kaos-tabular 0.1.0__tar.gz

kaos_tabular-0.1.0/.gitignore

@@ -0,0 +1,11 @@
+.coverage
+.mypy_cache/
+.pytest_cache/
+.ruff_cache/
+.venv/
+__pycache__/
+*.pyc
+dist/
+*.egg-info/
+.mcp.json
+.kaos-vfs/

kaos_tabular-0.1.0/CHANGELOG.md

@@ -0,0 +1,501 @@
+# Changelog
+
+All notable changes to `kaos-tabular` are documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+
+## [0.1.0] — 2026-05-20
+
+### Released
+
+- 0.1.0 GA — WU-L of GA plan. First stable release. Public API frozen.
+- Pin floor raised to `>=0.1.0,<0.2` across all kaos-* runtime and
+  optional dependencies. Refreshed `uv.lock` to pick up the 0.1.0
+  line of every upstream.
+
+### Internal
+
+- WU-L of the 0.1.0 GA plan
+  (`kaos-modules/docs/plans/2026-05-20-0.1.0-ga-plan.md`).
+
+
+## [0.1.0rc1] — 2026-05-20
+
+### Changed
+
+- Pin floor raised to `>=0.1.0rc1,<0.2` across kaos-* runtime and
+  optional dependencies (`kaos-core`, `kaos-content`, `kaos-mcp`).
+  Refreshed `uv.lock` to pick up the rc1 line of every upstream.
+
+### Internal
+
+- WU-J of the 0.1.0 GA plan
+  (`kaos-modules/docs/plans/2026-05-20-0.1.0-ga-plan.md`).
+  Release candidate; freezes the public API for `kaos-tabular`
+  ahead of 0.1.0 GA.
+
+
+## [0.1.0a5] — 2026-05-20
+
+### Changed
+
+- **kaos-core floor raised to `>=0.1.0a12`** (post-URI-redesign +
+  Capability type). WU-F.2 of the 0.1.0 GA plan; catch-up to
+  kaos-core 0.1.0a12. No public API changes in kaos-tabular.
+
+## [0.1.0a4] — 2026-05-17
+
+### Changed
+
+- **kaos-core floor raised to `>=0.1.0a10`** to pick up the URI
+  contract redesign. Pass-through for kaos-tabular file-input tools.
+  See `kaos-modules/docs/plans/uri-contract-redesign.md`.
+
+## [0.1.0a3] — 2026-05-17
+
+### Changed
+
+- **Both file-input MCP tools — `kaos-tabular-register` and
+  `kaos-tabular-read-file` — now route agent-supplied paths through
+  `kaos_core.path_resolver.resolve_input_path` instead of raw
+  `Path(p).exists()`.** Both tools accept four input shapes for the
+  `path` argument: an absolute filesystem path (CLI / tests), a
+  `kaos://artifacts/<id>` URI returned by a previous tool, a `kaos://`
+  VFS URI, and a session-scoped VFS-relative path (e.g. a CSV uploaded
+  through a host UI like `kaos-ui`'s single-user-chat SPA). Both
+  tools'`path` parameter schema now documents these four shapes so the
+  LLM can discover the feature without reading source. Implements
+  Stage 3 of `kaos-modules/docs/plans/vfs-blind-tools-audit-and-fix-plan.md`;
+  closes the kaos-tabular slice of the production NDA-hallucination
+  incident (session `01KRVYAEA3B1HG95DBAG6H0DJ3`) where file-input
+  tools could not see SPA uploads and the agent fabricated a
+  jurisdiction / term-length analysis citing the unreadable files. The
+  resolver performs the artifact-store / VFS reads inside an `async
+  with` context manager; bytes are materialised to a temp file just
+  for the duration of the eager DuckDB `CREATE TABLE ... AS SELECT *
+  FROM read_csv(...)` (register) or the synchronous `_read_file`
+  build (read-file), then cleaned up. On `InputPathResolutionError`
+  both tools return the resolver's agent-friendly three-part message
+  via `ToolResult.create_error(exc.to_agent_message())`. When the
+  input was itself a `kaos://artifacts/<id>` URI, the existing
+  artifact id round-trips into `structured_content` so downstream
+  tools and the SPA's `ArtifactCard` can re-resolve the same handle.
