agnes-the-ai-analyst/CLAUDE.md
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

36 KiB
Raw Blame History

AI Data Analyst

Open-source data distribution platform for AI analytical systems. Extracts data from sources into DuckDB, serves via FastAPI, and distributes parquets to analysts who use Claude Code for local analysis.

First-Time Setup

When a user opens this project for the first time, guide them through interactive setup:

Step 1: Gather Information

Ask the user for:

  1. Company domain (e.g., "acme.com") - used for Google OAuth
  2. Data source type: keboola / bigquery / csv
  3. Instance name (e.g., "Acme Data Analyst")

Step 2: Generate Configuration

  1. Copy config/instance.yaml.example to config/instance.yaml
  2. Fill in values from Step 1
  3. If Keboola: ask for Storage API token, stack URL, project ID
  4. Create .env from config/.env.template

Step 3: Register Tables

  1. Use the FastAPI admin API (POST /api/admin/register-table, then PUT /api/admin/registry/{id} for updates) or webapp UI to register tables
  2. Tables are stored in DuckDB table_registry with source_type, bucket, source_table, query_mode
  3. For migration from old format: python scripts/migrate_registry_to_duckdb.py

Step 4: Docker Deployment

docker compose up          # Start app + scheduler
docker compose --profile full up  # Include telegram bot

# HTTPS mode — Caddy + corporate-CA certs at /data/state/certs
docker compose -f docker-compose.yml -f docker-compose.prod.yml -f docker-compose.tls.yml \
    --profile tls up -d

See docs/DEPLOYMENT.mdTLS for cert provisioning + scripts/ops/agnes-tls-rotate.sh (daily refetch from TLS_FULLCHAIN_URL, SIGUSR1 reload on diff, no-op when unchanged). The infra repo's startup.sh installs this as a systemd timer automatically.

Project Structure

├── src/                    # Core engine
│   ├── db.py               # DuckDB schema (system.duckdb, analytics.duckdb)
│   ├── orchestrator.py     # SyncOrchestrator — ATTACHes extract.duckdb files
│   ├── repositories/       # DuckDB-backed CRUD (sync_state, table_registry, users, etc.)
│   ├── profiler.py         # Data profiling
│   └── catalog_export.py   # OpenMetadata catalog export
├── app/                    # FastAPI application
│   ├── main.py             # App setup, router registration
│   ├── api/                # REST API (sync, data, catalog, admin, auth)
│   └── web/                # HTML dashboard routes
├── connectors/             # Data source connectors (extract.duckdb contract)
│   ├── keboola/            # Keboola: extractor.py (DuckDB extension) + client.py (fallback)
│   ├── bigquery/           # BigQuery: extractor.py (remote-only via DuckDB BQ extension)
│   └── jira/               # Jira: webhook + incremental parquet → extract.duckdb
├── cli/                    # CLI tool (`agnes pull`, `agnes query`, `agnes admin`)
├── app/auth/               # Authentication (FastAPI-based providers)
├── services/               # Standalone services (scheduler, telegram_bot, ws_gateway, etc.)
├── server/                 # Legacy deployment infrastructure
├── scripts/                # Utility + migration scripts
├── config/                 # Configuration templates (instance.yaml.example)
├── docs/                   # Documentation + metric YAML definitions
└── tests/                  # Test suite (633 tests)

Architecture: extract.duckdb Contract

Every data source produces the same output:

/data/extracts/{source_name}/
├── extract.duckdb          ← _meta table + views
└── data/                   ← parquet files (local sources only)

Remote table support (_remote_attach)

Extractors with remote/passthrough tables (query_mode='remote') include a _remote_attach table in extract.duckdb so the orchestrator can re-ATTACH the external DuckDB extension at query time:

CREATE TABLE _remote_attach (
    alias     VARCHAR,  -- DuckDB alias used in views, e.g. 'kbc'
    extension VARCHAR,  -- Extension name, e.g. 'keboola'
    url       VARCHAR,  -- Connection URL
    token_env VARCHAR   -- Env-var name holding the auth token, OR empty for
                        -- extensions with built-in auth (e.g. BigQuery uses the
                        -- GCE metadata server — see `connectors/bigquery/auth.py`).
);

The orchestrator reads this table, installs/loads the extension, fetches the token (via token_env lookup, or via the extension-specific auth path when token_env=''), creates a session-scoped DuckDB SECRET when the extension requires one (BigQuery), and ATTACHes the external source. Views referencing kbc."bucket"."table" then resolve correctly. This mechanism is generic — any connector can plug in.

