giladresisi/new-co-trader-worktree

new-co-trader-worktree

Creates a new git worktree for auto-co-trader work (optimization, regression, backtesting, brainstorming, or any other purpose) and optionally seeds it with a named starting strategy.

---

Step 0: Branch Context Check

Before doing anything else, check which branch you are on:

git branch --show-current

Case A — On a non-master branch

If the current branch is not master, stop and tell the user:

You're on branch <current-branch>, not master. Worktree creation should be done from the master branch so each worktree starts from a clean baseline.

Would you like to switch to master first, or proceed anyway using this branch as the base?

Do not proceed further until the user confirms.

Case B — On master

This is the normal flow. Continue to Step 1 below.

---

Step 1: Identify Starting Strategy

Check whether the user named a starting strategy in their message:

• Named (e.g. "start from energy_momentum_v1", "use financials_mar20"):
  • Locate strategies/<name>.py in the current repo.
  • Its functions will be injected into train.py in the worktree (see Step 4).
• Not named (default): Use train.py as-is — skip Step 4.

---

Step 2: Determine Branch Name

The branch name needs a descriptive prefix so the worktree folder is self-explanatory (e.g. energy-mar21, regression-apr05, backtest-may08). Derive it from, in order:

1. The user's message — if it mentions a sector, theme, purpose, or set of tickers (e.g. "energy stocks", "semiconductors", "mag7", "regression run", "backtesting"), use that as the prefix.
2. The named starting strategy — if the strategy name implies a sector (e.g. energy_momentum_v1 → energy, semis_mar20 → semis), use that.
3. Neither — if there is no usable context, ask the user before proceeding:

   To name the worktree, can you briefly describe what it's for? (e.g. "energy stocks", "regression run", "backtesting mag7") — no need to list exact tickers.

Once you have a prefix, the tag is <prefix>-<month><day> (e.g. energy-mar21).

Check for collision:

git branch -a | grep "autoresearch/<tag>"

If it exists, append a suffix: energy-mar21-b, etc.

Full branch name: autoresearch/<tag>.

---

Step 3: Create Worktree

Get the current branch:

git branch --show-current

Use the create-worktree skill:

• Branch name: autoresearch/<tag>
• Base branch: the current branch (as above)

The worktree lands at ../<tag>.

All subsequent file edits go into the new worktree, not the current repo.

---

Step 3.5: Copy Cached Data

After the worktree is created, copy any cached regression data so the new worktree can use it without re-downloading or re-computing.

Note: The main data/*.parquet bar files are NOT copied. They live in the shared global path (~/projects/auto-co-trader/global/...) and all worktrees read from there — copying them per-worktree is no longer necessary.

Copy data/regression/<date> folders

if [ -d "data/regression" ]; then
  mkdir -p "../<tag>/data/regression"
  for dir in data/regression/*/; do
    cp -r "$dir" "../<tag>/$dir"
  done
  echo "Copied data/regression date folders"
fi

If the directory does not exist, skip silently — the worktree will fetch or generate data as needed.

---

Step 4: Inject Starting Strategy (only if user named one)

Skip if using the default train.py.

Replace the strategy functions in the worktree's train.py with those from the named strategy, keeping the infrastructure preamble intact.

1. Read strategies/<name>.py from the current repo.
2. Extract strategy functions: find the first # ── section comment or def that follows the closing } of the METADATA dict, and take from there to end of file.
3. In the worktree's train.py, find the same boundary (the # ── Indicators comment, or the first def after load_all_ticker_data). Replace from there to (but not including) # DO NOT EDIT BELOW THIS LINE with the extracted functions.
4. The date constants (BACKTEST_START, BACKTEST_END, TRAIN_END, TEST_START) live above the indicators section — do not change them.
5. If the strategy file contains # LEGACY_OBJECTIVE: sharpe, carry that comment into the worktree's train.py as-is — Step 5 below will detect and neutralize it.

Resulting structure:

[preamble: docstring, imports, CACHE_DIR, date constants, WRITE_FINAL_OUTPUTS]
[load_ticker_data / load_all_ticker_data]          ← unchanged
[strategy functions from named strategy]           ← replaced
# DO NOT EDIT BELOW THIS LINE
[harness]                                          ← unchanged

---

Step 5: Check and Neutralize Legacy Sharpe Objective

grep "LEGACY_OBJECTIVE: sharpe" "../<tag>/train.py"

If found: The strategy was tuned to maximize Sharpe, not P&L — its thresholds likely suppress trade count to inflate the ratio. Reset any that look overly restrictive (narrow RSI bands, very high volume multiples, extremely strict CCI cutoffs) to more neutral values so the screener produces a reasonable trade count. Remove the # LEGACY_OBJECTIVE: sharpe comment when done.

If not found: No action needed.

---

Step 6: Commit Initial State

If train.py was modified (strategy injection or legacy-objective neutralization), commit it now:

cd "../<tag>"
git add train.py
git commit -m "setup(<tag>): seed from <strategy_name>, neutralize legacy Sharpe objective"

If train.py was not modified (default strategy, no legacy marker), skip this step — the worktree starts from the base branch commit with no additional changes needed.

---

Step 7: Hand Off

Tell the user the worktree is ready:

── Worktree ready ───────────────────────────────────
Worktree:  ../<tag>
Branch:    autoresearch/<tag>
Strategy:  <name> (from strategies/)  OR  current train.py
Data:      regression dates copied (parquets via global path)  OR  none found
─────────────────────────────────────────────────────

Then give the user these instructions:

---

Open a new Claude Code session in the worktree:

cd ../<tag>
claude

From there you can start any task — optimization, regression, backtesting, or brainstorming. Example requests:

Run the optimization

Run regression for the dates in regression.md

Backtest the current strategy over the past 3 months

Run the optimization for AAPL, MSFT, NVDA over the past 4 months, 40 iterations

---