beats-pm-kit

Beats PM Kit (the Beats Agentic PM Harness) is an OfficeBeats project packaging a local-first, cross-runtime harness for evidence-backed product-management workflows. It archives scattered product context — meetings, chat threads, email, documents and notes — as plain Markdown locally, makes it searchable, and triangulates it into workstreams and tasks that an AI agent can act on. The README states that nothing is uploaded by the kit itself. Structure and concepts The kit uses a numbered vault taxonomy: 0. Incoming, 1. Company, 2. Products, 3. Meetings, 4. People, 5. Trackers, 6. Resources, 6. SOPs, 7. Partners, 8. Clients, plus .agent/ for canonical agents, rules, skills, templates and workflows, and system/ for scripts, tests and docs. Canonical task state lives in 5. Trackers/tasks/ as human-readable Markdown notes with stable IDs, evidence links, progress and decisions; TASK_MASTER.md and workstream notes are described as generated navigation over those files rather than a second source of truth. Single source of truth .agent/command-registry.json owns routing, aliases, execution profiles and runtime policy; .agent/workflows/ owns workflow behavior; .agent/skills/ owns reusable PM methods. Root adapter files (AGENTS.md, CLAUDE.md, GEMINI.md, CODEX_COMMANDS.md), .omp/config.yml and tool ignore files are generated views produced by system/scripts/sync_cli_adapters.py — editing .agent/ and regenerating is the intended path, since hand-edited generated files are overwritten by design. Commands and context Each slash command resolves through the registry to one workflow file and one execution profile (Fast, Balanced or Deep), and carries a per-runtime promotion level: promoted commands ship as native skills, guarded commands require confirmation, and the rest are dispatch-only. The documented execution loop is route, load bounded context, act with tools, checkpoint, verify, persist artifact; initial retrieval is capped and raw evidence remains hash-addressable locally, with compacted context traceable back to source files. Supported runtimes Entrypoints are listed for omp, Claude Code, OpenAI Codex CLI, Google Antigravity, Gemini CLI, GitHub Copilot, and an optional Obsidian UI over the same vault. The README says behavior is selected by positively detected capabilities rather than a provider hierarchy, model defaults are inherited from the runtime, and unknown capabilities fail closed. Getting started and workflows Setup is a git clone plus ./install.sh, a thin wrapper around python3 system/scripts/bootstrap.py --agent --non-interactive. Bootstrap verifies the repo, runs an upgrade-compatibility gate on existing workspaces, creates ignored local workspace folders, seeds templates, syncs runtime adapters, installs git hooks when possible, and runs privacy and adapter health checks. Users then open the folder in a runtime and run /start or /help. Workflows are grouped into intake and evidence (/paste, /find, /memory, /context, /chat and messaging integrations), meetings (/meet, /transcript, /prep, /boss), planning and strategy (/plan, /create, /deck, /discover, /interview, /intel, /prioritize, /sprint, /retro), tracking (/track, /day, /week, /handoff), comms and docs (/sop, /office-cli, /review, /challenge, /improve-plan), quality and maintenance (/accuracy, /regression, /build, /vibe, /maintain, /update, /vacuum, /archive) and setup/orchestration (/start, /help, /obsidian, /pack, /team, /fan-out). Natural language input is also supported. Local utilities Repeatable work is delegated to stdlib-only Python scripts in system/scripts/, including vault_query.py for bounded read-only queries over tasks, labels and quotes; context_router.py for indexed full-text retrieval across local evidence folders; task_store.py and task_intake_fast.py for canonical task-note writes and fast intake; transcript_pipeline.py for preparing, validating and processing meeting transcripts; pm_decision_router.py for classifying messy PM input before execution; obsidian_bridge.py for using the kit folder as an Obsidian vault; pack_manager.py for enabling dormant optional capabilities such as Trello; upgrade_compat.py for preflight and reversible migration of legacy configurations; and privacy_guard.py / adapter_guard.py for blocking private-content leakage and generated-adapter drift. Privacy and testing Privacy posture is local-first: personal vault folders are gitignored and tracked only through skeleton .gitkeep files, with local runtime state in an ignored .beats/. Guard scripts (privacy_guard.py --tree, adapter_guard.py --mode check) block private content, personal paths, token-like strings, transcripts and adapter drift from shared commits. WATCHDOG.md defines advisor rules such as no fabricated quotes, IDs or dates, explicit user approval for outbound sends, and keeping task state in canonical trackers. The README adds the caveat that the kit never syncs files to a cloud service, but the chosen AI runtime processes prompts and tool outputs under that provider's own settings. Tests live in system/tests/ in pytest style and cover the registry, adapters, task store, transcript pipeline, privacy and adapter guards, upgrade compatibility and vault invariants. Credits Beats PM Kit is built from daily product work; the action-first output policy is described as an independently worded approach inspired by the MIT-licensed i-have-adhd project, with no upstream plugin or hook code bundled.