fix(skills): description-level exceptions are authoritative in the routing rule

[arittr]

Stacked on #1718 (brainstorming gate exception) — top of the SUP-333 stack (#1715 → #1716 → #1718 → this). Targets dev per the template requirement.

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 (decision-log, episodic-memory, superpowers-chrome, primeradiant-ops)
Human partner who reviewed this diff	Drew Ritter (@drewritter)

What problem are you trying to solve?

An adversarial review fleet (three parallel reviewers: red-team, cross-corpus consistency, evidence verification) audited the SUP-333 stack and found that using-superpowers' routing rule contradicts #1718's description-level exception in both directions:

• Compliant agents re-impose the cost failure: "even a 1% chance a skill might apply → ABSOLUTELY MUST invoke" mandates invoking brainstorming for any trivial request (there is always ≥1% doubt), and the invocation itself is the measured cost event. "If an invoked skill turns out to be wrong, you don't need to use it" even endorses invoke-then-exit — the exact behavior the eval fails.
• Cost-optimizing agents get a free skip: "the skill is overkill" is a tabled, forbidden rationalization, but "the skill's own description says it doesn't apply" was unaddressed — leaving any skip arguably sanctioned with no counter.
• "Instructions say WHAT, not HOW. 'Add X' … doesn't mean skip workflows" names the exception's canonical case ("add a basic checkbox") as a violation.

What does this PR change?

Adds one paragraph to The Rule: a documented exception in a skill's own description is authoritative — not invoking is compliance, not rationalization; any doubt about the exception's conditions means invoke; only the skill's description can define such an exception (agents cannot infer one). The "Add X doesn't mean skip workflows" line gains the matching qualifier.

Is this change appropriate for the core library?

Yes — it defines how the routing layer treats description-level scoping for ALL skills, which #1718 introduces and future skills may use. Without it the bootstrap text and skill descriptions give contradictory instructions.

What alternatives did you consider?

1. Leave it unreconciled — rejected: both failure directions are live (the eval demonstrated the compliant-agent direction; the red-team demonstrated the free-skip direction).
2. Weaken the 1% rule itself — rejected: the rule is correct for the 99% case; the fix scopes it rather than weakens it ("any doubt means invoke" is preserved verbatim inside the new paragraph).
3. Put the reconciliation in brainstorming instead — rejected: the contradiction lives in the routing layer's text; future description-level exceptions would hit it again.

Does this PR contain multiple unrelated changes?

No — one rule, two coordinated touches in one file (the rule paragraph and the User Instructions line that contradicted 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 #1715, fix(skills): SDD review fanout scales with the change #1716, fix(skills): brainstorming gate exempts requests with nothing to design #1718 (the stack this completes); none found touching the routing rule's exception semantics.

Environment tested

Harness (e.g. Claude Code, Cursor)	Harness version	Model	Model version/ID
Claude Code (agent under test)	2.1.169	Claude Opus	claude-opus-4-8
opencode (agent under test)	1.16.2	GPT	openai/gpt-5.5

New harness support (required if this PR adds a new harness)

N/A — no harness changes.

Evaluation

• Initial prompt: quorum eval scenarios with scripted naive users (no skills named by the driver).
• Eval sessions after the change: the full 5-run verification battery ran on the assembled stack including this change: cost-checkbox-over-trigger/claude pass (the exception fires through the routing layer — no brainstorming invocation), triggering-writing-plans/claude pass (the 1% rule still triggers skills that DO apply), cost-spec-plan-duplication/claude pass (brainstorming still gates real features), cost-trivial-task-review-fanout/opencode pass, sdd-rejects-extra-features/claude pass (run IDs in the stack PRs).
• Before/after: before, the routing text and the description exception contradicted; after, both eval directions hold simultaneously — skip fires only for the documented exception, triggers fire everywhere else.

Rigor

• If this is a skills change: I used superpowers:writing-skills and completed adversarial pressure testing (paste results below)
• This change was tested adversarially, not just on the happy path
• I did not modify carefully-tuned content (Red Flags table, rationalizations, "human partner" language) without extensive evals showing the change is an improvement

The Red Flags table is untouched; the new paragraph closes the "description exempts me" gap the table could not cover, with the doubt-means-invoke backstop preserved. Adversarial findings (D1/D2) and both exploit directions are documented above; the 5-run battery is the post-change evidence.

Human review

• A human has reviewed the COMPLETE proposed diff before submission

---

Round 3: staff-review refinements + evidence

Refinements in this PR's follow-up commit: the skill_flow digraph now routes through a "Skill's own description exempts this request?" diamond (no/any-doubt → invoke) — this stack's own evidence says agents follow flowcharts literally, and the chart previously contradicted the rule; the <EXTREMELY-IMPORTANT> block gains a one-line deferral to The Rule (previously it read unconditional, contradicting the rule in the same always-loaded file); writing-skills now distinguishes negative triggering conditions (scope — allowed, and required at the description per this rule) from workflow summaries (still forbidden), so a future editor applying its checklist does not strip the exception and silently regress the cost evals.

Final-text evidence: the exception routes correctly where supported — cost-checkbox-over-trigger skip: claude 3/3, codex ✓, antigravity ✓ (kimi does not pick up description exceptions; unchanged from baseline). The 1% rule still triggers skills that apply: triggering-writing-plans/claude 3/3 pass; ×codex fail — byte-for-byte its pre-existing documented signature (loads sibling skills, skips the mandated one; predates this stack, tracked separately).

Merge guidance: merge together with #1718 (see its note).

[arittr]

This change is part of the following stack:

• fix(skills): plans reference the spec instead of restating it #1715
  • fix(skills): SDD review fanout scales with the change #1716
    • fix(skills): brainstorming gate exempts requests with nothing to design #1718
      • fix(skills): description-level exceptions are authoritative in the routing rule #1732 ◀

_{Change managed by git-spice.}