available Python tools (Skills).
- Coding Conventions: Guidelines for contributing code.
- Module Contracts: Module boundaries, IO, evidence artifacts.
- Research Standards: Acceptance gates and evidence standards.
Legacy notes:
- Execution scaffolding MVP: consume scheme artifacts and generate IDE checklist/workspace
- Skill registry + intent router for install/discover/on-demand invocation
- Runtime abstraction with Local/MCP-compatible backend interface
MIT