The solution card: what drove a run, recorded and named (without the submit gate)

[Ruqii]

Split out of #34 — the additive half, with the submit-time capability gate left behind.

What lands here

The solution card: the labels that decide what a run measured — shape, agent build, model, provider, the options that actually applied, skill, command template, setup line, timeout — plus a sha256 digest over the identity-bearing fields, from canonical JSON so the site can recompute it and refuse a mismatch.

• report.json records it under provenance.solution.adapter and provenance.solution.adapter_digest. Both additive and optional; no existing key changed (CLAUDE.md constraint 4).
• A live-synced run is named after it. identity.name, model.config and skills.installed come from the card, so a site-graded row reads claude-agent-acp@0.76.0 · sonnet · a/b@0123456 instead of "model not recorded".
• tp submit shows the command line before publishing it. The card's cmd and setup are echoed verbatim — never paraphrased, truncated or re-quoted — and unconditionally, including under --yes: that flag means the user pre-consented to skipping the prompt, not that they were never told what a carded submit makes public.

docs/reference/solution-card.md + tests/data/solution_card_vectors.json are the contract for the TypeScript side (W1/W2/W6/W7). Both worked examples and all four shared vectors recompute byte-for-byte from live code.

What is deliberately left out, and why

#34 also gated tp submit on a GET /api/v2/capabilities probe for features.solution_adapter.supported, withholding the solution repository when the server would not confirm it stores cards. Verified against the deployed site today:

GET https://trapstreet.run/api/v2/capabilities
features: local_live_progress, platform_grading, checkpoint_recovery, run_context

There is no solution_adapter key, so that gate resolves one way for every caller. Merging it now would mean every carded, anchored tp submit loses its repo and goes up unranked — which is exactly the built-in-shape runs, the path that exists so someone can measure an agent without writing a solution. Three things compound it:

1. No override. Nothing on the command line can opt back in.
2. A probe failure is indistinguishable from a server "no." capabilities() catches httpx.HTTPError and returns {}, which takes the same branch, and the user-facing message reads the same either way — so a flaky network silently demotes a run.
3. The secret scan that would make the verbatim display a net is not built (cmd/setup are shown, not checked), so the gate adds surface without adding protection.

None of that is an argument against the gate — it is an argument for landing it with the site work that makes it resolve both ways. The card model, card_digest and card_label here are unchanged from #34, so that half rebases onto this cleanly.

Gates

Gate	Result
ruff format --check	123 files already formatted
ruff check	All checks passed
scripts/lint_docs.py	docs lint OK
pyright src/	0 errors
pytest --cov	1336 passed, 15 skipped, 100% branch coverage

Built on top of #35, whose RUN_SETTLED changes are preserved intact.

🤖 Generated with Claude Code