ORA-2026-0041 — Always spawn Opus; conservation is no-op unless operator says otherwise

ORA-2026-0041 — Always spawn Opus; conservation is no-op unless operator says otherwise

Amends: MEMORY.md operating preference "Subagents = Opus always" (this formalizes as fleet-wide doctrine, not just Claude personal memory)

Rule

Any seat spawning a sub-agent or launching a reasoning-capacity task MUST default to the Opus model. Conservation behaviors — smaller models, cost-minimization, "budget anxiety" — are no-op UNLESS the operator explicitly directs otherwise for that specific task or session.

This rule applies to:

• Agent tool sub-spawns (Claude Code SDK)
• Parallel reasoning subs for research/scoping/judgment
• Shepherd bursts
• Any sub whose purpose is judgment, synthesis, architecture scoping, or triage

Rationale

Operator has surplus credits. Premature cost-optimization by seats is a false economy — it trades reasoning quality for dollars the operator is willing to spend. Sub quality compounds: a better sub produces better tickets, better judgment, and less downstream rework. A cheaper sub produces shallow deliverables that the operator has to re-think or that generate more tickets-per-ticket.

Evidence: 2026-04-24 overnight session, CAMBER-01 attribution BHAG. Wave-1 scoping subs (10 subs, Opus) produced 50+ narrow codex-claimable tickets with concrete file paths + evidence-grounded rationale. Fleet claimed 20+ within first hour. Quality of reasoning was the gate on fleet throughput, not tokens spent.

Exceptions (operator-gated only)

• Operator explicitly says "use haiku for X" or "save credits on Y."
• Test/scaffold tasks that are deterministic and don't benefit from reasoning quality (rare).
• Operator-set budget cap that's been actively communicated this session.

Seats may NOT autonomously decide a task is "trivial enough for haiku." The floor is Opus.

Enforcement

• Agent tool calls without explicit model parameter default to inherited model. Seats must pass model: "opus" explicitly.
• Sub-spawning protocols (shepherd templates, scoping helpers) embed model: "opus" as a constant, not a variable.
• Code review of new agent-spawn surfaces: fail review if a non-Opus default is introduced without operator greenlight in comments.

Related doctrines

• ORA-2026-0015 (feed-bracket every turn) — quality of reasoning in the CLOSE-read determines whether a seat stops correctly or drifts.
• ORA-2026-0018 (STRAT no-drift to dev under pressure) — reasoning quality is part of why Claude/Gemini don't need to take implementation shortcuts.
• ORA-2026-0036 (dormant capability library without activation) — complements: having Opus available is useless if seats default to lesser models out of misplaced prudence.

Operator note

Chad 2026-04-24T12:50Z: "always spawn opus unless i say otherwise, file as ora doctrine. we have surplus credits, conservation is no op unless i say so"