+- **`kaos-core` dependency floor raised from `>=0.1.0a4` to
+  `>=0.1.0a9`** to pick up `kaos_core.path_resolver` (released in
+  kaos-core 0.1.0a9).
+
+## [0.1.0a2] — 2026-05-15
+
+### Fixed
+
+- **Nine array parameters across the MCP tool catalog now declare
+  their element types.** Each was previously `type=array` with no
+  `items`, which OpenAI's strict JSON Schema validator rejected
+  with HTTP 400 `invalid_function_parameters`, taking down the
+  whole tool catalog for openai-provider turns. Per-tool fixes:
+  - `kaos-tabular-dedupe.columns`, `-correlate.columns`,
+    `-melt.columns`, `-join.on`, `-pivot.group_by`,
+    `-aggregate.group_by`, `-top-n.by` → `items: {type: "string"}`.
+  - `kaos-tabular-aggregate.aggregates` → typed object schema
+    `{func: enum[sum/avg/min/max/count/count_distinct/median/stddev/
+    variance/first/last], column: string, alias?: string}`. The
+    LLM now produces correct payloads on the first try instead of
+    trial-and-erroring across two ReAct iterations.
+  - `kaos-tabular-aggregate.order_by` → typed object schema
+    `{column: string, direction?: enum[asc/desc]}`.
+  kaos-core 0.1.0a7's defensive `items: {}` floor is belt +
+  suspenders.
+
+### Security
+
+- **Documented the SQL-quoting safety contract on six query sites in
+  ``engine.py`` (bandit B608).** ``TabularEngine`` builds SQL via
+  f-strings against table/column/path inputs and routes every dynamic
+  fragment through one of two validating quoters:
+  ``_quote_ident`` (validates + double-quotes the identifier) or
+  ``_q_lit`` (doubles single quotes for SQL string literals). Bandit's
+  static B608 heuristic can't see the quoter — it just sees an
+  f-string concatenating SQL fragments — so every call site is
+  flagged as a possible SQL-injection vector. Added inline
+  ``# nosec B608`` comments at each site with a one-line justification
+  pointing at the relevant quoter; the quoting contract itself is
+  unchanged. Files: ``kaos_tabular/engine.py``.
+- **bandit + vulture now run in both pre-commit and CI.** Two new
+  hooks in ``.pre-commit-config.yaml`` (bandit + vulture), mirrored
+  by two new jobs in ``security.yml`` (``bandit (static security)``
+  + ``vulture (dead-code scan)``). Pre-commit gives contributors fast
+  feedback before push; CI makes the scan publicly visible on every
+  PR. Skip lists justified inline. Mirrors the rollout from
+  kaos-core. **Depends on PR #1** (bandit B608 nosec justifications
+  in engine.py) — bandit will fail on this branch's first run until
+  #1 merges, then rebase clears it.
+### Changed
+
+- **uv.lock bumped to the current PyPI-latest of three kaos-* siblings:**
+  ``kaos-content`` 0.1.0a2 → 0.1.0a4, ``kaos-core`` 0.1.0a4 →
+  0.1.0a5, and ``kaos-mcp`` 0.1.0a1 → 0.1.0a2. All three bumps are
+  no-op for kaos-tabular's public API. 276 unit tests continue to
+  pass.
+
+## [0.1.0a1] — 2026-05-08
+
+### Added (structured shape tools + did-you-mean error suggestions)
+
+A second pre-tag pass reconsidered the "tools earn their weight when
+SQL is genuinely awkward" framing. The framing held for `pivot`,
+`unpivot`, `join`, and `correlation`, but was too narrow for the
+`GROUP BY` / `WHERE` / `ORDER BY ... LIMIT` trio: agents write that
+SQL correctly, yes, but typed wrappers buy validation at the boundary,
+structured-event audit (the call shows up in `engine.history()` as
+`aggregate:<table>` instead of an opaque `query:` string), and
+dialect-insulation if the engine ever grows a non-DuckDB backend.
