Ana içeriğe geç

Claude Code + Laneward Workflow v1

Status: approved architecture baseline
Decision date: 2026-08-03
Revised: 2026-08-05 (roadmap order, check ownership, cleanup, documented-versus-enforced boundaries)
Revised: 2026-08-07 (platform support, Git-boundary enforcement moved off the sandbox, Phase 2 probe measured)
Implementation status (2026-08-15): largely implemented. The plan, lane, gate, verification and dashboard surfaces described here exist and are covered by the suite. What remains unbuilt is recorded per decision as a Current state line and in the README's Limitations.

A Current state line in this document set states the date it was measured. A line without a date is a memory, not a measurement.

This document set supersedes an earlier gaps-and-recommendations note, which was retired once it had been fully absorbed here.

Purpose

This document set defines how Claude Code, Laneward, agent workers, project-local records, validation gates, Linux services, and the independent verification layer work together.

The system is intended primarily for:

  • developing new applications;
  • adding features and fixing defects in existing projects;
  • building and maintaining Linux services;
  • creating and operating automations.

A task is not complete merely because code was written or tests passed. It is complete only after the result is installed in its target environment and its real behavior is verified.

Core distinction

Claude Code is the cognitive orchestrator. It interprets the user's goal, performs research, asks questions, produces a plan, makes judgment calls, and integrates results.

Laneward is the deterministic control plane. It records approved plans, creates and schedules lanes, enforces dependencies and ownership, tracks execution, stores evidence, manages approvals, and exposes operational state.

Laneward must not become a second reasoning brain. Claude decides; Laneward records and enforces the approved decision.

End-to-end workflow, by actor Five lanes, one per actor. The user states a task and approves a plan; Claude researches, plans, reviews and commits; Laneward records the plan, dispatches lanes and holds the integration branch; Codex writes; CI gates the release candidate and verifies the deployment. An unapproved plan returns to Claude to revise, and a failed review returns the work to Laneward. User Claude Laneward Codex CI User task Claude research and plan User approval Laneward plan and lanes Codex write workers Claude review and tests Claude commit Integration branch Release candidate gate Deploy and runtime verification Done Revise Approved Pass Fail

Document map

  1. Decisions
  2. Target architecture
  3. Workflow lifecycle
  4. Plan and lane model
  5. Safety and Git policy
  6. Operations and notifications
  7. Skill changes
  8. Future ACOS integration
  9. Implementation roadmap
  10. Platform support

Active, deferred, and excluded systems

System v1 status Role
Claude Code Active Single user entry point and cognitive orchestrator
Laneward Active Central operational control plane
Agent Active Write worker; no Git authority. No default: LANEWARD_AGENT selects a preset (codex, claude)
Claude subagents Active Read-only research and review helpers
Dashboard Active target Operational inspection
Linux notifications Active target Attention and approval alerts
Windows Active target First-class platform alongside Linux (D-022)
macOS Excluded No path designed, documented, or verified
Verification layer Decided, unbuilt Independent clean run, then an advisory reader (D-028)
ACOS Deferred Not the audit layer; refactoring ACOS itself is a separate pilot (D-021)
GPT Pro handoff Disabled No active GPT Pro subscription
OMP Excluded Removed until separately measured and reconsidered

Non-goals

  • Laneward does not replace Claude's reasoning.
  • The agent does not create commits, branches, merges, or pushes.
  • Obsidian is not part of project state management.
  • OMP is not part of this architecture.
  • GPT Pro is not a dependency.
  • ACOS does not block the initial implementation.