Build reliable dev / staging / prod isolation for LangChain 1.0 services โ
Pydantic `Settings` + `SecretStr`, cloud Secret Manager in prod, per-env
prompt and model version pinning, env-specific checkpointer and observability.
Use when graduating from `.env`-in-dev to real prod infra, or debugging a
config that loaded the wrong values in the wrong env.
Trigger with "langchain multi-env", "langchain pydantic settings",
"langchain secret manager", "langchain env config", "langchain prod setup".
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Build reliable dev / staging / prod isolation for LangChain 1.0 services โ
Pydantic `Settings` + `SecretStr`, cloud Secret Manager in prod, per-env
prompt and model version pinning, env-specific checkpointer and observability.
Use when graduating from `.env`-in-dev to real prod infra, or debugging a
config that loaded the wrong values in the wrong env.
Trigger with "langchain multi-env", "langchain pydantic settings",
"langchain secret manager", "langchain env config", "langchain prod setup".
Designed for Claude Code, also compatible with Codex
LangChain Multi-Env Setup (Python)
Overview
A team ships a LangChain 1.0 service to staging with python-dotenv loading
.env.staging into os.environ. Security audits โ
docker exec STAGING-POD env prints ANTHROPIC_API_KEY=sk-ant-api03-... in
plain text. Anyone with kubectl exec, any sidecar, any core dump, any
error tracker that auto-captures process env sees the key. This is pain
P37: secrets loaded from .env in production containers leak via env.
A second failure chains. A developer runs the staging deploy from a shell
where LANGCHAIN_ENV=production was set hours earlier. The loader picks
the prod .env, staging answers with a prompt commit tuned only for the
prod model tier, latency doubles. Two root causes: no type-safe env gate,
no startup validation that would have caught the mismatched model id.
Both are one refactor:
# BAD โ dotenv populates os.environ; any process with container access sees itfrom dotenv import load_dotenv
load_dotenv(".env.production")
api_key = os.environ["ANTHROPIC_API_KEY"] # P37: leaks via `docker exec env`# GOOD โ SecretStr in a validated Settings object, pulled from Secret Managerfrom pydantic import SecretStr
from pydantic_settings import BaseSettings
classSettings(BaseSettings):
env: Literal["dev", "staging", "prod"]
anthropic_api_key: SecretStr
settings = build_settings() # pulls from GCP Secret Manager in prod
api_key = settings.anthropic_api_key.get_secret_value()
# repr(settings) prints `SecretStr('**********')` โ safe to log
This skill owns the per-env config plumbing โ Settings skeleton,
Secret Manager integration, per-env pinning, startup smoke test. It does
own the full secrets lifecycle (rotation, revocation, scope) โ
that belongs to .
SecretStr masks repr(settings) to SecretStr('**********') โ a routine
logger.info(settings) cannot leak the key. The only way to read plaintext
is .get_secret_value(), which greps like a sore thumb in review.
extra="forbid" catches typos (LANGCHIN_MODEL_ID) at import time.
HttpUrl rejects http:/otel:4318 before the exporter wastes 60s on DNS.
Step 2 โ Per-env config loading (file OR Secret Manager, never both)
import os
from pathlib import Path
defbuild_settings() -> Settings:
env = os.environ.get("LANGCHAIN_ENV", "dev")
if env == "dev":
# Local dev: .env.dev file, values checked into 1Password not gitreturn Settings(_env_file=Path(".env.dev"))
if env == "staging":
# CI / staging: env vars injected by the orchestrator# (GitHub Actions secrets, k8s envFrom: secretRef, etc.)return Settings() # reads os.environ directlyif env == "prod":
# Prod: pull from Secret Manager into memory ONLY
values = pull_from_secret_manager()
return Settings(**values)
raise ValueError(f"unknown LANGCHAIN_ENV: {env!r}")
Three loaders, one class. Dev touches a file on disk. Staging inherits env
vars from the orchestrator โ envFrom: secretRef is readable via
docker exec env, but the blast radius is bounded and rotation is weekly.
Prod is the P37 fix: pull_from_secret_manager() builds a dict and passes
kwargs to Settings(...). Values land in the instance attribute and
never touch os.environ. A subprocess will not inherit them.
from google.cloud import secretmanager
defpull_from_secret_manager() -> dict[str, str]:
client = secretmanager.SecretManagerServiceClient()
project = os.environ["GCP_PROJECT_ID"]
secret_names = ["ANTHROPIC_API_KEY", "OPENAI_API_KEY", "LANGSMITH_API_KEY"]
out: dict[str, str] = {}
for name in secret_names:
resource = f"projects/{project}/secrets/{name}/versions/latest"
response = client.access_secret_version(request={"name": resource})
out[name] = response.payload.data.decode("utf-8")
# Non-secret passthrough (model id, prompt hash, endpoints)for key in ["LANGCHAIN_ENV", "LANGCHAIN_MODEL_ID", "LANGCHAIN_PROMPT_COMMIT",
"LANGCHAIN_VECTOR_INDEX", "LANGCHAIN_CHECKPOINTER_URL",
"OTEL_EXPORTER_OTLP_ENDPOINT"]:
if key in os.environ:
out[key] = os.environ[key]
return out
No os.environ[k] = v line. The dict goes straight into
Settings(**values). Workload-identity IAM handles auth; no static key on
disk. For AWS / Vault see Secret Manager Integration.