+
+Three new MCP tools (14 → 17) and matching public engine methods:
+
+- **`kaos-tabular-aggregate`** + **`engine.aggregate(table, *, aggregates,
+  group_by=None, where=None, having=None, order_by=None, limit=None,
+  target=None)`**. Composed `GROUP BY`. `aggregates` is a list of
+  `(func, column[, alias])` tuples; `func` ∈ `{sum, avg, min, max,
+  count, count_distinct, median, stddev, variance, first, last}`.
+  Validates the table, every column, and every aggregate function
+  *before* SQL is generated, with did-you-mean suggestions on a miss.
+  `where` / `having` remain opaque DuckDB SQL fragments (predicate
+  shapes are unbounded). `order_by` items must reference either a
+  group_by column or an explicit aggregate alias; bare aggregate
+  expressions in `ORDER BY` are rejected at the wrapper.
+- **`kaos-tabular-filter`** + **`engine.filter(table, *, where,
+  limit=None, target=None)`**. Typed `SELECT * WHERE`. The table is
+  validated; `where` is opaque DuckDB SQL. Useful when the caller
+  wants the call to show up in the structured history log under
+  `filter:<table>` instead of inside an opaque `query:` event.
+- **`kaos-tabular-top-k`** + **`engine.top_k(table, *, by, n=10,
+  ascending=False, target=None)`**. `ORDER BY ... LIMIT N`. Defaults
+  to descending so "top N by units" reads naturally; pass
+  `ascending=True` for bottom-N.
+
+### Added (did-you-mean suggestions across the engine)
+
+Every error path that mentions a missing table or column now carries
+a `Did you mean '<closest match>'?` suggestion using
+`difflib.get_close_matches` with a 0.6 cutoff. The cutoff is high
+enough to avoid spurious matches on short identifiers (`id` / `ip`)
+but low enough to forgive single-character typos on typical 6+
+character column names.
+
+The mechanism is wired into `describe_table`, `sample`, `count`,
+`find_duplicates`, `correlation`, `join` (both sides + `on=`),
+`pivot`, `unpivot`, `export_table`, and the three new structured
+shape methods (`aggregate`, `filter`, `top_k`). The aggregate
+function whitelist also gets did-you-mean against the supported
+function names. Module-level `_suggestions` and
+`_did_you_mean_fragment` helpers are unit-tested in isolation against
+the cutoff edge-cases (empty universe, no near-match, plural form);
+the `TestExistingErrorPathsRetrofit` class pins the retrofit so a
+future refactor can't silently drop suggestions from the older
+analytical surfaces.
+
+Test count: 216 → 276 unit tests (60 new in
+`tests/unit/test_structured_ops.py`); coverage stays above the 70%
+`fail_under` floor.
+
+Quick benchmark (100k-row CSV, 5 distinct group keys): structured
+`aggregate` runs 7.5 ms median vs. 3.4 ms for the equivalent raw
+`execute` — ~4 ms validation overhead from two `information_schema`
+lookups per call. The overhead is constant regardless of data size,
+acceptable for interactive agent use; throughput-bound batch loops
+should reach for `kaos-tabular-query` instead.
+
+### Fixed (post-release-review pass before tag)
+
+External review found gaps the audit-01 sweep missed; all addressed
+before tagging:
+
+- **#1 P0: SQLite table-name SQL injection in `_register_sqlite`.**
+  `src_table` values from `sqlite_master` were interpolated raw into
+  the next `sqlite_scan('{path}', '{src_table}')` call. A crafted
+  SQLite file with a hostile table name could escape the literal and
+  execute injected DuckDB SQL. New module-level `_q_lit` helper
+  performs the standard `'` → `''` escape; both the path and the
+  `src_table` now flow through it. Adversarial test in
+  `tests/unit/test_sqlite_register.py::test_register_sqlite_hostile_table_name_does_not_inject`.
