Turminus-2

Imported from singhaganesh/Turminus-2 (AGENTS.md). Install upstream with npx skills add singhaganesh/Turminus-2. Copyright stays with the author.

AGENTS.md

TB3 Ed2. Unit: tasks/<task-name>/. Start: docs/ARCHITECTURE.md; platform bar: docs/edition2/; policy/rules: docs/HARD_BUT_FAIR_AUTHORING.md, .cursor/rules/, commands.md. Practical reviewer blockers: terminus_blockers_checklist.txt.

Must

1. No PASS/READY/APPROVED/SUBMIT claim without command output.
2. Long run: python3 scripts/tb3_doctor.py.
3. New-task gate: lint_spec -> spec_satisfiability -> task_gate -> check-task.sh -> oracle_idempotency_probe.py --strict -> harbor_gate --oracle --nop -> step2b_ready -> harbor_gate --oracle-repeat 10 -> package_task --validate -> Step 5 rubric + Step 6 submission notes (UI only, chat handoff) -> first-look -> approve_task.py --strict; exact commands in commands.md.
4. check-task.sh is only preflight; Harbor, first-look, zip, approval remain.
5. Task edits stale Step 2b evidence; approval checks .step2b-checksum.
6. Do not invent paths, task.toml fields, or commands; open source/commands.
7. Host Python is not container Python; task python3 needs Dockerfile support.
8. If terminal output disappears, run echo visible; else use a fresh shell.
9. Milestones: steps/milestone_N/{instruction.md,tests/,solution/}, version="2.0", matching [[steps]].
10. No new multi-container/UI starts; in-progress may finish.
11. All nine primary categories are open (debugging, software-engineering, data-processing, system-administration, build-and-dependency-management, games, machine-learning, security, scientific-computing). Classify honestly; do not relabel. See .cursor/rules/submission-category-blocklist.mdc and prompts/terminus_task_category_guide.md. Empirical difficulty is judged on Claude Opus 5 and GPT-5.6. Metadata and checker complexity must be medium or hard (floor medium; prefer hard). easy / too_easy do not ship.
12. New tasks must be strictly unique (≤ 15% similarity local+platform; target ≤ 10%). Never invent a similarity %. Run python3 scripts/uniqueness_probe.py inventory then probe + structure-plan before creating tasks/, then structure after the skeleton. Official dir check: python3 ci_checks/check-similarity.py tasks/<task> --include-structure --enforce-threshold. Instruction-only is not enough. Follow .cursor/rules/unique-task-generation.mdc and prompts/oneshot_unique_task.md. Never stash or move peer tasks under tasks/.peers/ or any tasks/.* path — only tasks/<slug>/ is valid.
13. Before new task drafting, task revision, packaging, or approval, check terminus_blockers_checklist.txt and remove any blocker it names. Also author to pass prompts/TASK_CHECKER.md via .cursor/rules/task-checker-ready.mdc and .cursor/rules/checker-lessons.mdc (P1–P5; R1–R9; no recurring ship-killers; prompts/AUTHOR_PRE_SHIP_PROBES.md for R8 twice / GAMEABLE traps; rubric.txt required; pytest in Dockerfile, not test.sh). After fixes, always re-run package_task.py --validate and verify zip fingerprints (especially Rust .cargo/ + checksums) — the checker scores the zip, not the chat tree.

Zip bans: output_contract.toml, quality_check_adjudication.json, construction_manifest.json, waivers.json, rubrics.txt (plural), .step2b-checksum, AI scaffolding (AGENTS.md, .cursor/, .claude/, .aider/, .continue/, skills.md). rubric.txt is required at the task and archive root (checker format). Do not put a [rubric] table in task.toml (package_task.py strips it; zip validation rejects it). Step 6 submission notes remain UI-only; see prompts/Step6.md.