fix(model_metadata): drop stale cache entries for qwen3.6-plus, grok-4-fast, grok-4.20

[AhmetArif0]

Summary

Three model families gained explicit DEFAULT_CONTEXT_LENGTHS entries after some users may have already cached lower (incorrect) values through earlier resolution paths. This PR adds stale-cache guards for each, mirroring the pattern used for Kimi (≤32K), MiniMax-M3 (≤204,800), and Grok-4.3 (≤256K) in prior PRs.

qwen3.6-plus (1M context)

The qwen3.6-plus entry was added on 2026-05-17 (d9abbe7fa). Prior builds resolved it via the generic qwen catch-all (131,072) and could have persisted that value via the context-overflow cache path. Guard threshold: cached ≤ 131,072.

grok-4-fast and grok-4.20 (2M context each)

Both entries were added on 2026-04-10 (b57769718). Before that date, these slugs had no explicit entry and fell through to DEFAULT_FALLBACK_CONTEXT (256,000) or lower probe tiers. Guard threshold: cached ≤ 256,000.

Changes

• agent/model_metadata.py: add _model_name_suggests_qwen3_6_plus() and _model_name_suggests_grok_4_fast_or_4_20() helper predicates; add three elif branches in the step-1 stale-cache guard block inside get_model_context_length().
• tests/agent/test_model_metadata.py: add TestQwen36PlusStaleCacheGuard and TestGrokFastAndGrok420StaleCacheGuard (12 tests total) — helper unit tests, stale-drop-and-re-resolve, correct-cache-preserved, no-clobber for sibling slugs.

Test plan

• pytest tests/agent/test_model_metadata.py::TestQwen36PlusStaleCacheGuard tests/agent/test_model_metadata.py::TestGrokFastAndGrok420StaleCacheGuard -x -q — 12 tests pass

[teknium1]

Thanks for the focused regression fix. The premise remains valid on current main: get_model_context_length() reads the persistent cache at agent/model_metadata.py:2110 and returns a non-exempt cache hit at agent/model_metadata.py:2173-2178, before it can reach the explicit qwen3.6-plus and Grok defaults at agent/model_metadata.py:277 and agent/model_metadata.py:311-312. Current main already uses the same targeted stale-cache-guard pattern for Kimi, MiniMax-M3, and Grok-4.3 at agent/model_metadata.py:2125-2158.

The PR is stale against its June base and will need a mechanical port into the moved guard/test sections, but its implementation approach and behavioral regression coverage remain appropriate.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Three PRs address the stale persistent context-cache path: #37068 and #38996 implement the same Grok-4.3 ≤256K eviction guard, while #37684 extends that established fix to qwen3.6-plus, grok-4-fast, and grok-4.20 with model-specific thresholds and regression coverage.

Related pull requests

• #37068 [closed] related — (+81/-0) — superseded by #38996: adds the Grok-4.3-specific ≤256K stale-cache guard and tests, directly correcting cache hits that otherwise bypass the newer 1M default; it remains relevant because its commit was cherry-picked with attribution into the merged implementation.
• #37684 related — (+214/-0) — keep open and mechanically port to current main: adds targeted eviction of stale ≤131,072 qwen3.6-plus entries and ≤256,000 grok-4-fast/grok-4.20 entries, while preserving correct caches and sibling model slugs. This agrees with the keep_open review on #37684, which confirms that the cache-first return path and explicit defaults still leave the reported cause present on current main.
• #38996 [merged] related — (+81/-0) — merged reference implementation: the cherry-picked successor to #37068 adds the same narrowly scoped Grok-4.3 guard and regression tests, resolving stale 256K cache entries without affecting plain Grok-4.

Duplicates

#37068 and #38996 are substantively identical; #38996 is the merged, current-main salvage of #37068. #37684 is complementary rather than duplicate because it covers different model families and stale thresholds.

Suggested consolidation

Merge #37684 after the documented mechanical port onto current main and targeted regression validation; its qwen3.6-plus, grok-4-fast, and grok-4.20 guards complement the pattern already merged in #38996. Keep #37068 closed as superseded by duplicate #38996; no additional PR should be closed as a duplicate of #37684.

Cross-PR triage: Reviewed 3 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 24 kB of PR diffs, 5 kB of issue/PR text, 2 kB of discussion (3 comments), 1 verify verdict. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[teknium1]

The guards you requested here (qwen3.6-plus, grok-4-fast, grok-4.20) are now on main via PR #85434 (merged 91e550b) — credited to you in the PR body. Rather than adding three more per-model predicates, we replaced the whole family with one generic _stale_pre_catalog_cache_entry() guard driven by a _PRE_CATALOG_STALE_KEYS set; all three of your models are listed and covered by TestGenericPreCatalogStaleGuard, including an E2E stale-drop-and-re-resolve for qwen3.6-plus.

Thanks for spotting the class — closing as absorbed by the generic fix.