Project
Workspace development graph workflow
A dependency-graph workflow for planning, implementing, testing, and human feedback across personal repositories.
- Role
- Author and maintainer
- Markdown
- Python
- Typer
- CLI
On 2026-07-29 I completed the workspace migration to a dependency-graph development workflow. The
agent-memory store (plan: specs/graph-workflow-plan.md) is authoritative for migrated work: new
work starts exclusively with /create-plan, which decomposes an intent into single-agent nodes, and
existing work is inspected with hub graph status and taken with the skill matching each node’s
status. The old specs/ ledgers are frozen — pointers and acceptance contracts, not work tracking.
Migrated tracks include VidSavant hardening and commercialization, PowerCoach v1, the AI roadmap’s
remaining phases, WeightLifting’s human items, PersonalSite’s post-launch phases, MemoryHub’s Phase
6 acceptance, and the workflow itself.
Node lifecycle: /implement → /test issues the verdict (done · rejected · needs-feedback ·
replan) → /replan replaces a node whose plan turned out wrong. Graph verdicts, not agent prose,
signal the next phase.
Human decisions are a two-step loop, so a blocked node never costs the user the whole analysis:
an agent preps it with /prep-feedback (needs-feedback → feedback-ready, writing a decision
payload of context, choices with a recommendation, and/or a playbook), then the user answers in the
Bridge’s Feedback view and the answer routes to the tier that can finalize it — accepted fix →
implement, playbook done → test, replan or custom text → replan, approve → done. /feedback is the
same loop without the console. The payload schema and routing table live in agent-memory/PROTOCOL.md.
Audit layer (added after the initial migration): /validate-intent re-reads a supernode that
reached done against the intent recorded in its own body — technical acceptance is not evidence
that what the user asked for was built — and /audit-run reads a finished orchestration’s ledger,
journal, and evidence and files each harness defect as an ordinary planned node for the next
window. Both extend the graph rather than reporting into prose.
Durable lesson: a summary is a question never answered. There is no human in the loop mid-run, so anything needing the user has to become a node with a decision payload the Bridge can show.