Phase 20f sandbox command policy - #2
Merged
Merged
Conversation
Make the first five minutes excellent and focus the product on four workflows. - Demo mode: deterministic, no-key `apeironcode demo` with todo-app, fix-test, improve-ui, and review scenarios. Writes real files and runs real local validation; never calls a provider, model, Ollama, or the network. - CLI help: beginner-focused `--help` (fix/build/improve/review/demo/doctor/ setup), with `help developer` and `help advanced` for the rest. Adds the four headline workflow commands. Every existing command stays registered. - Setup/doctor: intent-based `setup` menu (demo/local/cheap/best/BYO), new `local setup`, `local doctor`, `real setup`; doctor appends friendly next steps and points to demo when no real provider is configured. - Providers: local OpenAI-compatible endpoints (localhost/127.0.0.1/0.0.0.0/ ::1/host.docker.internal/.local) no longer require an API key; remote ones still require OPENAI_API_KEY and fail fast cleanly. - Docs: README four-workflows + no-key demo sections; new docs/demo.md and docs/getting-started.md; providers.md documents no-key paths and that mock is a test/demo stub, not a real coding model.
Phase 20B: make `apeironcode fix` a real repair loop and harden file-plan parsing. File plans (Task A/B): - New robust parser (src/agent/filePlanParser.ts) with multi-strategy extraction (direct/fenced/balanced/tolerant-repair/tool-call-envelope), a zod schema (filePlanSchema.ts), and shape coercion/repair (filePlanRepair.ts). No single strict JSON.parse failure kills the workflow; failures return exact validation errors plus a ready-to-send correction prompt. Rejects absolute/.. paths, .git, .env (by default), and bad operation/content combinations. - Legacy parseFilePlanResponse now delegates to the robust parser. Repair workflow (Tasks C–H) under src/repair/: - projectDetector / commandDetector / checkRunner: detect package manager, framework, scripts; ordered checks (build/e2e gated behind --all); redacted, truncated check output. - errorParser / failureClassifier: structured issues (typecheck, lint, test, build, module-not-found, syntax) with file/line and confidence. - contextBuilder: small, secret-free context packet with a selection trace. - approvalBundle: one bundled approval with risk level; high-risk commands (rm -rf, sudo, curl|sh, npm publish, git push) blocked. - checkpoint: snapshot before apply, rollback on failed apply (.apeironcode/ checkpoints, gitignored). - repairLoop + finalReport + repairPlanner: detect → run → propose minimal plan → one correction pass → bundled approval → checkpoint → apply → rerun, up to a mode-based limit. Safe/dry-run/all/until-green modes; --commit deferred. CLI: `apeironcode fix [--dry-run --safe --all --until-green --commit --yes]` wired via src/cli/fixHandler.ts; dry-run never constructs a provider or writes. Tests: file-plan parser, detectors, error parser, checkpoint, approval bundle, repair-loop (injected deterministic deps), and real-node fixtures under tests/fixtures/repair. Docs: docs/fix-code.md + README/getting-started updates.
Phase 20C: make `apeironcode new` / `apeironcode build` create small apps from deterministic templates + real validation — no API key or model required. App builder (src/appBuilder/): - requestParser: free-text idea → honest AppRequest (name/slug/kind/stack/style/ features). Auth/db/billing become mock/local UI; Next.js downgrades to Vite. - templateRegistry + templates: static-todo, static-landing, dashboard-static, vite-react. Shared premium design system (tokens, spacing/type scale, focus states, responsive, empty states). - templateRenderer: deterministic render + path safety (no absolute/.., .git, .env, no writes outside the app dir). - designQuality: static checklist (title, viewport, linked assets, button/focus states, labeled forms, responsive, no obvious overflow). - appValidation: files-exist, linked-assets, node --check, design checklist; Vite build/typecheck detected and reported as skipped (never auto-installed, never faked green). - appBuilder orchestrator + appReport: parse → select → approval bundle → write → validate → honest report. Non-empty dirs require --overwrite and existing files are never replaced. --dry-run writes nothing. CLI: `apeironcode new`/`build [--stack --style --dir --dry-run --overwrite --yes]` wired via src/cli/appBuilderHandler.ts; both are beginner commands. Approval is required (auto in bypass/trusted or --yes; non-interactive declines). Tests: requestParser, templateRegistry/renderer (+path safety), designQuality, appValidation (broken JS / missing CSS / skipped Vite), orchestrator (create/ dry-run/overwrite-protection/rejection), and docs/packaging honesty. Docs: new docs/build-apps.md (packaged) + README "Build an app" section.
Phase 20D: generated apps prove they actually render via an optional Playwright
browser smoke. No API key or model required; Playwright is optional and never
fake-passes.
