Implementator

Daily use

A normal day with Implementator

Once a project is connected, most days come down to a few sentences. Your agent does not remember the framework between sessions, so begin each one by naming it.

One change, start to finish Plan, code, tests, review, manual test, pull request, close.
  1. Start a session

    Use implementator. Where did we leave off?

    The agent loads the rules and shows every active workstream: its status, the newest journal entry, how old it is, and any pull request. If you added the pointer file during setup, a session opened in the notes repo finds its way there alone.

  2. Start new work

    Use implementator. Add a --version flag to tally that prints the version from package.json.

    Every change opens its own workstream and goes through the loop, small ones too. Asking it to skip the plan, or waiving a gate yourself, will not work. That is the point of the pack.

  3. Or start from a ticket

    Start work on TAL-1.

    If your agent has your tracker's connector, it reads the ticket and saves it unchanged as the brief, then writes the plan from it. Without a connector it asks you to paste the ticket. It never fills one in from the id.

    For Jira, add Atlassian's MCP server and sign in. You can grant read access only, which is all the pack needs:

    claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp
    codex mcp add atlassian --url https://mcp.atlassian.com/v2/mcp

    We tested this on a free Jira site in Claude Code. Each call uses Atlassian's Rovo credits.

    Work from a Jira ticket The ticket is read, saved unchanged, and planned from.
  4. Come back later

    The journal is the memory. A new session reads it, finds the last plan and test record, and carries on. Short work and long work are handled the same way.

  5. Run several sessions at once

    One workstream takes one agent at a time. A session claims its workstream before working on it, and a second session gets a clear refusal instead of quietly sharing it. A claim from a crashed session expires after four hours. Different workstreams can run side by side.

    With features.git.worktree on, each workstream gets its own git worktree, so parallel changes to one repo stay apart.

What else you can ask for

Grill me on this plan.

Rounds of questions, each with a recommended answer, until nothing is assumed. The agent looks up facts itself and asks you for decisions.

Diagnose this flaky test.

It builds a check that fails on this bug first, reproduces and narrows it, tests one explanation at a time, and fixes it behind a regression test.

Mock up a few options for this page.

Temporary visual edits in the running app, a screenshot of each, then everything reverted. You pick one.

Add before and after screenshots to the PR.

Shots of the changed surface, at the viewports you set.

Check tally/cli is ready.

Runs preflight on its own.

Refresh the skills for tally/cli.

Rescans the repo and shows you what moved.

From a mockup to shipped code Options on the running page, a pick, then the whole loop.

Looking after the notes