Skip to content

release-notes-v1.2.1.md

Latest commit

 

History

History
84 lines (63 loc) · 5.95 KB

release-notes-v1.2.1.md

File metadata and controls

84 lines (63 loc) · 5.95 KB
Raw

Release v1.2.1

Highlights

  • Model Variant Support Fix — Variant-suffixed models (e.g., codex/gpt-5.5-xhigh, codex/gpt-5.5-high) are now grouped under their base model ID instead of appearing as duplicate top-level entries. Includes synthetic base model creation, xhigh variant support, and getModelFamily() fix for provider-prefixed versioned models.
  • models.dev enrichment no longer fails on transient network slowness. The fetch pipeline now retries up to 3 times with exponential backoff and falls back to stale cached data if live refresh fails.
  • Default context limit corrected from 4096 to 128000 tokens to match OmniRoute API behavior.
  • Structured observability for enrichment failures with per-attempt diagnostics and fallback decisions.

What Changed

Reliability

  • fetchModelsDevData() now uses a bounded retry loop:
    • Maximum 3 attempts with 250ms / 500ms backoff.
    • Retries on: timeouts (AbortError), network errors, HTTP 429, and HTTP 5xx.
    • Fail-fast on: HTTP 4xx (non-429) and structurally invalid responses.
  • Stale in-memory cache fallback:
    • If cached data exists but TTL expired, live refresh is attempted first.
    • If all refresh attempts fail, the stale cached data is returned instead of null.
    • If no cache exists and all attempts fail, returns null (safe fail-open).
  • Timeout budget increased from 1000ms to 5000ms per attempt.
  • Failure classification: timeout, network, http_retryable, http_non_retryable, parse, invalid_structure.

Model Variant Support Fix

Problem: When OmniRoute lists variant-suffixed models separately (e.g., codex/gpt-5.5-xhigh, codex/gpt-5.5-high), each variant appeared as an independent top-level entry with incorrect generated variants ({low, medium, high}), causing duplicate/confusing model entries in OpenCode's model picker.

Solution:

  • Added groupVariantModels() in src/models.ts — a pure two-pass algorithm that:
    1. Categorizes models into real bases and variants using stripVariantSuffix()
    2. Builds result: real bases pass through unchanged; for each base with variants, merges all variants under the base model with a variants Record
    3. Synthetic bases: when only variants are returned (no explicit base), creates a synthetic base from the first variant, copying all fields and setting id/name to the stripped base ID
    4. Metadata merging: base inherits max contextWindow, max maxTokens, and the union of supported capability flags across all variants
  • Integrated into fetchModels() pipeline: normalizeModel → deduplicateModels → groupVariantModels → enrichModelMetadata → toProviderModels
  • Fixed toProviderModel() in src/plugin.ts to prioritize pre-populated model.variants over generated {low, medium, high} defaults
  • Added 'xhigh' to OmniRouteModelVariant.reasoningEffort type and generated variants

Edge Cases Handled:

Scenario Behavior
Only variants returned, no base model Creates synthetic base from first variant
Base model + variants both returned Uses real base; merges variant metadata (max limits)
Non-reasoning suffix (e.g., -preview) stripVariantSuffix() ignores it; no grouping
Mixed provider prefixes post-dedup Grouping operates on canonical IDs
Variant without supportsReasoning=true Still grouped; base becomes true if any variant has it

Fixes

  • DEFAULT_CONTEXT_LIMIT corrected from 4096 to 128000.
  • getModelFamily() for Provider-Prefixed Models — Fixed incorrect family extraction for versioned models with provider prefixes. getModelFamily('codex/gpt-5.5-xhigh') now correctly returns 'gpt' (was 'codex/gpt').

Code Review Fixes

  • Cache Isolation — modelsDevCache is now keyed by URL (Map<string, ModelsDevCache>) to prevent cross-config data leakage when different configs specify different modelsDev.url values.
  • JSDoc Accuracy — OmniRouteModelsDevConfig.timeoutMs JSDoc updated to reflect the new 5000ms default.
  • Lockfile Sync — package-lock.json version aligned with package.json (1.2.1).
  • Test Suite Speed — Eliminated real setTimeout sleeps from test/models-dev.test.mjs by using cacheTtl: 0 for stale-cache tests. Reduces test runtime and improves scalability.
  • Latency Documentation — Explicit JSDoc added on fetchModelsDevData() documenting worst-case cold-start latency (~15.75s) as an accepted reliability trade-off.
  • models.dev Structural Validation — Fetched models.dev payloads now validate provider entries and nested models records before accepting data, preventing malformed upstream objects from entering cache/index paths.
  • Variant Capability Union — Grouped variants now merge supportsVision, supportsTools, supportsStreaming, supportsTemperature, and supportsAttachment into the base model when any variant advertises those capabilities.

Testing

  • Added 9 focused tests in test/models-dev.test.mjs covering all retry, cache, fallback, and malformed-provider validation paths.
  • Added 1 unit test in test/models.test.mjs covering capability union across grouped variants.
  • Added 2 regression tests in test/plugin.test.mjs for variant grouping and synthetic base model creation.
  • Added cache isolation (clearModelCache(), clearModelsDevCache()) to test/plugin.test.mjs afterEach to prevent cross-test contamination.
  • Full regression suite: 54/54 tests pass (0 failures).

Documentation

  • Internal docs/superpowers/ planning/spec artifacts are kept local only and are excluded from the GitHub repository.

Verification

  • npm run prepublishOnly passes (clean, build, check:exports).
  • npm test passes: 54 tests, 0 failures.
  • TypeScript strict mode compiles cleanly.

Upgrade Notes

  • No breaking changes. Plugin behavior remains safe when models.dev is fully unavailable.
  • Existing modelsDev.timeoutMs and modelsDev.cacheTtl config options continue to work as before.