The shortest path to the right workflow.

Choose the workflow #

Use the smallest workflow that matches the uncertainty, impact, and proof required.

01

Plain Codex

Small, reversible work with an immediate oracle.

Use for: mechanical edits, explanations, localized fixes.

02

Forge

You have an idea, not yet a proposal.

Creates: one implementation-ready concept from five creator lenses.

03

Council

You have a proposal and want pressure, dissent, and blockers.

Returns: a ranked judgment and verification plan.

04

Mind

You need Forge → Council → authorized Hyper execution in one guided path.

Orchestrates: proposal, review, stop gates, and the authorized path through Hyper execution.

05

Hyper

Implementation workflow for authorized, complex repository changes.

Owns: contract → one writer → checks → falsification → evidence-backed close.

Fast rule. If a single test can settle a small change, stay plain. If the unknown is the idea, use Forge. If the unknown is the decision, use Council. Use Mind for the complete gated path. Use Hyper directly only for authorized implementation.

How it works #

Mind connects Forge → Council → Hyper; each workflow can also run alone. None of them turns consensus into proof.

Council: isolate, score, preserve dissent

01

First opinions

Members answer before seeing one another’s work, limiting direct cross-contamination.

02

Anonymous review

Candidates lose authorship. Review depth and count follow the selected mode; Standard scaffolds performance-impact and coverage roles.

03

Aggregation

Scores combine deterministically when valid reviewer JSON exists. Dissent and blockers stay visible.

04

Final Call

The Chairman returns approved, approved-with-risk, or blocked with confidence and verification.

What isolation can—and cannot—do. It is designed to reduce first-answer anchoring and expose disagreement. The roles are still correlated outputs from Codex, not independent model vendors, and accuracy must be established with task evidence.

Workflow boundaries

Inputs and outputs by workflow
Workflow	Starts with	Ends with	Does not do
Forge	Problem or opportunity	Unified proposal	Judge or implement it
Council	Concrete proposal	Final Call + dissent + blockers	Turn votes into proof
Mind	Idea plus constraints	Proposal + Council verdict + Hyper result when authorized	Hide execution inside deliberation authorization
Hyper	Authorized change or approved handoff	Checked code with evidence, or an explicit blocker	Re-run Forge or Council recursively

Frontend evidence gate

Leonardo reviews UX/UI. Bob executes browser evidence and does not vote. If the browser did not run, visual, keyboard, history, console, and overflow claims remain UNKNOWN.

Hyper’s six stages

ContractObservable done, scope, constraints, rollback, oracle.
ObserveRepository rules, worktree, baseline, bounded read-only investigation.
OrientMap acceptance criteria to evidence and reserve one writer.
ActImplement the smallest reversible path; check meaningful increments.
FalsifyChallenge the contract, raw diff, and checks. Relay adds a fresh verifier; Solo declares reduced independent coverage. Critical FAIL or UNKNOWN blocks completion.
CloseReconcile the contract, rerun the right checks, review the diff, and report residual risk.

Evidence Lab #

One completed public record, with replayable code evidence and stated provenance limits.

Completed

Task, frozen input, patch, checks, verifier verdict, and replay record are public.

Provenance partial

The code outcome is replayable; raw orchestration transcripts were not retained.

Prompt example

A reusable prompt structure, never presented as a completed outcome.

Strong run record

run ID, date, plugin version, and source commit;
worktree state or frozen snapshot checksums;
exact task, clarifications, baseline, workflow, and route;
resulting patch, exact commands, exit codes, and unedited stdout/stderr—or explicit omissions;
workflow provenance: saved agent/verifier transcripts, or a clear PARTIAL label;
fresh-verifier verdict against the accepted contract;
residual risk and usage marked unavailable when not observable.

Completed public record

Local Hyper evidence captured on 23 July and replayed on 24 July 2026
Run	Observed result	Status
Node scheduler · Hyper Relay	13/13 strict tests PASS, syntax PASS, fresh verifier PASS against the final clarified contract, and published patch replay PASS.	Outcome replay PASS Workflow provenance PARTIAL

This is one case study, not a benchmark or a general quality claim. It records a Hyper Relay implementation from the exact local build named in the evidence bundle.

