Step 1: Roadmap view. High-level initiatives with rolled-up status. Each row expands to show phases and children. Click any row to drill in.
user expands "Vision Tracker" then clicks the Specification phase
Step 2: Specification phase. User clicked "Specification" under Vision Tracker. All children of this phase shown as expandable rows — decisions, specs, tasks. Same uniform pattern. Click any row to expand inline.
user clicks D1
D5: Compose-loop reads tracker as work queueplanned›
Migrate vision-state schema to v2planned›
+ 6 more items...
Step 3: D1 expanded — shows options. Clicking D1 reveals: a Discuss button and artifact links at the top, then options A-D. Each option is expandable (D is open showing pros/cons). Compare options, then Approve one or Discuss the whole decision.
user approves Option D
AROADMAP.md only — agent reads/writes natural language›
BVision Tracker only — queryable, filterable, visual›
CBoth, kept in sync — manual sync will fail›
DTracker primary, markdown generatedrecommended›
Single source of truth in the tracker. Markdown snapshots generated on demand for git diffability. The agent's work queue IS the tracker — self-interest keeps it current.
Pros
Single source of truth
Agent self-maintains
Git-diffable snapshots
Queryable and visual
Cons
Markdown easier to hand-edit
Tracker must be reliable first
Generated markdown may drift
user approves Option D
D2: New fields — priority, effort, planLinkplanned›
D3: New entity typesapproved›
+ 9 more items...
Step 4: D1 approved. User expands D4. D1 collapses and shows "approved". Now D4 is expanded — same pattern: rationale, pros/cons, linked dependencies. The linked item "D8: Markdown export" has a "depends on" edge — D4's typed edges enable D8's export logic.
user clicks "D8: Markdown export" link
Rationale: Every connection between items needs a semantic type. Without typed edges, "this blocks that" and "this informs that" look identical — the graph can't drive scheduling or render meaningful dependency lines.
Pros
Dependency-aware scheduling
Graph renders different edge styles
Enables the informs deliberation pattern
Fixed set keeps complexity bounded
Cons
Fixed types may not cover future needs
More complex than generic connections
Migration required for existing data
Linked items
Product realignment — DAG data model
informs
D8: Markdown export
depends on thisclick to drill in
Implement edge validation
implements
+ 8 more items...
Step 5: Drilled into D8 via dependency link. User clicked "D8: Markdown export" inside D4's expansion. Breadcrumb grows. D8 has its own children, links, and actions. Same pattern at every depth. The "blocked by" link back to D4 shows the relationship.
user clicks Back (or "Board" breadcrumb)
Generate ROADMAP.md from tracker state on demand. Human runs a command (or it auto-generates on commit), producing a git-diffable snapshot. Depends on D4 (typed edges) for rendering dependency info in the export.
Linked items 3
D4: Typed edgesplannedblocked by›
D1: Tracker is primaryapprovedinforms›
Build export CLI commandplannedimplements›
Step 6: Back to roadmap. Breadcrumb jump returns to the top level. Vision Tracker's Specification phase now shows "3 approved" (D1 was just approved). Progress bar updated. The full flow: roadmap → expand initiative → click phase → see decisions with options → approve → follow dependency → drill deeper → jump back.