Full Openscience Redesign!

[Aayam Bansal (aayambansal)]

Builds the OpenScience science workbench in incremental, tested milestones. This PR remains a draft and must not merge to main yet.

Workbench UI

• Split-view workspace: conversation center, contextual right pane that exists only when something is open. Session tabs (reorderable, persisted) on the left; open files render as reorderable, closable tabs in the right pane, capped with honest eviction and restored across reloads.
• Compact Claude-Science-inspired chrome; the display serif and decorative glow/orbit/sparkle flourishes are retired; collapsed user messages show ~5 lines.
• Composer: plan|act toggle (plan is a mode, not a specialist), inference-source label (managed / BYOK / ChatGPT — only when provable, no guessing), effort chip when non-default, auto-accept folded into the model popover.
• A persistent plan panel above the composer shows the live step checklist for the whole run with one-click plan approval.
• The command palette is a real project search: plain-text matches over session titles, conversation text, and artifact files, with message-anchor navigation. No semantic claims.

Permissions and plan mode

• Five-scope permission cards: allow once / this conversation / this project / always / deny. Standing approvals persist, are listed in settings, and are revocable; the old in-memory "always" that vanished on restart is gone.
• Cards state exactly what is granted: read-only vs read-write folder access, exact network host.
• Spend-class permissions (atlas, websearch) never inherit blanket wildcard allows.
• A webfetch to a blocked domain asks instead of erroring; "always" lands in the visible Network allow-list. Allow-list copy states its true scope.
• Plan mode stays enforced fail-closed in the backend.

Provenance, artifacts, review

• Every execution path records the shared provenance envelope — kernels, compute jobs, artifact versions, and now every bash command (redacted, with message linkage).
• GET /file/lineage exposes producing runs and messages for any file; the artifact inspector's Run and Messages tabs render that data for real (they were hardcoded empty), with show-in-conversation jumps.
• Explicit save turns a file into a durable, versioned, immutable artifact (toolbar action + end-of-turn affordances on written files); durable versions survive cleanup instead of silently expiring after 7 days.
• The Evidence surface is fully removed; provenance lives on artifacts.
• The reviewer is a direct action: a header entry launches it on the session with its provenance tools restored via session-scoped permission rules and structured context. Findings carry a lifecycle — open, addressed, confirmed only by a later verifying pass — listed in the inspector with resolve actions. Optional auto-review after durable saves (off by default). Deterministic manuscript checks are renamed "publication preflight" everywhere user-facing.

Memory

• Hermes-style bounded core memory: per-scope character budget with a capacity gauge, duplicate rejection, and a consolidation wall; a memory tool for the primary agents (add / replace / remove / search); SQLite FTS5 full-text search over notes and past session text, in the tool and the settings panel; agent-written notes are badged. Stored JSON and the settings API stay backward compatible.

Compute

• Kernel cards show live CPU and resident memory sampled per refresh, plus target; anything unreportable reads "Unavailable", never 0.
• Job dispatch requires reviewing the exact command first; scheduler wrappers are described truthfully, never fabricated.
• Model endpoints are fully removed (routes included); Atlas Compute appears only as a deliberate coming-soon slot.

Truth pass

• Display-only surfaces removed (registry-actions permission grid, fake managed-provisioning copy, dead legacy composer/center-tab/drag-drop code, shadowed CSS generations).
• Relative markdown images resolve everywhere through the raw-file endpoint; .html files preview inside a fully sandboxed iframe; DOMPurify is never bypassed.
• Scratch workspaces of deleted sessions are swept after a quiet day.

Guardrails

• Harness untouched: agent registry, agent prompts, system prompts, and prompt routing are byte-identical to main.
• No push to main. Aayam-only authorship. npm test channel only.

Verification

• Full backend suite, frontend workspace (406 tests) and shared-UI suites, monorepo typecheck (7/7), SDK/OpenAPI regeneration byte-clean, repository Prettier check — green locally and in CI on every push.

