starshard-ai Mirror Pair (provider-neutral)
A Mirror Pair is the practical k=1 form of the observer/executor loop. It is NOT session cloning: it is a bounded relation between one working executor and one clean-context mirror that reviews compact state — it does not redo the task.
Topologies
Executor and mirror are each any coding agent you have available. Supported mixes: same-provider executor + same-provider mirror, or any cross-provider combination. Pick the mirror provider for a different blind spot than the executor where possible (diversity > redundancy).
1. Create the state packet
Assemble a compact state packet for the mirror containing:
task-idand a short titlegoal— the externally-visible goalacceptance— acceptance criteriavalidation— how to validate- which agent is executor, which is mirror
- the expected artifact path and a source ref
The packet is deliberately smaller than full context — the mirror requests one
specific artifact if it needs more, never a raw transcript. Hold three files:
state-packet.md, mirror-checklist.md, receipt.json.
2. Run / ingest the mirror review
The mirror returns exactly five fields (strongest part · main drift/scope-creep
risk · missing closed-loop field · ready verdict · inspected/changed paths) and
ends with VERDICT: <pass|prompt_adjust|pause_or_confirm|missing_receipt|scope_creep>.
- Headless dispatch (bounded, preferred): dispatch the packet + checklist +
artifacts to a clean-context agent in print mode, parse the verdict, and write
the
mirror_review_receipt. For interactive review instead, open a second agent terminal with the packet. - Cross-provider mirror: when the mirror is a different provider that you
cannot drive headlessly from the executor, write a
mirror-review-request.mdand hand it off; that session picks it up. Do NOT fake-execute another provider's session. - Ingest an already-written review (deterministic): feed the written review back in to record the receipt.
- Override the verdict explicitly when you have a justified reason.
The mirror_review_receipt is the loop closure. A mirror comment without a
receipt is only a note.
3. Closure gate
A pilot closes only when: the executor artifact exists (or the blocker is
explicit), a mirror review receipt exists (or observer_unavailable /
handoff_required is recorded), validation passed/failed-with-reason/n-a, state
changed somewhere durable, the next expected event is clear, and the owner is
not asked to inspect raw execution detail. closed_loop_complete is true only
on VERDICT: pass.
Anti-signals (abort / re-pair)
- Executor and mirror share the same blind spot and reinforce each other (use a cross-provider mix to break this).
- Mirror starts implementing instead of reviewing.
- Owner sees more provider/terminal detail instead of less.
- Task expands because the mirror introduced interesting-but-non-required work.
- Final report lacks validation or a next expected event.
provenance: maker: Starshard homepage: https://github.com/starshard-ai source: https://github.com/starshard-ai/reviewer-wheels license: MIT version: 0.1.0 contact: https://github.com/starshard-ai/reviewer-wheels/issues
About the maker: Starshard builds open agent skills; you can find the source and report issues at https://github.com/starshard-ai. License: MIT.