hedgefund

hedgefund

A four-agent "hedge fund" system for the Solana trading bot in git.noncepad.com/pkg/optimizer (specifically its multimodelv1 mode -- see optimizer/cmd/multimodel.go), built on eino: one agent for risk, one for P&L, one for research, and one fund manager (compose.Workflow) that fans all three into a single synthesized decision.

This is a standalone CLI, runnable today (see Usage below) -- not a library waiting on a caller. For the other, independently built hedge-fund implementation (gitlab.noncepad.com/eflam/wiki/client/hedgefund, compose.Workflow-based text-to-SQL risk/pnl agents, gated real trigger access), see Relationship to go-wiki/client/hedgefund below -- that one's README documents its own, different status.

Status

All four agent objects are built and tested (29 passing tests, go build/go vet clean):

The fund manager never has real trigger access, at all, full stop. FundDecision mirrors optimizer/cmd/multimodel.go's own CLI flags field-for-field so a human can translate it directly into the real command they'd run by hand -- but this package never calls that command or holds a multimodelv1.Hook. There is no gate to bypass here because there is no trigger tool in this package to begin with -- a stronger guarantee than "gated" (compare go-wiki/client/hedgefund's fund manager, which does hold real trigger tools behind a *TriggerGate). Every manager run prints a PROPOSAL ONLY banner ahead of whatever it proposes.

../contrib/hedgefund/Dockerfile is an unfinished stub (FROM registry.noncepad.com/eflam/solpipe-terminal/run:dev, nothing else yet) -- there is no working container image for this today, only the go run usage below. Go source lives here, in hedgefund/; contrib/hedgefund/ holds only the Dockerfile, per this repo's convention that contrib/ subdirectories are for Dockerfiles and other non-code files, never Go source.

Usage

cd optimizer/hedgefund
go run . -node risk     -prompt "What's our real risk exposure right now?"
go run . -node pnl      -prompt "How has our value changed over the last week?"
go run . -node research -papers-dir /path/to/papers -prompt "Any new model ideas?"
go run . -node manager  -papers-dir /path/to/papers -prompt "Should we do anything right now?"

-prompt is optional for every node -- each has a sensible default question (see main.go) if left blank.

Prerequisites

• A reachable Ollama server with a tool-calling-capable model pulled (default qwen3-coder:30b at http://localhost:21434). Override with -model/-ollama-url. A smaller model (e.g. qwen2.5:7b) will run, but live-verified this session to reason less reliably about the numbers its own tools hand it -- worth knowing before trusting a manager proposal from a weaker model.
• risk/manager additionally need a reachable catscope state endpoint (-state-url, required, tcp://ip:port or unix:///path) -- the same internal gRPC/geyser-backed state.Client graph the live trading bot itself reads through (git.noncepad.com/pkg/bot/state), not the public Solana RPC endpoint -- plus Jupiter price API access (no flag needed, a plain HTTPS call) and a wallet to inspect (-wallet, defaults to harness.DefaultWallet -- the real trading child wallet this whole session's work centered on; public key only, no private key ever needed since every tool is read-only).
• pnl/manager additionally need a real prefetch.db (-db, defaults to ~/.optimizer/prefetch.db) -- the same database optimizer watch-pnl/the live trading bot itself writes to. An empty or missing database isn't an error; pnl just reports "no positions recorded."
• research/manager additionally read a local directory (-papers-dir, defaults to ~/.optimizer/research-papers) of .pdf/.txt/.md files, flat (not recursive). An empty or missing directory isn't an error either; research just reports nothing to propose.

All four flags above (-state-url, -wallet, -db, -papers-dir) plus -model/-ollama-url/-timeout are shared across every -node value; each node only actually uses the ones it needs.

Example

$ go run . -node manager -papers-dir ~/research -prompt "Should we open any new positions?"
wallet: Hg2p3cfmg3dratEVy94JArTVM7KzEywgNdmFnrfhroh9
> Should we open any new positions?

=== PROPOSAL ONLY -- no real transaction has been sent ===

Rationale: <the model's reasoning over the real risk/pnl/research findings>

  PROPOSED: enable_directional_trading (target So1111...)

or, when the model decides nothing is warranted:

  (no action proposed)

Testing

Every node has its own _test.go (29 tests total). Agent construction and round-trip/error-propagation tests use fakeToolCallingModel (fake_model_test.go) -- deterministic, no live Ollama/network needed. Where real local infrastructure is cheap and worth exercising for real rather than mocking, tests use it directly: pnl_test.go seeds a real temp SQLite database (newTestStore, same pattern go-wiki/client/hedgefund's own tests use) with real pnl_position_snapshot rows; research_test.go writes real temp files, including a real path-traversal attempt against a real file placed just outside the configured directory, to prove resolvePaperPath's guard actually blocks it rather than just asserting the string check fires in isolation. manager_test.go builds BuildFundManagerWorkflow (the testable graph-wiring function -- see its own doc comment) with three stub agents and confirms the fan-in wiring is correct by giving each stub a unique marker response and checking all three reach the fund_manager node's prompt in the right order; this was verified to actually catch a broken mapping (a field-mapping swap was tried by hand once and confirmed to fail the test) before being reverted.

Live network/RPC/Ollama calls are never made in go test -- only via go run per Usage above.

Relationship to go-wiki/client/hedgefund

See that package's own README for the full comparison table. Short version: same four-agent shape, different choices throughout (live state.Client graph reads vs. persisted-snapshot data, ReAct-tool-calling vs. text-to-SQL agent shape, no trigger access at all here vs. hard-gated real trigger access there). Neither supersedes the other; nobody has yet decided whether or how to converge them.