Repository navigation
Choose the Onyx surface for the thin layer into Expected Parrot #130
Description
Question
Which Onyx surface becomes the thin layer into Expected Parrot?
Owner direction (2026-11-01): Onyx is a thin layer over Expected Parrot's own tools, code, skills and examples; stand those up instead of rebuilding them.
Facts found:
- Expected Parrot ships no MCP server. Its agent interface is the
epCLI (all output JSON;studyscaffolds a study,workflowholds approval gates such as plan-approved,reportvalidates reports,presentshows files), dozens of agent-first CLIs (golden, bower, umriss, zermelo, urteil, helmer, ...), and skill collections (ep-skills,skills) that tell a coding agent how to drive them. - Onyx Craft is that same shape: an OpenCode coding agent in a sandbox with a Python venv, MCP access, and skills pushed at session setup. Onyx can import skills straight from a GitHub repository (
POST /skills/github/import). Normal Onyx chat does not use skills. - Craft is off on agent-dev:
/api/build/sessionsanswers "Onyx Craft is not available". It needsENABLE_CRAFT=trueand a sandbox backend: Docker (the API mounts the Docker socket, root-equivalent on the host) or Kubernetes (a kind cluster). The agent's account has no Docker access on agent-dev.
Options:
- Craft route: enable Craft with the Docker sandbox on agent-dev (or a dedicated host), add
edslto the sandbox venv, importexpectedparrot/ep-skillsandexpectedparrot/skillsfrom GitHub, and let Expected Parrot's own instructions driveep. Retire the custom tool wrapper. Needs the owner to grant the Docker socket on a shared host or provide a dedicated one. - Chat route: keep the chat persona and MCP, but replace the 24 hand-built tools with a thin passthrough to the
epCLI, and load Expected Parrot's skill text into the persona prompt (ep-skills has no licence, so it is fetched at runtime rather than copied). Works on agent-dev today.
Activity
beca-oc commentedon Nov 2, 2026
Goal: Beca in Onyx is a thin layer over Expected Parrot. A person opens Onyx, names a market question, and Beca designs, runs, charts and explains the study using Expected Parrot's own CLI, skills, examples and plots, reliably.
Route (assumed; confirm or overrule on #130): Onyx Craft. Craft runs an OpenCode coding agent in a sandbox with a Python venv and skills, which is the shape Expected Parrot's tools are built for. Do not build new tools. Reuse, in this order: Expected Parrot's
epCLI and EDSL, its skill repos (expectedparrot/ep-skills, expectedparrot/skills), its study scaffold (ep study,ep workflowgates,ep report,plot_style.py,build_standard_plots.py), its agent-first CLIs where a study needs them (golden, bower, umriss, zermelo, urteil), then Onyx's own Craft, skills import, model gateway and egress proxy. Custom code is allowed only as glue with a named gap, and every gap is recorded on the map.Map: #119 ("Investigate a market from the Beca chat with Expected Parrot"). Decisions already made: respondents run on Claude Haiku 4.5 on Bedrock (#120); follow Expected Parrot's practice for audience, study type, plan approval then cost approval, and standard plots (#123); test-model placeholder fix (#122); live-model needs and costs (#121). Always pass
--repo Subconscious-ai/onyxto gh.Steps:
ubuntuuser to thedockergroup or names a dedicated host. Until then, do every step that does not need a running sandbox.ENABLE_CRAFT=true,SANDBOX_BACKEND=docker, per backend/onyx/server/features/build/sandbox/README.md andinstall.sh --include-craft. Proof:GET /api/build/sessionsstops answering "Onyx Craft is not available", and a Craft session starts for the non-admin test account in~/.config/ep-mcp/test-user.env.edsl(the version on agent-dev, 1.0.9 or newer) to the sandbox image'sinitial-requirements.in, rebuild the image, and confirmep --helpruns inside a session.openai_compatibleservice at Onyx's Craft model gateway so respondents run on Bedrock Haiku 4.5 through Onyx. If that fails, use the egress proxy's credential injection for Bedrock. Record which one works. Proof: one keyless test-model run and one live Haiku run of a single question inside a Craft session.POST /api/skills/github/import) for expectedparrot/ep-skills and expectedparrot/skills, at a pinned revision. Do not copy their text into this repository (no licence).Acceptance: three consecutive headless passes of the three investigations and the eleven sample flows through Craft on agent-dev, transcripts and charts saved, no empty answers, no invented tool output, total live spend recorded per run and under the cost each user approved. Report on #119 with links.
Boundaries: never print secrets (Doppler project
beca); never send external messages; never spend beyond the approved estimate; no change to auth, billing or production; merging follows the repository's merge policy, and enabling Craft on a shared host is the owner's call (step 1). Stop and report if a step needs a decision only the owner can make.aviyashchin commentedon Nov 2, 2026
Resolution (owner, 2026-11-01)
Craft route on agent-dev with the Docker sandbox backend. The owner preferred avoiding Docker, but Craft supports only Docker or Kubernetes, Docker is already installed there, and k3s would add more root-level machinery. Expected memory: about 2 GB per active sandbox (Onyx's default cap, freed after an hour idle); agent-dev has 39 GB available. The plan and goal script are in the comment above.