agnes-the-ai-analyst/app/instance_config.py
Vojtech d6ad08f107
Flea-market upload guardrails + soft delete + JOIN-based admin queue (#233)
* feat(store): flea-market upload guardrails + soft delete + JOIN-based admin queue

Adds an end-to-end guardrails pipeline for store uploads (manifest +
static-security + LLM review), persists blocked bundles for forensics,
introduces soft-delete (Archive) semantics, consolidates the legacy
/store/{id} surface into /marketplace/flea/{id}, and reworks the admin
queue so lifecycle filters read live entity visibility via LEFT JOIN
rather than a denormalized submission column.

Schema v29 → v35:
  * v29 store_submissions table + store_entities.visibility_status
  * v30 file_size, bundle_sha256, bundle_purged_at on submissions
  * v31 reshape store_submissions (drop legacy unique on entity_id)
  * v32 store_entities.archived_at/by + 'archived' visibility value
  * v33 drop store_submissions.retry_count (unused)
  * v34 ensure idx_store_submissions_entity exists post column-drop
  * v35 broaden visibility_status enum + JOIN architecture cutover

Pipeline (src/store_guardrails/):
  * Inline checks: manifest_check, static_scan, quality_check
  * LLM review configurable haiku|sonnet|opus (default haiku)
  * BackgroundTasks-driven async path with structured-output JSON
  * Per-submitter daily quota (default 50)
  * 30-day TTL purge job (POST /api/admin/run-blocked-purge)
  * Bundle SHA256 + size persisted; sha256 survives purge for forensics

Visibility model:
  * pending | approved | hidden | archived
  * _enforce_visibility returns 404 (no leak) for non-owner non-admin
  * Owner sees own non-approved entries via include_owner_id widening
  * Install refused with 409 entity_not_approved when not approved

Soft-delete (DELETE /api/store/entities/{id}):
  * Default = soft (visibility_status='archived'); existing installs
    keep getting served the bundle so users don't lose the plugin
  * ?hard=true admin-only: drops bundle + cascades user_store_installs
  * Hard-delete preserves entity_id on submission as tombstone so
    audit_log linkage survives for the activity timeline

Admin queue lifecycle (the JOIN refactor):
  * Verdict (store_submissions.status) is immutable forensic record
  * Lifecycle (store_entities.visibility_status) is live state
  * /admin/store/submissions Archived chip translates to
    `e.visibility_status='archived'` via LEFT JOIN — any path that
    flips visibility surfaces in the queue immediately
  * Detail page renders Status (verdict) and Entity lifecycle side by
    side so admins see "approved at review, now archived" at a glance

URL consolidation:
  * /store/{id} deleted (no redirect, stale bookmarks 404)
  * /marketplace/flea/{id} is the canonical detail surface
  * Three in-tree callers (upload-success, my-stack card, store
    listing card) updated to point at the new URL
  * Quarantine banner extracted to _quarantine_banner.html partial,
    self-guarded, included from both flea detail templates
  * Banner JS auto-refreshes when the verdict lands by polling
    /api/marketplace/flea/{id}/detail (visibility_status +
    submission_status — the latter is needed because blocked_llm
    keeps the entity at visibility_status='pending')

Audit log resource format:
  * runner.py emits prefixed `store_submission:{id}` (post-fix)
  * Detail-page timeline query handles three patterns: prefixed
    submission, helper-emitted `store_entity:{sub_id}`, and bare-id
    legacy rows — all surface in the activity timeline

UX fixes:
  * Owner sees Under review / Quarantined / Hidden banner with status
  * Install button gray-disabled (not blue) when non-approved
  * Owner cannot delete quarantined entries (403); admin can
  * Admin queue: filter chips, sortable columns, paging, page-size
  * Auto-refresh queue every 5s while pending rows are visible
  * Store upload page file picker no longer opens twice (label →
    input default action collided with explicit JS handler)

Tests: 168 passed across the guardrails suites (admin submissions,
store API, inline / LLM / purge guardrails, store repositories,
marketplace filter, schema version). New regression coverage
includes: archive surfaces via JOIN even when API path is bypassed;
deleted submission renders activity timeline (tombstone); flea
detail surfaces submission_status only for owner/admin; detail page
renders Entity lifecycle row; audit log resource format covers both
helper and runner paths.

