Skip to content
Merged
Changes from 2 commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
244 changes: 244 additions & 0 deletions .github/workflows/platform-preview-seed-fixture.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,244 @@
name: AutoGPT Platform - Preview Seed Fixture Bake

# Bakes a fully SYNTHETIC preview-database seed fixture and publishes it as a
# rolling GitHub release asset (tag: preview-seed-fixture). Preview environments
# restore this fixture BEFORE a PR's own migrations run, so schema changes are
# exercised against populated tables (catching NOT-NULL / unique / backfill
# failures that an empty DB would silently pass).
#
# SECURITY INVARIANT: this workflow holds NO cloud credentials and has NO read
# path to any real database. It runs entirely against a throwaway Postgres
# service container. The fixture contains ONLY Faker-generated synthetic rows
# plus the already-public marketplace agent exports checked into
# autogpt_platform/backend/agents/. No production data is ever touched.

on:
push:
branches: [dev]
paths:
Comment thread
ntindle marked this conversation as resolved.
- "autogpt_platform/backend/migrations/**"
- "autogpt_platform/backend/test/**"
- ".github/workflows/platform-preview-seed-fixture.yml"
schedule:
# Weekly backstop (Mondays 06:00 UTC) so the fixture never goes stale even
# if no migration/test change lands for a while.
- cron: "0 6 * * 1"
Comment thread
ntindle marked this conversation as resolved.
Comment thread
ntindle marked this conversation as resolved.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟢 low (performance/ci-efficiency)

Trigger path autogpt_platform/backend/test/** re-bakes the full fixture on any test-dir change, not just the 4 seeder scripts, producing many byte-identical fixtures. cancel-in-progress only coalesces overlapping runs, not sequential unrelated edits.

Suggestion: Narrow the path filter to the specific seeder files (test_data_creator.py, e2e_test_data.py, test_data_updater.py) plus migrations, relying on the weekly cron as the staleness backstop.

workflow_dispatch:

# Coalesce overlapping bakes: a newer commit cancels an in-flight bake since
Comment thread
ntindle marked this conversation as resolved.
# they would publish the same rolling asset anyway.
concurrency:
group: preview-seed-fixture-bake

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟢 low (architect/concurrency / atomicity)

cancel-in-progress can interrupt the non-atomic publish sequence (release edit -> asset upload --clobber -> tag retarget), leaving the rolling release with a new manifest but stale fixture.dump.gz, or updated assets but un-retargeted tag. Consumers restore the latest asset, so a torn update is a correctness hazard.

Suggestion: Set cancel-in-progress: false, or scope cancellation so an in-flight publish step always completes.

cancel-in-progress: true

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟢 low (security/least-privilege)

contents: write is granted at the workflow level, so all steps (including repo-defined seeders and the piped Poetry installer) run with the write token available, even though only the final publish step needs it.

Suggestion: Move the permissions block to the bake job or split publishing into a separate minimally-permissioned job to reduce blast radius.

permissions:
contents: write # required to create/update the rolling release + its assets

defaults:
run:
shell: bash
working-directory: autogpt_platform/backend

jobs:
bake:
runs-on: ubuntu-latest
timeout-minutes: 45

services:
# Throwaway Postgres. pgvector/pgvector:pg15 matches the platform's real
# Postgres major version (supabase/postgres:15.8.x used in dev/prod) and
# ships the `vector` extension (required by the docs-embedding migrations)
# plus the `pg_trgm` contrib module (creator-search index migration). No
# credentials of any kind — a disposable localhost container.
postgres:
image: pgvector/pgvector:pg15
Comment thread
coderabbitai[bot] marked this conversation as resolved.
env:
POSTGRES_USER: postgres
POSTGRES_PASSWORD: postgres
POSTGRES_DB: postgres
ports:
- 5432:5432
options: >-
--health-cmd "pg_isready -U postgres"
--health-interval 10s
--health-timeout 5s
--health-retries 10

env:
CI: "true"
PLAIN_OUTPUT: "true"
# Prisma connection to the throwaway container, using the platform schema
# exactly like the backend tests / local dev (.env.default).
DATABASE_URL: "postgresql://postgres:postgres@localhost:5432/postgres?schema=platform"
Comment thread
ntindle marked this conversation as resolved.
DIRECT_URL: "postgresql://postgres:postgres@localhost:5432/postgres?schema=platform"
Comment thread
ntindle marked this conversation as resolved.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟢 low (architect/comment-durability)

Comment uses change-relative language ('e2e_test_data.py no longer talks to Supabase'), which documents a past migration rather than a standing fact and has already misled reviewers.

Suggestion: Rewrite to the standing invariant: 'e2e_test_data.py does not talk to Supabase (users are created through Better Auth via raw Prisma inserts), but backend settings still read these values at import time, so they only need to exist.'

