Skip to content

v1.0.0 — Production-stable

Choose a tag to compare

View all tags
@theagenticguy theagenticguy released this 13 May 19:07
· 32 commits to main since this release
v1.0.0
6df7b62
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

v1.0.0 — Production-stable

First production-stable release. Cuts the 1.x train from the v1.0 windowed-pipelines milestone (#42) and the v1 vision + roadmap RFC (#43).

Highlights since v0.7.0

New analytics pipelines

  • Per-session windowed trajectory classifier — pair-keyed on (prev_uuid, curr_uuid) with a six-value transition_kind enum: frustration_spike, resolution, reset, drift, clarification, none. Each text turn produces exactly one row; the session-first turn gets a synthetic pair with prev_uuid IS NULL. Sonnet 4.6 sees up to 16 windows per chunk in one structured-output call; sessions longer than 16 text turns split into anchor-sharing chunks.
  • Pair-keyed conflicts — rekeyed on (turn_a_uuid, turn_b_uuid) with conflict_kind ∈ {disagreement, correction, reversal, impasse} and severity ∈ {low, medium, high}. The legacy empty=true sentinel is gone — sessions with no conflicts produce zero rows.
  • Friction SQL stamps — regex fast-path (status_ping, interruption, correction) runs before Sonnet; LLM-only for unmet_expectation / confusion / frustration where session context is required.

Data model

  • turn_window SQL view — LAG(uuid) OVER w with a single named window clause beats the prior correlated-min by ~3:1 on read_json views (every textual reference re-runs schema inference; LAG collapses to one FROM).
  • Ingest stamps — ingest_stamps.parquet carries approx_tokens (tiktoken cl100k × 0.78 Anthropic approximation), simhash64 (blake2b 64-bit over lowercased 3-grams), and canonical_uuid. Embed workers now skip near-duplicate rows via LEFT JOIN ingest_stamps with a canonical_uuid IS NULL OR canonical_uuid = mt.uuid guard.
  • is_compact_summary projection added; attachment-only rows filtered out of messages_text.

Plumbing

  • 1h prompt cache on every Sonnet 4.6 system block — cache_control: {"type":"ephemeral","ttl":"1h"}. Per-pipeline cache-stat accumulator (threading.Lock-protected) emits one summary line at INFO per pipeline run.
  • CLAUDE_SQL_HOME migration — caches move from ~/.claude/ to a dedicated parent dir resolved via XDG_DATA_HOME, macOS Application Support, or ~/.claude-sql/. Idempotent legacy-cache move on first connect.
  • analyze stale-connection bugfix — _rebind_vss runs after embed (to bind the newly populated Lance dataset) and between cluster/community stages. Drops VIEW-or-TABLE shape mismatch from register_vss's empty-namespace fallback.
  • Parallelized trajectory dispatch via anyio.create_task_group — sessions run concurrently; the CapacityLimiter inside classify_one is the single throttle (no double-acquire at the dispatcher).

Quality gates

  • Patch coverage lifted to 91%+ (trajectory_worker 100%, conflicts_worker 100%, cli.py 92%).
  • Two CodeQL py/catch-base-exception findings fixed — narrow BaseException → Exception so anyio CancelledError propagates.
  • CodeQL py/file-not-closed in a test mkstemp fake fixed by opening the fd per-call inside the closure.
  • Four CodeQL hygiene rules now documented in CLAUDE.md.

Breaking changes

  • Trajectory parquet schema — (session_id, prev_uuid, curr_uuid, prev_sentiment, curr_sentiment, delta, is_transition, transition_kind, confidence, classified_at). Legacy per-message shards are detected via pq.ParquetFile.schema_arrow marker-column probe (uuid + sentiment_delta) and deleted on first run. The sentiment_arc(sid) macro now joins on message_trajectory.curr_uuid and returns (ts, role, curr_sentiment, delta, transition_kind, is_transition, confidence) — callers relying on sentiment_delta must rebase to delta + curr_sentiment.
  • Conflicts parquet schema — (session_id, turn_a_uuid, turn_b_uuid, conflict_kind, severity, agent_position, user_position, confidence, detected_at). Any caller wanting every session in the result set must LEFT JOIN sessions and coalesce missing counts to 0 (no empty-sentinel row). Legacy shards carrying conflict_idx / empty columns are purged on first run.
  • Embeddings store location — ~/.claude/embeddings_lance/ is canonical; legacy ~/.claude/embeddings/part-*.parquet is migrated automatically on first connect and remains for rollback.

Install / upgrade

# end-user path
uv tool upgrade claude-sql

# dev
uv sync

Known follow-ups (v1.1 scope)

  • Conflicts prompt calibration — the v1.0 pair-keyed prompt returned 0 pair rows on the live 3,653-session corpus; v1.0 shipped storage-shape only.
  • Conflicts pair-scanner (RFC 0002 §4.2) to replace the whole-session prompt with one row per adjacent turn pair.
  • cluster_worker.run_clustering --dry-run should no-op cleanly when the Lance dataset is missing.
Assets 3
Loading