Agent Continuity

Privacy-first continuity notes for long-running AI coding sessions.

Agent Continuity is a small agent skill for projects that span more than one session. It gives an AI coding agent a lightweight way to initialize, resume, and close out work from project-local notes without turning those notes into a private memory dump.

The goal is simple: keep enough state to continue the work, and nothing more.

Why This Exists

Long-running agent work usually fails in one of two ways:

• The next session has no context and repeats discovery work.
• The project grows a "memory" folder full of private paths, raw logs, chat fragments, and stale personal notes.

Agent Continuity uses four short files instead:

• docs/START_HERE.md for the stable entry point
• docs/CURRENT_STATE.md for the current snapshot
• docs/PROGRESS.md for append-only work chunks
• docs/DECISIONS.md for durable project decisions

What It Creates

When starting a new long-running project, the skill asks what work should be tracked and creates a compact docs/ continuity set:

docs/
  START_HERE.md
  CURRENT_STATE.md
  PROGRESS.md
  DECISIONS.md

Install

Copy the skill folder into your Codex skills directory.

Copy-Item -Recurse .\skills\agent-continuity "$env:USERPROFILE\.codex\skills\agent-continuity"

On macOS or Linux:

cp -R ./skills/agent-continuity ~/.codex/skills/agent-continuity

Quick Start

Set up continuity notes for a project:

Use $agent-continuity to set up continuity notes for this project.

Resume existing work:

Use $agent-continuity to resume this project.

Close out a session:

Use $agent-continuity to close out this session and leave the next actions clear.

Workflow

flowchart LR
    A["Initialize or resume"] --> B["Identify project folder"]
    B --> C{"Continuity notes exist?"}
    C -- "No" --> D["Ask what work to track"]
    D --> E["Create docs/ notes"]
    C -- "Yes" --> F["Read existing notes"]
    E --> G["Do scoped work"]
    F --> G
    G --> H{"State changed?"}
    H -- "No" --> I["Skip note updates"]
    H -- "Yes" --> J["Update snapshot"]
    J --> K["Append progress"]
    K --> L["Record durable decisions"]

Loading

Privacy Model

Continuity notes are project state, not personal memory.

Store:

• current goal
• current implementation state
• files or areas being changed
• known blockers
• next actions
• durable decisions and why they matter

Do not store:

• API keys, tokens, passwords, private keys, or credentials
• .env values
• private URLs or internal dashboards
• client names, user records, customer data, or personal identifiers
• absolute local home paths
• raw chat transcripts
• raw logs that may contain secrets
• unreleased business metrics or private project history

See docs/privacy-model.md for the full model.

Repository Layout

skills/agent-continuity/
  SKILL.md
  agents/openai.yaml
  references/state-notes.md
examples/safe-project/docs/
docs/privacy-model.md
docs/comparison.md
scripts/privacy-scan.py

The installable skill stays small. Public documentation, examples, and validation scripts live at the repository level.

For positioning against other approaches, see docs/comparison.md.

Privacy Scan

Run the scanner before publishing or tagging a release.

python scripts/privacy-scan.py .

Add project-specific terms when preparing a public release from a private workspace. Use environment variables so the example command does not add the private terms to the repository itself.

python scripts/privacy-scan.py . --term "$PRIVATE_PROJECT_TERM" --term "$PRIVATE_CLIENT_TERM"

On PowerShell:

python scripts/privacy-scan.py . --term $env:PRIVATE_PROJECT_TERM --term $env:PRIVATE_CLIENT_TERM

License

MIT