+- **#2 P0: `save()` path SQL injection.** `EXPORT DATABASE '{p}'`
+  pasted the caller-supplied path directly. `save("'; ATTACH ...; --")`
+  could break out of the literal. Same `_q_lit` mitigation.
+  `export_table` (added in this release) was already correct but is
+  now consolidated onto `_q_lit` for consistency. Adversarial tests
+  in `tests/unit/test_path_injection.py`.
+- **#3 P0: `duckdb` minimum lifted from `>=1.0` to `>=1.4.2`.** 1.0.0
+  has no cp313 wheel; 1.1.1 was the first cp313 release; 1.4.2 was
+  the first cp314 release. Since we support both 3.13 and 3.14, the
+  floor must clear both — pre-1.4.2 made the lowest-direct CI job
+  build duckdb from source on cp314, which is why min-deps took
+  20+ minutes.
+- **#4 P0: MCP tool annotations now match real behaviour.** Pre-fix,
+  every tool used `_TABULAR_ANNOTATIONS` with `openWorldHint=False`,
+  including ones that genuinely reach the filesystem
+  (`Register` / `Query` / `ReadFile`); `ExportTool` used
+  `_TABULAR_WRITE_ANNOTATIONS` with `destructiveHint=False` despite
+  writing/overwriting files. Split into three classes:
+  `_TABULAR_READ_ANNOTATIONS` (closed-world catalog reads — `List` /
+  `Describe` / `Sample` / `Count`), `_TABULAR_OPEN_READ_ANNOTATIONS`
+  (open-world filesystem reads — `Register` / `Query` / `ReadFile`),
+  `_TABULAR_WRITE_ANNOTATIONS` (open-world destructive writes —
+  `Export`, now `destructiveHint=True`). Agents make auto-approval
+  decisions on these flags; getting them right is the largest
+  actual safety improvement in this commit.
+- **#5 P1: `_ENGINES` cache bounded with LRU + close-on-evict.**
+  Pre-fix, the per-session engine cache was an unbounded `dict`;
+  long-running streamable-HTTP servers leaked DuckDB connections
+  forever. Now an `OrderedDict` capped at
+  `_ENGINES_MAX_SESSIONS = 64`; the oldest engine is closed on
+  insert past capacity. TODO: replace with proper kaos-mcp
+  per-session lifecycle hook at 0.1.0a2. Coverage in
+  `tests/unit/test_session_engines.py`.
+- **#6 P1: stale integration assertion fixed.**
+  `tests/integration/test_mcp_tabular_pipeline.py` asserted the
+  pre-KTAB-007 error string `"Cannot infer format"`. Updated to the
+  current `"Cannot infer export format"`. CI doesn't gate the
+  integration tier today; raised as a separate platform tracker.
+- **#7 P1: SECURITY.md scope rewritten for kaos-tabular.** The
+  template carried over from kaos-mcp listed LLM/program-execution/
+  cache/provider concerns that don't apply here. New scope names:
+  the DuckDB SQL boundary, file registration paths, export/write
+  paths, MCP tool surface, the SQLite extension network fetch, the
+  transitive dep supply chain.
+
+### Added (post-Kelvin-comparison surface expansion)
+
+A pre-tag review against the legacy ``kelvin_tabular`` package
+(roughly 60 MCP tools across inspection / manipulation / statistics /
+quality / transformation categories) found that most of those tools
+were SELECT one-liners that don't earn their weight when the agent
+already has free-form SQL. The ones that *do* earn their weight are
+the SQL-is-genuinely-awkward cases — joins where column ambiguity
+catches agents writing `JOIN ON l.x = r.x` by hand, the
+``PIVOT`` / ``UNPIVOT`` syntax, long-form correlation matrices,
+provenance tracing — and those are the six we ported. The package
+explicitly does NOT ship Kelvin's full tree; SQL is the expression
+layer for everything else.
+
+Six new MCP tools (8 → 14) and matching public engine methods:
+
+- **``kaos-tabular-history``** + **``engine.history(*, last_n=20)``**
+  + ``EngineEvent`` exported on the public surface. Returns the
+  recent register / query / drop events for the session — provenance
+  for agents tracing back what's been loaded.