Next

• npm test-channel publish, full browser walkthrough of every surface, fix-and-republish loop until clean.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
openscience	Ready	Preview	Aug 1, 2026 7:07pm

[Aayam Bansal (aayambansal)]

Critical UI + isolation slice pushed at 42c8569.

Local verification:

• 36 focused backend isolation/kernel tests passed
• 16 focused model/compute UI tests passed
• full backend: 1,464 passed, 1 skipped, 1 known standalone SPA-fallback failure
• repository typecheck, production workspace build, Prettier, and diff checks passed
• live browser geometry verified for both model selectors, revised palette, Kernels, and Jobs

Fresh test-publish run for this head: https://github.com/synthetic-sciences/openscience/actions/runs/30551467789

No harness/workflow, memory/Hermes, Atlas Compute, or LLM provider-routing files were changed.

[Aayam Bansal (aayambansal)]

Checkpoint d1b48e9 pushed. New slices on the draft PR: compact provider-marked model picker; ChatGPT/Codex connection restored; files remain in the contextual pane; session-scoped read/write grants with traversal and symlink enforcement; fail-closed Plan-mode tool gate; lazy session-owned kernel lifecycle with real stop/crash/process identity; immutable artifact version history; and project/session-scoped UI state with compact navigation. Integrated verification at this checkpoint: 46 backend security/kernel/artifact tests passed, 292 frontend tests passed, backend and frontend typechecks passed. Memory/Hermes, Atlas Compute routing, harnesses, and workflows remain intentionally untouched. Continuing with the project-capability/folder abstraction audit and critical security review.

[Aayam Bansal (aayambansal)]

Project-capability checkpoint pushed at 7903ce3. Canonical frontend URLs now use opaque project IDs with transparent legacy base64-link replacement; SDK and session-scoped caches carry project identity. Pre-Instance repo/Atlas bridge/folder routes and all local Compute Jobs HTTP routes now resolve canonical server-owned project roots, preserve legacy clients, and return structured unknown/stale/mismatch errors. Focused additions: 47 pre-route tests, 13 settings-compute route tests, 12 compute lifecycle tests, and full frontend 301 pass / 12 skip / 0 fail with typecheck and production build. The previous head passed CI Test, Typecheck, builds, workflow lint, launch smoke, CodeQL, Secret Scan, npm publication, and all four OS package smokes; its only red was the three-file format issue corrected in this checkpoint. Fresh CI/CodeQL/Secret Scan/test-publish runs are active for 7903ce3.

[Aayam Bansal (aayambansal)]

Phase 0 package-isolation slice (DIST-003) is tested and ready to push.

What changed

• OPENSCIENCE_CONFIG_DIR and OPENSCIENCE_DATA_DIR are now authoritative runtime roots. An explicit config root prevents fallback discovery of normal XDG and ~/.openscience config, and an explicit data root takes precedence over a normal data-location pointer.
• Scanning a config directory no longer creates dependency metadata or runs bun add / bun install. An explicitly configured plugin still goes through the existing execution-config trust gate before the plugin loader can install it.
• Early synced-config/env loading and the synsci launcher now follow the explicit roots.
• openscience doctor reports executable, version, channel, native platform package, config root, data root, cache root, and state root.
• The release note now isolates npm prefix/cache/userconfig plus HOME, config, data, and all XDG roots. The native smoke packer gained a current-platform mode for local packaged validation.
• The canonical repair plan was committed separately and updated with verified evidence. No route or schema changed, so JavaScript SDK regeneration was not required.

