- Session class wrapping a clawdforge session_id; context-manager auto-close
- forge.session(agent=...) block form (preferred)
- forge.create_session() / forge.list_sessions() / forge.get_session() admin shapes
- TurnResult dataclass with .text() helper concatenating text events
- Idempotent Session.close() — safe in finally / __exit__
- tests/test_sessions.py: 16 tests covering block/manual/idempotency/exception/list/state/text-helper/v0.1-regression
- README "Multi-turn / Sessions (v0.2)" section
- pyproject version 0.1.0 -> 0.2.0; package __version__ matches
Architecture: matches the existing v0.1 client — sync, requests-based,
single Forge-owned requests.Session for connection pooling. Session holds
a back-reference to the Forge for HTTP work (no per-Session HTTP client).
This mirrors how Forge already exposes its own context-manager pattern, so
nothing about the threading/lifecycle story changes for callers.
v0.1 /run path unchanged — 49 existing tests still green, +16 new tests
for v0.2 (target was 9; covered the spec's 9 plus extras for empty-prompt
local validation, include_closed=false param, empty-id ValueError, and
s.state() round-trip).
mypy --strict src/clawdforge/ clean.
Spec: memory/spec-clawdforge-v0.2.md
Server core:
|
||
|---|---|---|
| .. | ||
| examples | ||
| src/clawdforge | ||
| tests | ||
| pyproject.toml | ||
| README.md | ||
clawdforge — Python SDK
Sync Python client for the clawdforge HTTP service. Wraps the bearer-token-gated REST API behind a typed dataclass surface so other Sulkta apps (cauldron, petalparse, johnny5) don't have to hand-roll requests calls.
Install
pip install -e clients/python/
(LAN-only, not on PyPI.)
Quickstart
from clawdforge import Forge
forge = Forge(base_url="http://192.168.0.5:8800", token="cf_...")
print(forge.healthz())
# {'ok': True, 'claude_present': True, 'claude_version': '1.x.y'}
r = forge.run(prompt='Reply with JSON: {"hello": "world"}')
print(r.result) # {'hello': 'world'}
print(r.duration_ms) # 4321
print(r.stop_reason) # 'end_turn'
Forge is also a context manager — preferred when you have a script-shaped lifetime:
with Forge(base_url="...", token="...") as forge:
forge.run(prompt="...")
Multi-turn / Sessions (v0.2)
For agent-shaped workflows that span multiple turns, use a Session. A session keeps acpx context across turn() calls — file reads, tool-call results, and conversational state persist across turns until the session is closed (or hits the server's TTL).
Block form (preferred — auto-close)
with forge.session(agent="claude") as s:
r1 = s.turn("Read README.md and summarize it")
print(r1.text()) # concatenated 'text' events from the model
r2 = s.turn("Now look at the auth flow")
print(r2.duration_ms) # 12345
for ev in r2.events:
print(ev["type"], ev.get("content"))
# Session is closed automatically here, even if an exception was raised
# inside the block.
Manual form (explicit lifecycle)
For callers that need a Session to outlive a single block (background workers, cross-handler state):
s = forge.create_session(agent="claude", meta={"caller": "petalparse"})
try:
r = s.turn("hello")
finally:
s.close() # idempotent — safe to call repeatedly, including on a
# session that was already closed by the server's TTL sweeper
Session.close() is idempotent: subsequent calls short-circuit without an HTTP request, and the server itself returns 200 (already_closed: true) on a repeat DELETE. Safe to drop in finally and __exit__ blocks without try/except.
TurnResult shape
@dataclass(frozen=True)
class TurnResult:
ok: bool
session_id: str
turn_index: int # 1-based, monotonic per session
events: list[dict] # 'text' | 'tool_call' | 'thinking' | ...
stop_reason: str | None # 'end_turn' | 'timeout' | 'error' | ...
duration_ms: int
def text(self) -> str: ... # concat of all 'text' events' content
events is passed through verbatim from the v0.2 server — see the server doc for the exhaustive event-type schema. text() is sugar for "give me the model's prose reply, ignore thinking and tool calls."
Listing & inspection
# All sessions for the calling token (newest first):
sessions = forge.list_sessions()
for row in sessions:
print(row["session_id"], row["turn_count"], row["closed_at"])
# Drop closed ones:
live = forge.list_sessions(include_closed=False)
# Fresh state for a specific session:
state = forge.get_session("sess_abc")
# {'session_id': 'sess_abc', 'agent': 'claude', 'cwd': ...,
# 'created_at': ..., 'last_turn_at': ..., 'turn_count': 3,
# 'closed_at': None, 'live': True, 'meta': {...}}
# Same thing from inside a block:
with forge.session() as s:
s.turn("hi")
print(s.state()["turn_count"])
Errors
The same exception classes apply as v0.1. Two cases worth calling out:
- Cross-token access →
ForgeAPIError(404). If token A asks for token B's session, the server returns 404 (not 403) so it can't be used to enumerate live session ids across token boundaries. - Session closed mid-flight →
ForgeAPIError(410). The TTL sweeper or another caller's explicitclose()may have ended the session between turns.
from clawdforge import ForgeAPIError
try:
r = s.turn("hello")
except ForgeAPIError as e:
if e.status_code == 410:
# Session expired or was closed — start a fresh one.
...
elif e.status_code == 404:
# Server forgot us, or we're asking for someone else's id.
...
raise
Construction
Forge(
*,
base_url: str,
token: str,
default_model: str = "sonnet",
default_timeout_secs: int = 120,
http_timeout_margin: int = 30,
session: requests.Session | None = None,
)
base_url—http://host:port. Trailing slash stripped.token— bearer. App token (cf_...) for/runand/files; admin bootstrap token for/admin/*.default_model— passed to/runwhen caller doesn't override. Server default is alsosonnetso this matches.default_timeout_secs—timeout_secsfor/runwhen caller doesn't override.http_timeout_margin— seconds added to the run subprocess timeout to derive the HTTP-level timeout. The HTTP deadline must outlast the server's subprocess deadline so we don't bail while clawdforge is still doing legitimate work for us. Default 30s — same pattern cauldron has been running inline.session— bring your ownrequests.Sessionif you want shared connection-pooling across multiple Forges. If omitted, the Forge owns one and closes it on exit.
