Load this skill whenever you touch fastWorkflow configuration: any LLM_* / LITELLM_API_KEY_* / SPEEDDICT_FOLDERNAME / INTENT_DETECTION_* / SESSION_STATE_* env var, the fastworkflow.env or fastworkflow.passwords.env files, a fastworkflow CLI subcommand or flag, LiteLLM proxy routing, or JWT server settings. Trigger symptoms: "env var not found" warnings, "SPEEDDICT_FOLDERNAME env var not found!", "DSPy Language Model not provided", shell exports being mysteriously ignored, refine failing with a missing-LM ValueError, or needing to add a new config knob. Do NOT use it for how to run/operate workflows end to end (fastworkflow-run-and-operate), recreating the dev environment and secrets (fastworkflow-build-and-env), or debugging non-config runtime failures (fastworkflow-debugging-playbook).
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Load this skill whenever you touch fastWorkflow configuration: any LLM_* / LITELLM_API_KEY_* / SPEEDDICT_FOLDERNAME / INTENT_DETECTION_* / SESSION_STATE_* env var, the fastworkflow.env or fastworkflow.passwords.env files, a fastworkflow CLI subcommand or flag, LiteLLM proxy routing, or JWT server settings. Trigger symptoms: "env var not found" warnings, "SPEEDDICT_FOLDERNAME env var not found!", "DSPy Language Model not provided", shell exports being mysteriously ignored, refine failing with a missing-LM ValueError, or needing to add a new config knob. Do NOT use it for how to run/operate workflows end to end (fastworkflow-run-and-operate), recreating the dev environment and secrets (fastworkflow-build-and-env), or debugging non-config runtime failures (fastworkflow-debugging-playbook).
fastWorkflow Configuration and Flags Catalog
Facts verified 2026-07-09 against v2.22.2 (commit c33b9a5). Every file:line below was
checked against the working tree on that date. If you are reading this later, run the
re-verification commands in "Provenance and maintenance" before trusting line numbers.
Jargon, defined once:
dotenv file — a KEY=value text file parsed by python-dotenv's dotenv_values(). It is
NOT automatically exported to the process environment.
LiteLLM — the multi-provider LLM client library; model strings look like
mistral/mistral-small-latest. LiteLLM Proxy — a separate gateway server; model strings
get the litellm_proxy/ prefix and calls go to LITELLM_PROXY_API_BASE.
DSPy — the LLM-programming library fastWorkflow uses for agent/planner/param-extraction
calls; dspy.LM wraps a LiteLLM model string.
LLM role — fastWorkflow assigns a separate model + API key pair per job (agent, planner,
parameter extraction, synthetic data generation, ...), so each role can use a different model.
When to use / when NOT to use
Use this skill when you need to:
Look up any env var: its default, consumer, and sharp edges (Section 3).
Understand why a shell export is ignored (Section 2 — the precedence trap).
Look up a CLI subcommand/flag, including the ones the --help text lies about (Section 5).
Add a new env var or CLI flag correctly (Section 7 checklist).
Passwords are merged AFTER the env file, so a key in the passwords file wins on conflict.
dotenv_values never touches os.environ; config lives in an in-process dict
(fastworkflow._env_vars, set at fastworkflow/__init__.py:174-176).
fastworkflow.init() must run before using framework classes (module globals like
CommandContextModel are None until then) and before importing modules that read env vars
at import time (see Section 4 traps).
init() also re-applies LOG_LEVEL from the dotenv dict (__init__.py:180-182).
When the CLI's env-file arguments are omitted, defaults resolve to
<workflow_path>/fastworkflow.env and <workflow_path>/fastworkflow.passwords.env
(cli.py:177-191). The --help text saying ".env in current directory" is WRONG — see Section 5.
For fetched examples, fastworkflow examples fetch copies the two templates to ./examples/
(not into each example folder), so you must pass them explicitly:
fastworkflow train ./examples/<name> ./examples/fastworkflow.env ./examples/fastworkflow.passwords.env.
2. get_env_var precedence — the code default SHADOWS the OS environment
fastworkflow.get_env_var(var_name, var_type=str, default=None) at
fastworkflow/__init__.py:211-227 resolves in this exact order:
_env_vars dict (the two dotenv files passed to init())
the default= argument at the call site, if one was provided
os.getenv(var_name) — consulted ONLY when no code default exists
None + a logged warning
Consequence (the trap): for any var whose call site passes a non-None default=, a shell
export VAR=... is silently ignored — step 2 returns before os.getenv is ever consulted.
You can only override such vars via the env FILES. The shell-unoverridable vars as of v2.22.2
(call sites verified): SESSION_STATE_STORE (default "disk", session_state_store.py:123),
INTENT_DETECTION_TINY_MODEL and INTENT_DETECTION_LARGE_MODEL
(model_pipeline_training.py:892-895). Put overrides for these in fastworkflow.env, never
in the shell.
