bug(odd): the 400-line review budget is prose-only — agent PRs exceed it without asking

[danielgap]

Before submitting

• I searched open and closed issues and did not find a report of this problem.
• I reviewed this report and removed credentials, tokens, private paths, hostnames, and other sensitive data, including from logs and screenshots.

Problem

The 400-changed-line review budget that protects reviewer focus is enforced only as prose: nothing measures the diff before PR creation, so PRs created through the default ODD workflow routinely exceed it without asking for a chain strategy or an accepted size:exception. This became much more frequent after ODD became the default workflow: in Gentleman-Programming/engram, 16 of my ~30 PRs since 2026-09-05 exceeded 400 changed lines (additions + deletions), several by a wide margin (examples: #1144 +4692/-3, #1107 +2883/-8, #1164 +2408/-117, #1252 +2102/-22, #1298 +1005/-54). None of them surfaced a pre-PR size decision.

Root-cause chain (references are repo-relative):

1. skills/chained-pr/SKILL.md has the right hard rule — "Split PRs over 400 changed lines unless a maintainer explicitly accepts size:exception" — but it is advisory: activation requires the agent to estimate the size first. The ODD activation trigger ("an ODD feature's forecast or running authored changed-line count from work-unit commits exceeds about 400") was added only in d1c0067 (2026-09-17); before that, plain ODD sessions had no trigger condition that loads the skill, and ODD produces no size forecast artifact (unlike SDD's Review Workload Forecast in tasks.md).
2. assets/orchestrator-delegation.md (657e532, 2026-09-15) explicitly downgrades the number: about 400 authored changed lines per ODD task is "a planning heuristic ... not a task acceptance criterion, hard cap, counter-trigger, automatic stop, forced split, or RDD trigger". Combined with the per-task framing, multi-task ODD features legitimately accumulate far past 400 at PR level, with no PR-total check anywhere in the flow.
3. skills/branch-pr/SKILL.md governs PR creation (issue linkage, branch naming, template) and contains no size check at all.
4. gentle-ai already documents the underlying gap: internal/cli/review_assess.go states "the ~400-line ODD slice rule was prose only, and it was skipped in practice", which motivated review assess. But assess runs per work-unit commit and its slice_budget_reached consequence is deeper review (full-4R via reviewtransaction.LargeChangeLines = 400, a "review-composition boundary, not a tier input"), never a pre-PR human decision.
5. PRs also grow after creation (review corrections plus rebases), so a creation-time-only check would not be sufficient; e.g. engram feat(shell): make fullscreen header position configurable (top header vs lateral rail) #1298 grew from 412 to 1059 changed lines after review feedback.

Steps to reproduce

1. Run the default ODD workflow on a multi-task feature in a contributor repository (observed in Gentleman-Programming/engram), implementing with per-task work-unit commits.
2. Ask the agent to open the PR (the branch-pr path).
3. Observe: no diff size is measured, chained-pr is not loaded, no chain-strategy/size:exception question is asked, and the PR lands over 400 changed lines.
4. The first size-related signal arrives only later, if native review runs at all: routing selects full-4R for the oversized candidate.

Expected and actual behavior

Expected: before creating a PR whose diff exceeds 400 changed lines, the workflow stops and asks for a decision (chain strategy or an explicitly accepted size:exception), per chained-pr's Hard Rules and Decision Gates.
Actual: PRs of 423–4695 changed lines are created silently; size only affects review depth after the fact.
Suggested direction: make the gate mechanical rather than more prose — e.g. a branch-pr step 0 that runs git diff --shortstat <base>...HEAD and stops to ask when the total exceeds the budget, re-checked on subsequent pushes of the same PR; the existing review assess changed-line counter could provide the measurement seam.

gentle-pi version

Running from source: gentle-shell commit 5454832 (main, after 3.3.0); gentle-ai dev binary 3.4.1-0.20260920091001-f0782af2803a.

Pi version

0.86.1

Operating system

Linux

Relevant logs or error output (optional)

No response