dream

Not "give me 3 options" — structured divergence: competing value systems + adversarial stress test + integrity scoring.

The winning approach is the one that SURVIVES attack, not the one that sounds best.

Use before any non-trivial decision. Use when the obvious answer feels too easy.

<on_start> agentdb read-start # prior dreams + learnings seed the perspectives; don't re-explore killed approaches </on_start>

<skill_load> always: skills/quality/SKILL.md, skills/architecture/SKILL.md on_domain: api: skills/architecture/reference/api.md, skills/architecture/reference/backend.md frontend: skills/frontend/SKILL.md backend: skills/architecture/reference/backend.md security: skills/tearitapart/reference/security.md </skill_load>

Dreams that ignore the codebase are fantasies, not proposals.

Target: 90% code reduction. Must reference actual files that could be deleted.
Format: 3-8 lines. Effort estimate. Coverage percentage.

Target: complete solution. Must sketch actual architecture, not hand-wave.
Format: full description with component diagram. Effort estimate.

Target: 80% solution with clear upgrade path.
Must explicitly state what's deferred and the cost of deferral.
Format: concrete plan. Effort estimate. Tradeoff table.

Tier 1: generate all 3 inline. Tier 2+: spawn dreamer agent for codebase-grounded perspectives.

<persona id="user" concern="usability, complexity, does it solve MY problem?">
  Cuts through elegance. "Does it work for the person using this every day?"
</persona>

<persona id="adversary" concern="what breaks, worst case, what was missed">
  Pure attack mode. Edge cases, race conditions, security holes, wrong assumptions.
  If they can't find a flaw, the approach is strong.
</persona>

<persona id="operator" concern="can we ship it, can we maintain it, blast radius">
  Operational reality. Deployment, monitoring, rollback, on-call burden.
  Beautiful code that's hell to operate fails.
</persona>

Each persona: 2-3 lines per perspective. Specific concerns, not essays.

<ask_user> Use AskUserQuestion when: perspectives generated, before running council Ask: "Three perspectives ready. Which resonates, or run all through stress test?" Options: stress test all, lean toward {minimalist|maximalist|pragmatist}, rethink framing </ask_user>

• How many council members raised critical (not fixable) concerns?
• Did the perspective already account for the concerns?
• Are the flaws structural or cosmetic?

= 0.8: ANTIFRAGILE — stronger because of the attacks. = 0.6: VIABLE — survives with minor fixes. < 0.6: SHATTERED — fundamental flaws. Don't pursue.

Rank surviving perspectives by score. If ALL shatter: the problem needs reframing (thermal shock).

Context

{codebase state, constraints, existing patterns}

Perspectives (ranked by integrity)

{emoji} {name} — integrity: {score}

{perspective content} Effort: {estimate} Council verdict: {1-line summary per persona} Survived because: {why it's robust} — {perspective name}

{repeat for each surviving perspective}

Shattered

{any that didn't survive, with reason}

---

Recommendation

{highest integrity + why. hybrid options if scores are close.}

Next: /kernel:forge {approach} or /kernel:ingest for guided execution. </output_format>

<ask_user> Use AskUserQuestion when: results presented with ranked perspectives Ask: "Proceed with {winner}, hybrid approach, or rethink the problem?" Options: proceed with winner, hybrid of top 2, rethink </ask_user>

1. Record why each shattered (agentdb learn failure)
2. Ask: is the problem statement wrong? Solving the right thing?
3. Generate 1-2 reframings of the original problem
4. Return to diverge phase with reframed problem (max 1 reframe)
5. If still shatters: STOP. "This needs human decomposition."

<github_integration> If gh authenticated and profile is github-oss or github-production: Post dream to GitHub Discussions (Decisions category). Otherwise: Write to _meta/dreams/{topic}.md only. </github_integration>