Step 4 โ Per-env model and prompt pinning
Dev, staging, and prod run different model ids and different prompt
commit hashes. Pinning happens at env-var level so app code is env-agnostic
(see the Env Matrix below for values). One function reads
settings.prompt_commit_hash and pulls from LangSmith
(cross-ref langchain-prompt-engineering):
from langsmith import Client
ls = Client(api_key=settings.langsmith_api_key.get_secret_value())
defget_prompt(settings: Settings) -> ChatPromptTemplate:
return ls.pull_prompt(f"triage-prompt:{settings.prompt_commit_hash}")
Prevents: staging loading a prod prompt commit. Pinning per env makes
promotion explicit โ dev โ staging โ prod moves one hash at a time. See
Per-Env Pinning.
Step 5 โ Per-env checkpointer selection
Checkpointer choice is per-env too:
from langgraph.checkpoint.memory import MemorySaver
from langgraph.checkpoint.postgres import PostgresSaver
defbuild_checkpointer(settings: Settings):
if settings.env == "dev":
return MemorySaver() # ephemeral, resets on restart# staging + prod: Postgres with env-isolated schema# cross-ref langchain-langgraph-checkpointing (P20) for schema migrationreturn PostgresSaver.from_conn_string(
str(settings.checkpointer_url)
)
Dev uses MemorySaver โ no infra dependency, no state between runs.
Staging and prod use PostgresSaver against separate databases (or
separate schemas). Never share a checkpointer DB between envs; P20 explains
โ schema migrations on a version bump corrupt cross-env threads.
Step 6 โ Startup smoke test (< 10 seconds budget)
Call validate_integrations(settings)before the HTTP server binds.
Failure aborts the deploy โ the readiness probe never goes green, the
rollout halts, the bad version takes no traffic. Budget: 10 seconds.
Past 10s an integration is degraded โ fail loudly rather than ship a 30s
cold start. See Startup Smoke Test.
Output
Settings class on pydantic-settings with SecretStr for keys, Literal env, HttpUrl endpoints, extra="forbid"
Env-specific loader (file โ dev; env vars โ staging; Secret Manager โ prod); values land in Settings only, never os.environ
Cloud Secret Manager integration (GCP / AWS / Vault) with IAM-bound auth; no static keys on disk
Per-env pinning for model_id, prompt_commit_hash, vector_index_name, checkpointer_url
Per-env checkpointer (MemorySaver dev, PostgresSaver on isolated DBs staging/prod)
Startup smoke test โ model / vector / checkpointer / observability under 10-second budget
Env Matrix
Dimension
dev
staging
prod
Secret backend
.env.dev file (git-ignored)
orchestrator env vars
cloud Secret Manager, memory only
os.environ holds keys
yes (local)
yes (sidecar visible)
no (P37 fix)
model_id
claude-haiku-4-6
claude-sonnet-4-6
claude-sonnet-4-6
prompt_commit_hash
WIP
canary
stable (1 week old)
temperature
0.7
0.2
0.2
Checkpointer
MemorySaver
PostgresSaver (staging DB)
PostgresSaver (prod DB)
Vector index
dev-index
staging-index
prod-index
OTEL sample rate
1.0
1.0
0.1
RPM limit
10
60
provider tier
Daily budget
$1
$10
$500-$5000
Smoke probes
model
model + checkpointer + OTEL
all four
Error Handling
Error
Cause
Fix
docker exec POD env shows ANTHROPIC_API_KEY=... in prod (P37)
dotenv / plain env injection in prod
Pull from Secret Manager into Settings(**values); never write to os.environ
Staging answers with prod prompts / wrong model
Loader defaulted or picked stale LANGCHAIN_ENV
Literal["dev","staging","prod"] on env; raise on unknown; no default
ValidationError: extra fields forbidden at startup
Typo (LANGCHIN_MODEL_ID)
Fix the typo โ extra="forbid" working as intended
Startup takes 30s before first request
Serialized probes or degraded integration
Enforce 10s budget; parallelize probes; fail the deploy
repr(settings) in a log leaks the API key
Plain str used, not SecretStr
Change field to SecretStr; repr masks to '**********'
Prod silently using MemorySaver
build_checkpointer defaulted when checkpointer_url was None
Require checkpointer_url in staging/prod via a model validator
Secret Manager auth fails in CI
SA not bound; google.auth fell back to ADC
Bind SA with roles/secretmanager.secretAccessor
Prompt hash rolled forward in staging without dev validation
Promotion skipped the dev gate
Enforce dev โ staging โ prod order in CI (see per-env pinning ref)
Examples
Graduating a .env-in-dev service to prod
Start: a single .env committed (or leaked via docker exec env). End:
Settings class, three loaders, Secret Manager in prod, smoke test under
10s. Three PRs โ (1) introduce Settings without changing loader behavior,
(2) add SecretStr and migrate call sites to .get_secret_value(),
(3) swap prod to Secret Manager and remove the prod .env from the image.
See Settings Skeleton and
Secret Manager Integration.
Wrong-env prompt loaded in staging โ postmortem
Staging inherited LANGCHAIN_ENV=production from a stale shell. The
Literal["dev","staging","prod"] field rejects production; CI promotion
sets LANGCHAIN_ENV explicitly; direnv pins it per-project. See
Per-Env Pinning.
Smoke test blocked a bad model id
A prod deploy went out with LANGCHAIN_MODEL_ID=claude-sonnet-4-7 (not yet
rolled out). The 1-token ping failed with model not found,
validate_integrations raised, the container crash-looped, the rollout
halted, the previous version kept taking traffic. Zero user impact; failure
budget stayed under 3s. See Startup Smoke Test.
Related skills in pack: langchain-security-basics (secrets lifecycle, owns rotation and revocation โ not duplicated here); langchain-langgraph-checkpointing (P20 schema migration); langchain-prompt-engineering (prompt pin / LangSmith pull workflow); langchain-reference-architecture (where Settings fits in the DI layer)
