--- name: computer-setup description: Prepare or resume the Starter Pack computer baseline with approved capability installs, cross-context command checks, private saving, file transfer, and actual phone-to-harness control. version: 1.1.0 updated: 2026-11-12 --- # Computer Setup The learner directs work in their selected agent platform. The editor is their way to inspect and edit files, especially Markdown plans and notes. Do not require an editor extension workflow or a coding lesson. ## Load and stage before changing anything 1. Read the [setup manifest](https://starter.uppercut-labs.com/setup/computer-setup.manifest.json), [canonical requirements](https://starter.uppercut-labs.com/agent/requirements.json), and [progress guide](https://starter.uppercut-labs.com/artifacts/progress/README.md). Use that registry's IDs, prerequisites and completion rules; the capability manifest supplies execution evidence, not a second phase-gate definition. 2. Load the learner's existing canonical `starter-progress.json`, portable handoff, and `setup/computer.json`. Preserve completed Phase 1 outcomes, unknown fields, notes, decisions, declines and customized Workbench. A private repository is not an entry prerequisite. Initialize the [setup state format](https://starter.uppercut-labs.com/setup/state.example.json) only when no setup state exists. Never reset real progress from an example. 3. Preserve the chosen computer harness separately from the phone companion. Reuse supplied packets; do not repeat intake or a failing fetch loop. Gemini chat remains outside the supported companion route; do not replace the bounded Claude fallback or require a paid upgrade. Antigravity is the Google computer route, not Gemini chat. 4. Make this Computer Setup skill and focused references available before their work starts. For a native install, inspect the selected harness's current official skill locations and discovery behavior, obtain the relevant approval, and verify discovery/invocation in the actual workspace. A supplied document can be used as context without being natively installed. Record name, source/version, scope/location and `native_discovered`, `context_only` or `not_tested` privately. Never spray copies into guessed global folders. 5. At the first computer session, make a short private intervention log available. Record significant attempts, observed results, cause hypotheses, approved repairs, verification, and beginner-unlikely expert rescues. Keep credentials and environment dumps out. Append corrections rather than rewriting historical reports. Full skill theory, a 25-skill bundle, orchestration and Phase 3 are not setup requirements. ## Keep the local viewer and email non-gating Follow the [Workbench lifecycle](https://starter.uppercut-labs.com/artifacts/progress-workbench/README.md) at entry: download the verified self-contained HTML beside canonical progress only when missing; preserve an existing customized file. The viewer is read-only and not a second progress database. With no filesystem, preserve portable state and hand it to the computer-capable harness. Ask once about missing development-account and notification-email preferences; explicit decline/deferral is enough to continue. Follow the progress guide for `choices.development_email` and `choices.service_auth`. Preserve legacy non-email overrides as sign-in notes. Never infer emails, change Git author identity, or grant Access merely from these preferences. A chosen private cloud-hosting route needs separate explicit Access identity confirmation. When Python is verified, start the local viewer independently of GitHub. Inspect port 8000 and reuse a listener only if its exact progress directory and loopback binding match; otherwise choose 8001 through 8010. Invoke the resolved interpreter with `-m http.server PORT --bind 127.0.0.1 --directory "ABS_PROGRESS_FOLDER"`. Record the managed process, binding, actual port and resume/stop action. Use a hidden window for a Windows background process. Never stop unrelated listeners or serve a broader workspace. Refresh after saves; direct-file mode needs the saved JSON selected again. Viewer failure never blocks other setup or graduation. A phone's localhost is not this server. Optional [Tailscale Serve access](https://starter.uppercut-labs.com/setup/private-workbench.md) requires separate approval and an isolated allowlisted snapshot, not a proxy to this repository-root listener. Later [Cloudflare Access hosting](https://starter.uppercut-labs.com/setup/workbench-publishing.md) is a different optional route, lower priority than instant-app deployment extra credit. Neither is required. ## Inspect, approve, install, verify, checkpoint 1. Inspect OS/version/architecture, native versus WSL context, selected harness, permissions, command resolution, installed versions, and prior results. Use [Windows](https://starter.uppercut-labs.com/setup/windows.md), [macOS](https://starter.uppercut-labs.com/setup/macos.md), or best-effort [Linux](https://starter.uppercut-labs.com/setup/linux.md). Classify each capability as working, missing, failed or not tested. Reuse the [Codex readiness checkpoint](https://starter.uppercut-labs.com/setup/codex-readiness.md); recheck only facts affected by a material context change. 2. Reuse healthy compatible installations. For missing tools compare the verified package-manager route with a supported vendor binary/installer, including provenance, signatures/checksums when published, updates, dependencies and repair. Resolve current supported versions at execution time. Avoid source compilation when a supported binary fits. Homebrew/WinGet are good options, not ceremony when direct tooling already supplies the baseline. 