fix: make sandboxed research execution reliable

[Ishaan Gangwani (ishaan1124)]

What does this PR do?

• Makes the research harness accept and canonicalize harmless legacy tool-call shapes without weakening strict schemas, including direct and batched compute_job calls.
• Enables low-friction routine terminal, kernel, shell, and local-compute work inside verified native sandboxing while keeping remote jobs, environment mutations, project code, and host execution behind project trust.
• Removes agent-guessed WebFetch download caps and the fixed Modal browser-download ceiling; downloads use live free disk minus a 512 MiB host reserve, stream with cancellation-safe cleanup, and fail with actionable errors.
• Hardens cross-process settings, data-root relocation, credential ownership, and kernel/process teardown across Linux, macOS, and Windows.
• Updates workspace copy, documentation, OpenAPI, generated SDK types, and bundled docs to match the shipped behavior.

How did you verify your code works?

• bun test --silent --timeout 15000 in backend/cli: 2217 passed, 20 platform skips, 0 failed across 273 files.
• Root bun run typecheck: 7/7 packages passed.
• Modal/jobs/settings focused gate: 98 passed, 0 failed; final adapter/browser regression rerun: 43 passed, 0 failed.
• Settings/documentation focused UI gate: 39 passed, 0 failed.
• bun run --cwd backend/cli build --single with required Bun 1.3.14: passed and built darwin-arm64.
• Privileged Linux lifecycle gate: 32 passed, 2 platform skips, 0 failed; repeated fork-storm containment: 3/3 passed with zero survivors.
• macOS lifecycle gate: 31 passed, 5 platform skips, 0 failed; musl supervisor gate: 5/5 passed.
• Targeted Prettier and git diff --check: clean. Gitleaks pre-commit hooks: clean.
• Independent pre-landing and design re-reviews: CLEAR, no P0/P1/P2.

Documentation

• README.md and SECURITY.md: describe the low-friction sandbox default and full project-trust boundary.
• frontend/docs: document sandbox behavior, download flow, commands, and security semantics.
• frontend/landing/public/docs: regenerated from the current docs source.
• CHANGELOG.md: records the harness reliability, download, compute, and sandbox changes under Unreleased.
• Generated OpenAPI and JavaScript SDK artifacts are synchronized with the server contract.

Checklist

• bun run typecheck passes
• bun test (in backend/cli) passes
• Scoped Prettier checks are clean
• Generated SDK/OpenAPI and bundled docs are current
• Security/reliability and design reviews are clear
• Linked issue — direct release-hardening work; no issue was supplied
• Screenshots/video — no new visual layout; settings copy and existing controls only

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
openscience	Ready	Preview	Aug 14, 2026 5:54pm