* fix(store-guardrails): PR #233 follow-up — prompt injection, atomic PUT, BG race, schema, reaper, sort whitelist

Addresses 9 of the 23 findings from the PR #233 review (spec at
docs/superpowers/specs/2026-05-09-pr233-guardrails-fixes-spec.md).
Merge-gate items #1-#6 plus high-value mediums #7, #9-#12, #23.
Architectural items (#8 enum split, #14 factory) and pure
maintainability (#15-#22) deferred to follow-ups.

Security:
* #1 prompt injection — SYSTEM_PROMPT now passed via the SDK's
  dedicated system= parameter; bundle wrapped in <bundle>...</bundle>
  sentinels declared data-only by the system prompt; literal
  sentinel strings in user content are escaped so an adversarial
  README can't forge a close tag.
* #6 static scan honesty — module docstring + admin copy + docs
  declare static scan as signal not gate; .md/.txt/.rst/.html/.json/
  .yaml/.yml/.toml skipped to avoid false positives on prose.
  AST mode for Python deferred (separate flag, FP comparison work).

Correctness:
* #2 PUT atomicity — bundles bake into plugin.staging-<rand>/
  alongside live, atomic-rename on success; failed checks leave
  live tree byte-for-byte intact.
* #3 BG-task race — set_visibility_if_pending guards verdict flips
  to the (pending, hidden) review window; admin archives during
  review survive; skipped flips audit-logged.
* #4 v35 NOT NULL/DEFAULT — schema v35→v36 re-applies them on
  store_entities.visibility_status. CHECK constraint enforced
  application-side (DuckDB ADD CHECK on existing column unsupported).
* #7 stuck-review reaper — reap_stuck_llm_reviews flips pending_llm
  rows older than guardrails.stuck_review_grace_seconds (default
  1800) to review_error. Scheduler runs every 15 min via new
  /api/admin/run-reap-stuck-reviews. Set knob to 0 to disable.
* #9 quota counter — count_blocked_for_submitter_since now counts
  blocked_inline + blocked_llm + review_error so a submitter
  triggering only LLM-blocked verdicts is bounded.
* #10 missing risk_level — surfaces as review_error with
  error='missing_risk_level' instead of silently defaulting to
  'medium' (which looked like a model-decided block).
* #11 archived_at clear — set_visibility nulls archived_at +
  archived_by when transitioning out of 'archived' so a future
  read doesn't show stale archive forensics on an approved row.

Maintainability:
* #12 FSM doc comment — accurate insert/transition/lifecycle
  description in src/db.py near store_submissions schema.
* #23 sort-key whitelist — admin queue rejects unknown sort keys
  with 400 invalid_sort_key; substring-replace footgun removed.

Deferred (separate PRs):
* #5 quota race — proper fix requires asyncio.Lock spanning the
  full pipeline; threading.Lock blocks event loop, DuckDB MVCC
  doesn't help. API-level slowapi bounds worst case for now.
* #6 part 3 (AST static scan), #8 (enum split), #13 (import
  bundle docs), #14 (factory consolidation), #15-#22 (maint).

Tests:
* New: tests/test_store_guardrails_prompt_injection.py (corpus +
  trust-boundary invariants), tests/test_store_put_atomic.py,
  tests/test_store_guardrails_reaper.py.
* Extended: test_store_guardrails_llm.py (system param, missing
  risk_level, BG race), test_admin_store_submissions.py (quota
  counter widening, sort whitelist 400), test_store_repositories.py
  (un-archive metadata clear), test_db_schema_version.py (v36).
* Full suite: 3738 passed; 17 pre-existing baseline failures
  unchanged (db migration tests, cli binary rename, catalog export,
  user mgmt v5 backfill — confirmed by stash + rerun on clean tree).
2026-05-09 17:32:53 +04:00

