Project continuity control surface for long-term AI engineering.
belong to the system Level 1 Product · Project & Agent Workflow · Windows local portable desktop application · Current product version v0.0.20 · Has been used in three real projects of AI Project Harness, Local Chat and personal website (see Engineering chain)·Latest verification: 2026-09-17
01 / PRODUCT PROBLEM
Long-term projects span multiple tasks, windows, and agents.
Harness takes the project truth out of the chat. Project Steward reads the structured Project HEAD, CTO and Review use their own limited control context, and the execution agent only takes the Mission Packet related to the current task.
02 / PRODUCT MODEL
Project ├── Objective + constraints ← Current goals and long-term boundaries ├── Section + Task ← Workflow and Acceptable Units ├── Decision + Blocker ← Finalized content and real blocking points ├── Checkpoint + Project HEAD ← Currently accepted project version └── Evidence ← File, diff, commit and running evidence Session ├── CTO ← Light project deployment ├── Review ← Independent review context └── Task run ← New session or bounded temperature session Storage ├── harness-state.json ← Atomic state snapshot ├── harness-state.jsonl ← Replayable event ledger ├── PROJECT-LEDGER.md ← Dynamic Project White Paper └── content-addressed archive ← Large transcript lossless archive
Task, Run, Session and Agent are four different objects.
03 / WORKFLOW
v0.0.20 has connected the desktop control plane, Codex Project and persistent state into an auditable vertical closed loop:
ADual entrance access
You can create a new blank project or pull it from the Codex official Project index.
BControl Session
Each project establishes two long-term control sessions, CTO and Review, and uses new sessions or bounded temperature sessions to perform tasks.
CMission Packet
The execution agent only receives the current task, acceptance conditions, relevant decisions, blocking and evidence pointers.
DCandidate review and promotion
The worker writes back candidate results with evidence.
The core principle is "lossless archiving, evidence-based compression, and on-demand checkout": the complete transcript is saved by content addressing, and the compression context only carries the decisions, constraints, and evidence pointers required for the current task.
04 / CURRENT BUILD
v0.0.20 has formed a runnable Windows x64 local desktop MVP, and verified project access, task distribution, auditing and evidence write-back in three real projects: AI Project Harness, Local Chat and personal website.
AReal Codex Project control surface
Create or locate the project through the official app-server Project API, inject the Project ID in thread/start, and establish readable CTO, Review and task dialogues; do not rewrite Desktop's own legacy global status file.
BProject Ledger and atomic persistence
Project status, running status, candidate results, review and Git snapshots are written to atomic JSON and replayable JSONL, and PROJECT-LEDGER.md is automatically generated; large transcripts are archived with SHA-256 content addressing.
CSection, tasks and review
Master control, reusable and disposable sections have a clear life cycle; task creation only enters ready and is dispatched after the user clicks.
DCross-Agent adaptation boundaries
Codex is connected through the background app-server; Claude Code and Hermes follow the Skill/MCP path actively called by the user.
05 / CURRENT LIMITS
AComplete import of old projects
The read-only positioning and import Agent turn of the old project have passed the acceptance; the next step is to complete the import candidate acceptance, idempotent backfill, and check the project sidebar and control thread after completely exiting the Codex Desktop.
BContinuous workflow and platform extensions
Complete the Delta Packet of the same workflow and session on a real project, and verify the Claude/Hermes user initiative relay; then evaluate the SQLite/CAS optimization and macOS adapter.
System/Module