プロダクト

Auto-Dev Universe

すべてのフェーズに責任ゲートを備えたエージェント駆動の開発。コードは統制パイプラインを通じて生成・レビュー・テスト・マージされます ── 魔法の杖ではなく。

開発サイクル

コードジェネレーターではない。開発システムである。

Even with ambiguous instructions, auto-dev reaches the completion line.

Process

Define Completion Line

Not perfect spec, but minimum success criteria that can be verified

Acceptance: verifiable

Small Iterations

Build, run, verify, fix — always recoverable, never destructive

Max cycles: 3

Auto-Complete

Continues until acceptance criteria met or stops with full context

False-done: 0%

Iteration Cycle

Build: Patch generation

Pass & Complete

Stop + Context

Why It Completes

Executor AI

Builds & Runs

separated

Judge AI

Verifies & Approves

No RunawayExecutor & Judge are structurally separate

No Goal DriftCompletion criteria locked at ticket creation

No False DoneOnly verifiable acceptance counts

Ideal for:Code improvementsTest additionsFeature buildsRefactoringPrototype drafts

Ambiguity in. Verified completion out.

AUTO-DEV UNIVERSE — セクション01 ── 開発パイプライン

仕様からリリースまで ── すべてのフェーズにゲート。

Speed is not the goal. Structured evolution that never breaks is.

Input Packs

Repo Pack

Code, deps, build

Spec Pack

Issue, PRD, ADR

Policy Pack

Gate YAML, rules

Quality Pack

Tests, SLO, lint

Work DAGtopological order

Plan→Design→Implement→Test→QA→Docs

Responsibility DAGfail-closed

Scope Gate→Risk Gate→Quality Gate→Release Gate

Action Gateway

git pushPR createDeploySecrets

require_ticket: true · require_trace: true

Trace Store

Event Sourcing→Immutable Records→Audit Replay

No trace, no side effects. Gaps are impossibilities.

Agents propose. Gates authorize. Gateway executes. Trace proves.

Universe Builder

開発Universeを構成する。

5 agents, each with sub-agents. Risk surfaces at every step.

Dev Chief

Ticket

Issue accepted

Scope

Blast radius estimated

Plan

DAG assembled

Execute

Builder + sub-agents run

Test

Coverage verified

Review

Schema + invariants checked

Security

Vuln scan + auth review

Release

Canary plan + deploy

Five agents. Fifteen sub-agents. Every risk scored before execution.

AUTO-DEV UNIVERSE — ワークDAG

判断を統制し、実行を並列化

3つのゾーン：判断、実行、検証。

Zone A

ImplementDONE

TestRUNNING

AnalysisREADY

SecurityBLOCKED

DocsBLOCKED

Zone B

ImplementDONE

TestDONE

AnalysisRUNNING

SecurityRUNNING

DocsREADY

Zone C

ImplementRUNNING

TestBLOCKED

AnalysisBLOCKED

SecurityBLOCKED

DocsBLOCKED

Formal Definition

Executable(n) ⟸ ∀ d ∈ deps(n) : status(d) = DONE

DAG_zone = (N, E) where E ⊆ N × N is acyclic

Parallel by default

Ready nodes across zones execute simultaneously

No cross-zone side effects

Zones are isolated; coupling only through Decision DAG

Deterministic replay

Same inputs, same execution order, same results

Parallelism is not a performance hack. It is a structural guarantee.

AUTO-DEV UNIVERSE — エージェント機能

5つのエージェント。明確な境界。重複なし。

各エージェントが開発ライフサイクルの1フェーズを担当。

Builder

Implementation-focused execution

Repo NavigationPatch CraftingBuild & TestTool Discipline

Test

Test generation and failure isolation

Test GenFailure LocalizationRepro StepsCoverage Analysis

Reviewer

Design integrity and diff quality

Schema DisciplineArchitecture InvariantsImpact AnalysisEvidence Production

Security

Vulnerability and permission audit

Dependency ScanAuth ReviewSecret DetectionThreat Model

Release

Deploy planning and rollback design

Canary PlanFeature FlagsRollback DesignStage Deploy

AExecution

BStructure Preservation

CRisk & Responsibility

DAudit Readiness

Risk Scoring Function

risk(v) = w₁·financial + w₂·auth + w₃·PII + w₄·blast_radius − w₅·rollback

Risk determines gate escalation thresholds. Higher risk requires stronger evidence and human approval.

