gfa-sdk 0.3.0

Opinionated Python client for gfa (Git for Agents) — smart routing, session cache, profile-aware hints.

gfa-sdk — Python SDK for gfa (Git for Agents)

Opinionated Python client for gfa. Picks the right endpoint for the access pattern, caches what is safe to cache, and emits hints when a strategy change would help. Designed for AI agents reading, mutating, and forking gfa-hosted repos.

Install

pip install gfa-sdk

Python >= 3.10 required.

Quickstart

import os
import gfa

client = gfa.Client(
    endpoint=os.environ["GFA_ENDPOINT"],
    token=os.environ["GFA_JWT"],
)

# Profile a repo (cheap, cached for the session)
profile = client.profile_repo("myrepo")
print(f"{profile.file_count_at_head} files, {profile.storage_bytes} bytes")

# Read one file
src = client.read_file("myrepo", "src/main.go")

# Read many files — SDK picks /file vs /files/batch vs partial-clone
files = client.read_files("myrepo", ["a.go", "b.go", "c.go"])

# Create a commit
sha = client.create_commit(
    "myrepo",
    branch="main",
    message="agent: refactor",
    files=[gfa.FileChange(path="a.go", content=b"// rewritten\n")],
)

# Workspace for isolated changes
with client.create_workspace("myrepo", base_ref="main", ttl_hours=2) as ws:
    ws.create_commit(branch="main", message="WIP", files=[...])
    ws.merge()

Smart routing

client.read_files(repo, paths) auto-routes by len(paths):

Count	Surface
1	GET /file
2–200	POST /files/batch
> 200 (default)	SuggestPartialCloneError — opt into auto_clone=True for automatic partial clone

Threshold tuning:

from gfa import Client, ClientConfig
client = Client(
    endpoint,
    token,
    config=ClientConfig(batch_read_threshold=500),
)

Hints

The SDK emits structured hints when access patterns suggest a strategy change. The default handler writes to stderr; override with a logging hook:

def my_handler(hint):
    logger.info("gfa hint", extra={"code": hint.code, "repo": hint.repo})

client = gfa.Client(endpoint, token, hint_handler=my_handler)

Silence entirely:

client = gfa.Client(endpoint, token, hint_handler=lambda _: None)

Auth

Either a pre-minted JWT string, or a TokenProvider that mints per-call:

provider = gfa.FileKeyTokenProvider("/path/to/preview-jwt-priv.pem", ttl_hours=1)
client = gfa.Client(endpoint, provider)

FileKeyTokenProvider mints ES256 JWTs with sub=<repo>, iat=now, exp=now + ttl_hours. Mirrors the Go tools/mint-token binary.

Diagnostics

client.read_file("myrepo", "x.go")
client.read_files("myrepo", ["a.go", "b.go"])
print(client.stats)
# ClientStats(request_count=2, cache_hit_rate=0.0, ...)

Partial clone

with client.partial_clone("myrepo", filter="blob:none", depth=1) as repo:
    content = repo.read("src/foo.go")

Requires git binary on PATH.

Troubleshooting

Symptom	Likely cause
UnauthorizedError	Token missing, malformed, or expired
ForbiddenError	JWT sub does not match the repo being accessed
SuggestPartialCloneError	Asked for too many files; pass auto_clone=True or use partial_clone
GitBinaryMissingError	partial_clone needs git on PATH
ConflictError on merge	Workspace and source diverged — inspect via conflict_surface()

License

MIT.