3. Explain the necessary batch and obtain approval, honoring existing authorization. Optional packs are separate. Execute dependencies in order using recoverable managed jobs; do not run conflicting installers together, stop OS installation daemons, bypass security, or reinstall everything to fix one command. 4. Prioritize Git and GitHub CLI when their prerequisites work. Follow [authentication and Git privacy](https://starter.uppercut-labs.com/setup/authentication.md). As soon as appropriately scoped GitHub write access exists, create or confirm the private progress repository and save an intended nonsecret commit. Verify remote files, revision and private visibility. An existing approved connector may save earlier, but does not waive baseline Git/gh. A failed save retains local/portable work and blocks claiming the private-save outcome, not independent setup. 5. Verify the selected harness in its intended workspace and the [required IDE](https://starter.uppercut-labs.com/setup/ide.md): Codex/Claude Code use VS Code; Antigravity agent application and Antigravity IDE are checked separately; Cursor's desktop editor is sufficient. Preserve existing editors, settings, extensions and workspaces. 6. Verify every applicable baseline CLI in the current shell, a newly launched terminal, the IDE terminal and a real harness subprocess/task. Include later selected CLIs. Use [command-context evidence](https://starter.uppercut-labs.com/setup/tool-troubleshooting.md#command-resolution-in-each-context), not installer exit status. Python is required up front; use uv to discover a healthy interpreter or provision a supported one. Project-local dependencies resolve through their project runner, never global `node_modules/.bin` PATH entries. 7. Complete [networking and actual harness control](https://starter.uppercut-labs.com/setup/remote-access.md) plus [both file-transfer directions](https://starter.uppercut-labs.com/setup/file-transfer.md). Record computer connection, phone connection, intended device/network identity, phone-to-computer receipt, computer-to-phone receipt/opening, and actual phone-to-harness action independently. A native remote message cannot complete Taildrop; Tailscale alone cannot complete harness control. Preserve already successful checks and honest learner-reported evidence. 8. After each outcome update `setup/computer.json` and only the supported canonical progress outcomes. Save nonsecret install source/scope/architecture/version/path, context results, evidence source/time and the exact next action. Pause for sign-in, administrative prompts and necessary restarts; save before relaunch and inspect again on resume. Do not persist secrets. 9. Review the final state diff, save it to the intended private remote, and verify its revision and files. Do not make a commit merely to record its own SHA in another commit: keep verification externally in the session or next substantive save. If saving fails, retain local successes with an explicit pending remote-save action; do not claim durable completion or start project planning. ## Setup closeout and handoff Reconcile required setup outcomes against the current registry and their evidence. Do not infer missing checks from an old aggregate boolean, a file arriving, or historical phase completion. Preserve older graduation with its original revision; propose targeted follow-up rather than silently revoking it. Personal-state repairs require a separately reviewed, approved diff. When setup outcomes and the private save pass, proactively state that setup is ready. This is not Phase 2 graduation: the approved project plan, durable source, live main behavior and final private save still follow the registry. Stage the current [Grill to Build skill](https://starter.uppercut-labs.com/skills/grill-to-build/current/SKILL.md) and its [planning templates](https://starter.uppercut-labs.com/artifacts/grill-to-build/README.md) before their first use; it is the standard Phase 2 planning workflow, with the course rules in the [Phase 2 planning reference](https://starter.uppercut-labs.com/phases/2/planning.md). Stage the current [Quick Build skill](https://starter.uppercut-labs.com/skills/quick-build/current/SKILL.md) and its [templates](https://starter.uppercut-labs.com/artifacts/quick-build/README.md) only when the learner explicitly asks for Quick Build. Verify native discovery or label context-only delivery honestly. Do not add Phase 3 machinery as a gate. Return a compact handoff with canonical progress/private location, selected harness, resource revisions, verified outcomes, unfinished work, optional choices and the exact next action. No secrets and no claim of automatic cross-provider chat transfer. Suggest a new project unless the learner asks to reuse their Phase 1 build, and preserve the explicit build go-ahead. ## Reuse Codex readiness at milestones Reuse the [Codex readiness checkpoint](https://starter.uppercut-labs.com/setup/codex-readiness.md) at Computer Setup entry and closeout. A readiness check never installs Codex, starts login, or runs inference. Preserve prior results when the target is unchanged; recheck only affected facts after a material change. Record readiness as the checkpoint describes. Keep an explicit decline of the Codex route and a deferral with one recovery action; continue to project planning either way. Selected Docker Desktop is conditional and installed last; OrbStack is excluded. For pnpm approvals, runtime/module errors, Docker or Rosetta, use [bounded troubleshooting](https://starter.uppercut-labs.com/setup/tool-troubleshooting.md); do not apply those fixes universally.