384 lines
15 KiB
Python

"""Instance configuration — loads instance.yaml and exposes to FastAPI."""
import logging
import os
from pathlib import Path
from typing import Any, Optional
logger = logging.getLogger(__name__)
_instance_config: Optional[dict] = None
def reset_cache() -> None:
"""Drop the in-process instance.yaml cache; the next ``load_instance_config``
call re-reads from disk. Used by `/api/admin/server-config` after a save.
Public alias so callers don't have to reach into the private global.
Also clears ``connectors.bigquery.access.get_bq_access`` so the v2 endpoints
pick up new BigQuery project IDs after an admin saves `instance.yaml` —
without this, `get_bq_access`'s `@functools.cache` would freeze the projects
at first call and require a container restart to pick up changes (Devin
ANALYSIS_0004 on PR #138). Lazy-imported so this module stays usable in
environments where the connectors package can't be imported (e.g. unit
tests of instance_config in isolation)."""
global _instance_config
_instance_config = None
try:
from connectors.bigquery.access import get_bq_access
get_bq_access.cache_clear()
except Exception:
# Connectors module not loaded yet, or BQ deps missing — both fine.
pass
def _deep_merge(base: dict, patch: dict) -> dict:
"""Deep-merge `patch` into `base`, returning a new dict.
Dict-into-dict recurses; everything else (scalars, lists, None) is
replaced wholesale. Used so the writable overlay can hold only the
sections an operator has touched, while everything else flows from
the static file unchanged. Same semantics as the helper in
`/api/admin/server-config`'s POST handler.
"""
out = dict(base)
for key, value in patch.items():
if isinstance(value, dict) and isinstance(out.get(key), dict):
out[key] = _deep_merge(out[key], value)
else:
out[key] = value
return out
def load_instance_config() -> dict:
"""Load instance.yaml as a deep-merge of the static file and the
writable overlay.
Resolution:
1. Static base: ``CONFIG_DIR/instance.yaml`` via ``config.loader``
(the source of truth for sections the editor doesn't expose —
``datasets``, ``corporate_memory``, ``openmetadata``, etc.).
2. Overlay patch: ``DATA_DIR/state/instance.yaml`` (written by
``/api/admin/configure`` and ``/api/admin/server-config``;
contains only the sections those endpoints accept).
3. Overlay wins per-leaf via deep-merge — operator edits persist,
static-only sections still flow through.
Pre-2026-04-28 this function returned the overlay verbatim when it
existed and only fell back to static when it didn't. That was a
silent footgun: the moment someone saved any section through the
new editor (which writes a narrow overlay by design), every
consumer of static-only sections (corporate memory page, dataset
list, OpenMetadata client) saw empty defaults. See PR #107.
"""
global _instance_config
if _instance_config is not None:
return _instance_config
import yaml
# Static base — strict validation lives in config.loader.
base: dict = {}
try:
from config.loader import load_instance_config as _load
base = _load() or {}
logger.info("Loaded instance.yaml base from config/")
except Exception as e:
logger.warning(f"Could not load static instance.yaml: {e}")
# Overlay patch from the writable volume. Best-effort — a corrupt
# overlay shouldn't take the app offline (we'd rather serve stale/base
# config than 500 every request), but log loudly with a traceback so
# the corruption surfaces in the operator's logs immediately. The
# write-side endpoints (POST /api/admin/server-config and /configure)
# refuse to overwrite a corrupt overlay with HTTP 500, so an admin
# noticing the saves break is the second line of defence.
#
# ${ENV_VAR} interpolation: ``config.loader.load_instance_config`` runs
# the static base through ``_resolve_env_refs`` already, but raw
# ``yaml.safe_load`` here would leave overlay strings like
# ``${ANTHROPIC_API_KEY}`` as literal placeholders. /api/admin/configure
# writes exactly that string into the seeded ai: block (#176), so we
# mirror the resolver here before the deep-merge — without it, the
# LLM factory receives the literal placeholder and rejects it as an
# invalid api key (#179 review fix).
# Resolve via _state_dir() so the path matches the writer in
# app/api/admin.py — under the flat-mount layout (STATE_DIR=/data-state)
# both the configure-endpoint and the server-config-endpoint write
# ``/data-state/instance.yaml``; reading from ``/data/state/...`` here
# would silently load stale config from the regenerable data disk.
from app.secrets import _state_dir
overlay_path = _state_dir() / "instance.yaml"
if overlay_path.exists():
try:
overlay = yaml.safe_load(overlay_path.read_text()) or {}
from config.loader import _resolve_env_refs
