Role

Write profile field in state.json. Emit new profile summary.

Invocation

/ultra-analyzer:set-profile <tier> [run-name]

Where tier is one of: small | medium | large | xl.

Protocol

Step 1: Parse arguments

• First token of $ARGUMENTS = tier. Validate against enum {small, medium, large, xl}. Reject others with an error listing the valid options.
• Second token (optional) = run-name. If absent, auto-detect single run per standard rules.

Step 2: Profile effects

Each tier declares downstream effects. Write them all into state.profile as a nested object so skills downstream can read without re-deriving:

{
  "tier": "large",
  "ultra_gate_tier": "--large",
  "worker_model": "sonnet",
  "worker_model_complexity_S": "haiku",
  "validator_model": "opus",
  "validator_model_complexity_S": "sonnet",
  "synthesizer_model": "opus",
  "topic_target_min": 45,
  "topic_target_max": 70,
  "redundancy_pair_rate_p1": 0.60,
  "suggested_parallel_terminals": "3-5"
}

small

tier: small
ultra_gate_tier: --small
worker_model: haiku (all complexities)
validator_model: sonnet
validator_model_complexity_S: sonnet
synthesizer_model: sonnet
topic_target: 15-25
redundancy_pair_rate_p1: 0.20
suggested_parallel_terminals: 1-2

medium

tier: medium
ultra_gate_tier: --medium
worker_model: sonnet (M/L), haiku (S)
validator_model: opus
validator_model_complexity_S: sonnet
synthesizer_model: sonnet
topic_target: 25-45
redundancy_pair_rate_p1: 0.40
suggested_parallel_terminals: 2-3

large (DEFAULT)

tier: large
ultra_gate_tier: --large
worker_model: sonnet (M/L), haiku (S)
validator_model: opus
validator_model_complexity_S: sonnet
synthesizer_model: opus
topic_target: 45-70
redundancy_pair_rate_p1: 0.60
suggested_parallel_terminals: 3-5

xl

tier: xl
ultra_gate_tier: --xl
worker_model: opus (M/L), sonnet (S)
validator_model: sonnet
validator_model_complexity_S: haiku
synthesizer_model: opus
topic_target: 70-120
redundancy_pair_rate_p1: 0.80
suggested_parallel_terminals: 5-10

Cross-model separation rule (C-3 hard constraint)

For EVERY tier, the following must hold:

• validator_model != worker_model (for M/L complexity topics)
• validator_model_complexity_S != worker_model_complexity_S (for S topics)

If you edit the tier tables above, re-verify this constraint. analyze-unit enforces a runtime fallback that bumps the validator one tier when this would otherwise be violated, but set-profile should get it right by default so the fallback is never silently invoked.

Step 3: Write the profile

bash ${CLAUDE_PLUGIN_ROOT}/bin/state.sh set <run-path> .profile <profile-json-object>

Step 4: Print confirmation

✓ Profile set: <tier> (was: <previous-tier>)
  Ultra gate tier: <value>
  Worker model: <value>
  Topic target: <min>-<max>
  Suggested terminals: <value>

Note: profile takes effect on the NEXT step. If workers are already running, they retain the model they were started with.
If you want to re-run topics with the new profile, use /ultra-analyzer:health to identify topics eligible for re-analysis.

Hard rules

• Reject tiers not in the enum with a clear error listing valid options.
• NEVER change profile mid-synthesize. If current_step == "synthesize", refuse and explain: synthesizer_model is already committed for this run.
• NEVER silently alter completed work. Profile applies forward only.
• Default profile for a new run is large (set by state.sh init).