agnes-the-ai-analyst/pyproject.toml
Vojtech fb6e930bc9
feat(store-guardrails): per-component description quality + plain-language UX (#276)
* feat(store-guardrails): enforce per-component description quality

Two-tier hard guardrail on flea-market submissions. Empty / placeholder /
single-word descriptions now block before any LLM call; vague-but-passes-
floor descriptions block on the substantive LLM review layer.

Tier 1 — inline mechanical check (src/store_guardrails/content_check.py).
Walks the baked plugin tree, evaluates each component (plugin manifest,
agents, skills, commands) plus the submission-level form description
against a 60-char / 25-char (commands) / 5-distinct-word / 200-char-body
floor with a placeholder denylist (TODO, TBD, {{var}}, etc.). Floors
calibrated against real ecosystem norms: Claude / superpowers /
compound-engineering skill packs cluster 150–220 chars, npm / Docker /
VS Code at 100–120. InlineResult.passed now ANDs in content.status.

Tier 2 — LLM review extension (prompts.py + llm_review.py). System
prompt gains a content-quality criterion; REVIEW_JSON_SCHEMA carries a
content_quality {verdict, issues[]} object alongside the existing
security findings. is_safe() requires content_quality.verdict == 'pass'.
Single LLM call covers both dimensions. MAX_RESPONSE_TOKENS bumped
2000 → 2500 for the extra payload. Verdicts missing content_quality
treated as pass (backwards compat with already-recorded rows).

Submitter UX:
- /store/new wizard now carries a "Before you upload — what passes
  review" collapsible disclosure on both step 1 and step 2 with the
  bar + patterns that work. Live char counter on the description
  field. Per-component preview table (green/red dots from the new
  summarize_for_preview helper) renders after the ZIP /preview round
  trip, scoping each finding to its file.
- New /store/examples page with rejected/passes pairs for skill /
  agent / plugin / command plus a "Why these limits" research table.
  Anchored sections (#skill / #agent / #plugin / #command) so the
  rejection banner can deep-link by component_type.
- Quarantine banner _content_findings.html groups findings by file
  (one "See <type> example ↗" per component, not per field) and
  translates field codes (frontmatter.description / body / etc.) to
  plain-English labels. _content_howto_fix.html surfaces a static
  "Re-upload as new version" + "See examples" action row beneath any
  content failure on the entity detail page.
- _parse_frontmatter moved to src/store_guardrails/_frontmatter.py so
  the new check module shares the parser without inverting the
  app → src dependency direction.

Tests:
- New tests/test_store_guardrails_content.py (29 cases) covering
  every failure code per component type plus submission-level checks
  and the summarize_components / summarize_for_preview helpers.
- Extended test_store_guardrails_inline.py for the new
  InlineResult.content field + aggregate behaviour.
- Extended test_store_guardrails_llm.py for the new
  content_quality verdict pathways (fail blocks, missing field passes).
- Backfilled fixture descriptions across test_store_api.py,
  test_store_entity_versions.py, test_store_put_atomic.py,
  test_admin_store_submissions.py, test_marketplace_api.py,
  test_marketplace_v32_endpoints.py so existing happy-path tests
  clear the new 60-char floor.

* fix(content-guardrail): align agents walker with preview + drop import-time .format()

Two cleanups from the takeover review on #276 (vr/guardrails-content).

1) `_iter_components` for agents now skips files lacking frontmatter
   (no `name` AND no `description`). Pre-fix the walker greedily
   evaluated every `*.md` under `agents/` — `agents/README.md` and
   helper docs got flagged as "frontmatter.description empty"
   rejections. Worse: `summarize_for_preview` for `type=agent` ALREADY
   filters the same shape, so the upload preview gave a green dot
   while the post-bake check gave a red rejection on submit. Two new
   regression tests in TestAgentsWalkerSkipsNonAgentFiles pin both
   shapes (README + _NOTES.md) so the preview/check parity stays
   aligned.

2) `body_too_short` hints now use the same runtime-kwarg substitution
   pattern as every other hint in the table. Pre-fix the skill +
   agent body_too_short hints called `.format(min_chars=_MIN_BODY_CHARS)`
   at module-load time, but the call site `_hint_for(type_,
   "body_too_short")` didn't pass `min_chars=`, so the format() was
   just baking the constant at import. Cosmetic inconsistency; pass
   `min_chars=_MIN_BODY_CHARS` at the call site instead and let
   `_hint_for` do the substitution like it does for `too_short`.

Verified end-to-end:
- New TestAgentsWalkerSkipsNonAgentFiles cases fail on the unfixed
  walker (verified by reverting to the pre-fix file and re-running);
  pass cleanly after the fix.
- Full content-guardrail suite: 25/25 (23 existing + 2 new).
- Full pytest: 4189 passed, 25 skipped.