Exact validation

• Focused real-process isolation regression: 1 passed, 0 failed, 9 assertions.
• Config suite: 53 passed, 0 failed, 97 assertions.
• Installation suite: 18 passed, 0 failed, 55 assertions.
• Full backend suite: 1,624 passed, 1 skipped, 0 failed, 4,487 assertions across 181 files.
• Backend typecheck: passed (tsgo --noEmit).
• Production frontend/native build: passed; Vite built 2,467 modules in 9.68s, embedded 483 web asset entries, and compiled @synsci/openscience-darwin-arm64.
• Launcher syntax check and Prettier check for every changed implementation/test/doc file: passed.
• Packaged npm flow: locally generated wrapper + Darwin arm64 platform tarballs installed with npm --ignore-scripts into isolated npm/config/data/XDG roots. Packaged openscience --version returned 0.0.0-aayam/kernel-science-workbench-202607310623; doctor reported the assigned roots and @synsci/openscience-darwin-arm64; normal decoy root tree and file hashes were unchanged; npm uninstall removed the packages while preserving the isolated config/data roots.

State and remaining blockers

• DIST-003 is implemented in source and locally packaged behavior. Exact published-prerelease installation and test dist-tag verification remain open acceptance work.
• Remaining Phase 0 work: authoritative cross-surface version metadata (DIST-001), the complete lifecycle-disabled platform contract (DIST-002), a redacted support bundle, registry prerelease verification, and first-load console verification (CONSOLE-001).
• No npm version was published or dist-tag moved. No Atlas, compute, provider, or production service was contacted. No migration or normal persistent user data changed; only isolated temporary test roots were created.
• Outgoing commits are 53f704e0, 855124a2, and aef55c89; all author and committer fields are exactly Aayam Bansal <aayambansal@gmail.com>, with no co-author or agent/bot attribution. They are not pushed yet.

[Aayam Bansal (aayambansal)]

Push completed to aayam/kernel-science-workbench at aef55c890ec6dbd6bace32699ffccd1699393588 (range 9d5f344c..aef55c89). The pre-push hook ran workspace typechecking: 7 successful tasks, 0 failed. The unrelated .openscience/openscience.jsonc modification and openscience-starters/ directory remain uncommitted. No push was made to main, and PR #240 was not merged or closed.

[Aayam Bansal (aayambansal)]

Phase 0 version-authority slice (DIST-001) is tested and ready to push.

What changed

• The production frontend now receives the same Script.version used to compile the CLI and native package instead of displaying the stale workspace-manifest version.
• The build emits embedded version.json metadata with version and channel. Settings and error diagnostics continue to consume platform.version, which now carries the release build value.
• openscience doctor reads the embedded frontend metadata, prints it, and warns if its version/channel differs from the compiled CLI.
• The native smoke packer now refuses native/UI version mismatches and produces matching local wrapper and synsci launcher tarballs for packaged validation.
• A production-build regression was added. It failed before the fix because generated assets did not contain the explicit release version, and passes after the fix.
• No route or API schema changed, so JavaScript SDK regeneration was not required.

Exact validation

• Pre-fix regression: 0 passed, 1 failed as expected because VITE_OPENSCIENCE_VERSION was absent from the production assets.
• Post-fix production-version regression: 1 passed, 0 failed.
• Frontend unit suite with an explicit unit-file list: 404 passed, 12 skipped, 0 failed, 2,375 assertions across 80 files.
• Release-installation suite: 20 passed, 0 failed, 61 assertions across 11 files.
• Full backend suite: 1,626 passed, 1 skipped, 0 failed, 4,490 assertions across 182 files.
• Workspace typecheck: 7 successful tasks, 0 failed.
• Prettier and diff checks: passed.
• Final production frontend/native build: 2,467 modules transformed, Vite build completed in 9.44s, 484 embedded web entries generated, and @synsci/openscience-darwin-arm64 compiled.
• One broad bun test src script attempt also auto-discovered a Playwright E2E spec and produced the repository raw-Bun/Playwright loader error; the supported explicit unit-file run above is green and excludes E2E specs.

