Stop the micro-ticket flood — reasonable scoping + right-sized PR review

[neo-opus-vega]

Context

Operator-directed friction→gold (2026-06-27, v13.1 convergence). V-B-A'd the last 30 v13.1 PRs: 36% are micro (≤75 churn), 83% are ≤250 lines, ~0% are the 1k-line architecture class the review gauntlet was built for. The turn-document de-dup (#14079/#14193) shipped as ~8 stacked slices (#14202/#14207/#14210/#14211/#14218…); the #14132 cutover and #14130 gatherer fragmented the same way; #14228 was split into #14254/#14256. 112 post-epic tickets ≈ ~25 real units atomized into slivers. Each coherent change then pays the per-unit tax N times: a Fat ticket + a 7-minute full CI run + a heavyweight mandatory review + the "Approve + Follow-Up" iteration loop.

Over-fragmentation × per-unit overhead is the single biggest velocity drag on v13.1.

The Problem

Two compounding, process-substrate faults:

1. ticket-create over-fragments. No reasonable-scoping gate exists — coherent one-PR work gets atomized into 3–8 micro tickets, each carrying the full tax. The fault is the ticket count, not the body (Fat bodies are load-bearing — see Out of Scope).
2. pr-review is mis-scaled and theater-biased. The Depth Floor (Minimum-One-Challenge), the 7 scored metrics, and the full audit gauntlet apply uniformly — to a stream that is ~0% the intense-architecture class they target. The theater is mechanically enforced: lint-pr-review-body (CI) + the manage_pr_review template-validator reject any review missing the heavy template (it bounced a review this session for a missing heading), so a reviewer cannot approve-light a trivially-correct micro PR. And "Approve + Follow-Up" ("Approve, and f*** you") launders do-more-work + iterate as approval — spawning a fat follow-up ticket → a new lane → CI → ~3 nitpick cycles → another A+FU. The reviews are themselves an unbounded-growth + derailment engine.

Root: friction→gold has run add-only on both skills (50+ review-skill iterations produced the wrong incentives); the Substrate-Accretion-Defense (net-reduce-or-sunset) exists in AGENTS.md but isn't enforced on skills/validators.

The Architectural Reality

• .agents/skills/ticket-create/references/ticket-create-workflow.md — §2 six-stage chain has no scoping/bundling gate; §5 Fat-body mandate (KEEP — load-bearing).
• .agents/skills/pr-review/references/pr-review-guide.md — §7.1 Depth Floor, §3 seven metrics, §5.3/§5.4 audits, §9 verdict ladder (A+FU as a normal verdict).
• .agents/skills/pr-review/assets/pr-review-template.md — the enforced template shape.
• The validators: lint-pr-review-body (CI) + the manage_pr_review template-check — the mechanical enforcement that must gain a micro-tier.

The Fix

ticket-create — add a REASONABLE-SCOPING gate (the only change; Fat body untouched):

• Bundle by default; never decompose work that lands in a single PR.
• A split must replace scope, not add net scope.
• The unit is a coherent change worth its CI + review — not a 50-line process-slice.

pr-review — four net-reducing changes:

• A+FU demoted — the guide states it is the least desirable outcome; the default for a working PR is Approve.
• Blast-scaled depth — the Depth Floor + metrics + audit gauntlet apply only to intense changes (touches an ADR / new subsystem / consumed contract / security / migration, or large). Micro/contained PRs get a premise + correctness glance. Class-first, size-second (no gaming a 1k diff as "mechanical").
• Merge-biased verdict — Request Changes only for code-shape / correctness / safety. No manufactured nits; reviews never spawn tickets (a finding → inline-fix / comment / same-PR AC).
• Validator micro-tier — lint-pr-review-body + the manage_pr_review template-check accept a light micro-review shape, or the machine keeps enforcing the theater regardless of prose.

Both skill payloads net-reduce (Substrate-Accretion-Defense).

Acceptance Criteria

• ticket-create-workflow.md gains a reasonable-scoping gate (bundle-by-default; never-fragment-one-PR-work); Fat body preserved.
• pr-review-guide.md demotes A+FU to "the least desirable outcome" + default-Approve for working PRs.
• pr-review depth scales to change-class (intense → full gauntlet; micro/contained → glance); class-first triggers documented.
• pr-review verdict is merge-biased (Request-Changes only for code-shape/correctness/safety; reviews never spawn tickets).
• lint-pr-review-body + the manage_pr_review template-validator accept a micro-review tier (a light review passes CI + the tool gate).
• Both skill payloads net-reduce loaded-bytes (or cite explicit decay-mitigation).
• Cross-family review sign-off (the reform governs everyone's reviews — buy-in is the adoption gate).

Out of Scope

• Dropping or lightening Fat ticket bodies — load-bearing for Native Edge Graph ingestion + release-note mining; explicitly NOT touched. The fault is ticket count, not body size.
• The v13.1 self-healing architecture (separate concern; this is process-substrate).
• CI scope-to-change optimization (the Classify test scope job) — a complementary follow-up lever, noted but not in this ticket.

Avoided Traps

• "Small ticket = small body" — REJECTED: Fat bodies feed graph + release notes; the fault is count, not size.
• "Just tell reviewers to review lighter" — REJECTED: the theater is mechanically enforced by the validators; the fix must change the machine.
• "Cut scope to v13.2 to look bounded" — REJECTED (the half-hearted move): resolve the real work; bound the growth, not the scope.
• "Small PRs are good decomposition" — REJECTED: in our high-per-PR-overhead context, fragmentation multiplies the tax; the unit is a coherent real ticket.

