adalloul0928/cyrus-workflows

Set up or repair self-hosted Cyrus end-to-end, including prerequisites, Claude authentication, webhook exposure, integrations, repository registration, and launch. Use for a new Cyrus installation or when one setup stage is incomplete.

Append a structured learning (bug, error, friction, surprise, or improvement idea) to the PUMPD "Agent Learnings" log in Obsidian. Use mid-workflow the moment something is wrong, ambiguous, slow, or could be better in a skill/agent/pipeline — or when the user says "log that / note that learning / capture this for the retro". Quick append only; /pumpd-retro reviews these later and turns them into improvements.

Turn an APPROVED PUMPD plan into Linear for HANDS-OFF Cyrus execution. Creates the Linear Project (the "epic"), attaches the PRD as a Linear Document, and creates ONE orchestrator parent issue (labeled orchestrator + graphite + pumpd-agent) whose body carries the approved decomposition. You delegate that parent ONCE; Cyrus self-creates the stacked sub-issues and builds the PR stack overnight (auto-Done cascade). Use after /pumpd-plan once the plan passed /pumpd-review. Hard-to-undo — gate on approval first.

Author a consistent PUMPD feature plan from a /pumpd-research note. Produces a standard plan note in the Obsidian vault using the canonical 12-section format (summary, problem, goals/non-goals, decisions, approach, FILE STRUCTURE, data model, DECOMPOSITION, risks, EARS acceptance criteria, milestones, open questions). Decides file structure BEFORE cutting tasks. Use after /pumpd-research and before /pumpd-decompose. The plan must be approved before decomposing to Linear.

Investigate and scope a PUMPD feature idea BEFORE planning. Explores the mobile app (apps/mobile, React Native/Expo) and the Supabase backend (apps/backend), the web (libraries/SDKs/docs), and the Obsidian vault; reads the in-repo constitution; asks clarifying questions one at a time; proposes 2-3 approaches with trade-offs; then writes a research note to the vault. Use when the user wants to think through a new feature or large piece of work (e.g. "add Sentry to the mobile app", "redo onboarding"). This is the GATE before /pumpd-plan — do not write a plan or code until research is done and an approach is chosen.

Review accumulated learnings (bugs, errors, friction, surprises logged during the pumpd-* skills, agents, and Cyrus runs) and propose concrete improvements — edit a skill, add a skill/agent, add a durable rule, or fix a recurring error — gated on human approval before any edit. Reads the append-only "Agent Learnings" log in Obsidian; optionally enriches from recent session transcripts. Use for "run the retro / review learnings / what should we fix / improve the workflow", or on a schedule. This feedback loop remains human-triggered until scheduled automation is deliberately enabled.

Adversarially review a PUMPD plan BEFORE it goes to Linear. Spawns a red-team panel (assumptions · failure modes · sequencing & the Graphite stack · security/privacy) plus a completeness/consistency critic and a cross-model Codex (gpt-5.5) lens — each told to assume the plan is flawed and find what breaks — then synthesizes a filtered punch-list of MATERIAL holes and gates handoff to /pumpd-decompose. Use after /pumpd-plan, or standalone on any plan note.