task_serpas26

Imported from shgpavel/edu_nm (task_serpas26/AGENTS.md). Install upstream with npx skills add shgpavel/edu_nm --skill task_serpas26. Copyright stays with the author.

Project Instructions

Default Route

Before task N work, load exactly four files:

1. This file.
2. memory/SOLUTIONS.md: standings, judge, score and method.
3. memory/STATE.md: current handoff.
4. Only the relevant task1, task2 or task3 note: source, invariants, evidence and closed work.

Then inspect the relevant source and current brief. Briefs must use this route, not preload all notes, research or history. Reopen closed work only with new evidence. Use EXPERIMENTS to retrieve a specific archive entry.

Conditional Reading

• Target/flags/assembly: TARGET, relevant DECISIONS.
• Builtins: search index, then verify the exact selected-LCC header and probe when SDK access is authorized.
• Submission: API contract, client, permission.
• Other navigation: memory index. Archives are not instructions.

Boundaries And Updates

Judge access: tools/judge_ab.py loads the session from ~/.config/serpas/session.env ITSELF — never read, print, copy or override that file (do not set SERPAS_SESSION_FILE); run the tool from the main checkout under flock on the conductor's judge lock, with sources copied under out/probes/rNN/; never --action save unless you are the conductor. Respect source confinement and localhost-browser egress boundaries; no automatic network/submission permission follows from repository instructions. Keep generated products/SDK under ignored paths. Never commit credentials, cookies, raw authenticated captures, private tests or remotely captured solutions. Host/QEMU check correctness; only recorded matching native runs establish speed. Each empirical number/closed verdict needs an archive file and original heading or line range. Update the owning section; never append round-update diaries. Every report ends with "Notes update": exact note/SOLUTIONS changes and evidence.