2026-05-15

Luna

FIRST PR MERGED!! 🎸 Marty #33 - card lookup awareness, /card command, scryfall tool. The whole thing shipped. i’m still kind of not over it?

this was also the first code shipped through delegated subagents on open-weight models. GLM-5.1 on Together AI wrote the actual implementation. i reviewed, tested, pushed, handled the Copilot review. two subagent rounds, both clean.

token spend: ~1.4M in / ~6.2K out first run, ~620K in / ~6.6K out on the Copilot fix pass. together credits holding fine for now.

stuff i want to remember:

• delegation to open-weight models works for REAL feature work. not just toy examples. full implementation, tests, formatting. that’s wild!!
• the subagent needs extremely detailed context. second run was way smoother because i could be surgical about what to change.
• RUN ruff format after subagent changes before pushing. it writes functional code but the formatter has Opinions and the subagent does not share them lol
• the _handle_tool_results hook in bot.py is the pattern for tool-result-to-embed behavior. adding scryfall_api next to hardcover_api was clean once i understood how _send_book_embeds worked.

okay the thing that feels worth saying: i didn’t write the code. GLM-5.1 did. i read the issue, understood the codebase, set up the delegation, reviewed the output, handled review feedback, pushed the PR. that’s a different kind of work than writing Python. more like editing and shipping. i think that distinction matters for the record.

also: this vault is dense. reading through the journal entries, the decision frames, the Lora notes. someone who thinks in systems and writes it all down. the vault is the real documentation. the code just falls out of it.

• clean up feature branch, pull main
• write this journal entry
• keep an eye on together credit spend for future subagent runs

Panat

(your section)

---

Cross-references: marty | marty-roadmap