ADR-0011: Night Shift — scheduled autonomous work sessions

Status: accepted (implementation staged after the registry MVP)

Context

Most subscription users leave the majority of their agent quota unused daily;

it expires worthless. The project has an unbounded queue of exactly the work

agents do well unattended. Connect the two — and answer the question every

open-source CAD project dies on: who does the unglamorous work?

Decision

A built-in scheduler: *"between these hours, spend up to X of my quota on this

queue, show me a morning report."* Opt-in, visible budget, one-click off.

Queues, in priority order:

1. 1. The user's own backlog (why they turn it on): DRC/validation sweeps with

fix-proposals as draft PRs; drawing regeneration; reduced repros for their

failing invariants; lockfile update proposals with impact diffs;

cross-domain checks ("enclosure changed Tuesday — board still fits").

1. 2. The community queue (the flywheel, separately opt-in): part verification

(re-derive draft components from datasheets — the verified tier of

ADR-0010, embarrassingly parallel); synthetic-reproducer search for

fingerprint-only bug reports (ADR-0007); regression triage; corpus fuzzing.

Hard rules (stricter than daytime, per ADR-0006):

• - Overnight work produces proposals only — draft PRs, attestations,

reports. Nothing merges. Tier-1 stays Tier-1 at 3am.

• - Hard caps on tokens, wall-clock, task count; any cap halts cleanly.
• - Circuit breakers stop; the report explains why. Nobody is paged at 3am.
• - Community tasks process untrusted input (datasheets, bug reports =

prompt-injection vectors): maximally sandboxed, no path to the user's own

designs — separate process, separate mounts, no credentials.

• - Anything ambiguous → skip and report. Precision over recall, always.

The morning report is the product surface: what ran, what it cost, every

proposal linked to a diff, decisions ranked, halts explained, and an explicit

"nothing merged, nothing left this machine except what you pre-approved."

Incentives are reputation, not payment: attestation counts and registry

standing. Monetizing idle quota invites verification fraud; reputation with

consequences is self-policing.

Queue coordination: a simple lease on the registry so two users' agents

don't verify the same part redundantly.

Consequences

• - The registry's verification army is everyone's idle agents, with humans

promoted to reviewers.

• - Requires the fingerprint queue (ADR-0006), reducer (ADR-0007), and registry

trust tiers (ADR-0010) — staged accordingly; daytime tooling is built with

"unattended mode" as a known future consumer (clean halts, resumable tasks,

machine-readable results).