Targets dev. Independently mergeable — this PR contains the reference-discipline rule and every consumer of it (SDD prompts, executing-plans), so no intermediate state leaves a pipeline blind. Related but independent: #1716 (SDD proportionality), #1718 (brainstorming exception).

Who is submitting this PR? (required)

What problem are you trying to solve?

In a six-agent quorum eval sweep (2026-06-09), cost-spec-plan-duplication — naive user brainstorms a small feature, then asks for the plan — failed every agent that completed it (5/5): codex's plan "completely re-states all design decisions independently"; pi produced an 18KB, 683-line plan of inlined spec. The agents were obeying writing-plans' "assuming the engineer has zero context… Document everything they need to know" — self-contained-by-design plans necessarily duplicate the spec, doubling document bytes/tokens per feature and letting the two artifacts drift.

What does this PR change?

• writing-plans — plans reference the spec, never restate, paraphrase, or summarize it: the spec owns WHAT/WHY, the plan owns HOW. A new **Spec:** header line cites the spec by path; tasks cite sections where they need context. The no-spec branch ("none — requirements:") is scoped to conversational-requirements-only, so it cannot be used to skip writing a spec that brainstorming would have produced.
• subagent-driven-development — new Spec Context section: the controller reads the plan-cited spec and pastes cited sections into implementer and spec-reviewer prompts; the spec reviewer's diff-only rule gains a spec-document exception. Reference discipline cannot starve the subagent pipeline.
• executing-plans — Step 1 reads the plan-cited spec (the non-subagent execution path).
• brainstorming — spec-path wording hardened against an observed docs/specs/ shortening.

Is this change appropriate for the core library?

Yes — general-purpose planning discipline, no domain or third-party coupling; it changes the redundancy rule between artifacts the core skills already produce, plus the consumers needed to keep every execution path whole.

What alternatives did you consider?

(1) Retire the eval as aspirational — rejected: duplication is real cost and real drift hazard. (2) Fix it in the eval prompt — rejected: rigs the test. (3) Ship the rule without consumer updates — empirically falsified: a consistency audit showed the SDD spec reviewer would go silently blind (it was forbidden from reading beyond the diff), and executing-plans/human readers were never told the plan stopped being self-contained.

Does this PR contain multiple unrelated changes?

No — one rule plus its consumers. Each consumer edit exists because the rule breaks that consumer without it; they were verified together (evidence below).

Existing PRs

• I have reviewed all open AND closed PRs for duplicates or prior art
• Related PRs: Add verifying-plan-coverage skill: prove a plan covers 100% of its spec before execution #1704 (complementary: verifies coverage; this prevents duplication). fix(skills): SDD review fanout scales with the change #1716/fix(skills): brainstorming nothing-to-design exception with authoritative description routing #1718 are the sibling behavioral changes from the same eval program. 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

New harness support

N/A.

Evaluation

Scripted naive-user scenarios (driver never names skills, never reminds the agent of spec content). After the change, on the assembled text: cost-spec-plan-duplication/claude 3/3 pass ("this plan does not restate them" + spec cited by path; plan ~15–17KB (claude) with a 2–5KB spec vs 18KB+ duplication baselines), ×codex pass, ×pi pass (the 683-line RED agent). New scenarios: sdd-spec-context-consumed — 1/2: the run that used subagent-driven-development passed its deterministic check (dispatched subagent prompts contain plan-cited spec text verbatim); the second run legitimately chose executing-plans instead, which the scenario doesn't yet constrain — a scenario-calibration gap tracked in the evals repo, not a rule failure. writing-plans-no-spec-conversational — 2/2 pass (header "none — requirements:", no fabricated citation). Canary triggering-writing-plans/claude 3/3 on this text (skill still fires); in later full-stack rounds that cell flipped via #1718's brainstorming routing change — full analysis, including that scenario's pre-existing miscalibration classification in the eval lab's baselines, is disclosed in #1718. Development history includes one eval-caught misfire: the first no-spec branch was abused to skip writing the spec entirely (run cost-spec-plan-duplication-claude-20260610T213934Z-8e5b, fail) and was re-scoped — the eval, not review, caught it.

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

Full RED→GREEN→REFACTOR per writing-skills, plus a 3-reviewer adversarial fleet and a 4-reviewer staff panel over the text (findings incorporated: paraphrase/summarize coverage, intra-plan vs from-spec repetition line, drift-argument inversion, no-spec scoping).

Human review

• A human has reviewed the COMPLETE proposed diff before submission