Skip to content

CURRENT_STATE.md

Latest commit

 

History

History
155 lines (130 loc) · 7.68 KB

CURRENT_STATE.md

File metadata and controls

155 lines (130 loc) · 7.68 KB
Raw

Current State

Snapshot

  • v0.15.1 is the latest published release. It is available from PyPI and its record is immutable. v0.13.1 is the prior published release, and GitHub; exact artifact digests are in docs/release/v0.15.1-checksums.txt. v0.13.1 shipped the Phase 4 core-roadmap work with the exact release headline Replicated benchmark, faster SQLite at scale, reference integrations. The v0.13.0 tag exists but was never published (superseded; see the release notes).
  • v0.12.0 is the prior published release; its records remain under docs/release/.
  • v0.12.0 shipped the Phase 3 structural refactor, with the exact release headline Structure only. Zero behavior change.
  • Phase 3 implementation and the bounded builder matrix completed on codex/v0120-phase3-structural-refactor, based on f342d45dabe127acca6231f29830ff11d98a340e. Each code increment received an independent GO with no remaining P0-P3 finding. The independent final verdict is owned only by the handoff's REVIEW_REPORT.md; the exact-SHA external release gates passed on the release commit.
  • Both governed version sources were cut to 0.12.0 by the release engineer after verifying the handoff.
  • The historical LongMemEval_s result is 79.4% (397/500) from one run on 2026-07-07. It is not a repeated estimate or a measurement of this release.
  • Alice remains public-alpha, pre-1.0, local-first, single-user, and self-hosted.

What v0.11.1 Shipped

v0.11.1 shipped the bounded Phase 2 debt sweep on the post-periphery-cut product surface. Its immutable release notes and checksums remain the authoritative description; Phase 3 does not rewrite that history.

  • Stable HTTP, MCP, CLI, and onramp failures, plus the migrated provider, response, scheduler, evaluation, doctor, and connector diagnostics, keep static public vocabularies. Intentional legacy-on proxy_execution.py business-result reasons remain dynamic.
  • The list_memories(query=...) and list_resume_memory_events(query=...) legs keep their ASCII case-insensitive literal-filter contract; generic alice_recall retrieval remains separate.
  • Coupled true redaction scrubs the governed memory/project-update graph but retains shared source/source-chunk evidence for separate source hygiene and does not roll back accepted project state.

What v0.12.0 Shipped

  • A thin alicebot_api.main:app assembly module with the HTTP handlers moved into domain routers. Default and gated OpenAPI registries remain exactly 182 and 231 operations, a delta of 49.
  • Corresponding PostgreSQL and SQLite vNext store seams, a surviving-domain legacy-store split, and stable store facades. Generated SQL text and store protocols are unchanged.
  • Domain contract modules behind the stable contracts.py facade.
  • Per-domain MCP implementations behind the stable mcp_tools.py facade, preserving the 11-core/65-legacy/76-total registry and flag behavior.
  • Per-domain CLI command modules behind the stable alicebot_api.cli import, preserving both alice and alicebot entrypoints; alice-memory remains routed through the onramp entrypoint.
  • Response-hygiene and coverage enforcement that follows moved modules rather than shrinking onto a facade. The 296 public-response call inventory and the router aggregate coverage floor remain pinned.
  • No route, tool, command, schema, migration, dependency, or runtime behavior change.

Verification Posture

  • Final code-carrier evidence passed 3,804 unit tests with 80.3777897% package coverage. Router coverage was 3,604/5,373 statements, 67.0761%, above the 45% floor.
  • PostgreSQL 16 plus pgvector passed 399 legacy-on integration tests with one expected skip. The separate flag-off smoke passed one executed test with the nonzero-test guard enabled.
  • LongMemEval passed 127 tests; the checked evidence replay passed seven arms; the focused vector/retrieval lane passed two tests.
  • The web carrier passed 217 unit tests, core and vNext coverage floors, typecheck, lint, build, bundle budgets, and the 17+1+1+1 browser matrix.
  • Final-carrier package reproduction and installed-artifact evidence is owned by the Phase 3 BUILD_REPORT.md. The builder matrix ran before the version cut; its locally produced artifacts were verification inputs only and were never uploaded anywhere.
  • No security or cybersecurity audit was performed in Phase 3.

Release Boundary

v0.15.1 is tagged, published, and immutable. Its authoritative records are:

  • docs/release/v0.15.1-release-notes.md
  • docs/release/v0.15.1-checksums.txt

v0.13.1 remains published and immutable; its records are docs/release/v0.13.1-release-notes.md and docs/release/v0.13.1-checksums.txt.

The v0.13.0 tag exists but was never published; the number was superseded by v0.13.1 because stable tags are immutable.

v0.12.0 is the prior published release; its records remain at docs/release/v0.12.0-release-notes.md and docs/release/v0.12.0-checksums.txt.

v0.11.1 is the prior published release; its records remain at docs/release/v0.11.1-release-notes.md and docs/release/v0.11.1-checksums.txt.

What v0.14.0 Shipped

The Phase 5 enterprise track: the single-tenant self-hosted deployment contract executed on a real public host with an owner deployment receipt, least-privilege operations proven under a non-superuser admin role, executed backup and restore evidence on both backends, the one-time origin-bound browser-clip capability, and the five deployment-guide fixes surfaced by the first real-host execution. Scope details are in the dated handoff packages and the v0.14.0 release notes.

What v0.13.1 Shipped

  • The replicated LongMemEval_s baseline (81.2% mean over three runs on the published v0.12.0 code) committed as per-question evidence.
  • SQLite vector-scale work: bit-identical vectorized scan, resident vector cache with transactional stamp invalidation (vector stage 385-465ms warm at 100k inside 754MB peak / 760MB steady, 1024MB default cap, off-switch), one additive bootstrap table, Postgres unchanged.
  • Two CI-smoked reference integrations: the MCP quickstart and the OpenAI Agents SDK function-tool example with real per-agent key auth.
  • Trace-only count-intent diagnostics behind the aggregation gate (is_answer=false); the multi-session benchmark-closure NO-GO is recorded and no synthesis uplift is claimed.
  • The deferred mcp-tools.md legacy-alias wording correction.
  • No MCP registry, OpenAPI, HTTP route, dependency, or Postgres schema change.

Product Boundaries

  • No hosted service, multi-tenant control plane, or SLA.
  • No managed OAuth consent/account-linking or automatic account polling.
  • No Telegram or other channel transport in the current runtime.
  • No public bundled chat/response product, chief-of-staff product, or model packs. Internal response jobs support retained provider invocation only.
  • No silent capture from arbitrary conversations.
  • No OCR or transcription execution; Alice only ingests extracted text.
  • Durable agent writes remain policy-checked, provenance-linked, and reviewable.
  • Coupled true redaction scrubs governed content copies in the memory/project- update graph; it does not undo accepted project state or erase shared source evidence, upstream systems, exports, backups, or external logs.

What v0.15.1 Shipped

v0.15.1 is the latest published release and remains the install, checksum, and baseline reference.

The v0.15.0 tag exists but was never published. The commit it points at carried a release-gate step that could not run on a CI runner, so the gate could never pass from that tag, and repository rules correctly refuse both deletion and update of stable tags. No GitHub Release or PyPI artifact was ever created for it.