Install from the Codex marketplace
Find TokenX in the Codex plugin marketplace and install it.
Guarded Codex routing
Match Codex depth to the work.
TokenX routes each Codex task to the cheapest profile that can do it, and holds the line on depth where it matters. Routing starts after install and hook trust. TokenX does not guarantee savings.
- Codex-only routing
- Install + hook trust
- Guard-enforced assignments
- No measured savings claim
Install
Three steps, then it routes.
Review and trust the hooks
Run /hooks and inspect the four TokenX hook commands. Routing does not activate until they are trusted.
Start a new session
Open a fresh Codex thread. Eligible work is routed from the first prompt.
Allocation policy
Premium reasoning is a budget, not a default.
TokenX evaluates task scope, risk, context, and separability. The model floor rises with the work, while routine tasks stay out of the most expensive reasoning lane.
| Route | Profile | Default effort | Allocation intent |
|---|---|---|---|
| Economy | Luna |
Low | Reads, docs, bounded edits, and repeatable checks |
| Standard | Terra |
Medium | Normal implementation and multi-file engineering |
| Deep | Sol |
Xhigh | Security, architecture, concurrency, and hard debugging |
Defaults only. Task profiles and safety floors can raise effort within a route. Explicit no-delegation stays with the parent; conversational prompts can remain passthrough. TokenX does not change the already-running parent model.
What it won't do
Safety sets the floor.
TokenX does not downgrade risky or coupled work merely to make a route cheaper.
Complex, coupled, high-risk
- Authentication, authorization, and security work
- Cross-service architecture and concurrency
- Destructive migrations and production incidents
- Deep debugging after a comparable attempt failed
Safety and structural complexity set a floor. TokenX never downgrades risky work merely to make the route cheaper.
Assignment contract
Current-turn tokens, guard-enforced profiles.
Dynamic assignment dispatch is enabled by default. For separable
work, TokenX advertises decision-scoped assignment tokens. The
parent dispatches with the advertised task_name and a
message, and omits agent_type, model, and effort. The
guard injects the resolved model, reasoning effort, and role
contract, pins child conversation context to a bounded fork, and
enforces the per-turn sibling limit.
Current-turn only
Assignments are valid only in the turn that advertised them. On every new turn, obtain and use a newly advertised assignment; never reuse an old token.
Five bounded roles
repo_reader, repo_reviewer, test_runner, file_editor, and implementer cover read, review, test, edit, and implement workstreams.
Sibling caps
Automatic siblings stay within economy 4, standard 3, and deep 2. Coupled work stays on one agent; the guard enforces the selected ceiling.
Enterprise economics
Move the spend curve without lowering the safety floor.
TokenX can reduce avoidable model spend by assigning eligible routine work to the least expensive capable profile and reserving higher-capability profiles for tasks where depth changes the outcome.
Routine, bounded, separable
- Repository reads and status checks
- Small documentation or configuration updates
- Mechanical edits with an established pattern
- Independent reviews split across low-effort agents
Four economical readers can be more efficient than one deep agent when the work separates cleanly.
Evidence for operators
Give finance a route mix, not a marketing percentage.
Deterministic routing handles explicit controls, safety floors, and clear cases. For uncertain eligible work, TokenX can make one isolated Luna/low assessment and falls back deterministically if it cannot obtain an enforceable result.
- Route mix
- Economy, standard, deep, parent, and passthrough allocation
- Agent budget
- Up to 4 economy, 3 standard, or 2 deep siblings per turn
- Outcome signals
- Prompt-free retries, sensitivity denials, and smart fallbacks
Measure before claiming.
TokenX records routing allocation and bounded outcome signals, not billing-grade token usage. Actual savings depend on workload, model pricing, plan terms, task success, and retries. TokenX does not guarantee savings.
A sharper default