Open the Evidence Lab

Prompt examples #

Copy the structure, replace the placeholders, and define evidence that can settle the task. These are examples, not product results.

Prompt anatomy

Prompt example

Use <workflow> for <decision or change>.
Context: <repo, files, users, current behavior>.
Constraints: <compatibility, safety, time, dependencies>.
Success: <observable acceptance criteria>.
Evidence: <commands, tests, browser, metrics, counterexamples>.
Do not: <out-of-scope actions or claims>.

Plain Codex3 templates

Mechanical edit

Prompt example

Update <exact files> to <mechanical result>. Preserve behavior and unrelated work. Run <formatter/test> and report the diff.

Diagnosis only

Prompt exampleDiagnose <symptom>. Trace the smallest reproducible cause with file:line evidence and commands. Do not implement a fix.

Localized test fix

Prompt exampleFix <failing test> in <module>. Keep the patch local, add the missing regression case, and run the focused suite.

Codex Forge6 templates

New product concept

Prompt example

Use $codex-council:codex-forge to turn <user problem> into one implementation-ready product proposal. Compare user value, interaction, architecture, delivery, and simplification.

System shape

Prompt example

Forge a system design for <capability> under <scale, latency, privacy, compatibility>. Return boundaries, data flow, failure modes, rollout, and smallest viable slice.

Developer tool

Prompt example

Forge a developer tool for <workflow pain>. Optimize for adoption, debuggability, local-first operation, and a measurable first release.

Workflow redesign

Prompt example

Forge a simpler workflow for <current process>. Preserve <hard controls>, remove avoidable handoffs, and define the migration path.

Migration path

Prompt example

Forge a staged migration from <old system> to <target>. Include coexistence, rollback, observability, ownership, and exit criteria.

Contrarian simplification

Prompt example

Forge the smallest alternative to <large proposal>. Include what not to build, the assumption it tests, and the condition that would justify expansion.

Codex Council12 templates

Architecture decision

Prompt example

Standard Council: judge <proposal A vs B> for <system>. Ground the recommendation in repository constraints, migration cost, failure modes, and verification.

Risky diff / pull request

Prompt example

Standard Council: review <diff or PR>. Find correctness blockers, compatibility risk, missing tests, rollback gaps, and the smallest safe merge condition.

Bug and regression RCA

Prompt example

Standard Council: evaluate the proposed root cause for <regression>. Separate observed facts from hypotheses and require a reproducer that can falsify the winner.

Performance-sensitive change

Prompt example

Deep Council: judge <performance proposal>. Require baseline workload, contention behavior, measurement plan, budgets, and rollback thresholds.

Database migration

Prompt example

Deep Council: review <schema/data migration>. Preserve compatibility, idempotency, resumability, auditability, rollback, and partial-failure recovery.

Public API change

Prompt example

Standard Council: judge <API change>. Map consumers, versioning, error semantics, migration aids, observability, and breaking-change evidence.

Refactor vs rewrite

Prompt example

Standard Council: decide refactor, strangler migration, or rewrite for <module>. Quantify risks, reversible milestones, and stop conditions.

Dependency upgrade

Prompt example

Standard Council: review upgrading <dependency>. Check current official migration docs, compatibility surface, lockfile impact, rollback, and smoke tests.

Frontend interaction

Prompt example

Standard Council with frontend review: judge <flow>. Leonardo reviews UX; Bob verifies keyboard, focus, responsive states, console errors, and screenshots. Unknown browser evidence blocks UI completion.

Release go / no-go

Prompt example

Deep Council: issue a go, revise, or stop call for <release>. Use deployed facts, open incidents, rollback readiness, owner coverage, and explicit blockers.

Incident postmortem

Prompt example

Standard Council: review <postmortem>. Challenge the causal chain, contributing conditions, detection gaps, action owners, and evidence that recurrence risk fell.

Privacy and sharing

Prompt example

Deep Council: judge <data-sharing feature>. Identify data classes, minimization, retention, deletion, access boundaries, audit needs, and claims that require a dedicated security review.