# Dummy, UNREACHABLE Supabase config. e2e_test_data.py constructs a
# Supabase client unconditionally (supabase-py raises on an empty URL), but
# its per-user try/except falls back to raw synthetic UUIDs when the GoTrue
# admin API call fails. Pointing at a dead localhost port lets that
# documented fallback run without any real auth backend or credentials.
SUPABASE_URL: "http://localhost:54321"
SUPABASE_SERVICE_ROLE_KEY: "synthetic-bake-no-real-key" # pragma: allowlist secret
# Non-production test key (identical to the one used in backend CI) so
# backend module imports that read settings.secrets don't error.
Comment thread
ntindle marked this conversation as resolved.
ENCRYPTION_KEY: "dvziYgz0KSK8FENhju0ZYi8-fRTfAdlz6YLhdB_jhNw=" # pragma: allowlist secret # DO NOT USE IN PRODUCTION
Comment thread
ntindle marked this conversation as resolved.

steps:

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟡 medium (security/supply-chain / unpinned actions)

Third-party actions (actions/checkout@v6, actions/setup-python@v5, actions/cache@v5) are pinned to mutable major-version tags rather than full commit SHAs, in a workflow that holds contents:write and publishes release assets restored into downstream preview databases. A moved/hijacked tag would execute attacker code with a write token.

Suggestion: Pin each action to a full commit SHA (e.g. actions/checkout@ # v6) per GitHub's hardening guidance for privileged workflows.

- name: Checkout repository
uses: actions/checkout@v6
Comment thread
cursor[bot] marked this conversation as resolved.
Comment thread
coderabbitai[bot] marked this conversation as resolved.
Comment thread
ntindle marked this conversation as resolved.
with:
# Always bake dev: schedule/dispatch runs execute from the DEFAULT
# branch (master), which would otherwise silently bake master's
# schema into a fixture that previews (built from dev) restore.
ref: dev
# Only the release-publish step needs a token, via env — don't
# persist credentials into the workspace git config.
persist-credentials: false

- name: Set up Python 3.12
uses: actions/setup-python@v5
Comment thread
ntindle marked this conversation as resolved.
with:
python-version: "3.12"

- name: Set up Python dependency cache
uses: actions/cache@v5
Comment thread
ntindle marked this conversation as resolved.
with:
path: ~/.cache/pypoetry
key: poetry-${{ runner.os }}-py3.12-${{ hashFiles('autogpt_platform/backend/poetry.lock') }}
Comment thread
ntindle marked this conversation as resolved.

- name: Install Poetry
Comment thread
ntindle marked this conversation as resolved.
run: |
HEAD_POETRY_VERSION=$(python ../../.github/workflows/scripts/get_package_version_from_lockfile.py poetry)
echo "Using Poetry version ${HEAD_POETRY_VERSION}"
Comment thread
ntindle marked this conversation as resolved.
curl -sSL https://install.python-poetry.org | POETRY_VERSION=$HEAD_POETRY_VERSION python3 -
Comment thread
ntindle marked this conversation as resolved.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟢 low (security/supply-chain / remote code execution)

Poetry is installed via curl -sSL https://install.python-poetry.org | python3 -, piping an unpinned remote script into the interpreter inside a job with a write-scoped token.

Suggestion: Install Poetry from a pinned installer/pipx or verify a checksum before executing; avoid piping unverified network content to python3.


- name: Install Python dependencies
run: poetry install
Comment thread
ntindle marked this conversation as resolved.

- name: Generate Prisma Client
run: poetry run prisma generate

- name: Apply database migrations
run: poetry run prisma migrate deploy

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟢 low (performance/ci-caching)

Cache stores only ~/.cache/pypoetry but not the project virtualenv, so poetry install still rebuilds/links the env on every run, adding avoidable wall-clock time.

Suggestion: Also cache the .venv/virtualenv path (as backend CI does) keyed on poetry.lock to skip redundant dependency linking.


# --- Seeders --------------------------------------------------------
# Order mirrors each script's documented invocation. test_data_creator is
# REQUIRED (a failure fails the bake). The other three are best-effort:
# a failure emits a LOUD ::warning:: but does not abort the bake.

- name: "Seed: test_data_creator (REQUIRED)"
run: poetry run python test/test_data_creator.py
Comment thread
ntindle marked this conversation as resolved.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟡 medium (architect/robustness / silent degradation)

load-store-agents is best-effort (::warning:: + exit 0) even though, unlike the two Python seeders, it has no queue/cache dependency. If it fails, the release still publishes with only user rows, passing the USER_COUNT>=1 gate — producing a fixture missing store/agent data, which defeats the PR's goal of exercising migrations against populated tables.

