00 / HERO
How I ship with agents.
I own intent and every gate; Hermes runs the rest.
Four phases, human gates
Each phase ends at a gate a human or a validator owns.
01 / PHASE_TRACK
OWNER: human + agent
Frame the problem, boundaries and vocabulary before any code exists.
- Interview the plan until open questions close
- Record vocabulary and constraints
- Write the approved Shaping handoff
SKILLS ginflow · grill-with-docs
- docs/plans/2026-10-29-workflow-page-design.md
GATE · gin approves Shaping handoff
Who owns each phase
02 / OWNERSHIP
System boundary
03 / BOUNDARY · lit = in play during Discovery
HUMAN gin
WORKSPACE herdr
ORCHESTRATION Hermes Agent · profiles gintary, ginb · Kanban · cron · memory
HARNESS gin-harness · ginflow skill, ginflow-gate, routing core, skills
WORKERS Hermes Kanban workers · one per card · isolated worktree
CONTEXT CodeGraph, drawio MCP, Open Design, Notion, NotebookLM, 9router
TARGET repo + GitHub
Gin harness
04 / HARNESS
lit = selected phase · amber = rejected, back to Execution
WHAT IT IS
gin-harness is the rule layer around Hermes agents: the ginflow skill, ginflow-gate and a routing core. It decides what an agent may do and what proof it must leave.
HOW THE WORKFLOW RUNS
Work moves Discovery, Shaping, Execution, Decision. Agents only write code inside a Kanban card, in an isolated worktree, inside the card scope. Gin approves between phases.
WHAT IT SOLVES
- Agents invent requirements: work needs a spec and card first
- Scope creep: a card fixes the scope, one active card per workspace
- Fake done: no completion without fresh lint and build evidence
- Stale docs: ginflow-gate blocks when linked artifacts drift
- Risky actions: agents never push; gin merges by hand
Questions people ask
05 / FAQ
Why not just prompt the agent and review the result?
A raw prompt does not fix the workspace, requirements, risk or proof. The agent fills those gaps with guesses. The phases and the card settle them before code is touched.
What stops an agent from saying it is done when it is not?
ginflow-gate. Workers finish with kanban_request_review, and the gate checks card fields, verification evidence, linked artifacts, baseline commit and drift. It checks again at kanban_complete. Missing evidence blocks the transition.
Can an agent push or merge?
No. The agent profile forbids remote push. I merge by hand after marking approved_for_mr in Decision.
What happens when a card gets blocked?
A recoverable block is retried up to 3 times. Config, persist and human-input blocks are never auto-resolved; they stay for me.
Does the harness live inside my project?
No. The harness stays in its own repo and the Hermes profile. A project only keeps a short local rules file and one canonical verification command.
What if the plan changes mid-way?
Decision can send work back: enhance and resume go to Execution, reshape goes to Shaping, and a foundational change goes to Discovery. Evidence is kept.
Is it finished?
No, it is early and contracts may change. A Decision review TUI, gateway notifications and merge request publishing are not built yet.