feat(envs): browse crowd-cast-envs task packages in the UI

[yllkryeziu]

A read-only Environments view (g n) for the RL task packages in crowd-cast-envs, modelled on the MiMo RL-Envs Explorer. This PR only adds browsing; it doesn't run anything.

Config

[repos]
crowd-cast-envs = "/fast/project/<group>/repos/crowd-cast-envs"

[envs]
repo = "crowd-cast-envs"          # must name a [repos] entry
# rev       = "origin/main"       # default: as last fetched
# tasks_dir = "environments/crowd_envs_v1/crowd_envs_v1/tasks"

If [envs] is missing, the view explains how to enable it. tasks_dir must be relative with no ...

How it reads

• Uses git plumbing at a commit: rev-parse, ls-tree, cat-file --batch and log -- <task>. It never reads the work tree and never fetches; run git fetch in the checkout to see new tasks. GIT_TERMINAL_PROMPT=0 and GIT_OPTIONAL_LOCKS=0 are set.
• ?rev= is validated: no leading -, no .. or :, length ≤ 200. Snapshots are cached by resolved sha (max 64), so a page is about 30–65 ms against the real repo.
• Grader side (scoring/, reference/) is only served with grader=1, and the server enforces this: grader files return 403 without it. Package-level provenance.json (built_from ids) is never served; a test asserts this.
• Files over 1 MiB are truncated. A NUL byte or invalid UTF-8 marks a file as binary. Shiki highlighting is skipped above 200 KB.

UI

• List: search across title, key, tags, authors and prompt, plus tag chips. Columns: World, Tags, Authors, Workspace (files · size), Agent time. Keys: j/k/Enter, /.
• Task page with a sticky outline:
  • The task: metadata and prompt.md.
  • Workspace: file tree plus viewer, README first.
  • Sandbox: runtime, timeouts, extensions, setup, captured artifacts.
  • History.
• Behind the grader toggle:
  • Grading: the reward formula; outcome, invariant and veto groups with points, latitude and out-of-bounds notes; the judge system prompt from scaffold/judge.py; the scorer source.
  • Provenance: recording, predicted difficulty, findings, checks, reproduce steps, builder notes.
  • Reference solution, opened on solve.sh.
• Markdown is rendered with marked and sanitized with DOMPurify, with a Rendered/Source switch. Relative links and images are neutralised.
• Revision picker: default, branches and tags, commits touching tasks, or any sha typed in. File reads are pinned to the commit the page loaded. Pre-contract commits (task.toml layout) show an explanation instead of an empty list, and they appear greyed out in a task's history.

Not in this PR

Rollouts, calibration results, a cross-task reward-design page, and generic RL-Env-Contract repos. Possible follow-ups: link runs to the env revision they trained on (via the RL repo's submodule sha), and add calibration results.

Testing

• cargo test --all-features: 15 new tests over a temp git fixture.
  • rev validation; unknown rev returns None.
  • Snapshot finds manifests only and reports skipped dirs; older revisions.
  • Grader opt-in; provenance never surfaces.
  • History, plus the manifest flag on history commits.
  • Text vs binary; truncation; refs skip */HEAD.
  • Config validation.
• cargo clippy --all-features --all-targets and cargo fmt --check are clean.
• npm test: 64 pass, including new envs.test.ts. npm run check is unchanged from main (35 errors / 9 warnings, all pre-existing).
• Manual: labctl serve on a copy of the Berlin config plus a local Postgres, against the real crowd-cast-envs checkout.
  • 18 tasks at d897b4c.
  • Grader 403s; traversal 404s; bad revs 400/404.
  • Older and pre-contract revisions.
  • A headless-Chromium walkthrough of every section in light and dark themes, with no console errors.

🤖 Generated with Claude Code