The SyncOrchestrator scans /data/extracts/*/extract.duckdb, ATTACHes each into master analytics.duckdb, and creates views.

┌──────────────┐  ┌──────────────┐  ┌──────────────┐
│   Keboola    │  │   BigQuery   │  │   Jira       │
│  extractor   │  │  extractor   │  │  webhooks    │
│ (DuckDB ext) │  │ (remote BQ)  │  │ (incremental)│
└──────┬───────┘  └──────┬───────┘  └──────┬───────┘
       │                 │                 │
       ▼                 ▼                 ▼
   extract.duckdb    extract.duckdb    extract.duckdb
   + data/*.parquet  (views → BQ)      + data/*.parquet
       │                 │                 │
       └─────────────────┼─────────────────┘
                         ▼
              SyncOrchestrator.rebuild()
              ATTACH → master views in analytics.duckdb
                         │
              ┌──────────┼──────────┐
              ▼          ▼          ▼
          FastAPI      CLI
          (serve)    (agnes pull)

Source modes:

  • Batch pull (Keboola, query_mode='local'): DuckDB extension downloads to parquet, scheduled
  • Remote attach (BigQuery, query_mode='remote'): DuckDB BQ extension, no download, queries go to BQ
  • Materialized SQL (BigQuery, query_mode='materialized'): scheduler runs admin-registered SQL through DuckDB BQ extension (via BqAccess from connectors/bigquery/access.py) and writes the result to /data/extracts/bigquery/data/<id>.parquet. Distributed via the same manifest + agnes pull flow as Keboola tables. Cost guardrail via data_source.bigquery.max_bytes_per_materialize (default 10 GiB; set 0 to disable — YAML null falls through to the default).
  • Real-time push (Jira): Webhooks update parquets incrementally

Configuration

Instance-specific config: config/instance.yaml (see example). Environment variables: .env (never committed). Table definitions: DuckDB table_registry table in system.duckdb.

Development

# Setup
python3 -m venv .venv && source .venv/bin/activate
uv pip install ".[dev]"

# Run FastAPI locally
uvicorn app.main:app --reload

# Run tests
pytest tests/ -v

# Trigger sync manually
curl -X POST http://localhost:8000/api/sync/trigger

# Docker
docker compose up

Local sync & Claude Code hooks

agnes pull is the canonical analyst-side distribution path: pulls the RBAC-filtered manifest from the server, downloads parquets whose MD5 changed (skipping query_mode='remote' rows), rebuilds local DuckDB views over them. agnes push mirrors it for the upload direction (sessions, CLAUDE.local.md).

agnes init writes two hooks into <workspace>/.claude/settings.json:

  • SessionStartagnes pull --quiet — pulls fresh parquets at the start of every Claude Code session
  • SessionEndagnes push --quiet — uploads session jsonl + CLAUDE.local.md to the server

Both pass --quiet so they don't pollute Claude Code stdout, and trail with || true so a server outage never blocks a session. Workspace-level (not user-home) so the hooks fire only when Claude Code opens this analyst workspace, not in unrelated sessions on the same machine.

Admin RBAC for auto-sync: query_mode IN ('local', 'materialized') plus a resource_grants row for one of the analyst's groups → table appears in their manifest → agnes pull downloads it. No per-user sync config; the admin layer is the single source of truth.

Business Metrics

Standardized metric definitions live in DuckDB (metric_definitions table). Import starter pack:

agnes admin metrics import docs/metrics/

For AI agents analyzing data:

Before computing any business metric, look up the canonical definition:

  1. agnes catalog --metrics — find the relevant metric
  2. agnes catalog --metrics --show revenue/mrr — read the SQL and business rules
  3. Use the SQL from the metric definition, adapt to the specific question

Never invent metric calculations — always use the canonical definitions.

Querying Agnes data — agent rails

When asked about ANY data in Agnes, follow this protocol.

Discovery first

Before writing ANY query against a table, run:

agnes catalog --json | jq <filter>     # know what's available
agnes schema <table>                   # learn columns + types
agnes describe <table> -n 5            # see real values for shape

NEVER write SELECT * FROM <table> blindly. For local-mode tables it's wasteful; for remote-mode tables it can blow up at 225M rows.