Codex Mind3 templates

Idea to verdict

Prompt example

Use $codex-council:codex-mind to forge and judge <idea>. Show the combined compact estimate first. Return the proposal, dissent, blockers, and Final Call. Do not implement.

Idea to optional build

Prompt example

Use $codex-council:codex-mind for <idea>. After an accepted estimate, run Forge then Council. Offer bundled Hyper only if the Final Call is build and I separately authorize implementation.

Constrained feature

Prompt example

Mind: shape and judge <feature> under <hard constraints>. Preserve incompatible options and dissent until Council decides; no implementation handoff while a blocker remains.

Codex Hyper6 templates

Concurrency contract

Prompt example

Use $codex-council:codex-hyper to implement <concurrency change>. Characterize legacy behavior, define cancellation and ordering semantics, use deterministic race tests, and close with a fresh counterexample search.

Public API implementation

Prompt example

Hyper: implement <approved API contract>. Map consumers, preserve compatibility, keep one writer, add contract tests, and use a fresh verifier to falsify error and migration behavior.

Persistent data change

Prompt example

Hyper: implement <approved data change>. Require idempotency, crash recovery, dry-run, rollback, audit evidence, and failure injection before close.

Build or deployment contract

Prompt example

Hyper: implement <build/deploy change>. Freeze the current artifact contract, test clean and incremental paths, preserve rollback, and verify the produced artifact—not only command exit codes.

Frontend with browser gate

Prompt example

Hyper: implement <approved UI change>. Bind acceptance to keyboard, focus, mobile/desktop screenshots, console errors, history/state, and overflow. Browser UNKNOWN blocks completion.

Multi-module refactor

Prompt example

Hyper: refactor <approved scope> across <modules>. Use read-only explorers for separate surfaces, one root writer, characterization tests, final diff review, and a fresh verifier.

Operations #

Run only what you can account for. Estimates are local heuristics, not billing data.

Modes and budgets are different controls

Council deliberation controls
Control	Values	Changes
Mode	fast · standard · deep	Deliberation topology and review depth.
Budget	compact · balanced · expanded	Context and output allowance.

Standard and Deep require an accepted estimate. Expanded requires explicit confirmation. Expanded permits larger outputs and references; it does not guarantee complete evidence or an audit trail.

Preflight discipline

Real CLI shape · help verified locally

python3 scripts/codex_council.py estimate \
  --topic "<proposal>" \
  --mode standard \
  --type implementation \
  --token-budget compact

Keep usage as unavailable when the runtime does not expose it. Never turn the estimate into a bill or observed consumption.

CLI quick reference

# Validate the local plugin.
python3 scripts/codex_council.py validate --plugin-root . --strict

# Estimate before dispatch.
python3 scripts/codex_council.py estimate --topic "<proposal>" \
  --mode standard --type implementation --token-budget compact

# Preview the opt-in router without dispatching.
python3 scripts/codex_council.py estimate --topic "<proposal>" \
  --router auto --panel auto --json

# Scaffold session artifacts after accepting the estimate.
# init does not dispatch agents.
python3 scripts/codex_council.py init --topic "<proposal>" \
  --mode standard --type implementation --token-budget compact \
  --confirm-estimate

# Forge uses the same estimator/scaffolder with --type forge.
python3 scripts/codex_council.py estimate --topic "<idea>" \
  --mode standard --type forge --token-budget compact

# Inspect a completed session.
python3 scripts/codex_council.py compile-context \
  --topic "<handoff>" --constraint "<constraint>" --json
python3 scripts/codex_council.py doctor --session <session-path>
python3 scripts/codex_council.py stats --session <session-path> --json
python3 scripts/codex_council.py dashboard

Alters

Preview bounded role tuning before configuring it. Alters are advisory and cannot remove guardrails. Bob is an evidence runner, not a tunable voter.

python3 scripts/codex_council.py alters list
python3 scripts/codex_council.py alters preview --role grace --strictness high
python3 scripts/codex_council.py alters show --role grace

