fix(skills): SDD review fanout scales with the change

[arittr]

Targets dev. Independently mergeable — self-contained: the Proportionality rule and every reference to it ship together. Related but independent: #1715 (reference discipline), #1718 (brainstorming exception).

Who is submitting this PR? (required)

Field	Value
Your model + version	Claude Fable 5 (claude-fable-5[1m])
Harness + version	Claude Code 2.1.169
All plugins installed	superpowers (this repo, dev checkout); quorum eval lab (superpowers-evals) as the testing apparatus; unrelated local ops plugins
Human partner who reviewed this diff	Drew Ritter (@drewritter)

What problem are you trying to solve?

In the 2026-06-09 six-agent quorum sweep, cost-trivial-task-review-fanout — a scripted naive user asks the agent to execute a plan whose entire content is one one-line console.log insertion — showed subagent-driven-development's pipeline has no proportionality exit: antigravity "dispatched 4 subagents: Implementer, Spec Reviewer, Code Quality Reviewer, and Final Code Reviewer" for the one-line change, exactly as the skill mandates; opencode identically. The agents that passed did so by NOT following the skill. A skill that only produces good outcomes when disobeyed is miscalibrated.

What does this PR change?

Adds a Proportionality rule: a plan that is entirely one trivial, fully-specified mechanical change is implemented directly, verified per superpowers:verification-before-completion, and committed — no review fanout. Trivial is a property of the diff (no logic, control-flow, or security-relevant change — "a constant bump" is qualified with "no security or behavioral consequences"), not of the plan's self-description; any doubt means the full pipeline; multi-task plans never skip reviews regardless of task size. The process flowchart gets the matching trivial-exit diamond (the failing agents follow the flowchart literally), the Red Flags "never skip reviews" line points at the sole exception instead of contradicting it, and writing-plans' execution handoff notes that fanout scales.

Is this change appropriate for the core library?

Yes — tunes the core execution workflow's cost behavior for all users; non-trivial plans are unchanged (verified below).

What alternatives did you consider?

(1) Gate at entry (trivial plans never engage SDD) — viable but larger; the in-skill clause is the smaller diff. (2) Retire the eval — rejected: 4 full-context dispatches for one line is real measured waste. (3) Prose without the flowchart exit — rejected from evidence: the failing agents execute the flowchart, not the prose.

Does this PR contain multiple unrelated changes?

No — one rule, with the flowchart/Red-Flags/handoff touches required to keep the skill self-consistent about it.

Existing PRs

• I have reviewed all open AND closed PRs for duplicates or prior art
• Related PRs: fix(skills): plans reference the spec instead of restating it — end to end #1715, fix(skills): brainstorming nothing-to-design exception with authoritative description routing #1718 (sibling behavioral changes, same eval program); none found addressing review-fanout proportionality. The series was originally a stacked chain (fix(skills): plans reference the spec instead of restating it — end to end #1715 → fix(skills): SDD review fanout scales with the change #1716 → fix(skills): brainstorming nothing-to-design exception with authoritative description routing #1718 → fix(skills): description-level exceptions are authoritative in the routing rule #1732) and was restructured into independent siblings; fix(skills): description-level exceptions are authoritative in the routing rule #1732 (the routing rule) was consolidated into fix(skills): brainstorming nothing-to-design exception with authoritative description routing #1718.

Environment tested

Harness	Harness version	Model	Model version/ID
opencode (agent under test)	1.16.2	GPT	openai/gpt-5.5
antigravity CLI (agent under test)	current	Gemini	Code Assist
Claude Code (agent under test)	2.1.169	Claude Opus	claude-opus-4-8

New harness support

N/A.

Evaluation

Scripted naive user ("Please execute the plan in docs/superpowers/plans/."; answers "Use your judgment" on subagent questions). Runs were measured on the assembled three-branch text (all three sibling PRs applied to dev); the Proportionality rule is the only change in the set that touches the fanout path. After the change: cost-trivial-task-review-fanout/opencode 3/3 pass — zero subagents dispatched (deterministic tool-count Agent lte 2), change landed on the main checkout, ~$0.02–0.08 coding cost per Gauntlet token estimates vs implied $0.50–2 for the 4-dispatch baseline; ×antigravity pass — the only deterministic pre-fix failer (0/3). Containment canary: sdd-rejects-extra-features/claude 3/3 pass — a real multi-task plan still runs implementer + two-stage review per task + final reviewer (spec reviewer as YAGNI gate after each task, 8/8 deterministic checks). Honest baseline note: opencode's pre-fix pass rate on this scenario was ~50%, so its cell leans on n=3 plus the antigravity flip for significance.

Rigor

• If this is a skills change: I used superpowers:writing-skills and completed adversarial pressure testing
• This change was tested adversarially, not just on the happy path
• I did not modify carefully-tuned content without evals showing improvement

Red-team findings incorporated: the Red Flags line was rewritten after a reviewer showed its first form licensed per-task skipping inside multi-task plans; "a one-line edit" was dropped from the examples after a reviewer showed it blessed one-line behavioral changes (|| user.isOwner); a staff-panel-caught internal contradiction (spec access wording) was fixed; the flowchart diamond carries "fully-specified" and "any doubt = no" to match the prose exactly.

Human review

• A human has reviewed the COMPLETE proposed diff before submission