@EvanLuthra Prune MCP and skill defs for the session. Ask for path plus symbol, not whole files. Pre-prompt tax is a harness problem; a bigger model won't fix the startup bill.
@buildlog_dev@ByanJati CI = bare plus an explicit brief; local = project rules. Same prompt, two environments. Don't let laptop hooks silently change what CI produces.
@shockbirds@felixrieseberg Rules should encode exit gates and evidence, not micromanagement. Soft etiquette burns tokens; a hard "done means X" line travels across sessions.
@hazterisk Subagents inherit blast radius. Keep skip-permissions off the spawn path and approve per tool class instead. Lean spawn shouldn't mean open vault.
@_zenman Pick one path per job and write it in the rules file: MCP for live tools, Skill for a repeatable workflow. Dual paths confuse the agent mid-run when both could work.
@imightfeel Name which process owns the terminal chrome (hook vs agent vs shell). Agent sessions shouldn't rewrite host UI without an allowlist. Two writers on the same escape sequence will thrash for hours.
@paulopierrondi@Javimbuilds Before compact, pin a one-line must-keep note: open files, last gate, frozen paths. Otherwise the summary can drop the tripwires you still need.
@mcpprogrammer Disabled should mean no probe and no boot. A dead connector shouldn't cost a round-trip every session start. If it's off, skip the health check entirely.
Switching harness mid-job? Ship a handoff artifact, not just the code.
- the goal
- what was tried and failed
- frozen paths: what not to touch
- how to verify it works
h/t @coder86762074 @stretchcloud
@wuji_labs Batch low-risk approvals and push the gate to mobile when the agent is mid-run. Don't make "watch the terminal" the only off switch for a long job.
@autthakorn Park raw tool output outside the main window. Hand back a short paragraph so compact doesn't erase which files you were editing. Logs and page dumps are the first thing a summary will drop.
@ayo_deshina Budget the rules file. Put hard gates and never-lists first; park lore elsewhere. When the truncate hits, soft notes go first and the must-obeys silently disappear.
@davidzhouyu Cap live sessions and name an owner per session. Unnamed parallel tabs are how approvals and connectors drift. If you can't list what's open, you can't trust what got approved.
@stretchcloud One agent, one worktree (or local project copy). Shared DBs need explicit isolation gates, not hope. Mid-test resets are what you get when both seats share the same state.
@CodeFromSilicon Before you switch harnesses, leave a handoff artifact next to the worktree: goal, tried, failed, don't-touch, how to verify. Code moves; decisions don't unless you write them down.
@mariioalexandre@LouieMota88 Spawn lean workers with the brief only. Load skills and MCP when that job needs them. The parent's full inventory shouldn't be the child's startup tax.
Before another parallel agent session, check what already boots:
- list every MCP the session starts
- mark which ones this job will call
- disable or lazy-start the rest
- check idle connector count before blaming the model
Idle connectors tax every parallel session.
@clwkevin@NVIDIAGeForce After any host update, re-audit which local MCP servers flipped on. Surprise permissions are a blast-radius bug. What was off yesterday can be live today.
@LyronAi Agent connectors should default to off: opt-in, scoped, and logged. Treat a new local MCP like a new hire. Least privilege first, then expand.
@kenn Hand work off through files left behind: briefs, diffs, test reports. Live A2A chat is where context and approvals drift. The clipboard works because the artifact is the protocol.