Package Exports

pulseed

Readme

PulSeed

PulSeed is a lifelong personal agent that remembers your goals, watches the world with you, and keeps helping move your life forward.

PulSeed is a lifelong personal agent for goals that take time. Tell it what matters, and it remembers the goal, watches what changes, delegates work to agents, verifies progress, and keeps moving your life forward.

Its current technical wedge is long-running goal orchestration, but that is not the final category. PulSeed is being built toward a lifelong companion agent with durable personal context, presence, proactive dialogue, and safe delegation. See Positioning.

The primary entry point is pulseed. The normal flow is natural language, not a menu of subcommands.

Get Started

PulSeed supports Node.js 22 or 24.

Quick install (macOS / Linux):

curl -fsSL https://raw.githubusercontent.com/my-name-is-yu/PulSeed/main/scripts/install.sh | bash

Quick install (Windows / PowerShell):

irm https://raw.githubusercontent.com/my-name-is-yu/PulSeed/main/scripts/install.ps1 | iex

Then start PulSeed:

pulseed

For pinned-tag installs, fallback npm installs, and installer flags, see Getting Started.

Then describe the goal in natural language:

Increase test coverage to 90%.
Show me the current progress.
Keep this goal moving in the background.

PulSeed will guide provider and adapter setup when needed.

Why PulSeed?

Lifelong personal-agent orchestration for goals that take time
Honest goal negotiation when a target is not realistic as stated
Bounded agent execution with verification around delegated work
Verified playbook memory that turns successful runs into reusable planning hints
Local persistent state under ~/.pulseed/
Multiple runtime surfaces: CLI, chat, TUI, daemon, and cron
Support for OpenAI, Anthropic, Ollama, and adapter-based execution paths

What It Does

DurableLoop keeps a goal moving and decides whether to continue, refine, verify, or stop
AgentLoop handles bounded tool-using work for task execution, chat, and selected runtime phases
Centralized AgentLoop profiles keep task execution isolated by default, narrow chat permissions, and run /review through a dedicated read-only review posture
Dream-backed playbooks can feed verified workflow hints into later task generation without auto-writing executable skills
State, reports, schedules, and local memory live under ~/.pulseed/
Software-level approval and verification gates protect delegated work

Common Surfaces

pulseed for the primary interactive workflow
pulseed tui for the terminal UI
pulseed start and pulseed stop for daemon control
pulseed schedule ... for schedule management
Lower-level commands for scripting, diagnostics, and compatibility

Achievements

PulSeed completed a 30-hour autonomous dogfood run on Kaggle Playground Series S6E4.

Area	Result
Long-running workflow	Ran continuously for about 30 hours, with operator intervention only for deadline-bound submissions.
Final Kaggle result	Finished with `0.97057` balanced accuracy, ranked `1,303 / 4,325` teams, approximately the top `30%`.
Local validation	Improved OOF balanced accuracy from early 0.45-level baselines to a 0.970+ candidate set.
Autonomous exploration	Generated, tested, and compared hypotheses across CatBoost variants, probability adjustment, class weighting, and post-OOF calibration.
Evidence preserved	Kept logs, metrics, OOF predictions, submission candidates, and follow-up engineering issues from the run.

Docs and Community

Start with the public doc map:

For project participation:

read Contributing before opening a pull request
use Issues for bugs and feature proposals
follow the Code of Conduct

Safety Boundary

PulSeed uses approval gates and verification around delegated work. Native agent_loop task execution can use isolated git worktrees, and supported CLI adapters can be wrapped with a Docker terminal backend. These reduce blast radius, but local backends and plugins still run with the user's privileges. See Security.

License

MIT

JSPM

pulseed