ecoseek

Alpha checklist

EcoSeek is pre-alpha. This checklist is the honest accounting of where things stand and the gate for the first public alpha. It is meant to be uncomfortable: nothing here is a marketing claim, and “mocked” and “unsafe” appear on purpose.

Last updated: 2026-05-18, after P0 stabilization (agenticSeek#33, agenticplug#74).

What works

• The documentation shell in this repository.
• The high-level architecture (three layers: client, gateway, intelligence) is articulated, consistent across docs, and backed by canonical architecture and 5 ADRs.
• The product modes (DIY / BYOK / Lab-managed) are defined.
• The relationship to upstream AgenticSeek is acknowledged, attributed, and license-aware.
• AgenticPlug gateway is functional — dual-layer auth (GitHub Device Flow → opaque session), role-based access control, scoped sessions, approval workflow for 6 risky capabilities, connector discovery API, persistent SQLite session store, rate limiting. 600+ tests across 26 suites.
• DeepSeek BYOK provider — Fernet-encrypted local keystore (cryptography>=42), OS keychain preferred with encrypted file fallback, fails closed without crypto library. 41 keystore tests.
• EcoCoder local provider — wraps Ollama with model validation and auto-resolution of generic model names to ecocoder.
• EcoCoder cluster provider — routes inference through AgenticPlug to remote clusters via OpenAI-compatible API.
• EcoAgent tool server — 30+ ecological tools exposed via HTTP (/v1/tools, /v1/tools/{name}/execute), discoverable by AgenticPlug connector manifest. Docker-packaged.
• Security hardening — path traversal jail on save_block (realpath + commonpath), remote symlink containment on HPC logs (readlink -f over SSH), safety.py unsafe command list fix, rate limiting, secret redaction in audit logs.
• Landing page — live at ecoseek.org.
• Full-stack threat model — 24 scenarios, 12 assets, 6 actor profiles. See security.md.

What is mocked or partial

• End-to-end DIY demo. Individual components work and are tested, but no single docker-compose up wires them all together yet.
• EcoCoder authoring loop. The runtime surfaces exist but the authoring scaffolding (write agent → run → inspect audit → iterate) is not yet pinned.
• Knowledgebase wiring. Not connected at runtime. Documentation references it as a read-only source.
• Reproducibility hooks. Designed, not implemented. Seeds, dataset pins, and environment captures do not yet land in any artifact.
• Lab-managed mode. Multi-user AgenticPlug sharing is a future target. Treat every install as single-user.

What is unsafe or not production

• No real secrets in this repo. Do not configure EcoSeek with a real API key, OAuth token, or production credential against untested code paths.
• No public exposure. Do not expose AgenticPlug or EcoAgent on a public network without additional hardening. The auth model is tested but not independently audited.
• No data handling guarantees. Any data you give EcoSeek may end up in logs, working directories, or — if you misconfigure a substrate — in cloud calls you did not intend.
• Client sandbox not independently audited. The security review (agenticSeek#28) identified 4 Critical + 4 High findings. P0 fixes landed in PR #33 (save_block jail, safety.py fix), but Python exec() in the code interpreter still runs with full builtins.
• License is undecided. See NOTICE.md. Until that is resolved, do not redistribute.

Minimum alpha demo

The bar for calling something an “alpha” is a single end-to-end path that a reviewer can run and inspect:

1. A reviewer clones the relevant repos and runs EcoSeek in DIY mode with no real secrets.
2. EcoAgent loads a tiny EcoCoder-authored scientific agent.
3. The agent makes at least one call through AgenticPlug to a local model substrate.
4. AgenticPlug emits an audit log entry that names the actor, the requested action, the policy decision, and the timestamp — without leaking any inputs that look like secrets.
5. The reviewer can re-run the agent and get a deterministic-enough result to compare against the previous run.

If any of the five fails, it is not alpha yet.

Before public alpha

Even after the minimum demo works internally, the following must be true before anything is published as a public alpha:

• Final license decision committed to the repository and reflected in NOTICE.md.
• All AgenticSeek-derived code in any EcoSeek component carries its GPLv3 headers and is identifiable.
• A written threat model in security.md. Done — full-stack threat model with 24 scenarios across 5 layers. See security.md.
• A documented BYOK flow exercised with at least one provider. Done — DeepSeek BYOK with Fernet keystore (agenticSeek PRs #23, #33). See deepseek-byok.md.
• An audit log format that a second person, not the author, can read and reason about.
• A “how to report a security issue” contact published in security.md.
• A clear public statement that EcoSeek is not affiliated with or endorsed by AgenticSeek or DeepSeek. Done — in README.md, NOTICE.md, UPSTREAM_CREDITS.md, and the landing page.
• At least one independent reviewer has run the minimum alpha demo and confirmed each step.

Until every box is checked, the project stays pre-alpha, regardless of how good the demo looks.

This site is open source. Improve this page.