01
The graph holds memory
Plans, specs, decisions, tasks, and epics are linked Markdown. The context stays readable without GitDocket and moves with the code.
Project memory for humans and coding agents
Give your agent the whole project, not just the next prompt.
GitDocket keeps plans, decisions, documentation, and work in one linked Markdown graph. Your agent can understand what matters, carry an outcome through completion, and leave the project easier for the next person—or agent—to pick up.
Local-first · plain Markdown · reviewed through Git
The context gap
Agents can write code. Projects are harder.
The hard part is knowing why the work exists, what depends on what, which decisions still apply, and what “finished” means. That context is usually scattered across a tracker, docs, chat, and someone’s head.
- 01 Return to a project without reconstructing it from scratch.
- 02 Delegate an outcome without giving up control of the process.
- 03 Catch stale work and documentation drift before they become archaeology.
A concrete mental model
Memory, facts, judgment, and proof.
GitDocket adds a small project-operating layer to the repository you already have. Each part has one job.
02
The engine derives facts
Readiness, status transitions, search, indexes, and activity come from deterministic commands—not an agent’s best guess.
03
Workflows guide judgment
Versioned procedures tell an agent how to pick up, supervise, verify, groom, and close work without hiding the important choices.
04
Git keeps the proof
Commits link implementation to tasks. Outcomes and reconciled docs make the repository the durable record of what actually happened.
Talk at the right altitude
Commands for facts. Requests for outcomes.
Deterministic CLI
docket overview
docket ready
docket search "release gate"
docket lintUse narrow commands when you need a trustworthy answer or a precise state change. JSON output gives agents the same facts.
Agent workflows
“Run the launch epic to completion.”
“Groom the backlog.”
“Refresh the state of play.”
Ask for the real outcome. The checked-in workflow supplies the procedure, guardrails, verification, and stopping condition.
More than task tracking
The project can explain how to operate itself.
Epic supervision
Hand over the outcome, not every next step.
Ask an agent to run a named epic through completion. It reads the dependency graph, selects only ready work, isolates parallel changes when that is safe, and verifies each result before moving on.
It stops only when the epic is proven complete or one concrete blocker needs you—and returns a single receipt of the work.
- 1Read the graphAcceptance criteria, dependencies, current Git state
- 2Run ready workParallel only when isolation and write boundaries are safe
- 3Verify and integrateOne accepted result at a time
- 4Close with evidenceOutcomes, reconciled docs, commits, final receipt
Backlog grooming
Find work that no longer tells the truth.
Audit stalled tasks, broken dependencies, incomplete outcomes, and inconsistent epic links. GitDocket proposes the cleanup first; mechanical fixes can be applied without silently reprioritizing or closing work.
Result A backlog you can trust before asking what is next.
Documentation freshness
Ask what recent work made untrue.
Sweep commits since the last review, inspect high-risk changes that bypassed a tracked task, and rotate through older reference docs. The watermark makes the next review bounded instead of endless.
Result Documentation that stays connected to reality.
One project, two useful views
Readable as files. Useful as a system.
Humans can browse the project locally. Agents can consume the same state through structured commands. Both work from the Markdown and Git history already in the repository.
docket init --agent codex
docket overview
docket serve
Practical trust
Automation you can inspect and interrupt.
No second source of truth
Markdown and Git stay authoritative. The local cache can be deleted and rebuilt.
No hidden project policy
Agent procedures live in the repository, where the team can read, review, and version them.
No fake certainty
Facts come from the engine. Judgment stays explicit. Blockers are reported instead of worked around invisibly.
Public preview
Make the repository easier to return to—and safer to hand off.
Start with a project you already have. Initialization is additive and leaves existing Markdown untouched until you review it.