* release: 0.53.5 — content guardrail (flea-market submitter UX) + catalog ENTITY column + BQ hint dispatch

Bundles three threads landed in [Unreleased]:
- Vojta's flea-market content guardrail (two-tier mechanical + LLM)
- Zdeněk's `agnes catalog` ENTITY column replacement for FLAVOR
- Zdeněk's `/api/query` remote_estimate_failed hint dispatch fix

Plus the takeover hygiene from #276 review (agents walker preview/check
parity + body_too_short hint runtime kwarg consistency) and the
backslash-escape fix follow-up to v0.53.4 #275.

No DB migration; no API change. Patch upgrade lands transparently.
Upload form's new "Before you upload" disclosure + per-component preview
table appear on the next dev-VM auto-pull. Quarantine banner now groups
findings by file with "See <type> example ↗" deep-links to the new
/store/examples reference page.

---------

Co-authored-by: ZdenekSrotyr <zdenek.srotyr@keboola.com>
2026-05-12 21:48:27 +02:00

182 lines
7.9 KiB
TOML
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

[project]
name = "agnes-the-ai-analyst"
version = "0.53.5"
description = "Agnes — AI Data Analyst platform for AI analytical systems"
requires-python = ">=3.11,<3.14"
license = "MIT"
readme = "README.md"
dependencies = [
# Core database
# 1.5.2 fixes a FK-dependency regression that affected ALTER TABLE on
# tables referenced by other tables — broke the test_db migration
# ladder replay on 1.5.1. CI runs 1.5.2; local devs need it too.
"duckdb>=1.5.2",
# Web framework (FastAPI)
"fastapi>=0.115.0",
"uvicorn[standard]>=0.32.0",
"python-multipart>=0.0.27",
"jinja2>=3.1.0",
"starlette>=0.41.0",
# Authentication
"PyJWT>=2.8.0",
"itsdangerous>=2.1.0",
"authlib>=1.6.11",
"argon2-cffi>=23.1.0",
# HTTP client. `h2` enables HTTP/2 multiplexing for the persistent
# CLI client used by `agnes pull` (one TCP connection serves N
# concurrent parquet streams + range chunks). `cli/client.py`
# gracefully falls back to HTTP/1.1 if h2 is missing, so this
# extra is for performance, not correctness.
"httpx>=0.27.0",
"h2>=4.1.0",
# CLI
"typer>=0.12.0",
"rich>=13.0.0",
# Configuration
"python-dotenv>=1.0.0",
"pyyaml>=6.0",
# Data processing
"pandas>=2.0.0",
"pyarrow>=12.0.0",
"pytz>=2024.1",
# SQL parsing — server-side WHERE validator for /api/v2/scan (app/api/where_validator.py)
# Minimum 30.x — older versions had walk() yielding (node, parent, key)
# tuples instead of expression nodes, which would silently bypass the
# WHERE-validator structural checks (isinstance(tuple, exp.Subquery)
# is always False). 30.x yields nodes directly.
"sqlglot>=30.0.0",
# Data source connectors
"google-cloud-bigquery>=3.0.0",
"google-cloud-bigquery-storage>=2.0.0",
# Google Workspace Cloud Identity / Admin SDK (Workspace group membership sync)
"google-api-python-client>=2.0.0",
# Profiler visualizations
"matplotlib>=3.8.0",
"numpy>=1.24.0",
# Claude Code marketplace endpoint — pure-Python git server mounted in FastAPI
"dulwich>=0.22.0",
"a2wsgi>=1.10.0",
# In-process TTL cache for marketplace etag (transitively present via
# google-auth, declared explicitly here because we depend on it directly).
"cachetools>=5.3.0",
# Per-IP rate limiting on auth endpoints (#45). In-process counters by
# default — fine for single-replica deploys. Multi-replica rollouts can
# swap the storage backend via slowapi's `storage_uri` (Redis, Memcached).
"slowapi>=0.1.9",
# LLM provider SDKs — core (not dev) because connectors/llm/*_provider.py
# is imported by services/{corporate_memory, verification_detector} which
# the scheduler drives in production. Promoted from [dev] in #176 to fix
# ModuleNotFoundError boot loops on default Compose deploys.
"anthropic>=0.30.0",
"openai>=1.30.0",
# Keboola Storage API SDK — used by:
# - `connectors/keboola/client.py` for admin-side bucket / table list
# (consumed from `app/api/admin.py` discover-and-register, table
# metadata refresh).
# Extraction itself uses the lightweight `connectors/keboola/storage_api.py`
# module (export-async + signed-URL download) which talks to Storage API
# directly via `requests` — no SDK dependency on the data-path side. The
# SDK stays for the metadata reads.
#
# NOTE: kbcstorage moved to the [server] extra below — see the rationale
# in [project.optional-dependencies].server. CLI wheels installed via
# `uv tool install` deliberately ship without it.
"sse-starlette>=2.0",