overlay = _resolve_env_refs(overlay)
base = _deep_merge(base, overlay)
logger.info("Merged overlay from %s", overlay_path)
except Exception:
logger.exception(
"instance.yaml overlay at %s is corrupt — falling back to "
"static base config; saves through the editor will refuse "
"until the file is repaired", overlay_path,
)
_instance_config = base
return _instance_config
def get_value(*keys, default=None) -> Any:
"""Get nested value from instance config."""
config = load_instance_config()
current = config
for key in keys:
if isinstance(current, dict):
current = current.get(key)
else:
return default
if current is None:
return default
return current
def get_data_source_type() -> str:
return os.environ.get("DATA_SOURCE", get_value("data_source", "type", default="local"))
def get_home_route() -> str:
"""Path that ``/`` redirects to for an authenticated user.
Resolution order: ``AGNES_HOME_ROUTE`` env var (Terraform-friendly,
overrides everything) > ``instance.home_route`` in instance.yaml >
default ``/dashboard``. The env-overrides-yaml shape mirrors
:func:`get_data_source_type` (precedent in this file) so operators
can flip a fork to ``/home`` per-deployment without forking the
YAML.
Validated to start with ``/`` and not ``//`` so a misconfigured
value can't pivot the root redirect to an external host.
"""
raw = os.environ.get("AGNES_HOME_ROUTE") or get_value(
"instance", "home_route", default="/dashboard"
)
route = (raw or "").strip()
if not route.startswith("/") or route.startswith("//"):
return "/dashboard"
return route
def get_gws_oauth_credentials() -> dict:
"""Pre-configured Google Workspace CLI OAuth client (client_id + secret).
When set, /home renders a connector prompt that tells the analyst (and
Claude) to export `GOOGLE_WORKSPACE_CLI_CLIENT_ID` and
`GOOGLE_WORKSPACE_CLI_CLIENT_SECRET` and skip the "create your own GCP
project" walkthrough — the operator has already provisioned a shared
OAuth app for the instance. When unset, the prompt falls back to the
manual `gws auth setup` flow.
OAuth client_id + secret here are app identifiers for an installed
"Desktop app" OAuth client, not a per-user secret. They're rendered
into the public /home page on purpose — they identify the OAuth app,
and the redirect-URI / scope guardrails on the GCP-side OAuth client
are what enforce safety. Treat them like a publishable bundle ID,
not a credential.
Resolution order (env-overrides-yaml, mirrors :func:`get_home_route`):
- ``AGNES_GWS_CLIENT_ID`` env > ``instance.gws.client_id`` YAML > None
- ``AGNES_GWS_CLIENT_SECRET`` env > ``instance.gws.client_secret`` YAML > None
- ``AGNES_GWS_OAUTHLIB_INSECURE_TRANSPORT`` env > ``instance.gws.oauthlib_insecure_transport`` YAML > "1"
(kept as "1" by default because the gws CLI binds an HTTP loopback
on 127.0.0.1:8080 for the OAuth redirect, and Google's oauthlib
refuses non-HTTPS redirects without this flag).
Both id and secret must be set for the configured branch to engage;
a half-configured instance falls back to manual setup with a warning.
"""
cid = os.environ.get("AGNES_GWS_CLIENT_ID") or get_value(
"instance", "gws", "client_id", default=""
)
secret = os.environ.get("AGNES_GWS_CLIENT_SECRET") or get_value(
"instance", "gws", "client_secret", default=""
)
insecure = os.environ.get("AGNES_GWS_OAUTHLIB_INSECURE_TRANSPORT") or get_value(
"instance", "gws", "oauthlib_insecure_transport", default="1"
)
project_id = os.environ.get("AGNES_GWS_PROJECT_ID") or get_value(
"instance", "gws", "project_id", default=""
)
cid = (cid or "").strip()
secret = (secret or "").strip()
project_id = (project_id or "").strip()
# Derive project_id from the client_id when not explicitly set. Google's
# OAuth client_id format is "<numeric-project-number>-<random>.apps.
# googleusercontent.com"; the numeric prefix is required by the
# client_secret.json schema (gws CLI's Rust struct treats it as
# non-Option). Falls back to "" when the client_id is empty or
# malformed; the configured branch in the template degrades gracefully.
if not project_id and cid and "-" in cid:
project_id = cid.split("-", 1)[0]
return {
"client_id": cid,
"client_secret": secret,
"project_id": project_id,
"oauthlib_insecure_transport": str(insecure).strip() or "1",
"configured": bool(cid and secret),
}
def get_home_automode_visibility() -> bool:
"""Whether /home renders the "Step 3 — turn on auto-accept mode"
install-block. Auto-accept mode is the recommended middle ground
between default per-action prompting (slow) and full YOLO
(`--dangerously-skip-permissions`, broad blast radius).
Cautious-rollout instances can hide the section by setting
``AGNES_HOME_SHOW_AUTOMODE=0`` so users learn the permission flow
first; the same content stays available on /setup-advanced.