Decision Runtime

The runtime is experimental and shadow-only. Legacy session artifacts and the Chairman verdict remain authoritative. Cells, frontier, recovery, and replay improve local traceability; they do not change the verdict.

python3 scripts/codex_council.py cells project --help
python3 scripts/codex_council.py cells apply --help
python3 scripts/codex_council.py cells plan --help
python3 scripts/codex_council.py cells doctor --help
python3 scripts/codex_council.py cells recover --help
python3 scripts/codex_council.py cells rollback --help
python3 scripts/codex_council.py cells purge --help
python3 scripts/codex_council.py cells replay --help
python3 scripts/codex_council.py cells fault-test --help

Governance and privacy

Local by default. Keep session artifacts and alter configuration local unless the user authorizes sharing.
Minimize prompts. Do not hand subagents secrets, credentials, raw logs, PII, or unrelated repository content.
One writer in Hyper. Explorers and verifiers remain read-only; evidence decides disagreement.
Security boundary. Council can judge trade-offs. Repository security scans and fixes belong to Codex Security workflows.

Reference #

Names, verdicts, limits, and the mistakes that should stay visible before a run.

Council roles and gates

Ada Lovelace

Boundaries, integration, maintainability, migration risk.

Grace Hopper

Failure modes, tests, rollback, observability.

Hypatia

Secrets, permissions, privacy, provenance, policy.

Florence Nightingale

Workflow fit, docs, adoption, operational friction.

Alan Turing

Hidden assumptions, simpler alternatives, overengineering.

Seymour Cray

Latency, throughput, memory, cost, scale, measurement.

Leonardo da Vinci

UX/UI gate when frontend review is enabled.

Bob

Browser evidence runner. Reports observed PASS / FAIL / UNKNOWN; never votes.

Competency packs

Optional lenses cover implementation, regression, performance, context discipline, governance, operator UX, frontend UX, and contrarian simplification. They sharpen a run; they do not create independent model diversity.

Read the output in this order

Final Callapproved, approved-with-risk, or blocked.
Blockersmust-fix conditions before approval.
Dissentminority risks that survived aggregation.
Verificationwhat ran and what remains UNKNOWN.
Residual riskwhat is accepted, reversible, or still exposed.

Common mistakes

Using the council for everything. Small, reversible, checkable work does not justify Council overhead; use plain Codex.
Treating consensus as proof. Agreement is advisory; the verification must run.
Calling a template a run. No task + patch + checks + verdict means it was not executed.
Declaring UI complete without a browser. Browser-only claims stay UNKNOWN.
Confusing modes with budgets. One changes topology; the other changes allowance.
Assuming init dispatches agents. The CLI scaffolds artifacts; the Codex skill orchestrates the run.

FAQ

How is this different from prompting Codex twice?

Two ordinary prompts do not enforce isolation, blinded attribution, a shared rubric, or preserved dissent. Council does. Whether that improves an answer must be established by task evidence, not assumed.

Does it call other model providers?

No. The workflow stays inside Codex. Role diversity is useful but correlated.

Will it spend a lot of tokens?

Standard and Deep pause on a local estimate; compact is the default. Expanded needs explicit confirmation. Estimates are not billing telemetry.

Do I need the CLI?

No. Skills orchestrate everyday runs. The CLI provides reproducible estimates, scaffolding, scoring, validation, health, and stats.

When can Mind offer Hyper?

Only after Council returns build, no blocker remains, and implementation is explicitly authorized.

Does Decision Runtime change the verdict?

No. It is a shadow projection; the legacy session and Chairman verdict remain authoritative.

Glossary

ChairmanMain Codex agent that synthesizes the saved record.

Candidate A–FAnonymized Council outputs during scoring.

BlockerCondition that prevents approval.

DissentPreserved minority position.

PreflightEstimate and explicit acceptance gate.

Mission ContractHyper’s observable done, scope, constraints, rollback, and oracle.

Evidence runnerExecutes checks and reports observations; does not vote.

Coverage partialSome expected artifacts are missing; the report is incomplete.

Decision CellExperimental shadow record for a claim, option, risk, or decision.

QuarantineState that must not be used because it is unknown, corrupt, or policy-invalid.