Capability is not intelligence. It is accountable execution within governed boundaries.

並列開発、ゲート付きマージ

エージェントは同時に稼働。マージにはゲート承認が必要。

Each agent spawns specialized sub-agents. Risk scores aggregate upward. Quality loops never stop.

Builder

Repo Nav

Patch Craft

Build Run

Patch Set

blast_radius: 3

Test

Test Gen

Failure Loc

Repro Steps

Coverage Report

coverage_gap: 0

Reviewer

Diff Check

Impact Eval

Invariant

Review Verdict

schema_drift: 1

Security

Dep Scan

Auth Review

Secret Det

Security Report

vuln_surface: 2

Release Gate Convergence

risk(v) = w₁·blast + w₂·vuln + w₃·drift ≤ threshold

Verified Release Package

Built, tested, reviewed, scanned — with canary plan

Parallel Agents

Sub-Agents Active

Risk Score (max)

Gate Violations

Parallel execution. Aggregated risk. Gate-enforced release.

AUTO-DEV UNIVERSE — 品質ゲート

コードがリリースされる前に4つのゲート。

Gates are not monitoring. They are execution boundaries. If the gate does not open, the action cannot happen.

Agent→Scope→Risk→Quality→Release→Ticket→Gateway→Trace

AgentAction GatewayDirect execution: impossible

Scope Gate

Trigger: files_changed > 25, sensitive paths touched

Action: BLOCK or ESCALATE

Risk Gate

Trigger: risk >= 3

Action: REQUIRE_HUMAN_APPROVAL

Quality Gate

Trigger: Evidence incomplete (tests, typecheck, lint, security)

Action: BLOCK

Release Gate

Trigger: Missing canary_plan or rollback_plan

Action: BLOCK_IF_MISSING

Gate Policy YAML

principles:
  fail_closed: true
  no_ticket_no_side_effects: true
  no_trace_no_side_effects: true

action_gateway:
  require_ticket: true
  require_trace_prewrite: true
  verify_signature: true

No Ticket, No Side Effects

If the gate did not issue a ticket, the action cannot execute. Enforcement by construction.

No Trace, No Side Effects

If trace cannot be written synchronously, execution is denied. Gaps are not bugs, they are impossibilities.

Enforcement is not detection. It is structural impossibility.

AUTO-DEV UNIVERSE — 収束

生成からデプロイまで ── 1つのパイプライン。

4つの状態変数がパイプライン全体のコード成果物を追跡。

G0→Evaluate→Propose→Approve→G1→Evaluate→Propose→Approve→G2

Self-rewrite within same generation: FORBIDDEN

State Variables

q_k

Quality

Tests, static analysis, SLO compliance

e_k

Uncertainty

Unverified scope, low repro rate

r_k

Risk

Change area, permissions, data impact

d_k

Diff Size

Lines, files, dependency reach

Convergence Criterion

Converged ⟸ ∀ k ∈ [t−N, t] :
q_k ≥ q_target ∧ e_k ≤ e_target
∧ critical_alerts = 0
∧ replay_match = 100%
∧ d_k ≤ d_threshold

Scope Bound

d_k <= d_max

Auto-split if exceeded

Risk Escalation

r_k >= r_human

BLOCKED until approval

Evidence Required

EvidenceComplete(v)

No downstream without it

Convergence Meters

Quality↑ 85%

Uncertainty↓ 20%

Diff Size↓ 15%

Replay Match↑ 100%

Divergence Detection

Same failure repeated M times triggers spec-request or human escalation. Never retries blindly.

Auto-Dev is not about building faster. It is about converging with certainty.