Skip to content

Latest commit

 

History

History
56 lines (41 loc) · 3.29 KB

architecture.md

File metadata and controls

56 lines (41 loc) · 3.29 KB

Architecture

request → dev-cycle (router)
            ├─ Feature → dev-feature-cycle
            │    intake (gap round) → specify → [design] → [tasks] → execute → verify-feature
            │    → mr-guided-review → qa-guided-review → [fix-reviews] → ship
            └─ Bug → dev-bug-cycle
                 intake (gap round) → reproduce → debug → fix-bug → verify-bug
                 → mr-guided-review → qa-guided-review → [fix-reviews] → ship

direct  → dev-feature-agent | dev-bug-agent
            same steps, decisions instead of gates, stop before ship

dev-cycle does not do heavy work and does not enter agent mode. Each orchestrator loads only the current skill. The main session owns gates, hyperlinks, and state. Fresh workers own execute/verify.

Artifacts live in the target project, not this repo:

  • .dev/STATE.md — flow, mode (manual when absent), last gate, branch, next step
  • .dev/features/<slug>/ — intake (## Decisions, visual contract notes), spec, design, tasks, validation, reviews/round-NNN.md, reviews/readiness.md
  • .dev/bugs/<slug>/ — intake, reproduction, investigation, hypothesis, fix-plan, validation, reviews/, optional postmortem

Canonical reviews stay in chat and include a Completeness profile. round-NNN.md only after confirmation (manual) or after the agent recorded the evidence (agent mode). Inside a flow, each review also rewrites its verdict, open-row count, and reviewed commit in readiness.md; dev-ship refuses a receipt that is not approved or is older than the code.

Gates

Manual mode: Gate briefing in chat (confirming, artifact path, summary, confirm vs reject), then the one-line question prompt. If the host has a structured question tool, use it (allow_multiple: false); otherwise numbered labels after the briefing. A finished manual step ends with Next skill: \dev-...``.

Agent mode: one intake gap round, then decide from ## Decisions and surrounding evidence. Stop before dev-ship. Gate S stays manual.

id prompt
gate-flow Gate 0 — Confirm flow
gate-feature-context Gate F-A — Confirm context
gate-feature-spec Gate F-B — Confirm requirements
gate-feature-design Gate F-C — Confirm design
gate-feature-tasks Gate F-D — Confirm execution
gate-feature-debug Gate F-X — Confirm task diagnosis
gate-bug-context Gate B-A — Confirm symptom and severity
gate-bug-reproduction Gate B-B — Confirm reproduction
gate-bug-hypothesis Gate B-C — Confirm root cause
gate-bug-fix Gate B-D — Confirm minimal patch
gate-ship Gate S — Confirm remote actions

QA keeps its own Gates A/B/C. The Gate briefing still applies.

UI work uses the visual contract in dev-qa-guided-review (visual.sizing FIXED, absolute position, complex fills, effects, stroke, opacity, text). A spec-sheet is not a screen.

Permissions

No-escalation: the current skill does not inherit the next skill’s writes.

  • Read-only MCP: router, orchestrators, intakes, specify/design/tasks, reproduce, debug, verify, reviews.
  • Working tree: dev-execute, dev-fix-bug, dev-fix-reviews.
  • Remote: only dev-ship after Gate S (gitlab_save_merge_request; Jira if selected).