+- **``kaos-tabular-find-duplicates``** + **``engine.find_duplicates(table, *, columns=None)``**.
+  Returns rows that share their key with at least one other row,
+  via DuckDB ``QUALIFY COUNT(*) OVER (PARTITION BY …) > 1``. Default
+  ``columns=None`` uses every column (full-row duplicate detection).
+- **``kaos-tabular-correlation``** + **``engine.correlation(table, *, columns=None)``**.
+  Pairwise Pearson correlation between numeric columns, returned as
+  long-form ``(col_a, col_b, corr)`` rows. Default auto-selects
+  every numeric column from the catalog.
+- **``kaos-tabular-join``** + **``engine.join(left, right, *, on, how="inner", target=None)``**.
+  Wraps DuckDB's ``USING (col)`` clause so the join key appears
+  once in the result. ``how`` ∈ ``{inner, left, right, outer, semi,
+  anti, cross}``; ``target`` materializes via
+  ``CREATE OR REPLACE TABLE`` and registers.
+- **``kaos-tabular-pivot``** + **``engine.pivot(table, *, on, using,
+  aggregate="sum", group_by=None, target=None)``**. Wraps DuckDB
+  ``PIVOT``. ``aggregate`` ∈ ``{sum, avg, min, max, count, first}``.
+- **``kaos-tabular-unpivot``** + **``engine.unpivot(table, *, columns,
+  name_column="variable", value_column="value", target=None)``**.
+  Wraps DuckDB ``UNPIVOT``.
+
+Each tool declares its own per-tool ``ToolAnnotations`` literal
+(closed-world for catalog-only ops, open-world for arbitrary SQL,
+destructive-write for ``export``). Engine methods emit 3-part
+errors via ``EngineError`` and the MCP layer forwards them through
+``ToolResult.create_error``. New unit-test file
+``tests/unit/test_analytical_methods.py`` covers all five engine
+methods + their tool wrappers — 27 tests, including round-trips
+(pivot then unpivot), edge cases (empty columns list, missing
+column, invalid ``how``), and tool-side error translation.
+
+Test count: 189 → 216 unit tests; coverage stays at ~75% above
+the 70% ``fail_under`` floor.
+
+### Refactored (post-review code-quality pass)
+
+A self-review against `docs/python/{boundaries,modules,errors,
+dry-abstraction}.md` flagged five items worth addressing before tag.
+All landed; none change the public API:
+
+- **Item 3: `_ENGINES` global → `EngineRegistry` class.** New
+  module `kaos_tabular/_session.py` owning the bounded LRU.
+  `EngineRegistry(max_sessions=..., engine_factory=...)` lets tests
+  build isolated registries and inject a `_CountingEngine` factory
+  to spy on `close()` without monkey-patching module state. The
+  process singleton `SESSION_REGISTRY` keeps live MCP-session
+  behaviour identical. `tools._get_engine` is now a thin async
+  wrapper that delegates to the registry (with the same
+  `context is None` ephemeral-engine policy).
+- **Item 4: `cast(Literal[...], fmt)` → typed inference helpers.**
+  New `_coerce_export_format(value: Any) -> ExportFormat | None`
+  and `_infer_export_format_from_extension(ext: str) -> ExportFormat | None`
+  return literal types directly so ty sees the narrow without a
+  `cast`. ExportTool's `execute` gets simpler too.
+- **Item 5: brittle eviction test → `_CountingEngine` subclass.**
+  Replaced the `engine.close = lambda: ...` monkey-patch with a
+  real `TabularEngine` subclass that bumps a counter. Bonus:
+  asserts the evicted engine's DuckDB connection actually raises
+  `duckdb.ConnectionException` post-eviction.
+- **Item 6: focused `_q_lit` unit tests.** New
+  `tests/unit/test_engine_helpers.py` pins six properties + a
+  parametrized 7-input round-trip through real DuckDB
+  (`SELECT {_q_lit(s)}` → `s`). The adversarial tests still cover
+  the engine-end-to-end path; this catches contract drift before it
+  reaches them.
