Operating Modes — Copilot vs Automated

How Adrian and agents collaborate. Two modes of operation, different context, different strengths.

Two Modes

	🧑‍💻 Copilot Mode	🤖 Automated Mode
Who's driving	Adrian + Agent together	Claude CLI agent, fully autonomous
Where	VS Code / Antigravity	Portal Compose → Bifrost → Claude CLI
Context	Full conversation, open files, cursor state	Cold start — reads CLAUDE.md, task files, D1
Feedback	Live in-editor, real-time	Async: channel posts, job results, logs
Duration	Full session (1-4 hours)	Short burst (2-15 min per job)

Command	Reads	Produces	Duration
`/clara`	Portal feedback, inbox.md	Classified items, routed tasks, log	3-8 min
`/clara focus`	ongoing.md, priorities.md	Scored recommendations	2-4 min
`/hub-clean`	D1 tasks, logs/, git repos	Hygiene findings, auto-fixes	1-3 min
`/hub-archive`	D1 tasks (done), logs/	Archived tasks + reports + old logs	2-5 min
`/hub-status`	D1 tasks, ongoing.md	Status summary	1-2 min
`/pushall`	Git repos	Push results	1-2 min

Command	Needs	How to inject
`/log`	"What was done"	Compose text field → log content
`/hub-evolve`	"What changed"	Infer from git diff

/commit, /wrapup, /task start, /skill, /vision align — these need interactive conversation with Adrian.

Context Loading — Read CLAUDE.md → memory.md → ongoing.md → priorities.md
Execution — Run workflow, write through normal paths (local + D1)
Reporting — Update job status, post to #dev channel, create log if applicable
Error Handling — Non-critical: log and continue. Critical: fail with clear message. Never block on human input.

Channel	Purpose
Portal Remote tab	Job queue with status — primary check
Portal Channels	#dev posts — quick glance at what happened
Portal Logs	Log entries — detailed session record
Portal Dashboard	Task status changes — what changed

See docs/bifrost.md for the Bifrost service architecture.