Choose the right tool

Tables in agnes catalog have a query_mode:

  • local: data is on the laptop as parquet (synced via agnes pull). Query directly with agnes query "SELECT … FROM <table>".

  • remote (typically BigQuery): the parquet does NOT exist on the laptop. You MUST either:

    1. agnes snapshot create a filtered subset → query the local snapshot, OR
    2. agnes query --remote for one-shot server-side execution. Works on all query_mode='remote' rows regardless of upstream BQ entity type (BASE TABLE → Storage Read API with predicate pushdown; VIEW / MATERIALIZED_VIEW → BQ jobs API, no pushdown). Cost-guarded by a 5 GiB scan cap (configurable in /admin/server-config). Direct bq."<dataset>"."<table>" paths are registry-gated — unregistered paths return 403 bq_path_not_registered.
    3. agnes query --register-bq for hybrid joins (rarely needed).

agnes snapshot create workflow (preferred for remote tables)

# 1. estimate first
agnes snapshot create web_sessions_example \
    --select event_date,country_code,session_id \
    --where "event_date >= DATE_SUB(CURRENT_DATE(), INTERVAL 30 DAY) 
             AND country_code = 'CZ'" \
    --estimate
# → "estimated_scan_bytes: 4.2 GB, result: ~250k rows, 12 MB locally"

# 2. if reasonable, fetch
agnes snapshot create web_sessions_example ... --as cz_recent

# 3. query the local snapshot
agnes query "SELECT event_date, COUNT(*) FROM cz_recent GROUP BY 1 ORDER BY 1"

Heuristics for agnes snapshot create

  • ALWAYS list specific columns in --select. Avoid implicit SELECT *.
  • ALWAYS include a --where for remote tables; otherwise add --limit.
  • ALWAYS run --estimate first when:
    • You're not sure of the data shape
    • The table has partition_by or clustered_by set (per agnes schema)
    • The fetch could plausibly exceed 1 GB local bytes
  • Reuse agnes snapshot list before fetching — if a snapshot covers your query already, skip the fetch.

BigQuery SQL flavor for --where

For source_type=bigquery (per agnes catalog):

  • Date literal: DATE '2026-01-01' (NOT '2026-01-01'::date)
  • Timestamp literal: TIMESTAMP '2026-01-01 00:00:00 UTC'
  • Now: CURRENT_DATE(), CURRENT_TIMESTAMP()
  • Date arithmetic: DATE_SUB(CURRENT_DATE(), INTERVAL 30 DAY)
  • Regex: REGEXP_CONTAINS(col, r'pattern') (raw string!)
  • NULL: col IS NOT NULL (standard)
  • Cast: CAST(x AS INT64) (NOT INT)

For source_type=keboola / source_type=jira (local), use DuckDB SQL flavor in your agnes query calls — there's no --where on local since fetch is implicit.

Snapshot hygiene

  • Reuse snapshots across questions in the same conversation.
  • Use descriptive names: cz_recent, orders_q1_us, sessions_today.
  • Drop with agnes snapshot drop <name> when done with a topic.
  • agnes disk-info to see total cache size.

When NOT to use agnes snapshot create

  • Single aggregate on remote BASE TABLE (SELECT COUNT(*) FROM remote): use agnes query --remote "SELECT COUNT(*) FROM web_sessions_example". Storage Read API pushes the COUNT into BQ — cheap, no materialization.
  • Single aggregate on remote VIEW/MATERIALIZED_VIEW: same syntax works (#160), but the BQ jobs API can't push WHERE/COUNT into the view body. Cost guardrail (default 5 GiB) catches expensive scans → 400 remote_scan_too_large with agnes snapshot create suggestion. Pivot to agnes snapshot create <id> --where '<predicate>' if the cap is hit.
  • Throwaway exploration: agnes query --remote "SELECT … FROM <registered_id>". Direct bq."<dataset>"."<table>" paths are now registry-gated — register first or use the catalog id.
  • Cross-table JOIN with both tables remote: combine agnes snapshot create for one side + agnes query --remote for the other; full cross-remote JOIN requires more thought (see #101 for design space).

Marketplace Repositories

Admin-managed git repos cloned nightly to ${DATA_DIR}/marketplaces/<slug>/ so FastAPI can read their contents from disk.

