v0.1.4 — ultra prompt hardening + DX sweep + OE_OPENAI_MODEL override

0.1.4 — 2026-05-28

Patch release bundling 4 user-facing improvements: 2 DX fixes from the v0.1.3 post-release sweep, the OE_OPENAI_MODEL / OE_ANTHROPIC_MODEL override unblocking self-hosted users, and a substantial synthesizer-prompt hardening for oe ultra. All caught from real-API testing.

Added

• OE_OPENAI_MODEL and OE_ANTHROPIC_MODEL env vars — defaultModelFor now honors these as overrides. Previously oe ultra and oe evolve hardcoded gpt-4o-2024-11-20 and threw 404 against any self-hosted OpenAI-compatible endpoint (vLLM / Ollama / LM Studio / etc.). Per-node YAML model: overrides still take precedence; the env vars apply when no per-node model is set.

Changed

• oe ultra synthesizer prompt — substantial hardening. Live-tested against a self-hosted MiniMax-2.7-w8a8 endpoint; the first synthesis attempt produced YAML with 3+ validation errors. Six iterations of prompt improvements landed:

  1. Top-level shape rule — nodes: and edges: MUST nest under graph:, not at root
  2. YAML escaping rule — always quote strings containing :, #, {}, [], etc.
  3. Per-kind required-field cheatsheet (one block per node kind, dataset/cli-agent fully spelled out)
  4. Edges have ONLY from/to/when — no description:, no label:
  5. state.schema type values restricted to string | number | boolean | object | array | null (no integer, no int)
  6. Dataset semantics — kind: dataset is for tabular rows (JSON/JSONL/CSV/Parquet/SQLite) only; for single-file loads use kind: tool with readFileSync; writes: MUST be exactly one field
  7. Defensive tool stubs — oe run MUST succeed before any user wiring; every state read uses ?? fallback; the first node populates everything downstream needs from a fixture
  8. Explicit 11-item self-check before returning

  Reduced failure rate from "100% need manual YAML fixes on MiniMax-2.7-w8a8" to "schema-clean and runtime-mostly-clean on first try". The remaining hard case is cross-field consistency (writes: matching state.schema names) — best solved by auto-retry on validation failure, planned for the next release.

• The clone destination for subpath-based gh: installs uses the last meaningful path segment as the name, so gh:org/monorepo/examples/x and gh:org/monorepo/examples/y co-exist in .openexpertise/experiences/x/ and …/y/ instead of colliding.

Fixed

• oe validate <non-existent-path> now prints a structured experience.yaml not found at <path> message instead of a bare ENOENT propagating from the file-system read. Matches the error style of every other command. (Reported by the v0.1.3 patch sweep.)
• oe install gh:owner/repo/subpath/path now parses the trailing path components as a subpath instead of silently dropping them. gh:jane/monorepo/examples/digest@v0.2.0 clones jane/monorepo at v0.2.0 and installs only examples/digest/ into .openexpertise/experiences/digest/. (Sweep finding #2.)