+- **Item 7: shared annotation constants → per-tool literals.**
+  Removed `_TABULAR_READ_ANNOTATIONS` / `_TABULAR_OPEN_READ_ANNOTATIONS`
+  / `_TABULAR_WRITE_ANNOTATIONS`. Each of the 8 tools now declares
+  its own `ToolAnnotations(...)` literal in its `metadata` property,
+  matching the kaos-reference / kaos-citations pattern. Eliminates
+  the misclassification-via-shared-constant risk that motivated
+  review #4 in the first place.
+
+Tests: 173 → **189** unit tests, 32 integration tests still green,
+coverage 75% → 73% (more code under coverage tracking; gate still
+above the 70% floor).
+
+### Deferred to next release (tracked, not blocking 0.1.0a1)
+
+- Make `INSTALL sqlite` / `LOAD sqlite` opt-in via a settings flag
+  (post-release-review #8). Currently the actionable error path is
+  in place (KTAB-010); making the network fetch opt-in is a real
+  API change worth doing in a settled release.
+- Include `SECURITY.md` in the sdist (post-release-review #9). Cheap
+  to do at the cross-package level alongside other sdist policy.
+- Pin GitHub Actions and gitleaks Docker image references to SHAs
+  for stronger supply-chain posture (post-release-review #10). Best
+  done as a platform-wide sweep across all kaos-* repos at once.
+
+## [0.1.0a1-original] — superseded entries below
+
+The remainder of this entry documents the pre-review release
+preparation; left intact so the audit-01 / OSS Phase A trail is
+preserved.
+
+First public alpha. DuckDB-powered tabular data engine with 8 MCP
+tools for register / query / describe / list / sample / count /
+export / read-file workflows. Closes every finding in
+`docs/audit-01/kaos-tabular.md` (KTAB-001..KTAB-010).
+
+### Removed (dep minimization)
+
+- **`polars` dropped from required dependencies.** A pre-release
+  audit confirmed nothing in `kaos_tabular` source or tests imports
+  polars; the DuckDB bridge in `kaos-content` doesn't need it
+  either (the polars bridge lives behind kaos-content's own
+  `[polars]` extra, which kaos-tabular never pulled). Result: the
+  resolved tree shrinks 56 → 54 packages and the install no longer
+  fetches the polars + polars-runtime-32 native binaries (~30 MB
+  combined). The `polars` keyword and the README polars mentions
+  are also dropped.
+
+### Compliance
+
+- **License audit (50 distinct deps in the resolved tree).** Every
+  inbound license is on the `docs/oss/10-licensing-legal/dep-license-policy.md`
+  allowlist: MIT, Apache-2.0, BSD-2/3-Clause, ISC, MPL-2.0 (certifi,
+  weak-copyleft permitted), PSF-2.0 (typing-extensions). Zero
+  matches against the denylist (GPL family, AGPL family,
+  Commons-Clause, SSPL, BUSL, anyone else's proprietary). Audit
+  evidence: `uv tree --no-dedupe` × per-PyPI license metadata.
+
+### Added
+
+- **`LICENSE`, `NOTICE`, `CHANGELOG.md`** seeded for the public release.
+  License flips from `LicenseRef-Proprietary` to Apache-2.0 via PEP 639
+  (`license = "Apache-2.0"`, `license-files = ["LICENSE", "NOTICE"]`).
+  `License ::` classifier removed (PEP 639 supersedes).
+
+- **`TabularEngine.export_table(table_name, output_path, format=...)`**
+  — public engine method that owns DuckDB COPY, format mapping, and
+  path quoting. ExportTool MCP and `kaos-tabular export` CLI now call
+  it instead of reaching into `engine._con` and importing the private
+  `kaos_content.bridges.duckdb._quote_ident`. Closes audit-01 KTAB-003.
