Set up PhaseDrift in this project, then prove that reporting works on a real task.

Read https://phasedrift-site.jppotess-main.workers.dev/llms.mdx/docs/getting-started/content.md and https://phasedrift-site.jppotess-main.workers.dev/llms.mdx/docs/harnesses/content.md when needed. This request authorizes local installation, project instructions and supported project hooks. Preserve unrelated work, custom instructions, existing databases and running trackers. No commits, pushes or global hooks. If I have supplied a TypeSafe key to this project, this setup also authorizes sending saved report prose to TypeSafe for Jev classification with a new allowance of 10 requests per UTC day, one per batch and concurrency one. Preserve an existing explicitly configured analysis allowance. Never print the key.

1. Confirm the target Git repository and branch. Check phasedrift --version. If missing, install from authorized repository access or a supplied release archive using the installation guide. If neither is available, report that exact prerequisite; do not invent an npm package or public app download.
2. From the target project run phasedrift init, phasedrift status and phasedrift doctor. Init installs the CLI-matched operator/reporting skills, the managed AGENTS.md block, Gemini entry point and supported project integrations. Reuse an existing connection and preserve its settings. Refresh stale integrations with phasedrift session setup --upgrade --hooks. Confirm what was actually installed.
3. Verify that this agent loads the project's PhaseDrift instructions. Read .agents/skills/phasedrift-session-review/SKILL.md and .agents/skills/phasedrift/SKILL.md; follow references only when relevant. If setup cannot supply the helper skills, download https://phasedrift-site.jppotess-main.workers.dev/downloads/phasedrift-agent-skills.zip and verify its SHA-256 against https://phasedrift-site.jppotess-main.workers.dev/skills/manifest.json before extracting without overwriting customized files. The kit is instructions, not the application.
4. Run phasedrift session start --task "Connect PhaseDrift and verify reporting" --area setup --host <actual-host>. Keep your returned attempt. Without a native identity, use the returned closeout command with --attempt; never share another agent's handle. Follow the installed skill to report the genuine setup result, checks, friction and learning. Disposition each delivered guidance item. Use suggestions: [] unless a concrete suggestion is warranted.
5. Verify the saved report using phasedrift learning reports and the project dashboard. Check native activity separately with phasedrift harness status; a report invents no native tokens or runtime. Include Jev setup: inspect phasedrift learning status. If the running tracker has a supplied key, enable automatic report classification within the authorized limits using phasedrift init --analysis when listed in CLI help. For an older CLI, use phasedrift learning analysis policy --stdin with {"enabled":true,"perRun":1,"perDay":10,"concurrency":1}; preserve existing explicit limits. If the key was added after launch, restart only the verified managed project tracker and recheck credential presence. If no key is supplied, tell me to put TYPESAFE_API_KEY in this connected project root .env.local and keep reporting working. Never print credentials or invent provider success. Read https://phasedrift-site.jppotess-main.workers.dev/llms.mdx/docs/jev-analysis/content.md for details. Give me the project name, exact Agent work URL, saved report ID, scan status, Jev status/limits and any exact blocker. A configured flag is readiness; a succeeded job and saved classification prove a provider request completed.
6. Tell me to finish one ordinary task with the updated project instructions loaded, then look for that task in Agent work → Reports. Reporting is proven by a saved report, not by a populated chart. After relevant reports accumulate, inspect one recurring obstacle and propose a focused improvement through the PhaseDrift operator skill. Do not implement a suggested project change without authorization.

After setup, use the reporting skill at each meaningful task entry and completion, pause or handoff. Reports name what changed, actual checks, friction, useful learning and concrete suggestions. Keep ordinary authorized work moving if the tracker is temporarily unavailable.