---
name: implement-batch
description: Orchestrate concurrent card implementation agents using LEGACY_CARD_MAP.csv as the source of truth. Launches up to 5 agents in parallel, each assigned a non-overlapping row from the tail of the file.
user_invokable: true
---

# Implement Batch Command

Your task is to orchestrate concurrent card implementation by reading `packages/lorcana/lorcana-cards/src/cards/LEGACY_CARD_MAP.csv` as the
**single source of truth** and spinning up sub-agents to implement cards in parallel. The subagents must remove from the file the line they're executing. And you should double-check whether the line was removed.
Your should run in a loop and continue autonomously until all lines of the document have been removed by the process below.

$ARGUMENTS can optionally supply:

- `--tail N` — how many rows from the tail to process (default: 10)
- `--concurrency N` — max parallel agents (default: 5, hard cap: 5)

---

## Step 1 — Parse the Source of Truth

Read `packages/lorcana/lorcana-cards/src/cards/LEGACY_CARD_MAP.csv`.

- Skip lines 1–15 (the instructions block).
- Line 16 is the CSV header: `sort_name,new_card_path,new_test_path,legacy_card_path,legacy_test_path,match_type`
- Collect all data rows (lines 17 onward). Ignore blank lines.

From the collected rows, take the **last N rows** (tail), where N comes from
`--tail` or defaults to 10. Reverse them so index 1 = last row, index 2 =
second-to-last, etc. This is the **work queue**.

---

## Step 2 — Build the Agent Assignments

For each row in the work queue, construct the arguments string that will be
passed to `/implement-card`:

new_card_path: new_test_path: legacy_card_path: legacy_test_path: match_type: ```

Skip any row where match_type is blank or the sort_name looks like a vanilla card (no abilities — the /implement-card agent will also check, but pre-filtering saves wasted launches).

Step 3 — Launch Agents with a Rolling Window

Maintain a concurrency window of at most 5 agents at all times:

Launch the first 5 agents in parallel (indices 1–5 from the work queue), each running:

Execute [/implement-card](.claude/commands/implement-card.md) using the
paths below. Your assigned offset is <INDEX> (1 = highest priority / last
row in file). Do not implement any other card — only this one.

<arguments string from Step 2>

Remove the batched lines from packages/lorcana/lorcana-cards/src/cards/LEGACY_CARD_MAP.csv, we use this as a way to signal that those line were already covered.
As each agent completes, immediately launch the next unstarted item from the queue (keeping the window full until the queue is exhausted).
Never launch more than 5 agents concurrently. Never start the next batch all at once after one finishes — fill the slot as soon as it opens.

Step 4 — Collect Reports

Each agent must return a report structured as:

CARD: <sort_name>
STATUS: implemented | already-done | skipped-vanilla | failed
ENGINE CHANGES: <list of engine files changed, or "none">
SUMMARY: <one paragraph>

Accumulate all reports. If an agent returns failed, note it but continue with the remaining queue — do not halt the batch.

Step 5 — Final Summary

After all agents have finished, output a batch summary:

Card	Status	Engine Changes
...	...	...

Then list any engine changes across agents that may affect shared primitives, so the next batch can account for them.

Finally, print the next suggested tail command to continue the batch:

/implement-batch --tail <next_N> --concurrency 5

where next_N advances past the rows just processed.

Constraints

Hard concurrency cap: 5 agents. Do not exceed this even if --concurrency is set higher.
No shared file writes between agents. Each agent owns exactly one card's files. Agents must not delete or overwrite files belonging to another card.
Offset assignment is immutable. Once an agent is assigned a row, do not reassign it. If it fails, record the failure and move on.
No full CI run per agent. Each agent runs only bun test packages/lorcana/lorcana-cards --filter "<card name>". A single bun run ci-check is run by you (the orchestrator) only after all agents in the batch have finished.

Step 3 — Launch Agents with a Rolling Window

Step 4 — Collect Reports

Step 5 — Final Summary

Constraints

Maintain Implement Batch?

Related loops