Keep the project workflow when the agent changes

Installing a skill does not make every agent read it on every task. A new agent or conversation can repeat old research or edit the wrong checkout.

Initialization gives the project an explicit starting policy: where facts live, which skills to read, how to isolate changes and how to verify the result. It helps agents share knowledge; it cannot guarantee that a model obeys every instruction or that every stored fact is correct.

Project harness creates a shared guide and managed pointer blocks in AGENTS.md, CLAUDE.md and GEMINI.md. Existing content is preserved. Leave the harness option off to avoid changing those files.

Memory offers three modes. Read and maintain each task instructs agents to retrieve relevant context and verify, correct and record facts. Ask before use instructs the agent to seek your answer first. No connection adds no automatic Memory use through this setup and does not disable Global Memory for other projects.

Worktree workflow uses .worktree through the same GUI and CLI creation policy. Browser integration connects the chosen browser skill when needed; it does not install an extension, grant permissions or authorize external messages.

New setup preselects five skills: CLI, Memory, worktree workflow, browser and computer-use. Selected skills are copied into the project. Computer-use can be deselected in the skill list, and existing saved choices do not silently receive the new default.

Computer use connects the installed VibeLink Windows desktop provider for tasks that structured CLI or browser operations cannot express. Startup guidance requires runtime and provider checks first. Selection alone does not control an app, grant permissions or install another automation provider.

Orchestration is an optional skill for supervising worker assignments, task dependencies, results and merge decisions. It is not preselected and is used when supervision is requested. Selecting it alone does not launch a team; role creation and ChatGPT connection likewise require the corresponding task request.

.vibelink/initialize.json records your decision, options, ownership information and setup status. WORKFLOW.md and BOOTSTRAP.md in that directory hold the shared and startup instructions. .vibelink/skills/<skill>/SKILL.md contains selected skill copies. .vibelink/workflow-log.json holds a bounded local activity log.

In Git projects, owned local setup files and the selected .worktree location are added to the private .git/info/exclude file. Existing .gitignore and unrelated .vibelink configuration stay intact. Native instruction pointer changes remain reviewable in Git.

No account settings or global agent configuration is changed. If you edit an owned block or generated file, setup reports the difference instead of overwriting it. Reapply Workflow settings when you want project skill copies updated from the current bundle.

Unmodified built-in Claude Code, Codex and OMP launch commands receive the startup instruction as a literal initial argument. Supported task, handoff and automation paths also connect the workflow. Startup prose is never pasted into a general-purpose shell.

Other agents and customized launch commands depend on their own project-instruction discovery. Check whether they read the generated pointer and selected skills. Custom commands, resume arguments and SSH commands are not rewritten, and already-running conversations are not retroactively changed.

Agents launched outside VibeLink use their own instruction-file discovery. Direct Memory CLI use requires the VibeLink daemon. Explicit workspace selection takes precedence; otherwise the CLI identifies the project from its current working directory. A remote host’s path is not initialized as a local PC folder.

Start with memory context for a small relevant packet: by default up to six items, about 800 estimated tokens and 4KiB of serialized context. Fetch needed originals separately instead of injecting the entire store into every conversation.

Age alone does not make a fact wrong. review-due finds items worth checking; update keeps the original ID and revision history. A confirmed outdated fact can be marked stale or connected to an active replacement. Default recall includes active facts rather than proposals, stale entries or superseded records.

Agents should record evidence and the reason for a correction. Competing design decisions may need your judgment. This setup connects that maintenance workflow; it does not silently invent updates or certify freshness.

Global Memory lives in the user home. Project Memory lives under the Git common directory or a plain folder’s .vibelink/memory. Linked worktrees share the project store. Initialization does not enable cloud synchronization. Migrating an older store or importing another agent’s records uses a separate preview and confirmation.

The + button and CLI use the same native worktree creation and registry path, including opted-in .worktree storage. An optional initial prompt requires an unmodified built-in Claude, Codex or OMP profile and is passed as an agent launch argument; it is never automatically typed into a plain shell. Registered tasks retain checkout, branch, start-commit and worktree identity. An agent can continue in a checkout it already owns for the task instead of creating another.

Creating a worktree does not verify or merge its code. The agent works from an exact committed base, runs relevant checks and commits. If the target branch advances, it compares changes and updates its own checkout before a safe local fast-forward merge. Other people’s edits are preserved; push and deployment permission are not implied.

Open Workspace settings → Workspace log to see creation stages, errors, startup instructions and Git reflog merge observations. Git observations are separate from VibeLink events. A prompt delivery is not labeled as task completion or merge success. Plain folders have no Git merge history.

Both Yes and No are saved choices. Restarting the app or opening another linked worktree will not show the automatic setup prompt again. Closing without a choice postpones the prompt for this app session; a later restart can offer it again. Existing users with no decision record receive the same offer.

A partial setup failure preserves the decision. Open Workflow settings to inspect and retry it manually. A damaged decision record needs review because the choice cannot be established. Optional setup failures do not prevent an ordinary terminal from opening.

Choosing No later disables owned guides and removes unchanged managed pointer blocks. User edits are not silently deleted. Existing memories, worktrees and conversations are kept. Start a new conversation to use the changed policy.

Create or open a workspace for your project folder. Review the options and choose Yes or No. Start a new agent after setup, then inspect Workspace Log from Workspace settings.

A child project can inherit an initialized parent workspace’s policy. Its own saved decision takes precedence. Linked worktrees resolve the original project’s setup. The dialog shows the effective policy location.