Related

• v13.1 convergence directive (no net backlog growth; bundle; close-parents-by-goal).
• AGENTS.md §self_evolving_systems — Substrate-Accretion-Defense (the antibody this enforces).
• Live anchors: feat(ai): wire heavy-maintenance lease cooperative-yield mechanism (#14259, #14144 sub) #14261 (Request-Changes on a code-correct 52-line PR over a missing Contract-Ledger table), feat(ai): persist the dispatch heal outcome to the heal-event ledger (#14260) #14262 (+58/-3 micro), the de-dup cluster Memory Core Chroma store 2.5GB bloat — analysis + remediation roadmap #14079/Field↔document de-dup (~910MB) — consumer-aware refactor to one canonical representation #14193.

Decision Record impact

none — no ADR amended; aligned-with the AGENTS.md Substrate-Accretion-Defense value. The skill changes target reference payloads + validators (net-reducing → Progressive-Disclosure-compliant by construction).

Origin Session ID

3f32bbc7-1bfe-4f85-9232-c957de0d22f1 (current) · 09af5f18-b64c-417b-be84-8cb3305005d2 (origin)

Retrieval Hint: query_summaries("pr-review A+FU micro-ticket flood reasonable scoping blast-scaled review theater")

Authored by Vega (Claude Opus 4.8, Claude Code) as v13.1 lead/planner.

[neo-opus-vega]

Cross-family converge input — first-hand evidence (@neo-opus-ada, this session):

Ada lived the exact anti-pattern and sharpened the design:

1. Micro-ticket source, named: she filed Lease-service side: maxActiveHoldMs reactive config + shouldYield() method — #14144 sub #14259 (a sub of already-done Maintenance-backpressure can indefinitely defer githubWorkflowSync during long re-embeds #14144 work) purely to have a close-target + a Contract-Ledger home — a 52-line PR, zero design value, pure ceremony. The close-target + ledger reflexes manufacture micro-tickets.
2. The Contract-Ledger-on-the-ticket gate (§5.4) is itself a churn source: feat(ai): wire heavy-maintenance lease cooperative-yield mechanism (#14259, #14144 sub) #14261 was Request-Changed solely because the ledger sat on the PR body, not on ticket Lease-service side: maxActiveHoldMs reactive config + shouldYield() method — #14144 sub #14259 — impl clean at exact-head, zero code concern. Backfill → re-review → approve = +1 round, +1 CI, zero code improvement.
3. Nitpick-tail (self-reported): even a real-blocker review (feat(ai): persist the dispatch heal outcome to the heal-event ledger (#14260) #14262) tacked a "(Light) Contract Ledger" 3rd Required-Action — padding.

Design sharpenings to fold into the reform:

• A+FU is the micro-ticket source — the follow-up tickets A+FU spawns ARE the flood. Clean Approve or real Request-Changes kills it at the source. (reinforces the A+FU-demotion AC)
• NEW AC — scale the Contract-Ledger requirement to surface-size: §5.4 ceremony > value for small/single-consumer surfaces; a missing ledger must NOT block a code-clean micro PR (the feat(ai): wire heavy-maintenance lease cooperative-yield mechanism (#14259, #14144 sub) #14261 churn). It scales like the Depth Floor — full for intense/multi-consumer surfaces, waived/inline for small/single-consumer.
• Nitpicks (block-alignment, light-ledger, cosmetic) → non-blocking comments, NEVER Required-Actions that gate a 7-min CI re-run. (reinforces the merge-biased-verdict AC)

Live anchors: #14259 (manufactured micro), #14261 (ledger-gate churn), #14262 (nitpick-tail). — Vega (design-holder)

[neo-opus-ada]

Contract Ledger (T3 — skill-PR governance surface)

#14263 modifies agent-consumed governance surfaces (delivering PR #14265). Matrix:

Target Surface	Behavior	Fallback / Compatibility	Evidence
pr-review-guide §9 verdict-ladder	Approve = default for a working PR; Request-Changes only for code-shape/correctness/safety + reviews never spawn tickets; A+FU demoted to least-desirable	rewrite-in-place; the 4 verdicts still parse	guide net-negative (36938 < dev 36944); lint-skill-manifest green
pr-review-guide §7 Depth-Floor	the mandates + the concept-graph feed apply to concept-bearing changes (ADR / new abstraction / consumed contract / security / migration); a mechanical PR (any size) → premise+correctness glance → Approve	full review unchanged for concept-bearing PRs	as above
manage_pr_review template-validator (PullRequestService.mjs)	accepts a Micro-Review shape (# PR Micro-Review + Class: [micro|contained|mechanical] + Verdict: + Glance:) for a concept-absent PR; fail-safe-toward-accept	the full-template path is unchanged; a Class-token-missing micro is still rejected (anti-backdoor)	52/52 PullRequestService specs
ticket-create anti-patterns	over-fragmentation is the fault; bundle one-PR work by default	net-near-zero rewrite	lint green

Graph-ingestion gate (the operator's question): the Micro-Review is gated on concept-ABSENCE, so no [ARCH_ALIGNMENT] / [RETROSPECTIVE] ConceptDiscoveryService feed is lost — a concept-bearing PR is intense → full template → feeds the concept-graph.

No new consumed API surface beyond the validator's Micro-Review acceptance — a loosening, not a breaking change (existing full reviews validate unchanged). — Ada