  • Register via /admin/marketplaces (admin UI) or POST /api/marketplaces.
  • Scheduler calls POST /api/marketplaces/sync-all (admin-only, authed via SCHEDULER_API_TOKEN) at daily 03:00 UTC. Routing through HTTP keeps the app the sole writer to system.duckdb — the previous in-process call from the scheduler container raced the app's long-lived DB handle and 500-ed on Could not set lock on file.
  • Manual re-sync from the UI ("Sync now") hits POST /api/marketplaces/{id}/sync.
  • PATs for private repos persist to ${DATA_DIR}/state/.env_overlay (chmod 600) as AGNES_MARKETPLACE_<SLUG>_TOKEN. DuckDB stores only the env-var name (token_env), never the secret.
  • Registry lives in DuckDB table marketplace_registry (schema v9).
  • After each successful sync, src/marketplace.py parses .claude-plugin/marketplace.json from the cloned repo and caches the plugin list in marketplace_plugins (keyed by (marketplace_id, plugin_name)).
  • src/marketplace.py handles clone/fetch/reset with token redaction in any surfaced error message.

Access control (v13)

Two layers, no role hierarchy. Full reference: docs/RBAC.md.

  • user_groups — named groups. Two seeded as is_system=TRUE at startup: Admin (god-mode short-circuit on every authorization check) and Everyone (auto-membership for every user).
  • user_group_members(user_id, group_id, source). source ∈ {admin, google_sync, system_seed} so each writer only manipulates its own rows; Google's nightly DELETE+INSERT does not clobber admin-added members.
  • resource_grants — generic (group, resource_type, resource_id) triple. Replaces plugin_access from v12; the same shape now covers any future entity-scoped grant (datasets, knowledge categories, …).

Resource types are an app.resource_types.ResourceType StrEnum paired with a ResourceTypeSpec registered in RESOURCE_TYPES — adding a new one is one enum member, one list_blocks(conn) delegate (projects domain tables into the (block → items) shape the /admin/access tree renders), and one spec entry. No DB migration, no second wiring step. Endpoints gate with either require_admin (app-level) or require_resource_access(ResourceType.X, "{path}") (entity-level), both from app.auth.access.

Admin UI: /admin/access. CLI: agnes admin group {list,create,delete,members, add-member,remove-member} and agnes admin grant {list,create,delete}.

Claude Code marketplace endpoint

Agnes serves a single aggregated Claude Code marketplace over two channels, both gated by PAT auth and filtered per caller:

  • GET /marketplace.zip — deterministic ZIP download with ETag / If-None-Match (304 when content unchanged). Consumed by a client-side SessionStart hook.
  • GET /marketplace.git/* — git smart-HTTP (dulwich via a2wsgi). Registered in Claude Code once, then Claude Code owns the clone/fetch cycle.

Auth: ZIP uses Authorization: Bearer <PAT>. Git uses HTTP Basic where the password field carries the PAT (https://x:<PAT>@host/marketplace.git/) — git CLI does not speak Bearer.

Content: filtered via src.marketplace_filter.resolve_allowed_plugins which joins resource_grants ↔ marketplace_plugins (matching mp.marketplace_id || '/' || mp.name = rg.resource_id) scoped to the caller's user_group_members. Admin is treated as a regular group here — no god-mode shortcut for the marketplace feed, so admins curate their own view by granting plugins to the Admin group (or any group they belong to). On-disk layout in the served ZIP / git tree uses a slug-prefixed directory (plugins/<slug>-<plugin>/) so two marketplaces shipping a same-named plugin don't overwrite each other's files. The synth marketplace.json's name field, however, is the plugin's authoritative name from its own .claude-plugin/plugin.json (with a fallback to the upstream marketplace.json name) — Claude Code's /plugin UI resolves a loaded plugin back to its catalog entry by plugin.json name, so the catalog entry's name must match. Same-named plugins from two upstream marketplaces therefore collide in the catalog by design; admin RBAC (which grants survive the filter) decides which one wins, identical to how Claude Code behaves when a user adds two upstream marketplaces with overlapping plugin names directly. /marketplace/info exposes both name and prefixed_name so operators can disambiguate.

Cache: content-addressed bare repos at ${DATA_DIR}/marketplaces/git-cache/ keyed by sha256(filtered content). Two users with the same RBAC view share one repo; content change → new repo next to the old one. No TTL / prune yet.