# Optional observability — pure-Python, no compilation. Lazily initialized
# in src/observability/posthog_client.py and only emits events when
# POSTHOG_API_KEY is set in the environment. With the key unset the
# integration is fully off (no network, no init). See docs/observability.md.
"posthog>=3.7.0",
# Rust-backed (ammonia) HTML sanitizer for admin-edited rich content
# (news intro + body, curated marketplace-metadata.json descriptions).
# Allowlist-based with per-tag attribute scoping; closes the bypass
# shapes the legacy regex sanitizer in src/welcome_template.py was
# vulnerable to. Pre-built wheels published for all supported
# (mac/linux/windows × arm64/x86_64) targets.
"nh3>=0.2",
# CommonMark markdown renderer for curator-authored marketplace-metadata.json
# rich content (plugin description / sample_interaction.assistant). Pure
# Python, no compilation. Rendered output is funneled through nh3 above.
"markdown-it-py>=3.0",
# Cross-platform advisory file locking for the `agnes push` single-instance
# guard. Wraps fcntl.flock on POSIX and msvcrt.locking on Windows behind
# a uniform API; OS releases the lock automatically on process exit (no
# stale-lock detection required). Used by cli/lib/push_lock.py.
"filelock>=3.13,<4",
# Transitive dependency hardened directly to dodge 5 dependabot advisories
# (4 high, 1 medium) flagged on urllib3<2.7.0: cross-origin sensitive
# header leak on proxied low-level redirects, decompression-bomb bypass
# + unbounded decompression chain on the streaming API, redirects-when-
# retries-disabled. The `[server]` extra below adds kbcstorage which
# transitively caps urllib3<2.0.0; `[tool.uv] override-dependencies`
# forces 2.7+ in workspace installs (Dockerfile + dev). Wheel consumers
# who install only the CLI (`uv tool install <wheel>`) get no kbcstorage
# and no conflict.
"urllib3>=2.7.0",
]
[project.optional-dependencies]
# Server-side connectors. The CLI wheel does NOT need these — analysts who
# `uv tool install` the wheel never reach a kbcstorage import. Splitting it
# out keeps the wheel's METADATA `Requires-Dist` set free of the
# `kbcstorage<=0.9.5 → urllib3<2.0.0` cap that conflicts with our
# `urllib3>=2.7.0` security pin under any fresh resolver context (where
# `[tool.uv] override-dependencies` does NOT apply — see comment on
# [tool.uv] below). Server install pulls it in via Dockerfile's
# `uv pip install --system --no-cache .[server]`.
server = [
"kbcstorage>=0.9.0",
]
observability = [
# Already in base dependencies — listed here so operators who want to
# be explicit can `pip install -e ".[observability]"` and signal intent.
"posthog>=3.7.0",
]
dev = [
"pytest>=9.0.0",
"pytest-timeout>=2.0.0",
"pytest-xdist>=3.0.0",
"faker>=24.0.0",
# jsonschema validates the corporate-memory extraction-tool golden fixtures
# under tests/test_corporate_memory_v1.py (extraction.json, correction.json,
# confidence_calibration.json). Production code does not depend on it.
"jsonschema>=4.0.0",
# FastAPI debug toolbar — gated behind DEBUG=1 env var in app/main.py.
# Provides per-request panels (headers, routes, timer, profiling, etc.)
# for local development. Never loaded in production (no DEBUG=1 there).
"fastapi-debug-toolbar>=0.6.3",
]
[project.scripts]
agnes = "cli.main:main"
[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"
[tool.hatch.build.targets.wheel]
packages = ["app", "src", "connectors", "cli", "services", "config"]
[tool.ruff]
line-length = 120
target-version = "py313"
[tool.uv]
dev-dependencies = [
"pytest>=9.0.0",
"pytest-timeout>=2.0.0",
"pytest-xdist>=3.0.0",
"faker>=24.0.0",
"anthropic>=0.30.0",
"openai>=1.30.0",
"fastapi-debug-toolbar>=0.6.3",
]
# Override the urllib3<2.0.0 ceiling kbcstorage 0.9.5 declares (upstream
# hasn't relaxed it as of 2026-05-12 but the SDK works fine against
# urllib3 2.x in practice — we only use `Client` + `Tables` from it and
# both go through `requests`, which natively supports both lines). Lets
# the resolver pick a urllib3 line that closes Dependabot advisories
# CVE-2024-37891 / CVE-2025-{xxx}. See `urllib3>=2.7.0` in [project]
# dependencies above for the security rationale.
override-dependencies = ["urllib3>=2.7.0"]