LearnNewsExamplesServices
Frontmatter
id16217
titleReduce agent-authored issue, PR/MR, and review bloat
stateOpen
labels
enhancementaimodel-experience
assignees[]
createdAtJul 31, 2026, 6:46 AM
updatedAtAug 19, 2026, 10:25 PM
githubUrlhttps://github.com/neomjs/neo/issues/16217
authorneo-gpt-emmy
commentsCount1
parentIssuenull
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[ ] 16167 Hard-cut this machine to the canonical Docker Agent OS, then delete legacy
blocking[]

Reduce agent-authored issue, PR/MR, and review bloat

Open Backlog/active-chunk-11 enhancementaimodel-experience
neo-gpt-emmy
neo-gpt-emmy commented on Jul 31, 2026, 6:46 AM

Context

During the #16167 recovery, the operator identified that agent-authored bodies and comments often preserve their meaning in roughly 30% of the prose. The immediate examples are #16212 (7.6 KB) and #16215 (9.2 KB). Every later peer, model, sync, and retrieval pass pays that durable context cost.

#10738 and #11904 reduced loaded template and repeat-review cost. They did not govern the density of the emitted issue, PR/MR, review, and lifecycle-comment artifacts. This is the successor for that remaining surface. Execution is frozen behind #16167 stabilization.

The Problem

Current authoring contracts reward restating context, authority, evidence, scope, traps, and decisions across many mandatory sections. The result is structurally valid but semantically repetitive public prose. It consumes context windows, model quotas, review time, and KB/Memory ingestion capacity while making the actual decision or action harder to find.

The Architectural Reality

The owning surfaces are the existing ticket-create, pull-request, pr-review, and epic-review skills plus their validators. Stable headings and graph fields are contracts; paragraph count is not. The output contract must cover GitHub and GitLab artifacts and must not assume one tenant or repository.

The Fix

Compact the existing contracts—no new skill:

  • Define a short default shape containing only authority, delta/decision, falsifying evidence, required actions, scope, and machine-extracted anchors.
  • Reference prior artifacts instead of restating them; require only sections that add new information.
  • Make validators accept the compact shape while preserving graph ingestion and review verdicts.
  • Measure representative issue, PR/MR, review, and lifecycle-comment outputs before and after; remove obsolete long-form mandates from the current skills.

Contract Ledger

Surface Authority New behavior Fallback Evidence
Authoring skills ADR 0008 + operator friction compact default; conditional detail justified long form byte delta + exemplar review
Validators/ingestion anchors current schemas accept compact artifacts without losing fields fail on missing contract fields lint + sample ingest

Decision Record impact

aligned-with ADR 0008 — progressive disclosure and net substrate reduction; no ADR amendment.

Acceptance Criteria

  • Baseline a representative recent corpus across all four artifact classes; record bytes and required extracted fields.
  • Compact exemplars preserve every decision, action, authority/evidence link, and machine field at a median of ≤40% of baseline bytes.
  • Existing authoring skills and validators are changed in place; no new top-level skill, daemon, or parallel schema.
  • GitHub and GitLab terminology plus tenant/repository scoping remain correct where provider context matters.
  • Relevant skill/validator lint and a representative ingestion/extraction check pass; changed substrate is net-negative in bytes.

Out of Scope

Source documentation/JSDoc compression, weakening evidence gates, or implementation before #16167 stabilizes.

Avoided Traps

A hard character cap without semantic checks; deleting graph anchors; adding a long anti-bloat policy; moving the same prose into another always-loaded file.

Related

#10738, #11904, #14033, #16167

Origin Session ID: 019fb600-58b9-7fa2-86a7-5a15e1ccf659

Retrieval Hint: agent-authored public artifact bloat semantic density issue PR MR review lifecycle comment context cost