User registration inside Claude Code:

# ZIP channel (typically via a SessionStart hook that unpacks into ./marketplace/)
curl -H "Authorization: Bearer $AGNES_PAT" https://agnes.example.com/marketplace.zip

# Git channel — one-time registration. Two paths; pick the first that works.

# (a) Direct registration — preferred when it works.
/plugin marketplace add https://x:$AGNES_PAT@agnes.example.com/marketplace.git/

# (b) Two-step fallback — required when (a) fails. Bun-compiled `claude` on
#     macOS / Windows ignores the OS trust store and CA env vars on the
#     marketplace HTTPS path, so direct add can fail with TLS errors against
#     a private-CA Agnes instance even when system tools work fine. System
#     `git` honors GIT_SSL_CAINFO + the OS trust store, so cloning manually
#     and pointing Claude Code at the local clone sidesteps the Bun TLS path
#     entirely.
git clone https://x:$AGNES_PAT@agnes.example.com/marketplace.git/ ~/agnes-marketplace
claude plugin marketplace add ~/agnes-marketplace
# Optional hardening: strip the PAT from the cloned repo's origin so it
# doesn't sit in plaintext at ~/agnes-marketplace/.git/config — re-clone via
# the dashboard's setup flow when the PAT rotates.
git -C ~/agnes-marketplace remote set-url origin https://agnes.example.com/marketplace.git/

The dashboard-served setup payload (see app/web/setup_instructions.py) already branches between (a) and (b) automatically based on platform when a private CA is in play. The block above is the manual equivalent for users registering outside that flow (e.g. operators bringing up a new instance, or analysts whose first attempt failed and need to retry by hand).

Hybrid Queries (BigQuery + Local)

For tables too large to sync locally, use hybrid queries that JOIN local data with on-demand BigQuery results:

agnes query --sql "SELECT o.*, t.views FROM orders o JOIN traffic t ON o.date = t.date" \
         --register-bq "traffic=SELECT date, SUM(views) as views FROM dataset.web WHERE date > '2026-01-01' GROUP BY 1"

The --register-bq flag executes a BigQuery subquery, loads the result into memory, and makes it available as a DuckDB view for the final SQL. Multiple --register-bq flags can be used for multiple BQ sources.

For complex SQL, use stdin mode:

echo '{"register_bq": {"traffic": "SELECT ..."}, "sql": "SELECT ..."}' | agnes query --stdin

Extensibility

Data Sources (extract.duckdb contract)

New connector = connectors/<name>/extractor.py producing extract.duckdb + data/. Must create _meta table with columns: table_name, description, rows, size_bytes, extracted_at, query_mode. Orchestrator ATTACHes it automatically.

Authentication

Auth providers in app/auth/ (FastAPI-based):

  • Google: OAuth via Google (Workspace group memberships pulled at sign-in — see docs/auth-groups.md for the GCP setup checklist + the security label gotcha)
  • Email: Email magic link (itsdangerous token)
  • Desktop: JWT for API

RBAC

See Access control (v13) above and docs/RBAC.md for the full reference. TL;DR for module authors: gate endpoints with Depends(require_admin) for app-level mutations or Depends(require_resource_access(ResourceType.X, "{path}")) for entity-scoped grants. Add a new resource type by extending the ResourceType StrEnum and registering a ResourceTypeSpec (with a list_blocks projection delegate) in app/resource_types.py.

Release & deploy workflows

Two separate release.yml-style workflows produce GHCR images. Pick the one that matches what you're shipping.

release.yml — auto-build on every push

Runs on every push to every branch.

  • Push to main:stable, :stable-YYYY.MM.N (CalVer).
  • Push to non-main <prefix>/<branch>:dev, :dev-YYYY.MM.N, :dev-<branch-slug>, and (when prefix isn't a Git Flow convention) :dev-<prefix>-latest alias.