Packaged flow reproduced

• Built and packed local version 2.0.2-test.local.202607311445.
• Installed both locally generated npm tarballs with lifecycle scripts disabled and with isolated HOME, npm prefix/cache/userconfig, OpenScience config/data, and XDG roots. Public registry access was forced to 127.0.0.1:9, so package resolution could not use the external npm registry.
• The installed npm records, openscience --version, synsci --version, doctor, /global/health, and served /version.json all reported 2.0.2-test.local.202607311445.
• Doctor reported Frontend version: 2.0.2-test.local.202607311445 (test) and @synsci/openscience-darwin-arm64.
• Uninstall removed both npm packages and preserved the isolated config/data roots.

State and remaining blockers

• DIST-001 is implemented for current runtime surfaces and local package acceptance. A dedicated redacted support-bundle export and exact registry-published prerelease verification remain open.
• DIST-002 cross-platform lifecycle-disabled native-package behavior, test dist-tag resolution, and CONSOLE-001 first-load verification remain before Phase 0 is complete.
• No npm version was published and no dist-tag was moved. No normal persistent user data or migration state changed. GitHub was contacted only for branch fetch and this PR update; no Atlas, compute, provider, or production service was contacted.
• Outgoing commits are 9f0b3c7c8b17129eb71e11ff863ef284f39dbd5c and 4bee503255ea3bc29c9f24785c3df5a0c9ed9332; both author and committer identities are exactly Aayam Bansal <aayambansal@gmail.com>, with no co-author or agent/bot attribution. They are not pushed yet.
• The unrelated .openscience/openscience.jsonc modification and openscience-starters/ directory remain uncommitted.

[Aayam Bansal (aayambansal)]

Push completed to aayam/kernel-science-workbench at 4bee503255ea3bc29c9f24785c3df5a0c9ed9332 (range aef55c89..4bee5032). The pre-push hook ran workspace typechecking: 7 successful tasks, 0 failed. No push was made to main; PR #240 remains open and unmerged; no npm package or dist-tag changed. The unrelated .openscience/openscience.jsonc modification and openscience-starters/ directory remain uncommitted.

[Aayam Bansal (aayambansal)]

Pre-push repair update — DIST-002 native npm lifecycle independence

Outgoing commits:

• 4405d9ee — wrapper/native install implementation and regressions
• dde98ec3 — canonical-plan verification record

What changed:

• Native package resolution is authoritative at wrapper launch and no longer depends on lifecycle hooks.
• postinstall is advisory (|| exit 0).
• The wrapper no longer deletes ~/.openscience manifests, locks, modules, or attributes.
• Missing native packages no longer fall through to unrelated Homebrew, /usr/local, or legacy-home installations; the error names the expected scoped package.
• A canonical 11-target table now drives the build and a real npm tarball selection matrix for macOS arm64/x64, Linux x64 glibc/musl, Linux arm64 glibc/musl, and Windows x64.

Exact verification:

• Focused wrapper/manifest/matrix regressions: 10 passed, 0 failed.
• Release installation suite: 23 passed, 0 failed, 80 assertions across 12 files.
• Full backend suite: 1,629 passed, 1 skipped, 0 failed, 4,509 assertions across 183 files (74.18 s).
• Workspace typecheck: 7/7 packages passed.
• Production web/native build: 2,467 modules, 484 embedded web entries, Darwin arm64 compile passed; exact local version 2.0.2-test.local.202607311451.
• Fresh isolated local-tarball installs passed with lifecycle scripts disabled, npm default policy, and explicit allowlisting. Wrapper package, native package, CLI, launcher, health, and embedded web metadata agreed on the version; normal-home sentinels survived; uninstall preserved assigned config/data roots.
• Registry was forced to 127.0.0.1:9 during local package validation; no external package or token was used.
• Two preliminary manual checks were discarded rather than counted: one accidentally narrowed PATH below npm/node, and one used npm ls against local absolute file dependencies, which npm reports as invalid despite a runnable install. The three corrected fresh-install runs above are the acceptance evidence.