UI smoke (src/uiSmoke/):
- targetDetector: static / Vite / unknown via .apeironcode/app.json metadata,
index.html, or package.json. staticServer: tiny local file server with
path-traversal protection.
- SmokePage/BrowserLauncher abstraction so unit tests use a deterministic fake;
playwrightRunner dynamically imports playwright (no hard dependency) and
returns null → "skipped" when Chromium is unavailable.
- interactionChecks: template specs (static-todo add-task, dashboard cards/rows,
landing hero/CTA/nav) + generic checks. uiReport: text/JSON, redacts secrets,
saves .apeironcode/ui-smoke/{latest.json,latest.png,runs/<ts>/...}.
- runUiSmoke orchestrator: detect → serve → browser → checks/interactions →
screenshot; honest skip for no-browser / Vite (deps) / unknown targets.
CLI: `apeironcode test-ui [path] [--json --screenshot --port --timeout]`
(src/cli/testUiHandler.ts); exits nonzero only on real failure.
App builder integration: writes .apeironcode/app.json; `new`/`build` gain
--ui-smoke / --no-ui-smoke / --require-ui-smoke. Reports show UI smoke
passed/skipped/failed honestly; --require-ui-smoke fails the build when the
browser is unavailable/fails.
improve: attaches the latest UI smoke report as context (Task F).
Tests: targetDetector, staticServer (traversal), runner (passed/failed/skipped
via mocked launcher), report (redaction/latest.json), interactions, app-builder
integration, improve context, docs/packaging — plus an opt-in real-browser e2e
(APEIRONCODE_RUN_BROWSER_E2E=1 npm run test:ui-smoke). Docs: docs/ui-smoke.md.
Phase 20E: make fix/build/new/test-ui/improve/review feel premium with a shared progress board, friendlier approvals, diff preview, command log, rollback, and a beginner root screen — without rewriting the Ink app or breaking non-interactive mode. - Progress model (src/progress/): UI-agnostic ProgressRun (tasks/files/commands/ approvals/artifacts) + ProgressStore + compact plain-text formatter. Symbols ✓→○✗–; failure excerpts short in normal mode, full in verbose. - Approval bundle 2.0 (src/safety/): shared risk classifier (high-risk commands never bundled/auto-approved), compact bundle builder with Files-only option, and a redacted diff preview (create/modify/delete/rename, binary/huge summary). - Rollback (src/cli/rollbackHandler.ts + checkpoint list/load): `apeironcode rollback list | --last | <id>`, confirmation unless --yes, only touches files in the checkpoint, graceful when none exist. - Review (src/review/): deterministic, provider-free git-diff review (bugs/tests/security/maintainability/next steps); routes to the agent for a deeper review only when a real provider is configured; no-diff/non-repo handled. - Root home (src/ui/rootHome.ts): beginner "What do you want to do?" menu shown at interactive startup (TTY only; non-interactive unaffected). - App report surfaces UI smoke screenshot/report paths + a `test-ui` next step. Tests: progress transitions/formatting, risk/approval/diff, rollback round-trip, deterministic review heuristics, root menu, docs/packaging. Docs: docs/terminal-ux.md (packaged) + README. Updated help test for review/rollback.
Phase 20E.5: make the Phase 20E premium UX real in live workflows (not just helper tests), without rewriting the Ink app or breaking non-interactive mode. App builder (new/build): - Approval shows the bundle 2.0 format (compact summary + risk + validation + options) plus a redacted diff preview of the files to be created. - Build report renders a Plan checklist (parse → template → approve → write → validate → UI smoke) derived from a real ProgressRun, plus the existing completion panel (files/validation/artifacts/next, incl. test-ui). Fix: - Repair approval bundle now includes a redacted diff preview (old vs planned content) via safety/diffPreview, with [Approve all]/[View diff]/[Reject]. - High-risk commands are still stripped from the applied plan even under --yes. - After the run, fix prints a completion/failure panel (src/cli/fixProgress.ts + progress/completionPanel.ts) with next actions (fix --until-green, rollback --last, test-ui) and surfaces the checkpoint. Shared: new src/progress/completionPanel.ts renders honest Done/Stopped panels. Tests: app-builder bundle 2.0 + diff + Plan checklist + reject-changes-nothing; fix repair-bundle diff preview + failure/green completion panels; repair loop strips high-risk commands under approval. Docs: terminal-ux/fix-code/build-apps note the live wiring.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Changes
Validation
npm run typecheckpassesnpm run lintpassesnpm testpassesnpm run smoke:dogfood)npm run check:file-sizepassesSafety checklist
provider.chat();provider.stream()onlyToolRegistry