Call sites that pass default=None (e.g. SESSION_STATE_REDIS_URL/REDIS_URL at
session_state_store.py:127-129, LITELLM_PROXY_API_KEY at utils/dspy_utils.py:58) do NOT
trigger the shadow — the guard is if default is not None — so those fall through to the OS
environment normally.
Type coercion: var_type=bool accepts only true/1/false/0 (case-insensitive), else
ValueError (__init__.py:232-238). A provided default is returned as-is, uncoerced.
Masking hazard: importing litellm runs load_dotenv() over ./.env in the CWD
(verified: site-packages/litellm/__init__.py:29). The repo root has a generated .env
(make gen-env), so import-time env bugs that would crash a customer deployment are invisible
when developing in-repo. This masked the v2.21.0-2.21.3 regression fixed in 79e6986.
3. Which subcommands load env files (and which ignore them)
Subcommand
Env files loaded?
Evidence
Consequence
build
NO — fastworkflow.init(env_vars={})
build/__main__.py:326,349
LLM vars must be OS-environment exports (step 3 fallback works because call sites pass no default)
refine
NO — init(env_vars={})
refine/__main__.py:32
Same; no env-file CLI args exist at all
train
YES (both files)
train/__main__.py:269-271
—
run
YES (both files)
run/__main__.py:125-128
—
run_fastapi_mcp
YES (both files, in lifespan)
run_fastapi_mcp/__main__.py:259-264
Plus a pre-lifespan read of LOG_LEVEL for uvicorn (:1612-1617)
So before fastworkflow build / fastworkflow refine you must:
export LLM_COMMAND_METADATA_GEN=mistral/mistral-small-latest
export LITELLM_API_KEY_COMMANDMETADATA_GEN=<your key> # note: no underscore in COMMANDMETADATA
Nothing in --help tells you this; the failure mode is
ValueError: DSPy Language Model not provided. Set LLM_COMMAND_METADATA_GEN environment variable.
(raised from utils/dspy_utils.py:44-45 via build/genai_postprocessor.py:239).
File-presence probes at startup:
run HARD-FAILS if LITELLM_API_KEY_SYNDATA_GEN is absent (run/__main__.py:131-132) even
though run never generates synthetic data — it is used purely as a "did the passwords file
load?" probe. This breaks Bedrock/proxy users who legitimately have no such key; workaround:
put a dummy value in the passwords file.
train only WARNS for the same missing key ("OK if this is Bedrock",
train/__main__.py:276-277) but hard-fails on missing SPEEDDICT_FOLDERNAME (:273-275),
as does run (run/__main__.py:129-130).
4. Env var catalog
Legend: prod = production, load-bearing. dead = zero consumers in code.
"Default" = code default at the consumer (template value shown separately).
4a. LLM role variables (all resolved via dspy_utils.get_lm unless noted)
Model var (template value: mistral/mistral-small-latest)
Key var missing from the cli.py:136-139 fallback passwords stub
LLM_PARAM_EXTRACTION
LITELLM_API_KEY_PARAM_EXTRACTION
utils/signatures.py:252-255
prod
Bypasses get_lm: direct dspy.LM(model, api_key=...) at signatures.py:255, values cached in module globals after first read. Because no api_base is passed, litellm_proxy/ routing very likely does NOT work for this role (code-read inference, not runtime-verified)
Key doubles as run-time file-presence probe (Section 3)
LLM_CONVERSATION_STORE
LITELLM_API_KEY_CONVERSATION_STORE
run_fastapi_mcp/conversation_store.py:331
prod (server)
Key var missing from the cli.py fallback stub
LLM_COMMAND_METADATA_GEN
LITELLM_API_KEY_COMMANDMETADATA_GEN
build/genai_postprocessor.py:239 (used by build and refine)
prod (build-time)
(1) NOT in the packaged templates — only in docs/genai_postprocessor_readme.md and repo-local env/.env + passwords/.env; (2) key spelling is COMMANDMETADATA (no underscore) unlike every other key; (3) must be an OS export (Section 3)
LLM_RESPONSE_GEN
LITELLM_API_KEY_RESPONSE_GEN
none (rg over fastworkflow/ + tests/*.py: zero hits)
dead
Templated at examples/fastworkflow.env:8, examples/fastworkflow.passwords.env:9, and written by the cli.py:138 fallback stub — pure drift. Whether it is reserved for a future response-gen stage is an open question; do not delete without change control
LiteLLM Proxy routing (utils/dspy_utils.py:42-69): if a model string starts with
litellm_proxy/, get_lm requires LITELLM_PROXY_API_BASE (raises ValueError if unset,
:50-55), optionally uses LITELLM_PROXY_API_KEY (default None, no-auth proxies allowed,
:58), and IGNORES the per-role LITELLM_API_KEY_*. The [server] extra is NOT needed for
proxy routing. Full recipe: references/litellm-proxy-and-local-dev.md.