VMs that pin to a floating tag (:dev, :dev-<prefix>-latest) auto-upgrade within ~5 min via the cron in agnes-auto-upgrade.sh. Convenient for per-developer dev VMs; footgun for shared dev VMs (last pusher wins, regardless of who).

keboola-deploy.yml — tag-triggered, explicit deploy only

Runs only on git tags matching keboola-deploy-*. Publishes:

  • :keboola-deploy-<git-tag-suffix> — immutable, tied to the exact commit
  • :keboola-deploy-latest — floating alias the consumer pins to

Operator workflow:

git checkout <commit-or-branch>
git tag keboola-deploy-<descriptive-name>
git push origin keboola-deploy-<descriptive-name>
# → workflow builds + publishes both tags
# → VM cron picks up :keboola-deploy-latest within ~5 min
# → manual cron trigger (skip the wait): sudo /usr/local/bin/agnes-auto-upgrade.sh on the VM

Use this when the consumer (e.g. a customer dev VM) needs deploy-when-I-decide semantics — no surprise rollouts from upstream branch pushes by other contributors. The infra repo pins image_tag = "keboola-deploy-latest" on the relevant VM.

Module versioning

The customer-instance Terraform module under infra/modules/customer-instance/ is published as infra-vMAJOR.MINOR.PATCH git tags (separate from app CalVer tags). Bump on any module-API change; downstream infra repos pin to the tag in their source = "github.com/keboola/agnes-the-ai-analyst//infra/modules/customer-instance?ref=infra-v1.X.Y".

After merging a module change to main:

git tag infra-vX.Y.Z origin/main
git push origin infra-vX.Y.Z

Replacing a VM after a startup-script change

Module sets lifecycle { ignore_changes = [metadata_startup_script] } on google_compute_instance.vm so normal terraform apply doesn't churn running VMs. To propagate a startup-script update, trigger the consumer's apply workflow manually with the VM resource address — typical workflow_dispatch input is recreate_targets='module.agnes.google_compute_instance.vm["<vm-name>"]'.

Key Implementation Details

DuckDB Schema (src/db.py)

