Connect your coding agent
Setup instructions and available measurements for each host.
Connect the project first
Run in the repository where the agent works:
phasedrift init
phasedrift openReload the project instructions afterward. Choose your host below.
Codex
Start the next task with the project's AGENTS.md instructions loaded. The reporting skill asks Codex to save what changed, what it checked and what got in the way.
Recorded activity: tool calls, token usage and time for complete runs, when present in the local logs.
Check the connection:
phasedrift session statusAfter a completed task, open Agent work → Reports and find its report. Use Effort to inspect the recorded run.
Verify a real task
Finish one ordinary task. Find its saved result in Agent work → Reports, then inspect the matching activity in Effort when available. These are separate checks: collected tool calls do not mean a report was saved.
phasedrift doctor
phasedrift harness status
phasedrift learning reportsNo activity for a host? Confirm it has local records for this repository and read the collection status before assuming the adapter is broken.
Refresh an existing setup
phasedrift session setup --upgrade --hooksReload the project instructions. This updates supported project integrations, preserving unrelated hooks and settings. It installs no global hook and makes no paid analysis call.
Compare similar work within the same host and model. Token accounting differs across hosts, and unavailable fields remain missing.
Custom instruction filenames
Codex and Cursor use the shared AGENTS.md block. Claude can load it alongside custom Claude files using claude-md-and-agents-md in /config; verify with /context.
Gemini's GEMINI.md entry point loads the shared skill. For custom filenames, context.fileName in .gemini/settings.json can include both AGENTS.md and GEMINI.md. Preserve other entries and check /memory show.
Muse or another reporting host
Start a task with the actual host name:
phasedrift session start --task "Your objective" --area verification --host museKeep the returned attempt. Close using its exact review command. Each concurrent agent keeps its own handle.
This records task feedback; it does not create native runtime or token data. Windsurf, Cline/Roo Code and Aider do not have native activity adapters today.