Thorbit Codex ADW

Runtime Model

Three layers, one closed loop.

Planning skills produce durable artifacts. Execution agents turn artifacts into managed work. Fresh local state records every move. Codex hooks gate the whole thing so nothing ships unverified.

Planning Layer

architecture workflow task-orchestration

architecture → workflow → task-orchestration

Execution Layer

builder reviewer tester debt-triage

builder → reviewer → tester → debt-triage

Fresh Local State

tasks.db audit.db deposition.db

SQLite, initialized per target repo, never inherited

Codex Guardrails

runtime policy comment guard managed-plan stop gate task completion ping

Pre-tool and stop hooks enforce policy and capture completions

Skill Orchestrations

Four skills, each a bounded pipeline.

Every skill has one job and writes evidence to disk. Planning skills emit a universal spec, execution converts it to DB-backed work, verification proves it before completion.

architecture

Audits a codebase through a wide inventory wave, a diagnosis wave, synthesis into remediation tracks, a principles audit, then a product-impact enrichment. Outputs a read-only audit report or a universal spec.

inventory wave (36 agents)→ diagnose wave (28 checks)→ synthesize 6 tracks→ issues.jsonl→ principles audit→ audit / spec

workflow

The design layer. Turns a feature or refactor request plus architecture context into a universal implementation blueprint: an exact file manifest and grep-verifiable acceptance criteria, before any task is created.

read context→ use repo patterns→ draft spec→ file manifest→ acceptance criteria→ universal-spec.md

task-orchestration

Converts a universal spec into DB-backed tasks, dispatches builders, gates reviewer outcomes, runs test assessment, and converges through debt triage. Loops follow-up tasks back into the DB until triage is clean.

convert-to-tasks→ tasks.db→ builders→ reviewers (gate)→ test assess→ debt-triage→ complete-if-clean

execution-integrity

Audits runtime reliability after static quality is understood: tool handoffs, queue state, job lifecycles, UI readiness, and recovery paths. Reuses architecture artifacts when present, otherwise runs a fresh runtime investigation.

reuse / fresh→ inventory runtime→ diagnose failure classes→ synthesize→ spec→ tasks

Agent Roster

23 Codex agents, four jobs to be done.

All agent configs ship in every install. Builders mutate code, verifiers prove it, auditors find the risk, spec roles turn findings into plans.

Build

builder

Executes one task with required reads and mutation logs

forge-agent

Ingest, architect, build, infra, verify when briefed

test-resolver

Fixes test files only, escalates source or infra

live-runner

Starts live infra, runs system tests, tears down

Verify

reviewer

Gates task outcomes before completion

tester

Unit, integration, scenario, UX, UI evidence

test-creator

Detects gaps, creates missing test tasks

adw-tester

Live system tests, captures real HTTP output

task-reviewer

Reviews flagged findings for accuracy and severity

trace-scout

Validates file:line claims, returns JSON

Audit

auditor

Architecture diagnostic checks, read-only findings

inventory-auditor

Writes one inventory artifact per brief

diagnose-auditor

One architecture diagnostic artifact

principles-auditor

Verifies remediation coverage

log-auditor

Audits error-path logging quality

silent-failure-hunter

Finds silent failures and weak fallbacks

debt-triage

Turns task notes into follow-up records

researcher

Orient, explore, evaluate research loop

Spec & Synthesis

architecture

Spec-ready findings for one analysis task

spec-writer

Converts artifacts into universal-spec.md

synthesizer

Combines artifacts into plans and issues

generate-expert

Generates expert skills from audit artifacts

scenario-writer

Structured scenario files for user flows

The Discipline

What it never copies.

The installer is intentionally conservative. A fresh project means fresh runtime state. State from one codebase would only confuse the next.

.agents/data/*

SQLite state belongs to one codebase only.

.agents/runs/*

Old audits and plans would confuse new work.

active-plans.json

Replaced with an empty registry.

.env*

Secrets stay local. Never travel.

.vercel

Deployment linkage must be target-specific.

build output

Noise and stale artifacts, dropped.

Why It Matters Here

This is your doctrine, made executable in Codex.

Thorbit Codex ADW is a concrete, Codex-native build of the same agentic engineering principles the fleet already runs. The vocabulary lines up almost one to one.

• Planning skills emit a universal spec before any code
• Grep-verifiable acceptance criteria on every task
• Reviewer gate, then evidence-backed tester
• Debt-triage loops follow-ups until clean
• Hooks enforce policy and capture completions
• Fresh DBs, no inherited state

Fleet Doctrine

• Plan first: 80% of success is the plan
• Close the loop: validators embedded in the task
• PDCA: Plan, Do, Check, Act with a QA veto
• Petey: triage learnings into the next cycle
• Hooks and stop gates: deterministic enforcement
• Presence down: autonomous, verified execution

The agentic dev runtime, packaged as one install.

Three layers, one closed loop.

Planning Layer

Execution Layer

Fresh Local State

Codex Guardrails

Four skills, each a bounded pipeline.

architecture

workflow

task-orchestration

execution-integrity

23 Codex agents, four jobs to be done.

What it never copies.

This is your doctrine, made executable in Codex.

Thorbit Codex ADW

Fleet Doctrine

What was produced.

This breakdown →

Excalidraw diagram ↓

Infographic →