∞skylakegrep

release notes · v0.5.1

skylakegrep 0.5.1 — auto-trigger lazy on cold-start (no flag needed)

0.5.0 shipped a --lazy flag and called the architecture done. Wrong. The whole premise of lazy is that the user does not know whether they're in the right folder, or whether their query happens to align with the code's vocabulary — so the user cannot be expected to decide whether to add a flag. 0.5.1 flips lazy to auto-trigger.

What changed

skygrep "<query>" (default, no flag) on a never-indexed project now auto-decides between three tiers:

ripgrep cold-start           ←  ~100 ms, always runs
  ↓ if rg paths cover ≥ 2 distinct query tokens (camel/snake split)
  →  return rg-only          ←  user gets answer in 100 ms
  ↓ otherwise (vocabulary mismatch / wrong folder)
  →  also fire lazy semantic ←  ~5 s wait, LLM-routed entry-points
                                 + bge-m3 batch embed, σ-validated
                                 → merge with rg → return both

--no-lazy opts out (e.g. for benchmarking pure rg). --lazy is now the default (kept for explicit testing).

The "rg is strong" gate is intentionally strict: it requires ≥ 2 distinct query tokens (after PascalCase / snake_case splitting) to land in result paths. A single common token like schema matching five unrelated schema.py files in different backends does not count as strong — that's the vocabulary-mismatch trap that fooled 0.5.0's earlier opt-in design.

Verified end-to-end on a fresh, never-indexed Django checkout

$ skygrep "cache invalidation strategy"

╭─ django/middleware/cache.py                                       0.537
╭─ django/contrib/sessions/backends/cached_db.py                    0.518
╭─ django/templatetags/cache.py                                     0.514
╭─ django/core/cache/backends/redis.py                              0.514
╭─ django/template/loaders/cached.py                                0.509

[8.144 s · ripgrep cold-start + lazy auto (σ=0.010, conf=medium,
 25new/0cached) · intent=lexical · 0 filename + 5 rg + 5 lazy ·
 index building in background]

Six scenarios from CLI (not python API) on /tmp/oss-bench/django, fresh DB each time, no skygrep index . ever ran:

# Query rg paths align? Lazy fires? Latency Verdict
A ModelForm rg returned only docs/* YES 6.4 s ✓ lazy delivered model_forms/*.py
B schema synchronization across releases only "schema" matched YES 8.0 s ✓ semantic answer
C B with --no-lazy n/a NO (opt-out) 0.2 s ✓ pure rg respected
D auth backends tokens didn't co-occur in path YES 9.2 s ✓ semantic answer
E cache invalidation strategy vocabulary mismatch YES 7.3 s ✓ correct cache modules
F how does Django render template inheritance render + template aligned NO (rg strong) 0.3 s ✓ instant rg keyword

Pytest 201/201, no regressions.

Why this matters

Per the user's vision (and corrected this iteration):

"本身 lazy 的模式我们就不知道我们是不是在错误的这个 path 下面, 所以根本就 user 不可能知道要去加显式这种东西。就是在我们没有 搜到或者是没有完全 index 加载的情况下,就是因为我们不知道是不是 在错误的文件下或者 path 下提问,我们才会去 trigger 这个所谓的 lazy 的模式的对吧。"

0.5.1 implements that exactly: the user types skygrep "...". The system decides for them. When rg is path-token-strong, user gets the instant keyword answer. When rg is weak (the case the user can never judge a priori), the system spends 5 s on the lazy semantic tier so the user gets a correct answer instead of a wrong-fast one.

Compatibility

  • --lazy flag still accepted; semantics unchanged from 0.5.0 (now equivalent to default behavior). Use for explicit testing.
  • --no-lazy is the new opt-out; needed for benchmarks that want pure rg cold-start measurement.
  • Indexed projects (post skygrep index .) are unaffected — the full σ-adaptive cascade still owns the warm-query path.

What's not yet done (carried from 0.5.0)

  • Themed HTML pages for the 0.5.x release notes (only .md exists).
  • proactive.filename_extend integration of lazy_explore_cross_folder to replace its hardcoded ~/Downloads/~/Desktop/~/Documents defaults — module API is exposed, wiring deferred to 0.5.2.

Verified

  • pytest tests/ — 201 / 201 pass
  • 6 CLI scenarios on fresh Django (table above)
  • --no-lazy path measured at 0.2 s, footer correctly reads --no-lazy
  • --lazy/--no-lazy flag round-trips through click correctly