fulcrum
By moabualruz
Fulcrum Agent OS — Codex variant (MCP filtered mode + 33 skills + openai.yaml sidecars + hooks for UserPromptSubmit / PermissionRequest / SessionStart / PostToolUse / Stop).
Component Overview
Install
npx claudepluginhub moabualruz/fulcrum --plugin fulcrumComponent Details
Skills (33)
Call block_agent_run instead of guessing when you cannot proceed. Applies on missing info, failing dependency, ambiguous requirements, or permission denial.
As chief_of_staff, end every response with structured Status / Work Completed / Next Steps / Risks block. Every CoS turn, no exceptions.
Finalize agent run with meaningful summary + changed-file list. When work finishes, handoff, or stopping a run you started.
Generate structured implementation plan from PRD or issue.
Concise status summary of workspace activity for standup.
Structured debugging of failing test or unexpected behavior.
Create + assign task to another agent role.
Escalate blocked run to chief_of_staff with clear reason.
Emit heartbeat_agent_run every ~30s during any run >60s. Applies to long script runs, large-file edits, multi-step analysis.
Regular heartbeats keep long-running agent runs alive.
As integration_worker, verify review + test artifacts exist before processMergeQueue. Every merge attempt.
Only chief_of_staff may call invoke_team. Non-CoS roles escalate instead of attempting.
Discover available agent roles + capabilities before assigning work.
Compact + consolidate session memories before context window fills.
Verify action permitted by current workspace policy before executing.
Query Fulcrum memory before writing new code, docs, or architectural decisions. Applies before producing novel output on a topic project may have prior context on.
Structured code review of pull request or diff.
Run multi-step repeatable processes as registered workflows, not ad-hoc. Applies to named processes like grill-me, write-a-prd, prd-to-plan, prd-to-issues.
Execute named Fulcrum workflow instead of improvising a process.
Search agent memory before starting work — surface prior knowledge.
Never include credentials, API keys, tokens, passwords in tool inputs or memory writes. Every tool call accepting free-form text.
Complete or hand off Fulcrum agent run at end of session.
Initialize Fulcrum agent run at start of every working session.
Start new agent run for specific role + task.
Register agent run before touching code. Applies before first Write/Edit/MultiEdit/Bash in session, or when targeted by team invocation.
When + how to use Fulcrum task lifecycle (create_task / update_task).
Invoke team template to orchestrate multi-agent workstream.
Check status of running team instance.
Fetch workspace status at session start — running agents, WIP budget, blocked runs. Before starting new run or responding to user request.
Provision git worktree for isolated branch-based development.
Merge completed worktree branch back to base branch.
Record architectural or implementation decision to memory.
Persist memory after any task involving decision, trade-off, or surprising finding. Applies after complete_agent_run when work produced durable knowledge.
MCP Servers (1)
README
Similar Plugins
claude-mem
Persistent memory system for Claude Code - seamlessly preserve context across sessions