+
+- **`docs/security.md`** — canonical statement of the trust contract
+  (DuckDB is in-process; SQL has filesystem access matching the running
+  process; deployments wanting stricter isolation should run
+  kaos-tabular in a constrained working directory or container; the
+  strict-isolation alternative is `kaos_content.bridges.duckdb.create_safe_connection`,
+  which cannot register files). Closes audit-01 KTAB-001 alongside the
+  description honesty fix.
+
+- **`kaos_tabular/py.typed`** marker so the `Typing :: Typed` classifier
+  is honored by downstream type checkers. Closes audit-01 KTAB-004.
+
+- **`benchmark` pytest marker** registered in `pyproject.toml`. Wall-
+  clock performance tests relocated from `tests/unit/test_adversarial.py`
+  → `tests/benchmarks/test_engine_perf.py`. Bounded unit gates can now
+  exclude them with `-m "not benchmark"`. Closes audit-01 KTAB-006.
+
+- **`tests/unit/test_sqlite_register.py`** — positive (real SQLite
+  fixture) and negative (forced INSTALL/LOAD failure) coverage for the
+  new SQLite registration error path. Closes audit-01 KTAB-010.
+
+- **`tests/unit/test_serve.py`** — argparse + import-error coverage for
+  `kaos_tabular.serve.main`, lifting `serve.py` from 0% to ~55% and
+  total coverage from 63% (audit baseline) to 73%.
+
+- **`fail_under = 70` coverage gate** in
+  `[tool.coverage.report]`. Locks the new floor against regression.
+  Closes audit-01 KTAB-005.
+
+### Changed
+
+- **`QueryTool.metadata.description` is now honest** about the trust
+  contract: "Execute arbitrary DuckDB SQL against the session's
+  in-process engine ... SQL has filesystem access matching the running
+  process — for stricter isolation, run kaos-tabular in a constrained
+  working directory or container." Previously the description claimed
+  "queries against registered tables" while the engine accepted
+  arbitrary DuckDB SQL including `read_csv_auto('...')`. Closes
+  audit-01 KTAB-001.
+
+- **`_register_sqlite` now raises `RegistrationError` with a 3-part
+  message** when DuckDB's `INSTALL sqlite` / `LOAD sqlite` fails. The
+  message names the install command, the offline workaround
+  (pre-bundled extension), and the fallback (export tables to CSV /
+  Parquet first). Closes audit-01 KTAB-010.
+
+- **MCP error messages standardized to the what / how-to-fix /
+  alternative-tool shape** across `tools.py`. The audit explicitly
+  flagged the sample (`tools.py:359`) and read-file (`tools.py:489`)
+  errors as incomplete; both rewritten plus the file-not-found, no-
+  tables-registered, and register-failed paths. Closes audit-01
+  KTAB-007.
+
+- **Stale comment in `tests/unit/test_tools.py`** removed. The module
+  docstring claimed "Several tools have a bug where _get_engine(context)
+  is called without await" — current source awaits correctly. Closes
+  audit-01 KTAB-009.
+
+### Removed
+
+- **`[xlsx]` extra and `_register_xlsx` method dropped.** Both
+  introduced an undocumented sideways
+  `kaos-tabular -> kaos-office` extraction-module dependency that the
+  architecture DAG explicitly forbids. Callers wanting XLSX support
+  parse the file with `kaos_office.parse_xlsx(path)` (in kaos-office,
+  which is the right home for OPC reading) and pass each `Table` to
+  `engine.register_table(table, name=...)` (already public). The
+  workspace dependency on `kaos-office` is removed; `[tool.uv.sources]`
+  drops the kaos-office editable entry. Closes audit-01 KTAB-002.
+
+### Notes (audit findings already resolved)
+
+- **KTAB-008** — `kaos_tabular/__init__.py` `__all__` is already
+  alphabetically sorted under Python's default ordering (uppercase <
+  underscore < lowercase per ASCII). No change needed; documented here
+  as verified against `sorted()`.
+
+[Unreleased]: https://github.com/273v/kaos-tabular/compare/v0.1.0a1...HEAD
+[0.1.0a1]: https://github.com/273v/kaos-tabular/releases/tag/v0.1.0a1