Suggestion: Make load-store-agents required, or gate publish on row-count sanity checks for the tables each best-effort seeder populates (not just platform.User).

- name: "Seed: load-store-agents (public marketplace exports)"
Comment thread
ntindle marked this conversation as resolved.
Comment thread
ntindle marked this conversation as resolved.
run: |
poetry run load-store-agents || {
echo "::warning title=seed-soft-failure::load-store-agents seeder failed; continuing without it."
exit 0
}

- name: "Seed: e2e_test_data (GoTrue-less raw-UUID fallback)"
run: |
poetry run python test/e2e_test_data.py || {
echo "::warning title=seed-soft-failure::e2e_test_data seeder failed; continuing without it."
exit 0
}

- name: "Seed: test_data_updater (mutates existing rows)"

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟢 low (quality/duplication)

The soft-failure wrapper || { echo "::warning title=seed-soft-failure::..."; exit 0; } is copy-pasted across three seeder steps with near-identical text, so any change to the soft-failure contract must be edited in three places.

Suggestion: If a 4th best-effort seeder is added, factor the invocation into a matrix step keyed on {name, cmd, required} or a small bash helper; acceptable to leave as-is for the current three.

run: |
poetry run python test/test_data_updater.py || {
Comment thread
ntindle marked this conversation as resolved.
Outdated
echo "::warning title=seed-soft-failure::test_data_updater seeder failed; continuing without it."
exit 0
}

# --- Dump + manifest ------------------------------------------------
# pg_dump / psql run INSIDE the Postgres service container via docker exec
# so the client version always matches the server exactly (the runner's
# bundled client can lag the server major and refuse the dump). This is

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟢 low (quality/magic-value)

timeout 300 is duplicated at two seeder steps with no shared constant, so the two timeouts must be kept in sync manually.

Suggestion: Optionally hoist into a job-level env var (e.g. SEED_TIMEOUT: 300) and reference it in both steps.

# the same pattern platform-fullstack-ci.yml uses.

- name: Resolve Postgres service container
run: |
Comment thread
ntindle marked this conversation as resolved.
set -euo pipefail
PG_CID=$(docker ps --filter "ancestor=pgvector/pgvector:pg15" --format '{{.ID}}' | head -n1)
Comment thread
ntindle marked this conversation as resolved.
Comment thread
ntindle marked this conversation as resolved.
if [ -z "${PG_CID}" ]; then
echo "::error title=no-pg-container::Could not locate the Postgres service container."
docker ps
exit 1
fi
echo "PG_CID=${PG_CID}" >> "${GITHUB_ENV}"
docker exec "${PG_CID}" pg_dump --version

- name: Dump platform schema fixture (custom format, gzipped)
Comment thread
ntindle marked this conversation as resolved.
Comment thread
ntindle marked this conversation as resolved.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟢 low (discussion/deferred-concern)

Reviewer (High) noted the workflow publishes fixture.dump.gz without ever restoring it, so a corrupt/truncated archive passes all checks and only fails in every preview. Author acknowledged and deferred restore validation to the follow-up restore-side PR.

Suggestion: Ensure the follow-up infra PR adds a round-trip restore validation, or track it as an explicit follow-up issue so the deferral is not lost.

run: |
set -euo pipefail
# Loud sanity check: the required seeder must have populated users.
USER_COUNT=$(docker exec "${PG_CID}" psql -U postgres -d postgres -tAc "SELECT count(*) FROM platform.\"User\";" | tr -d '[:space:]')
echo "Seeded platform.\"User\" rows: ${USER_COUNT}"
if [ "${USER_COUNT:-0}" -lt 1 ]; then
Comment thread
ntindle marked this conversation as resolved.
Comment thread
ntindle marked this conversation as resolved.
Comment thread
ntindle marked this conversation as resolved.
Comment thread
ntindle marked this conversation as resolved.
Comment thread
ntindle marked this conversation as resolved.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟡 medium (testing/weak-fixture-verification)

Fixture completeness is gated only by platform."User" >= 1. If the three best-effort seeders all soft-fail, a fixture populated by test_data_creator alone is still published, undermining the goal of exercising migrations against realistic populated tables (store agents, executions, review stats).

Suggestion: Add row-count assertions for tables the best-effort seeders populate (e.g. AgentGraph, AgentGraphExecution, store listings) and fail or emit a distinct error when they are unexpectedly empty.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 🟢 low (discussion/declined-advisory)

Reviewer noted the only completeness guard (platform.User >= 1) is tautological and best-effort seeders (load-store-agents, e2e_test_data) can soft-fail, publishing a fixture missing store/agent rows while still reporting success. Author declined as advisory.

