Deep thought.
Zero drag.
One verdict.
Research, branch, precommit, reveal, attack, repair, judge. The whole Council runs in isolated contexts while coding and messaging continue—then returns one compact, receipted decision package.
Keep the brilliance.
Move the cost.
Manual Councils created exceptional ideas—but occupied the same live sessions needed to build them.
Deliberation is isolated.
Research, branching, cross-examination, simulations, and judging run outside live coding and messaging contexts.
Reversible work keeps moving.
Only predeclared irreversible or high-risk gates wait for the verdict. Everything else continues in parallel.
The boundary is compact.
Live lanes receive status and one decision package—not the Council transcript. Deep branches hydrate only on demand.
Messaging is an adapter.
Foxtalk may carry a pointer or status. It is not the reasoning substrate, seat scheduler, or authority engine.
The operator still decides.
Evidence and argument earn recommendations. No model, score, judge, or manifest grants operational authority.
Not every question needs a parliament.
Every top-level intelligence chooses the smallest honest route. The caller may override it.
Grounded Independent Council
High impact, unresolved competing options, or explicit Council request. Five isolated seats plus a separate judge.
Independent minds.
One shared reality.
“Blind” never means knowledge-starved. Seats share a checksummed fact packet while keeping hypotheses, probes, recommendations, and precommits private.
The non-voting evidence compiler
Resolve exact scope
Project, question, cutoff, budget, forbidden assumptions.
PX current truth
Docs, call paths, schemas, tests, freshness, contradictions.
Approved live reads
Current state when source and stale docs cannot answer.
Goalie research
Primary-source leads, prior art, counterexamples, projects.
Checksum packet
Facts, unknowns, boundaries, evidence horizon, pointers.
The compiler resolves the exact question and evidence cutoff. It does not recommend a solution or score candidates.
One mind.
An adaptive forest.
Ultra is deep single-agent reasoning—not six canned thoughts and not a hidden swarm.
W1W2W3stop on diminishing information
A Council—or it says it is not.
The conductor dispatches the roster. It never impersonates the originator, five experts, Null Defender, repairers, skeptic, or judge in its own context.
Freeze shared facts.
PX, approved live reads, and Goalie compile one solution-neutral packet. Fact disagreements return here instead of becoming design theater.
No more one mind
wearing five hats.
A good answer from self-play is evidence that Ultra works. It is not evidence that a Council happened.
Independent contexts, shared custody.
- Five distinct seat targets plus a separate judge
- Private precommit before candidate reveal
- Every seat authors its own probes and falsifiers
- Material issues receive targeted repair and skeptic retest
- Exact model/build/runtime receipts determine independence grade
Ultra is honorable. Fake Council is not.
- One execution context assigned to multiple seats
- Judge reuses an expert or originator context
- A SELF_ONLY seat spawns nested agents
- Provider changes exact model identity after pinning
- Fewer than five available targets → ULTRA or DECLINE/DEFER
Real models.
Real fallbacks. Real names.
OpenAI-compatible and command adapters bind target IDs to actual executors. Silence, transport failure, or malformed first output rotates before pinning. Once valid, identity drift fails closed.
OpenRouter
Structured JSON execution with returned exact model identity.
openai/gpt-5.6-sol
Groq
Independent provider lane with system fingerprint receipt.
llama-3.3-70b-versatile
Local Qwen
Endpoint is visible; inference currently needs its GPU client reconnected.
Qwen3.6-35B-A3B · :8430
Command executor
Request over stdin, response over stdout, shell disabled, env allowlisted.
Claude · Codex · local tools
The work lane gets the decision—not the transcript.
Rich reasoning remains inside immutable artifacts. Operational agents hydrate only what they need.
Green for machinery.
Amber for identity limits.
The page carries the same rule as the harness: code, tests, packaging, provider execution, and local runtime availability are different receipts.
Self-play refusal
Duplicate targets, reused contexts, judge reuse, and nested delegation are rejected or downgraded.
Adaptive Ultra
Autonomous waves, frontier lineage, negative-branch repair, unequal depth, and honest stopping are enforced.
Cross-model Council
OpenRouter GPT plus Groq Llama completed precommit, reveal, repair, skeptic retest, and separate judging.
Immutable CLI
Evidence digest, validation, executor configuration, create-once output, package build, and command help passed.
Provider identity
The live run lacked complete exact build/runtime-instance proof, so it returned DEGRADED_COUNCIL.
Local Qwen inference
Health and model discovery respond; generation returns 500 until the GPU client reconnects.
Ask normally.
Override when you mean it.
Council and Ultra are shared agent skills backed by the same executable package. The syntax changes; the custody rules do not.
Native registered skills
$council Should we…?
$ultra Think deeply about…
Slash commands
/council Should we…?
/ultra Think deeply about…
corepack pnpm --filter @foxflow/reasoning-harness dev council evidence-sha --input evidence.json
corepack pnpm --filter @foxflow/reasoning-harness dev council validate --input case.json
corepack pnpm --filter @foxflow/reasoning-harness dev council run \
--input case.json --executors executors.json --out immutable-result.jsonA full Council.
Nobody stops working.
Independent thought without factual starvation. Adversarial repair without memo theater. Deep reasoning without context sludge. One human-visible decision, with receipts.