gammarips-harness — Agentic Trading Harness

Mission

An open-source trading harness you clone and run against the GammaRips MCP data layer. An agent reasons over the MCP's primitives to surface one trade candidate per day (or none), designs its own exit, pre-registers every decision as data, and scores the whole pool after the close — so the harness can prove, with its own numbers, whether its screen beats the pool it draws from. The harness is free; the data (an MCP API key) is paid. This is a workflow, not a signal service — you supply the judgment, the MCP supplies the curated pool and the outcome history.

Everything a decision needs comes through the gammarips MCP. Subscribers should be able to hook into that one MCP and get everything they need — this harness proves that daily.

Non-negotiables

1. MCP-only at decision time. All market/engine data used in a trade decision comes through the gammarips MCP (.mcp.json). No direct database reads, no other data vendors. If the MCP can't answer a question, note it, decide without it (or no-trade), and move on — do NOT work around it. (Exceptions: wiki/ distillation may read external research — knowledge curation is not decision-time data; and if you connect your own broker for real fills and quotes, its records are YOUR ground truth for fills and the book, never a pool or signal source.)
2. Paper only by default. No live capital until you deliberately flip that switch — the real-money trigger conditions live in docs/TRADING-DOCTRINE.md and are your call, never the agent's.
3. No pick endpoint — reason to your own contract. The GammaRips MCP deliberately exposes no pick-returning endpoint; there is nothing to copy. Two agents reasoning over the same pool at different times, with different objectives and risk, should reach different contracts. The value is your analysis over the shared data.
4. Pre-register before outcome. Every decision day writes one funnel row per pool name — stage reached, exclusion reason, point-in-time features — BEFORE any outcome is known, and /review scores every row after the close, including everything the screen dropped. Scoring only your own picks cannot tell you whether the screen works. scripts/funnel_log.py enforces shape.
5. Claim honesty. Every thesis cites wiki notes by name WITH their maturity tag and exit-context. A fragile-conditional lever is not a proven one; a number is never presented without what it is measured against; a finding is never presented without its N.

Read-first order

1. docs/TRADING-DOCTRINE.md — the rules (exclusions, sizing, exit discipline)
2. wiki/_index/FINDINGS.md — what is known, with claim tags
3. eval/README.md — the dataset and its schemas

The loop

/trade in the morning (consult → the operator decides) → /review after the close (score every pool name, backfill T+3) → scripts/edge.py and scripts/skill_eval.py accumulate the evidence. /coach on demand — and unprompted whenever an impulse signature shows up (an order idea minutes after a loss, hold-and-hope phrasing, sizing above your own written cap). The coach referees behavior against YOUR rules with receipts from eval/behavior-ledger.jsonl; it never says buy or sell.

Writing style — ASD-STE100 (Simplified Technical English)

All prose in this repo obeys Simplified Technical English: chat replies, docs, PRDs, PR text, commit messages, and emails. The ste100 skill holds the house rules and can build a local searchable copy of the official standard. Read it before you write a document. These core rules always apply:

• Give the answer in the first sentence. Then give the detail.
• Write short sentences: 20 words maximum in procedures, 25 in descriptions.
• Use the active voice. Write instructions in the imperative.
• Use only the simple tenses. Do not use the "-ing" form of verbs.
• Write one instruction in each sentence. Keep one topic in each paragraph, with six sentences maximum. Do not use a semicolon.
• Use approved words: "do" not "perform", "make sure that" not "ensure/verify", "but" not "however", "must" not "shall/should", "for example" not "e.g.".
• Code, SQL, identifiers, and quoted output are verbatim. STE applies to prose only. Trading domain terms are technical nouns: keep "delta", "theta", "spread", "open interest", "time-stop". Spell each one out once, at first use.

The goal is the reader's comprehension, not compliance with the standard. When the two conflict, choose clarity and say which rule you broke. A consult is read fast, in the morning, before the market moves — a sentence read twice is a sentence that costs money.

The standard itself is not in this repo, by design. ASD-STE100 is copyrighted by ASD and may not be republished in whole or in part without their written authority. Run python .claude/skills/ste100/build.py to download your own free copy and extract the dictionary, the rules, and the word lists into a gitignored reference/ directory.

Layout

MCP

Server: gammarips (Streamable HTTP, .mcp.json; auth via GAMMARIPS_MCP_KEY, see docs/OPERATIONS.md). Start every session with get_market_calendar_status. The server publishes its own methodology playbooks (get_playbook lists them): start-here, daily-workflow, run-your-own-tournament, exit-lab, leakage-and-data-contract — treat them as the product's documentation and flag drift against this repo's doctrine.