Ease of use & getting started

Why. A first-timer today faces a wall: services to start in the right order, agent CLIs to install and authenticate, sandbox images to build, env vars to guess — and the docs assume you have everything. Worst case is the partial setup: someone with only codex (no claude, no glm) gets failures deep inside a run instead of a clear answer up front. Every hour a new user spends fighting installation is an hour they conclude the kit isn't for them. The kit can't become a factory for other people's projects if only its builders can switch it on.

Where it stands. dev.sh and start_services.sh work for us, on this machine, with all five providers authenticated. There is no installer, no preflight check, no "you have codex only, here's what works" answer, no quickstart that proves the loop on a toy project in ten minutes.

Ideas.

• Doctor first. One command (odin doctor or hk doctor) that checks everything — services, CLIs, auth, sandbox runtime, disk, memory — and prints a table: what works, what's missing, what exactly to run to fix each line. The error deep in a run becomes a red line before any run.
• Providers are optional, gracefully. The kit runs with WHATEVER agent CLIs exist. One provider is enough to start; routing, reviews, and merges pick from what's installed and say so. A feature matrix answers "I only have codex — what do I lose?" honestly.
• Ten-minute quickstart. Install, doctor, then one tiny built-in sample spec that runs a real task end to end — dispatch, sandbox, review, merge — so the new user SEES the loop work before pointing it at their code.
• Assumption audit. Sweep the codebase for hardcoded provider assumptions (claude-only paths, model names in defaults, auth expected on disk) and turn each into a capability check with a plain failure message.

When an idea here becomes real work, it gets its own section below this line (or its own file) with the design, the tasks, and the before and after numbers. The WHY above travels into every task description.

Assumption audit — task #252

Swept odin/src and taskit-backend for hardcoded provider assumptions. Fixed the HIGH blast-radius sites (planning/exec/summarize pre-flight CLI check, doctor base_agent probe, provider-agnostic reflection reviewer defaults). Remaining lower-priority items for a future pass:

• MergeAgentConfig.agent = "claude" (models.py:175) — latent trap if LLM-based merge resolution is implemented; currently deterministic
• AdvisorConfig.agent = "claude", model = "claude-sonnet-5" (models.py:192–193) — advisor disabled by default; needs the same derive-from-available treatment when enabled
• DEFAULT_ADVISOR_MODEL/AGENT constants (advisor.py:34–35) — same
• REFLECTION_PREFERRED_AGENTS = ["gemini", "codex", "claude"] (views.py:117) — ordering only (auto-path picks any available agent), but the list should include all providers for consistent preference
• ALLOWED_FORCED_PROVIDERS = {"gemini"} (forced_provider.py:12) — intentional trial restriction; revisit when the trial generalizes