Resolution: env var > ``instance.home.show_automode`` YAML > True.
Mirrors :func:`get_home_route` shape so Terraform overrides work
the same way.
"""
raw = os.environ.get("AGNES_HOME_SHOW_AUTOMODE")
if raw is None:
raw = get_value("instance", "home", "show_automode", default=True)
if isinstance(raw, bool):
return raw
return str(raw).strip().lower() not in ("0", "false", "no", "off", "")
def get_instance_name() -> str:
return get_value("instance", "name", default="AI Data Analyst")
def get_instance_subtitle() -> str:
return get_value("instance", "subtitle", default="")
def get_sync_interval() -> str:
"""Human-readable refresh cadence shown in the analyst welcome prompt."""
return get_value("instance", "sync_interval", default="1 hour")
def get_allowed_domains() -> list:
domain = get_value("auth", "allowed_domain", default="")
if domain:
return [d.strip() for d in domain.split(",") if d.strip()]
return []
def get_datasets() -> dict:
return get_value("datasets", default={})
def get_theme() -> dict:
return get_value("theme", default={})
def get_auth_config() -> dict:
return get_value("auth", default={})
def get_corporate_memory_config() -> dict:
return get_value("corporate_memory", default={})
def get_guardrails_config() -> dict:
"""Flea-market upload-guardrail config (see docs/STORE_GUARDRAILS.md).
Returns the ``guardrails:`` block from instance.yaml, or an empty dict
when not configured. Call site: ``src/store_guardrails/runner.py``.
"""
return get_value("guardrails", default={})
def get_guardrails_review_model() -> str:
"""Resolved Anthropic model ID used for the LLM security review.
Reads ``guardrails.review_model`` (one of ``haiku``, ``sonnet``,
``opus``, or a concrete ``claude-*`` model ID) and returns the
concrete model ID. Defaults to Haiku — the cheapest tier — when the
operator hasn't set the key. Override per-instance for higher-stakes
review at proportionally higher cost.
"""
from connectors.llm.factory import resolve_model_tier
raw = get_value("guardrails", "review_model", default="haiku")
return resolve_model_tier(raw)
def get_guardrails_blocked_quota_per_day() -> int:
"""Per-submitter cap on `blocked_inline` rows in the trailing 24h.
Defaults to 50. Set to 0 in instance.yaml to disable the quota
entirely (useful for trusted single-tenant deployments). Bounds the
worst case where a bot loops on malformed ZIPs and fills disk +
the admin queue with noise.
"""
val = get_value("guardrails", "blocked_quota_per_day", default=50)
try:
return max(0, int(val))
except (TypeError, ValueError):
return 50
def get_guardrails_blocked_bundle_ttl_days() -> int:
"""How many days to keep a blocked bundle's bytes on disk.
Default 30. The submission row + sha256 + size always survive — only
the bundle bytes get removed. ``bundle_purged_at`` is stamped so the
detail UI renders *"Bundle purged on …"*. Set to 0 to disable the
TTL purge entirely (bundles persist indefinitely until manual
Delete).
"""
val = get_value("guardrails", "blocked_bundle_ttl_days", default=30)
try:
return max(0, int(val))
except (TypeError, ValueError):
return 30
def get_guardrails_stuck_review_grace_seconds() -> int:
"""How long a submission may stay at ``status='pending_llm'`` before
the reaper flips it to ``review_error``.
The BackgroundTasks worker normally writes a verdict within a few
seconds. If the worker crashes between status flip and verdict
write, the row would otherwise sit at pending_llm forever — admin
queue surfaces it indefinitely; submitter never gets a verdict.
Default 1800s (30 min) comfortably exceeds the Sonnet/Opus p99
wall time for the configured ``MAX_REVIEW_BYTES`` payload. Set to
0 to disable the reaper entirely.
"""
val = get_value("guardrails", "stuck_review_grace_seconds", default=1800)
try:
return max(0, int(val))
except (TypeError, ValueError):
return 1800
def get_guardrails_enabled() -> bool:
"""Master kill-switch for the guardrail pipeline.
Defaults to True. Operators can disable by setting ``guardrails.enabled:
false`` in instance.yaml — useful for local development against the
UI without burning Anthropic tokens. Inline checks always run; this
flag only gates the LLM step (and skips the pending → approved hold).
Auto-fallback: when the YAML says enabled but no ANTHROPIC_API_KEY /
LLM_API_KEY is set in the environment, behave as disabled. This
keeps the test suite + first-boot operator experience sane — uploads
auto-approve until the operator wires up an LLM provider rather than
silently piling up in ``review_error``.
"""
if not bool(get_value("guardrails", "enabled", default=True)):
return False
if os.environ.get("ANTHROPIC_API_KEY", "").strip():
return True
if os.environ.get("LLM_API_KEY", "").strip():
return True
return False