Review Loop Version

Astryx review loop version

Current: 3.2.1 · effective 2026-08-31

This is one version for the whole review system: reviewer brief, critic rules, loop mechanics, review presentation, and delivery contract. The files do not version independently.

Every new review presentation records:

LOOP VERSION: 3.2.1

The version is the one active when gate 1 starts. A review already in progress keeps its starting version through its critic passes; a fresh rerun after a kit change takes the new version. Historical reviews before this file are unversioned (pre-1.0); do not relabel them retroactively.

Bump policy

• Major (3.0.0) — trust or authority changes: auto-post/auto-merge policy, who may decide API/design, or what the loop is allowed to mutate.
• Minor (1.1.0) — judgement can change: a new/removed slot, verdict gate, blocking rule, evidence requirement, author bucket, architecture/system rule, or changed severity/disposition.
• Patch (1.0.1) — same judgement, clearer operation: wording, citations, examples, command fixes, link corrections, or a detection recipe that does not change what passes.

Bump before the changed rule is used. Update this history, propagate all kit mirrors in the same turn, and include the new version in every new Review-* record. Metrics compare reviews within a version; cross-version changes are reported separately rather than treated as reviewer drift.

History

Version	Date	What changed
3.2.1	2026-08-31	Corrected duplicate 3.2.0 history produced by concurrent publication; no judgment change.
3.2.0	2026-08-31	Removed author identity and owner-roster membership from review eligibility, verdict, severity, and publication. The caller/task now owns PR scope for every entry point; all authors receive the same current-spec, evidence, risk, and impact gates. Review scope still never authorizes merge. The shared two-attempt rolling-hour budget and exact-head post reuse remain.
3.1.1	2026-08-31	Corrected 3.1.0 attempt accounting to the intended contract: at most 2 substantive PR review attempts per rolling 3600 seconds across all entry points; reserve once atomically at attempt start; the public GitHub review action proves and reuses the same reservation/idempotency identity without consuming another slot. Re-reviews and substantive retries are new attempts; transport-only retries reuse identity. Post still rechecks roster eligibility and exact head. Draft-spec blockers route mentions from that draft's owners frontmatter, but drafts remain evidence only and cannot settle judgement; unresolved mention mappings fail closed. Added the private Ruby mention-mapping execution requirement without making it ownership or decision policy.
3.1.0	2026-08-31	Restricted scheduled and on-demand review eligibility to the exact union of .github/ENGOWNERS and .github/DESIGNOWNERS; introduced a shared review-attempt budget and atomic reservation controls, but incorrectly described review starts and public actions as separately counted reservations (corrected in 3.1.1); fixed the one-blocker GChat handoff; and added exact-head maintenance for our open spec PRs. Roster membership authorizes review only, never merge. The 3.0 current-spec authority rule is unchanged, and this guidance does not resume any Automation.
3.0.0	2026-08-31	Replaced review decision authority: only committed facebook/astryx specifications declaring authority: current can settle judgement. Fork-wiki rulings, decision-window/auto-ratified outcomes, merged code, component docs, issues, and prior reviews are evidence only. Missing current specs force novel-human, one private owner question, AUTHOR CAN PROCEED: no, no contributor-facing REVIEW, and no post/merge.
2.0.0	2026-08-30	Changed source authority and decision routing: progressive current-context loading; four candidate classifications; drafts as context only; separate private reasoning and public voice; boundary-only durable records; required regression evidence and explicit docs obligation; and public-API meaning, predictability, and capability gates.
1.9.1	2026-08-27	Added a literal pre-critic/pre-post template gate after the #5608 test review replaced required slots with a custom summary. Architecture comments must state derivable contraction criteria; “where should this live?” is an unresolved human decision and therefore cannot be posted as request-changes.
1.9.0	2026-08-27	ARCHITECTURE now requires a problem-derived complexity budget, one authoritative source per domain fact, per-round burden deltas, and an architecture-reset trigger after repeated model-level findings or growing representations. Public architecture blocks must name the model defect, consequence, and contraction direction rather than the latest guard or test count.
1.8.0	2026-08-27	Every new feature review must state a concrete person, task, current failure, demand evidence, and experienced impact. Contributor-authored additions to maintainer-owned vibe-test infrastructure are human-blocked until that case is owner-verified.
1.7.0	2026-08-27	Human rulings now require a matching public disposition: request-changes must be posted as a GitHub review, while owner closure uses a closure comment. New feature/system PRs are routed to human review, and template-derived gaps must be revalidated before implementation.
1.6.0	2026-08-27	Added the on-demand visual REMEDY SEARCH checkpoint: five-minute cutoff, existing-pattern/human-review/owner-requested-extension outcomes, mandatory human-handoff fields, explicit requester for extensions, and no speculative author advice.
1.5.0	2026-08-26	PROBLEM now requires three whys plus an evidence-based severity class; SOLUTION must show burden proportionality, with high-burden machinery for polish becoming a human gate; an inline multi-Effect state machine or one without focused browser tests is an architecture block; every added/changed Effect requires justification and measured cost; VISUAL EVIDENCE must embed durable pixels.
1.4.1	2026-08-26	Made review execution explicitly self-contained: reviewer, critic, and evidence are phases of the one spawned agent, which must never recursively spawn or delegate any pass. Judgement and evidence standards are unchanged.
1.4.0	2026-08-26	Added a 5–10 minute fast lane for one-decision, low-risk changes with no new surface/state/effects/architecture and one decisive check. Any uncertainty or promotion trigger switches to full review. Evidence and verdict standards are unchanged; fast uses one critic pass and compact evidenced closures for untriggered slots.
1.3.1	2026-08-26	Screenshot receipt operation tightened without changing obligation/severity: mutated arms bank an inspected exact diff, expectations are authored before observation and never loosened to match output, and semantic state reads avoid hidden measurement copies. Same judgement as 1.3.0.
1.3.0	2026-08-26	Every review makes an evidence-based screenshot obligation decision: manual frames, exact-head visual gate with named endpoint-story coverage, or not applicable with non-rendering path proof. Visual checks may be minimal, but may never be silently omitted or guessed from labels/source.
1.2.0	2026-08-26	Visual evidence now requires fail-closed screenshot sensor receipts: build, story, rendered theme/mode, direction, viewport/media, semantic state, geometry, settled render, and capture errors. Before/after must match on every sensor except Build; missing receipts make the visual slot not looked at.
1.1.0	2026-08-25	Human-owned decisions are a hard pre-post gate; AUTHOR CAN PROCEED separates missing direction from proven outcome-based asks; advice must be cited or driven; private merge controls and private-residue handling moved to the local-only companion and are excluded from public review text/records.
1.0.0	2026-08-25	First versioned baseline: plain-language SOLUTION model; per-slot verdicts; ARCHITECTURE routing and driven seam checks; behavior-unit/lifecycle ownership; GOAL, DISPOSITION, ADVICE, and WORST OUTCOME gates; narrow Needs-human safety valve; teammate review-presentation links; reviewer→critic loop with up to three passes.
pre-1.0	2026-08-16–2026-08-25	Unversioned calibration and rule development. Existing review records remain unversioned.