Methods
healthz() -> dict
forge.healthz()
# {'ok': True, 'claude_present': True, 'claude_version': '1.x.y'}
run(prompt, *, model=None, system=None, files=None, timeout_secs=None) -> RunResult
r = forge.run(
prompt="Sterilize this ingredient line: 'about 2 cups of cooked white rice'",
model="sonnet", # optional, defaults to forge.default_model
system="You are a precise recipe parser. Always reply with valid JSON.",
files=["ff_..."], # optional file tokens from upload_file()
timeout_secs=60, # optional, 5..600, defaults to forge.default_timeout_secs
)
r.ok # True
r.result # parsed JSON when claude returned valid JSON, else str
r.duration_ms # 4321
r.stop_reason # 'end_turn' | 'timeout' | 'error' | ...
upload_file(path_or_fileobj, *, ttl_secs=3600, filename=None, content_type=None) -> FileToken
Either path-like or a binary file-object:
ft = forge.upload_file("/path/to/recipe.png", ttl_secs=3600)
# FileToken(file_token='ff_...', ttl_secs=3600, size=12345)
# Or from a buffer:
import io
ft = forge.upload_file(io.BytesIO(b"..."), filename="snippet.txt")
r = forge.run(prompt="Extract recipe data", files=[ft.file_token])
The server enforces 60 <= ttl_secs <= 86400 and returns 400 for out-of-range values, surfaced as ForgeAPIError.
session(*, agent="claude", meta=None) -> ContextManager[Session] (v0.2)
Block-form session — auto-closes on __exit__. See the "Multi-turn / Sessions (v0.2)" section above.
create_session(*, agent="claude", meta=None) -> Session (v0.2)
Manual-lifecycle session. Caller is responsible for s.close() (typically in finally). close() is idempotent.
list_sessions(*, include_closed=True) -> list[dict] (v0.2)
List sessions for the calling token. Pass include_closed=False to skip closed rows.
get_session(session_id) -> dict (v0.2)
Fetch fresh server-side state for a session. Raises ForgeAPIError(404) if the id is unknown or belongs to another token.
Admin (admin-bootstrap-token gated)
admin = Forge(base_url="http://192.168.0.5:8800", token=ADMIN_BOOTSTRAP_TOKEN)
t = admin.create_token("cauldron", ip_cidrs=["172.24.0.0/16"])
# AppToken(name='cauldron', token='cf_brandnew_xxx', ip_cidrs=['172.24.0.0/16'], ...)
# t.token is the plaintext bearer — store it now, the server only keeps a hash.
admin.list_tokens()
# [AppToken(name='cauldron', token=None, ip_cidrs=['172.24.0.0/16'], created_at=..., last_used=..., enabled=True), ...]
admin.revoke_token("cauldron") # True on success; raises ForgeAPIError(404) if no such token
Errors
Everything the SDK raises descends from ForgeError:
ForgeError
├── ForgeTransportError — connection/TCP-timeout failures (server never responded)
└── ForgeAPIError — server returned 4xx/5xx; .status_code, .body available
└── ForgeAuthError — 401 or 403 specifically
ForgeAPIError carries:
.status_code: int.body: dict | str | None— parsed JSON if available, else text.message: str— short summary
Typical pattern:
from clawdforge import Forge, ForgeAuthError, ForgeAPIError, ForgeTransportError
try:
r = forge.run(prompt="...")
except ForgeAuthError:
# Re-mint your token, or check the IP allowlist.
raise
except ForgeAPIError as e:
if e.status_code == 502 and isinstance(e.body, dict):
# Subprocess failed/timed out. e.body has 'error', 'stderr',
# 'duration_ms', 'stop_reason'.
log.warning("clawdforge run failed: %s", e.body.get("error"))
raise
except ForgeTransportError:
# Network blip — clawdforge is down or unreachable. Caller decides
# whether to retry; the SDK does not retry.
raise
No retries by default
Clawdforge runs are not idempotent — each /run spawns a real claude -p subprocess that costs tokens and time. If you want retries, wrap calls with whatever retry policy fits your call-site (tenacity, backoff, hand-rolled). The SDK won't double-charge you behind your back.
Tests
cd clients/python
pip install -e .[test]
python -m unittest discover tests
Tests use responses to intercept HTTP calls — no live network, no live clawdforge.
Notes
- The
resultfield ofRunResultmatches whatever the server'srunnerparsed out ofclaude -p --output-format json: a dict when the prompt asked for JSON and the model complied, a string otherwise. The SDK passes it through as-is — your prompt determines the shape. - The server's
GET /admin/tokensreturnsip_cidrsas a comma-joined string (it's stored that way in SQLite). The SDK normalizes that tolist[str]onAppToken. The create response already returns a list, so both shapes round-trip cleanly. revoke_tokenon a non-existent name raisesForgeAPIError(404)rather than returningFalse. That matches the server's contract and lets callers distinguish "didn't exist" from "was disabled".