Suggestion: Consider asserting minimum row counts for other critical tables or recording per-seeder success in manifest.json so a degraded fixture is detectable by the restore side.

echo "::error title=empty-fixture::Seeded fixture is empty (platform.User has 0 rows); refusing to publish."
exit 1
Comment thread
ntindle marked this conversation as resolved.
fi
# Custom-format dump of the platform schema ONLY (schema + data,
# including platform._prisma_migrations so the restore side can run
# only the PR's delta migrations). -Z0 disables pg_dump's internal
# compression so the outer gzip is the single compressor; restore is
# therefore `gunzip -c fixture.dump.gz | pg_restore ...`.
docker exec "${PG_CID}" pg_dump -U postgres -d postgres \
--format=custom -Z0 --schema=platform \
| gzip -9 > "${GITHUB_WORKSPACE}/fixture.dump.gz"
Comment thread
ntindle marked this conversation as resolved.
ls -lh "${GITHUB_WORKSPACE}/fixture.dump.gz"

- name: Write manifest.json
run: |
set -euo pipefail
# baked_at comes from the checked-out commit (NOT wall-clock), so the
# manifest is reproducible for a given dev SHA.
BAKED_AT="$(git log -1 --format=%cI)"
# NOT GITHUB_SHA: on schedule/dispatch runs that is the default
# branch's commit; the checkout above is pinned to dev.
DEV_SHA="$(git rev-parse HEAD)"
MIGRATION_HEAD="$(docker exec "${PG_CID}" psql -U postgres -d postgres -tAc \
"SELECT migration_name FROM platform._prisma_migrations WHERE finished_at IS NOT NULL ORDER BY finished_at DESC LIMIT 1;" \
Comment thread
ntindle marked this conversation as resolved.
| tr -d '[:space:]')"
echo "baked_at=${BAKED_AT} dev_sha=${DEV_SHA} migration_head=${MIGRATION_HEAD}"
if [ -z "${MIGRATION_HEAD}" ]; then
echo "::error title=no-migration-head::Could not read a migration head from platform._prisma_migrations."
exit 1
fi
jq -n \
Comment thread
ntindle marked this conversation as resolved.
--arg baked_at "${BAKED_AT}" \
--arg dev_sha "${DEV_SHA}" \
--arg migration_head "${MIGRATION_HEAD}" \
'{baked_at: $baked_at, dev_sha: $dev_sha, migration_head: $migration_head}' \
> "${GITHUB_WORKSPACE}/manifest.json"
cat "${GITHUB_WORKSPACE}/manifest.json"

# --- Publish rolling release ---------------------------------------

- name: Publish rolling preview-seed-fixture release
env:
GH_TOKEN: ${{ github.token }}
run: |
set -euo pipefail
TAG="preview-seed-fixture"
FIXTURE="${GITHUB_WORKSPACE}/fixture.dump.gz"
MANIFEST="${GITHUB_WORKSPACE}/manifest.json"
MIGRATION_HEAD="$(jq -r .migration_head "${MANIFEST}")"
DEV_SHA="$(jq -r .dev_sha "${MANIFEST}")"
Comment thread
ntindle marked this conversation as resolved.
NOTES="Rolling, fully SYNTHETIC preview-database seed fixture baked from dev@${DEV_SHA:0:12} (migration head: ${MIGRATION_HEAD}). Preview environments restore fixture.dump.gz (custom-format, gzipped, platform schema only, including _prisma_migrations) BEFORE running a PR's own migrations, so schema changes run against populated tables. SECURITY INVARIANT: this bake holds no cloud credentials and has no read path to any real database — it runs entirely against a throwaway Postgres container and contains only Faker-generated synthetic rows plus the already-public marketplace agent exports checked into autogpt_platform/backend/agents/. This asset is regenerated on every eligible dev push and weekly; treat it as disposable."
if gh release view "${TAG}" >/dev/null 2>&1; then
echo "Updating existing rolling release ${TAG}"
gh release edit "${TAG}" --title "Preview seed fixture (rolling)" --notes "${NOTES}" --prerelease
Comment thread
ntindle marked this conversation as resolved.
gh release upload "${TAG}" "${FIXTURE}" "${MANIFEST}" --clobber
Comment thread
cursor[bot] marked this conversation as resolved.
Outdated
# Retarget the rolling tag at the baked dev commit so release
# metadata matches the published assets (manifest.dev_sha is the
# authoritative record either way).
gh api -X PATCH "repos/${GITHUB_REPOSITORY}/git/refs/tags/${TAG}" \
-f sha="${DEV_SHA}" -F force=true >/dev/null
else
echo "Creating rolling release ${TAG}"
gh release create "${TAG}" "${FIXTURE}" "${MANIFEST}" \
--title "Preview seed fixture (rolling)" \
--notes "${NOTES}" \
--prerelease \
Comment thread
ntindle marked this conversation as resolved.
--target "${DEV_SHA}"
fi
Loading