  • Schema v35 with auto-migration v1→…→v35 (v5 adds users.active, v6 adds personal_access_tokens, v7 adds personal_access_tokens.last_used_ip, v8/v9 added the legacy internal_roles/role-grants tables, v10 added view_ownership for cross-connector view-name collision detection (issue #81 Group C), v11 added marketplace_registry + marketplace_plugins + user_groups + plugin_access, v12 added users.groups JSON + user_groups.is_system, v13 replaces internal_roles/group_mappings/user_role_grants/plugin_access with user_group_members + resource_grants and drops users.groups JSON, v14 adds FK constraints on user_group_members + resource_grants after orphan cleanup, v15 adds knowledge_items context-engineering columns + contradictions + session_extraction_state, v16 adds verification_evidence, v17 adds knowledge_item_relations, v18 drops stranded non-google memberships from google-managed groups, v19 drops legacy dataset_permissions, access_requests tables and users.role, table_registry.is_public columns — table access is now exclusively per-group via resource_grants(resource_type='table'), v20 adds source_query TEXT to table_registry to back query_mode='materialized' (BigQuery scheduled-query parquet path), v21 adds welcome_template singleton table backing the Agent Setup Prompt admin override (/admin/agent-prompt), v22 reserves the setup_banner table — feature dropped mid-development; table retained for forward compatibility with already-migrated instances, v23 adds claude_md_template singleton table backing the Agent Workspace Prompt admin override (/admin/workspace-prompt), v24 rewrites materialized BQ source_query from DuckDB-flavor bq."ds"."t" to BQ-native `<project>.ds.t` so the new wrapping path accepts them; idempotent + warns when project unconfigured, v25 adds store_entities + user_store_installs + user_plugin_optouts backing the /store and /my-ai-stack pages — the served marketplace is now (admin_granted opt_outs) store_installs, v26 unifies Keboola query_mode='local' rows into 'materialized' — the old local mode (DuckDB Keboola extension's COPY through QueryService) is replaced by the new Storage API export-async path which works regardless of project flags; existing local Keboola rows are flipped, NULL source_query means full-table export, v27 adds 7 columns to table_registry for Keboola per-table sync-strategy support: incremental_window_days, max_history_days, incremental_column, where_filters, partition_by, partition_granularity, initial_load_chunk_days. Layered on top of v26: admins can opt specific tables back to query_mode='local' (via the Direct extract Edit-modal radio) to enable the new dispatcher. The pre-existing sync_strategy column (default 'full_refresh') is reused — pre-v27 it was inert catalog metadata; post-v27 the Keboola extractor dispatches off it (full_refresh | incremental | partitioned). All new columns NULL on existing rows; meaningful only when paired with the matching strategy., v28 introduces explicit-install (Model B) for curated marketplace plugins — served set flips from (rbac user_plugin_optouts) to (rbac ∩ subscriptions). The user_plugin_optouts table+columns are reused (no DDL rename) so existing operator instances skip migration churn; row PRESENCE flips meaning from "excluded" to "subscribed", and the migration wipes existing rows so the inverted reading starts from a clean baseline. Also adds marketplace_plugins.created_at (per-plugin newest-first sort on /marketplace), backfilled from parent marketplace_registry.registered_at so existing plugins get a sensible date until the next sync overwrites with CURRENT_TIMESTAMP., v29 adds store_submissions table backing flea-market upload guardrails (manifest + static-security + LLM-review verdicts) plus store_entities.visibility_status (pending | approved | hidden) — entity visible in flea browse only when visibility_status='approved'. Existing rows backfilled to 'approved' so live flea content stays visible., v30 adds store_submissions.{file_size, bundle_sha256, bundle_purged_at} so blocked-inline bundles persist for forensics + admin rescan/override (instead of the prior rmtree-on-reject); SHA256 survives the 30-day TTL purge, bundle_purged_at flips on at purge time so detail page can render "purged on YYYY-MM-DD", v31 reshapes store_submissions (drops legacy unique on entity_id so multiple submissions per entity work — re-uploads/rescans land as new rows; idempotent table rebuild), v32 adds store_entities.{archived_at, archived_by} columns plus broadens visibility_status enum to include 'archived' for soft-delete; DELETE /api/store/entities/{id} is now soft (archive) by default, hard delete moves to ?hard=true (admin-only), v33 drops store_submissions.retry_count — counter mixed automatic LLM retries (capped) with admin-initiated rescans (unbounded), no useful semantics; admin Rescan button + audit_log carry the operational signal, v34 ensures idx_store_submissions_entity exists after the v33 column drop (DuckDB rebuilds the table sans index when dropping a column referenced by an index), v35 broadens store_entities.visibility_status enum to include lifecycle value beyond 'archived' already added in v32 — column-rebuild migration to register the new value with DuckDB's CHECK constraint, so set_visibility('archived') works against the constrained column. Also marks the architectural cutover from denormalizing 'archived'/'deleted' onto store_submissions.status to LEFT-JOINing store_entities at query time: verdict (sub.status) becomes immutable forensic record, lifecycle (entity.visibility_status) becomes the live source of truth that the admin queue's Archived chip filters by. — see CHANGELOG and docs/RBAC.md)
  • table_registry: id, name, source_type, bucket, source_table, query_mode, sync_schedule, etc.
  • sync_state, sync_history: track extraction progress
  • users, audit_log: account state + audit trail. RBAC lives in user_groups + user_group_members + resource_grants.
  • System DB at {DATA_DIR}/state/system.duckdb
  • Analytics DB at {DATA_DIR}/analytics/server.duckdb

SyncOrchestrator (src/orchestrator.py)

  • rebuild(): scans extracts dir, ATTACHes all, creates master views, updates sync_state
  • rebuild_source(name): single source (used after Jira webhooks)
  • Thread-safe via _rebuild_lock

Connector Pattern

  • Keboola: connectors/keboola/extractor.py uses DuckDB Keboola extension, fallback to client.py
  • BigQuery: connectors/bigquery/extractor.py uses DuckDB BQ extension (remote-only, no download)
  • Jira: connectors/jira/webhook.pyincremental_transform.pyextract_init.py updates _meta
  • connectors/keboola/client.py: legacy Keboola Storage API wrapper (kept as fallback)

Config Loading

  1. config/loader.py loads instance.yaml
  2. app/instance_config.py exposes get_data_source_type(), get_value()
  3. Table config lives in DuckDB table_registry (not markdown files)

Files NOT to modify (stable infrastructure)

  • connectors/jira/file_lock.py - Advisory file locking
  • connectors/jira/transform.py - Core Jira transform logic
  • services/ws_gateway/ - WebSocket notification gateway

Vendor-agnostic OSS — no customer-specific content

This repo is the public OSS distribution. Nothing customer-specific belongs in code, configuration defaults, comments, docs, commit messages, PR titles, or PR bodies. That includes:

  • Specific deployments or brands (private VM names, internal product brands, organization names that aren't already public sponsors).
  • Cloud project IDs, internal hostnames, runbook paths from a particular install (/opt/<deployment>, <host>.<internal-domain>, prj-<org>-…, internal SA emails).
  • Cross-references to private repos (<private-org>/<private-repo>#NN). Describe the integration in generic terms or link to public examples instead.

