Skip to content

HANDOFF.md

Latest commit

 

History

History
122 lines (105 loc) · 6.54 KB

HANDOFF.md

File metadata and controls

122 lines (105 loc) · 6.54 KB
Raw

Session Handoff

What I Was Doing (2026-06-15)

Big session: zero-tech-debt sweep (PR #127), v0.28.0 Hardening Sweep release, residual closure (#129: llms-full generator, Supabase search verdict, skill toolchain), the v1.0 beta→stable promotion pass (#130), and now cutting v0.29.0 — Stability Promotions (this handoff). v1.0 still baking ~mid-July; the Python 3.9 drop (#99) is done; 1.0 freeze next.

Current State

  • Version: 0.29.0 in src/selectools/__init__.py and pyproject.toml. CHANGELOG ## [0.29.0] - 2026-06-15 — Stability Promotions synced to docs/CHANGELOG.md. Release = push tag v0.29.0 (CI publish-pypi job fires on refs/tags/v*).
  • Suite: 7,796 tests collected; full non-e2e run 7,637 passed / 0 failed. 115 examples, 115 source-verified models.
  • Quality gate: ruff format + check clean, mypy fully clean (0 errors, 180 files), bandit clean. Architecture stability gate green.
  • 0.28.0 themes: execute_shell shell=False (behavior change), SSRF guard on browser/webhook (consolidated _ssrf.py), session branch() namespace + consistent round-trippable list() keys across all backends, public-API type tightening (AgentResult/AgentConfig off Any), embedding timeout/retry, new [cache] extra, doc/count reconciliation.
  • Docs: README (What's New v0.28 + badges/counts), ARCHITECTURE, COMPATIBILITY, QUICKSTART, EVALS, index, CONTRIBUTING, landing/index.html, og assets, llms.txt, and CHANGELOG all reconciled this release.

Merged to main since v0.26.0 (2026-06-13, all [Unreleased])

  • #108 v0.26.0 performance benchmarks published (docs/modules/BENCHMARKS.md)
    • post-1.0 backlog reconciled in ROADMAP.
  • #109 recall tool — completes the agentic-memory pair (toolbox/memory_tools.py).
  • #110 Toolbox +4 categories (Discord, S3, browser, image-gen): 48 → 56 tools.
  • #111 UnifiedMemory AgentConfig wiring — MemoryConfig(unified=True, ...); also delivered episodic-retention config (add_turn auto-prunes).
  • #112 Cache-rate cost for OpenAI + Gemini (calculate_cost_with_cached_input, cached_prompt_cost); 24 rates re-verified live; gemini-embedding-2 documented as GA/recommended-for-new, default stays -001 (incompatible space).
  • #113 Cron/scheduled agents (scheduler.py: AgentScheduler, cron, every).
  • #114 Reasoning tools (toolbox/reasoning_tools.py: think/analyze, bounded).
  • #116 MongoSessionStore; #117 DynamoDBSessionStore (sessions now 6 backends).
  • #118 PromptInjectionGuardrail — heuristic injection/jailbreak detection.

Next Up — roadmap is decision-gated

All autonomously-buildable Future/Watch items are shipped. The rest need a product call (see ROADMAP "needs a product decision"):

  • Firestore session backend — only if there's real demand (adds a dep).
  • Model-based guard (beyond the heuristic #118) — hosting decision.
  • Multi-channel bot gateway — in-repo vs separate package.
  • Learning system — needs a concrete spec.
  • Shadow git checkpoints — only if steering toward coding agents.

Sheriff #303 (the prior "next up") is DONE — merged + deployed 2026-06-12, prod logs clean on selectools 0.26.0.

The v1.0 Bake Window Continues — July Tag Unchanged

All v1.0 code work remains merged and baking. v0.26.0 is a mid-bake patch, not a new feature wave: the @beta surface got real-world mileage, the bake hunt caught a real safety bug, and the fix shipped. The July 1.0 tag plan is unchanged.

July 1.0 Tag Checklist (in order)

  1. Drop Python 3.9 — DONE (dedicated PR): requires-python = ">=3.10", 3.9 removed from the CI matrix, aclosing backport replaced with contextlib.aclosing (the only true 3.9-only shim), ruff retargeted (target-version = "py310"). NOTE: the stability.py Protocol workaround is for a 3.9–3.11 bug, NOT 3.9-only — it STAYS until the floor is ≥3.12. This is the only remaining breaking change and it lands BEFORE the tag.
  2. Promote-after-bake review — sweep the @beta surface for anything that baked cleanly through the window (v0.24 + v0.25 betas; check issue tracker for API-shape complaints). Promotions are deliberate, not automatic.
  3. Classifier flip at tag — Development Status :: 5 - Production/Stable in pyproject.toml, version 1.0.0, CHANGELOG entry, tag. The release commit for 1.0 should be boring.

Next Arcs (post-1.0 or parallel)

  • Sheriff adoption: DONE — caching, parser, knowledge, and the pending store/sanitizer deletion all merged + deployed (Sheriff #300-#303).
  • UnifiedMemory config wiring: DONE (#111).
  • Remaining backlog: prompt registry/versioning, durable execution, code sandbox, Bedrock-native provider (LiteLLM covers it meanwhile), P3 items. Plus the decision-gated Future/Watch items above.

Watch Out For

  • Venv: .venv now has ruff and mkdocs installed (this session); pymongo is NOT installed (the Mongo backend tests inject a fake via sys.modules, the same pattern as the Redis/Mongo/Dynamo backend tests). Run bandit via python -m bandit if the shim shebang is broken. The editable .pth points at the main repo's src; pytest's pythonpath = ["src", "."] makes worktree runs use the worktree's own source.
  • AgentConfig(hooks=...) now raises TypeError (#94). Migration mapping lives in docs/MIGRATION_1.0.md. Don't resurrect it for a "quick fix" — observers cover every hook point.
  • Stability gate is parametrized per symbol (tests/test_architecture.py::test_every_public_symbol_has_stability_marker). Any new public symbol without a marker fails CI by design; mark it, don't exempt it.
  • Protocol classes + markers: stability markers must not become structural members of runtime-checkable Protocols (py3.10-3.11 regression fixed in #95) — keep markers in the registry for Protocols, not as attributes.
  • tests/rag/test_property_based_rag.py::test_full_metadata_filter_always_matches flaked once during release prep (hypothesis draw), passed on re-run and in isolation. If it flakes again, pin and minimize the failing example instead of rerunning.
  • RedisPendingStore needs Redis >= 6.2 (GETDEL claim); tighten_ttl is an id-pinned atomic Lua rewrite — covered by real-Redis smoke tests (#93), keep a Redis running locally to exercise them.
  • Tool._serialize_result re-injects kind for ToolResult subclasses (#72) — ClassVars never survive asdict().
  • Gemini flash-lite + tools is unreliable upstream (docs/COMPATIBILITY.md) — don't re-litigate.