Truline
Build in public · Free skill

We stopped letting our best model write code.

Fable 5 plans, shards, gates, and judges. A $0.76 per MTok worker builds, and this week that worker tied our flagship three for three on real agentic coding slices. The system that runs it is one file, and it is free.

truline mission control · /mission-continuelive run

Context slices

ISSUE-702adone · 100%
ISSUE-702bdone · 100%
REQ-302running · 64%
M2 gatequeued

Dispatch feed

fable 5 packet approved, dispatching ISSUE-702a
worker handoff, GLM 4.7, 44s, validator guarded
fresh context validator, 1509/1509 tests
codex review, no blocking findings
advancing to boundary, orchestrator idle

Judgment is premium. Keystrokes are a commodity. Fable writes the contract and grades the work. The worker just types, and the validator does not care who typed.

The receipts

Measured, not simulated.

Three bounded coding tasks, run by real headless agents with a tool loop, an external validator, and git isolated repos. The same harness we dispatch production work with.

3/3
First pass, both models
2.6x
Faster on the mid tier
18x
Cheaper per MTok
44s
Fastest slice
Eval scorecard: GLM 5.2 and GLM 4.7 both pass three of three slices first try, with 4.7 faster on every task

the rematch · agentic workers · external validator · n=3 sharded slices

Cost per MTok by model, from the flagship at $14.00 down to the worker at $0.76

what the worker seat costs · the orchestrator seat stays premium, that is the whole point

The relay

How a slice moves.

Every issue travels the same relay. Premium judgment at the ends, commodity execution in the middle.

01

Fable 5 orchestrates

Interviews you, writes the validation contract, shards bounded issue slices, pins scope, and judges every result.

02

GLM executes

44s

Validator guarded slices on the cheapest capable backend. Measurably equal on sharded work, escalation only on failure.

03

Codex reviews

GPT 5.5 sits in a different vendor family, so the worker never grades its own homework. Blocking findings send it back.

04

Gemini sees

Screenshots and visual verification route to a model that can look. Text models stay blind by policy.

What you get

The actual skill, unedited.

Not a blog post about the idea. The two command files we run in Claude Code, plus install notes.

/mission, the planning half

A structured interview that ends in a reviewed packet: a validation contract with stable VAL assertions, a PRD, tracer bullet issue slices, adversarial and engineering reviews, and a hard approval gate. Nothing executes from planning.

/mission-continue, the execution half

Runs the approved packet one bounded context slice at a time: dispatch to the cheapest capable backend, structured worker handoffs, fresh context validators, and stop at boundary discipline.

The routing doctrine, embedded

The orchestrator seat stays premium. Workers route down and off vendor by default and escalate only on validator failure. The eval on this page is the doctrine paying rent.

Honest install notes

These files reference paths and gate tooling from our machine. The README marks every line you will adapt, roughly fifteen minutes to wire into your own Claude Code setup.

Free

Get the /mission skill.

Name and email. The download appears right here, no hoops.

Collected by Truline. You get the skill file and, at most, occasional build in public notes on this system. One click unsubscribe on anything we send. No resale, no ad lists.