When you motivate a change, frame it abstractly ("behind a TLS-terminating reverse proxy", "in containerized deploys") rather than naming a specific operator. When you show examples, use placeholders (example.com, <your-host>, <install-dir>). When config has reasonable defaults pulled from one deployment's habits, generalize them or surface them as documented examples — not hard-coded assumptions.

Customer-specific automation, hostnames, and identities live in private infra repos that consume this OSS. The OSS describes capabilities, defaults, and configuration knobs — not how a specific operator wired them up.

Changelog discipline — non-negotiable

Every PR that adds, removes, or changes user-visible behavior MUST update CHANGELOG.md in the same PR. No exceptions, no follow-ups, no "I'll do it after merge". User-visible = anything an operator, end-user, or downstream integrator can observe: CLI flags / output / exit codes, REST endpoints / payloads / status codes, web UI, instance.yaml schema, env vars, extract.duckdb contract, Docker / compose / Caddyfile knobs, default behaviors, breaking changes, security fixes.

How:

  • Add a bullet under the topmost ## [Unreleased] heading (create one if missing — it sits above the latest released version).
  • Group by ### Added / ### Changed / ### Fixed / ### Removed / ### Internal (Keep-a-Changelog sections).
  • Mark breaking changes with **BREAKING** at the start of the bullet — operators grep for that string before bumping the pin.
  • Reference the relevant doc/runbook if one exists (e.g. see docs/auth-groups.md), don't restate it.
  • Internal-only changes (refactors, test additions, dependency bumps without behavior change) go under ### Internal — still log them, just keep them terse.

When you cut a release:

  • Rename ## [Unreleased]## [X.Y.Z] — YYYY-MM-DD.
  • Append a new empty ## [Unreleased] section at the top so the next PR has somewhere to land.
  • Bump version in pyproject.toml to match X.Y.Z.
  • Tag the merge commit as vX.Y.Z and push the tag.

If you find yourself opening a PR without a CHANGELOG entry, stop and add one before requesting review. Reviewers should bounce PRs that touch user-visible behavior without a changelog update — same way they'd bounce a PR with no test changes for new logic.

Run tests before every push — non-negotiable

Before git push, run the full pytest suite locally. CI runs the same command (.github/workflows/ci.yml:29pytest tests/ -v --tb=short -n auto); a failure that surfaces in CI was discoverable in 90 seconds locally. Pushing first and watching CI fail wastes operator time, slows the PR, and trains everyone to ignore CI badges.

Command (matches CI):

.venv/bin/pytest tests/ --tb=short -n auto -q

-n auto parallelizes across CPU cores; the suite runs in ~90s on a modern laptop. Local-only env (no instance.yaml, dev defaults) is fine — fixtures use fresh_db per-test isolation.

When tests fail:

  • Failures in code you touched → fix before pushing. Update test expectations when the behavior change is intentional and documented (e.g. template restructure that changes assertion strings).
  • Failures unrelated to your diff → confirm with git stash && pytest <failing-test> && git stash pop. If they reproduce on a clean branch, they are pre-existing — note them in the PR body but don't block your push on them. Don't silently skip; flag them so someone owns the fix.
  • Flaky tests → re-run once. Two consecutive failures = real failure, fix or quarantine with a tracked issue.

Smoke shortcuts (when full suite is too slow during iteration):

  • pytest tests/ -k <pattern> -q for area-scoped checks while iterating
  • pytest tests/test_X.py tests/test_Y.py -q for the modules you touched

But the full pytest tests/ -n auto runs once before push. No exceptions.

If a CHANGELOG entry, doc edit, or pure-formatting commit genuinely doesn't touch any code path the tests exercise, you can skip the full run — but be honest with yourself about whether that's actually the case.

Git Commits & Pull Requests

  • Keep commit messages clean and concise
  • Do not include AI attribution in commits or PRs
  • Before opening a PR, scan the diff and the PR body for the customer-specific tokens listed above (grep -niE '<token1>|<token2>|...'). If anything matches, generalize or remove it.