Roots ALL disk state; hard-required by run and train (Section 3)
SESSION_STATE_STORE
"disk"
not templated
session_state_store.py:123
prod (server)
redis value can ONLY be set via env file (default shadows shell, Section 2)
SESSION_STATE_REDIS_URL / REDIS_URL
None
not templated
session_state_store.py:127-129
prod (server, redis only)
ValueError if SESSION_STATE_STORE=redis and neither is set
INTENT_DETECTION_TINY_MODEL
google/bert_uncased_L-4_H-128_A-2
commented (fastworkflow.env:30)
model_pipeline_training.py:892-893
prod (train-time)
Env-file-only override (Section 2). Defaults chosen for transformers 4.48+/5.x compat
INTENT_DETECTION_LARGE_MODEL
distilbert-base-uncased
commented (fastworkflow.env:31)
model_pipeline_training.py:894-895
prod (train-time)
Same
SYNTHETIC_UTTERANCE_GEN_NUMOF_PERSONAS
none (int)
4
train/generate_synthetic.py:14
prod (train-time)
Import-time read trap: all three are read as module globals when generate_synthetic is first imported. Import before fastworkflow.init() captures None (a warning is logged; downstream failure inferred, not runtime-verified). Never import train modules before init()
THREE paths: utils/logging.py:53 (OS env at import, invalid value raises ValueError); __init__.py:180-182 (dotenv, via reconfigure_log_level); run_fastapi_mcp/__main__.py:1612-1617 (dotenv pre-read for uvicorn)
prod
Put it in fastworkflow.env; that covers paths 2 and 3. Shell export covers path 1 only
Set 0 to silence unserializable-artifact warnings; becomes a hard rejection in v3.0 per the docstring (turn.py:147-154)
PYTEST_RUNNING
—
—
set by tests/conftest.py:16; zero readers in fastworkflow/
dead
Safe to ignore; do not build logic on it
Not fastWorkflow config, despite appearances: repo-root config.yaml is a Dolt SQL server
config for the beads issue tracker; repo-root .env, env/.env, passwords/.env are the
local dev convention (see references/litellm-proxy-and-local-dev.md).
4c. Known drift summary (candidates for cleanup, gated by change control)
LLM_RESPONSE_GEN + key: templated but unconsumed (dead).
LLM_COMMAND_METADATA_GEN + key: consumed but untemplated, inconsistent key spelling.
cli.py:136-139 fallback passwords stub lists only 4 keys (SYNDATA_GEN, PARAM_EXTRACTION,
RESPONSE_GEN, AGENT) — omits PLANNER, CONVERSATION_STORE, COMMANDMETADATA_GEN; includes the
dead RESPONSE_GEN.
run's hard requirement on LITELLM_API_KEY_SYNDATA_GEN (probe misuse, Section 3).
5. CLI subcommands and flags
Entry point: fastworkflow = "fastworkflow.cli:main" (pyproject.toml:31). Six subcommands
(verified live with fastworkflow --help on 2026-07-09).
Subcommand
Positional args
Flags
Notes
examples list
—
—
Lists bundled examples
examples fetch
name
--force
Copies example to ./examples/<name>, copies/creates the two env templates in ./examples/ (cli.py:104-139)
Re-execs python -m fastworkflow.run_fastapi_mcp as a subprocess (cli.py:453-472); requires the [server] extra (cli.py:384-413)
Help-text lie: the env-file positionals for train/run/run_fastapi_mcp claim
"(default: .env in current directory, or bundled env file for examples)"
(cli.py:239,245,259,265,288,294). The actual default is
<workflow_path>/fastworkflow.env + <workflow_path>/fastworkflow.passwords.env
(find_default_env_files, cli.py:177-191). Trust the code.
--keep_alive sharp edge: defined with default=True and no type= (cli.py:270), so
any value you pass arrives as a string — --keep_alive False yields the truthy string
"False". Effectively this flag cannot be turned off from the CLI (empty-string workaround
untested). Downstream signature is keep_alive: bool (chat_session.py:150).
