session-flow
A workflow for coding agents, with persistent work records and a generated backlog.
- Status
- v3.2.0
- License
- MIT
- Language
- Python
What is session-flow?
Session workflow orchestration for Claude Code: a complete development lifecycle chain, from research through release, with dependency-aware parallelization and collaborative brainstorming at every stage.
Writing
What my Claude Code plugin adds over using the agent with nothing installed, and the cases where it is not worth the overhead.
A Python runtime now manages work records and generates the backlog. The chain has sixteen skills, including a migration path for existing projects.
- Why session-flow is a chainSep 2026
A Claude Code plugin that writes down the instructions I was otherwise retyping every session.
Releases
Task sessions work in isolated source workspaces and hand off exact commits. One owner per repository integrates and releases.
Verify reuses evidence that still applies and runs only the checks that are missing, failing or stale. Simplify becomes an add-on, and a project can name a runbook for delivery.
3.0.0 has breaking changes to the runtime's payloads. The runtime refuses payloads it used to accept, and claim and record-result now move tasks through their lifecycle.
Deferred and cancelled work leaves SEQUENCE.md too. The generated backlog contains only open work, and retired identities stay taken.
Done items leave SEQUENCE.md while their records remain available. The backlog stays shorter as work is finished.
Repair can now survey an older backlog before its namespace or work root exists. The plan includes the setup it needs, with no writes until you approve it.
Review agents now label every finding by confidence instead of quietly dropping the uncertain ones. Two conversational skills join the chain, one agent leaves it, and a backlog can now be shared across repos.
Glossary
A pass that cites a file and line, a command's output or a test name for every claim, and treats missing evidence as failure.
Acceptance tests written before implementation by a different agent, which the implementers must satisfy and may never edit.
Review agents label every finding high, medium or low confidence instead of withholding the uncertain ones, and the skill that called them decides what to act on.
The generated backlog view: one line per open work item, with a stable identity and either a breakdown link or a marker that it needs one.
A unit of work small enough for one agent session to finish: one to five files, clear acceptance criteria, one commit.
The set of task records under one work item, each self-contained enough to hand to an agent on its own.
The arrangement of session-flow's skills so that each stage produces the artifacts the next one consumes.