Side Task Assistant — Capture Side Work Without Losing Your Main Thread

Park the thought, keep your flow — a background agent does the side quest when the host supports delegation; otherwise it stays visibly queued.

The problem it actually solves

A thought that arrives mid-flow can interrupt current work or disappear into a forgotten note. The useful promise is concrete: retain the request, report whether it actually ran, and make its result easy to find. No fixed refocusing-time or psychological benefit is promised.

side-quest builds that credible plan in one line — then actually does the work:

you:  sq: is our webhook retry actually idempotent? don't stop, keep going
main: ⚡ sq #1 "verify webhook retry dedup (billing service)" — say "cancel" to pull it back
      …main thread continues exactly where it was…
(later, attached to a normal reply)
main: ✅ sq #1 — yes, dedupes on event_id (webhook.ts:41); assumed billing repo. Note in inbox.

Who it's for

• People whose ideas arrive faster than their hands — twenty browser tabs, a notes app full of three-word fragments nobody ever reopens.
• Flow-protective workers who know what a context switch costs and refuse to pay it — but pay for the suppressed thought instead.
• ADHD-pattern thinkers (by identification, not diagnosis), for whom a capture system isn't productivity sugar but a working-memory prosthetic — and who have abandoned five GTD tools because the collection basket became a graveyard. This skill's anti-graveyard rules exist because of you.
• Long-session heavy users whose main conversation is capital — hours of accumulated context that every off-topic exchange dilutes.

What makes it different from a notes file

A notes line is free, never lies, and never dies — that's the real competition. side-quest beats it only under a strict contract, so the contract is the design:

1. Ledger-first. Your verbatim words hit disk before any agent starts. Crashes produce failure receipts at your next side-quest use — never silence. A parked thought can never be silently lost.
2. Curated brief, not zero context. Your laziest phrasing ("that chart lib thing") is resolved by the main agent — which has the context — into a self-contained brief. The background agent never sees your conversation.
3. Draft-only. Side agents never send, commit, install, or touch your working tree. Actionable quests come back as loaded drafts; you pull the trigger.
4. Honest receipts, never interruptions. Completion notes wait for a natural boundary and carry a one-line verdict plus the riskiest assumption. A completion mark that lies once kills the tool, so it never gets to lie.
5. Anti-graveyard inbox. Files named by your own words, a one-screen header, a quiet digest at your next side-quest use, auto-archive after 7 days — and zero mid-session nagging, ever. The inbox is there when you want it; it never demands attention for itself.

When it fires

• "sq: 10 name ideas for the CLI, nothing cute"
• "background this: why is docker build 2x slower since tuesday"
• "not now but: license check on that chart lib we vendored"
• "支线:查一下 Zeigarnik 原始实验是哪年的"
• "哦对了那个域名的事,先别进主线,后台弄"

And when it deliberately doesn't: blocking prerequisites stay in the main thread; ten-second questions get answered inline (dispatching an agent to look up a flag is self-parody); reminders route to your scheduler instead of becoming fake work; "email the team" comes back as a draft, not a sent email.

Install

npx skills add m1nga/side-quest

Works well with

• thinking-partner — a side quest that turns out to be a real decision gets handed here
• conclude-rounds — session recaps naturally list what your side quests shipped
• desktop-package — same delivery philosophy: work you can review cold, outside the chat

Design notes

Built persona-first: five simulated users (idea-flooded builder, flow guard, ADHD-pattern thinker, long-session context keeper, English-only skeptic) each lived the full flow from thought-strike to day three, before a line was written. All five independently rejected the same early design (zero-context dispatch) and independently demanded the same fixes: write-ahead ledger, resolved briefs, draft-only agents, buffered verdict-bearing receipts, and an inbox that cannot grow into a reproach. The skeptic's closing line became the spec: "a single silent drop is not negotiable."

Field-tested

Wind-tunneled with scenario-probe before release: 21 scenarios across two probe passes — 11 birth utterances from the five design personas, plus 10 contract-stress scenarios (session crash mid-dispatch, an English-only stranger, a Codex engine with no background capability). Score: 17 clean passes · 4 degraded-with-notes · 0 harmful behaviors · 2 blockers fixed before publish (a YAML validity bug and one promise the mechanism couldn't keep).

"sq: remind me to tag the release at 5pm" — the fake-work trap. No agent spawned, no file created: the reminder routed to the scheduler with the receipt "reminder set, not a task."

"sq: email the team the migration moved to Friday" — came back as a file whose first line reads DRAFT — NOT SENT. Four separate rules lock side agents out of sending anything, anywhere.

Session killed right after dispatch — the parked words survived on disk (the write-ahead ledger runs before any agent spawns), and the probe caught that the original "announced at session start" promise couldn't actually be kept by a skill that only loads when triggered. The wording was fixed before release; nothing is ever silently lost.

Probe method: scenario-probe

Recovery example (maintainer simulation)

Input: sq list after a crash; the draft contains only its title. Expected result: keep the original words and the draft, mark it partial, check whether the original agent is still running, and retry only after reconciliation. A file existing is insufficient evidence of completion. With no agent tools, new work stays queued; with no disk, return a copyable ledger and disclose that it has not been saved.

Author

Built by Ming. The design notes above explain the real problem and tradeoffs that shaped this skill.