--expect_encrypted_jwt is unreachable from the wrapper: it exists only on the module
server (run_fastapi_mcp/__main__.py:358-359, action="store_true", default False), and the
CLI wrapper never forwards it (cli.py:454-469). Therefore:
# JWT signature verification DISABLED (trusted-network mode) — the only mode the wrapper offers:
fastworkflow run_fastapi_mcp ./examples/hello_world ./examples/fastworkflow.env ./examples/fastworkflow.passwords.env
# JWT signature verification ENABLED — must bypass the wrapper:
python -m fastworkflow.run_fastapi_mcp --workflow_path ./examples/hello_world \
--env_file_path ./examples/fastworkflow.env \
--passwords_file_path ./examples/fastworkflow.passwords.env \
--port 8000 --expect_encrypted_jwt
Note the argument style changes: the module form uses --workflow_path/--env_file_path
flags (load_args, run_fastapi_mcp/__main__.py:346-360), not positionals.
6. JWT constants and jwt_keys/ — hardcoded, not env vars
run_fastapi_mcp/jwt_manager.py:22-26 hardcodes (comment admits "can be made configurable via
env vars" — acknowledged unfinished work):
Constant
Value
JWT_ALGORITHM
RS256
JWT_ACCESS_TOKEN_EXPIRE_MINUTES
60
JWT_REFRESH_TOKEN_EXPIRE_DAYS
30
JWT_ISSUER / JWT_AUDIENCE
fastworkflow-api / fastworkflow-client
Keys live at ./jwt_keys/private_key.pem (mode 600) and public_key.pem (mode 644),
relative to CWD (jwt_manager.py:29-31, save at :88-100) — launch the server from a
stable directory or you'll silently generate a fresh keypair. jwt_keys/.gitignore excludes
the PEMs; never commit them.
Module default EXPECT_ENCRYPTED_JWT = True (jwt_manager.py:38) is ALWAYS overwritten at
startup by set_jwt_verification_mode(ARGS.expect_encrypted_jwt)
(run_fastapi_mcp/__main__.py:267-268), i.e. effective default is False/unverified.
Unsigned mode issues alg="none" tokens (jwt_manager.py:203).
Deployment security posture (unauthenticated /admin endpoints, open CORS) is covered in
fastworkflow-run-and-operate.
7. How to add a config axis (checklist)
Work through ALL of these; a var added to code but not the templates (or vice versa) becomes
the next LLM_RESPONSE_GEN / LLM_COMMAND_METADATA_GEN drift entry.
Consumer: read via fastworkflow.get_env_var("MY_VAR", ...) — never a scattered
os.environ read (sanctioned exceptions: LOG_LEVEL at logger import, and
FW_EAGER_ARTIFACT_VALIDATION). Never call it at module import time — read lazily inside
functions/properties (the 79e6986 lesson; litellm's cwd-.env load will hide the bug in-repo).
Decide the precedence tier consciously: passing default= to get_env_var makes
the var shell-unoverridable (Section 2). If ops must be able to override via shell, pass no
default and enforce requiredness at the CLI entry instead.
LLM role? Add BOTH LLM_<ROLE> and LITELLM_API_KEY_<ROLE> and resolve through
dspy_utils.get_lm("LLM_<ROLE>", "LITELLM_API_KEY_<ROLE>") so litellm_proxy/ routing
works. Keep the _ spelling consistent (do not imitate COMMANDMETADATA).
Templates: add to fastworkflow/examples/fastworkflow.env (settings) or
fastworkflow/examples/fastworkflow.passwords.env (secrets), with a comment block.
CLI fallback stub: if it's an API key, add it to the stub writer at cli.py:134-139.
build/refine reachability: if the var is consumed on the build/refine path, document
that it must be an OS export (those subcommands init with env_vars={}).
This catalog: add a row to Section 4 (or 4a) with consumer file:line and sharp edges,
plus a re-verification grep in the Provenance section.
Docs: update docs/genai_postprocessor_readme.md / relevant doc of record if applicable
(see fastworkflow-docs-and-positioning).
Tests: remember tests init from repo-local ./env/.env + ./passwords/.env
(tests/test_command_executor.py:24-27); add your var there if tests need it, and re-run
make gen-env mentally — it merges every *.env in the tree into root .env.
Change control: renaming/removing an existing var is a breaking change
(deployed env files reference old names — e.g. the COMMANDMETADATA spelling is frozen by
existing installs until proven otherwise). Route through fastworkflow-change-control.
Do NOT commit or push any of this without the developer'sexplicit request in that turn.
Run scripts/audit_env_catalog.sh (read-only) after any config change: it re-derives consumed
vs templated vars and prints drift.
Provenance and maintenance
Facts verified 2026-07-09 against v2.22.2 (commit c33b9a5). Re-verification one-liners (run from repo root):
Import-time reads: sed -n '14,16p' fastworkflow/train/generate_synthetic.py; sed -n '19,23p' fastworkflow/_workflows/command_metadata_extraction/parameter_extraction.py
Intent model defaults: sed -n '890,896p' fastworkflow/model_pipeline_training.py
CLI flags + help-text lie + env defaults: sed -n '177,302p' fastworkflow/cli.py and fastworkflow <subcommand> --help