Release state:

• No npm version was published. No dist-tag was moved. latest was not read or changed.
• The protected .openscience/openscience.jsonc and openscience-starters/ remain uncommitted.

Remaining Phase 0 blockers: exact published @test install/tag resolution, cross-host native execution (selection is covered; execution is only verified on Darwin arm64), redacted support bundle, and browser first-load console verification. Next I will install the exact test package with isolated prefix/config/data roots and browser-drive model setup, Atlas-optional flows, and UI/console/network behavior.

[Aayam Bansal (aayambansal)]

Post-push confirmation: 4405d9ee..dde98ec3 is now on aayam/kernel-science-workbench. The push hook reran workspace typechecking: 7/7 packages passed. No npm package or dist-tag was changed.

[Aayam Bansal (aayambansal)]

Pre-push update — DIST-004 packaged macOS folder-access warning repair

Commits ready to push:

• 2fc12006 — fix(ui): make folder access warning truthful
• 806a8c95 — docs(plan): record packaged folder access repair

What changed:

• Shared CLI/server probe now classifies only EACCES and EPERM as protected-folder denials; missing/empty Desktop, unrelated errors, and non-macOS hosts fail quiet.
• Removed automatic System Settings launch, raw filesystem errors, and the hard-coded /opt/homebrew/bin/openscience instruction.
• Corrected the permission principal to the terminal/desktop app that launched OpenScience, with which openscience only for direct binary launches.
• Replaced the acronym-only FDA chip and oversized modal with a compact responsive Folder access recovery panel, numbered remediation, explicit settings link, focus states, and reduced-motion behavior.
• Updated the canonical repair plan with the public test reproduction and final packaged verification.

Exact validation:

• Focused backend regression: 4 passed, 0 failed.
• Full backend: 1,633 passed, 1 skipped, 0 failed across 184 files (4,517 assertions).
• Full frontend: 396 passed, 12 skipped, 0 failed across 78 files (2,362 assertions).
• Backend typecheck: passed (tsgo --noEmit).
• Frontend typecheck: passed (tsgo -b).
• Prettier check and git diff --check: passed.
• Production frontend/native build: passed (2,468 frontend modules; 484 embedded web assets; darwin-arm64 native binary).
• JavaScript SDK regeneration: passed; generated SDK remained unchanged.

Packaged behavior:

• Built and installed local npm test package 2.0.2-test.local.202607311515 into fresh isolated binary/config/data/cache/state/npm-config/npm-cache/home roots.
• openscience --version and synsci --version both reported 2.0.2-test.local.202607311515.
• Fresh browser history with no ~/Desktop: warning stayed absent after project loading.
• Fresh browser history with an actual unreadable ~/Desktop: Folder access appeared; the compact dialog stayed open after a denied Recheck, exposed no raw error or assumed install path, and opened no settings tab automatically.

Release status / remaining blocker:

• No npm package was published and no dist-tag was moved.
• Registry authentication is unavailable in this environment, so the public test tag remains 2.0.2-test.42.1; exact published-test-build validation remains open.
• Next after this coherent slice: resume the plan’s remaining critical issues and continue the packaged browser audit of setup/models/Atlas surfaces.

[Aayam Bansal (aayambansal)]

Post-push verification — DIST-004

Pushed 2fc12006 and 806a8c95 to aayam/kernel-science-workbench; remote branch tip now exactly matches 806a8c9581b5024b567d9cb93513a7fa23ab8b2d.

Push hook result: 7/7 typecheck tasks passed. Both outgoing commits retain exact author and committer Aayam Bansal <aayambansal@gmail.com> and contain no attribution trailers.

The protected unrelated .openscience/openscience.jsonc modification and openscience-starters/ directory remain unstaged and were not pushed. No npm package was published, no test tag was moved, and no production/latest action was taken.

DIST-004 is complete in source and local packaged-browser validation. Public test-package validation remains blocked only by unavailable npm registry authentication; public test remains 2.0.2-test.42.1.