Files
interactive-story/backend/app/context/encoding.py
T
JesseMarkowitzandClaude Opus 5 c1a73b3d77 M1: make the first story turn work with no Internet
Phase 0B ran the upstream application on a network with no route out and
the first turn died in tiktoken, which downloads its BPE table the first
time anything counts a token. The browser separately fetched three font
families from Google on every page load. Neither is visible on a machine
that has been online once, which is why both now have tests.

The tokenizer table is vendored at
backend/app/context/vendor/cl100k_base.tiktoken and
backend/app/context/encoding.py builds the encoding from it directly,
verifying its SHA-256 against the digest tiktoken itself pins for that
URL. No code path in the tokenizer can reach the network any more —
not a warm cache, not an environment variable a deployment could forget.
The encoding was checked token for token against tiktoken's own.

The three font families are self-hosted as variable fonts under
frontend/public/fonts/ (343 KiB, Latin and Latin Extended), declared in
frontend/src/styles/fonts.css, and re-vendored by
frontend/tools/vendor_fonts.py. Their OFL licences ship beside them.
With no remote asset left, the CSP drops both Google hosts and gains
object-src, base-uri and form-action; woff2 also gets its real media
type, which Python's table lacks on a slim image.

A trusted-LAN Ollama turned out not to work at all over HTTPS. httpx
verifies against the certifi bundle, so an endpoint whose certificate
comes from a CA the user installed on their own machines — a StartOS
server's Ollama, for one — was refused with CERTIFICATE_VERIFY_FAILED
while curl and the browser on the same host accepted it.
app/tlstrust.py builds one context that unions the platform CA store
with certifi's, and all four outbound clients use it. A union rather
than a swap, so an image with an empty system store cannot start failing
on endpoints that worked before. Verification itself is untouched:
CERT_REQUIRED, hostname checking on, and no insecure escape hatch.

The storyteller listener is now loopback by explicit statement rather
than by inheriting uvicorn's default: start.sh, start.ps1, and
docker-compose.yml, which publishes to 127.0.0.1 rather than every
interface. Reaching an Ollama on another machine is outbound and needs
none of that inbound exposure.

backend/requirements.lock pins the exact tested closure;
requirements.txt keeps the ranges. DEVELOPMENT.md covers setup, the
same-host and trusted-LAN Ollama configurations, and how to re-run the
offline proof. PROVENANCE.md records the upstream commit, the MIT terms,
and both vendored assets.

Verified, not just compiled. On an --internal Docker network with
1.1.1.1 unreachable and no name resolving, a campaign was created and
played for six turns through same-host Ollama, restarted, and resumed.
A second run played ten turns through Ollama on a separate physical
machine on the LAN over verified HTTPS, summaries and embeddings
included, with the storyteller's default route deleted so the LAN was
reachable and the Internet was not. Its capture: 893 packets to the
approved host, 730 loopback, zero anywhere else, and zero DNS queries.
Two induced model failures left the accepted story bit-identical. The
inherited SPA was opened in a browser and a campaign read back from it.
Evidence is in planning/reports/M1-BASELINE-REPORT.md, along with the
findings that did not belong in this change.

648 backend tests pass, up from the inherited 632; frontend lint and
build are clean; the image builds. No M2 work is included: the hosted,
cloud, analytics, Postgres and scripting surfaces are untouched.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017foPNqFjAJa2Ngebf5mEfL
2026-09-02 02:40:28 -04:00

91 lines
3.5 KiB
Python

"""`cl100k_base` without a first-use download.
Upstream called `tiktoken.get_encoding("cl100k_base")`, which fetches the BPE
table from `openaipublic.blob.core.windows.net` the first time it is used and
caches it under the system temp directory. That download is invisible on a
developer machine that has already made it once, and fatal on a machine with
outbound Internet blocked: the context builder counts tokens on *every* turn,
so the first story turn died with a `ConnectionError` instead of narrating
(Phase 0B offline report; acceptance tests A01 and H11).
The table is vendored beside this module and the `Encoding` is constructed from
it directly, so no code path inside the tokenizer can reach the network — not a
cache that happens to be warm, and not an environment variable a deployment
could forget to set.
The vendored file's SHA-256 is verified against the digest `tiktoken` itself
pins for that URL. A truncated checkout or a substituted table then fails
loudly, rather than silently changing every token count the context budget is
computed from.
"""
import base64
import functools
import hashlib
from pathlib import Path
import tiktoken
ENCODING_NAME = "cl100k_base"
#: Where the table came from, recorded so the vendored copy can be re-derived.
SOURCE_URL = "https://openaipublic.blob.core.windows.net/encodings/cl100k_base.tiktoken"
#: The digest `tiktoken_ext.openai_public.cl100k_base()` pins for SOURCE_URL.
BPE_SHA256 = "223921b76ee99bde995b7ff738513eef100fb51d18c93597a113bcffe865b2a7"
BPE_PATH = Path(__file__).resolve().parent / "vendor" / "cl100k_base.tiktoken"
# Both copied from `tiktoken_ext.openai_public.cl100k_base()`. They are part of
# the encoding's identity: the same merge table with a different pattern is a
# different tokenizer, so they are pinned here rather than imported, and the
# round-trip test asserts this build agrees with tiktoken's own.
_PAT_STR = r"""'(?i:[sdmt]|ll|ve|re)|[^\r\n\p{L}\p{N}]?+\p{L}++|\p{N}{1,3}+| ?[^\s\p{L}\p{N}]++[\r\n]*+|\s++$|\s*[\r\n]|\s+(?!\S)|\s"""
_SPECIAL_TOKENS = {
"<|endoftext|>": 100257,
"<|fim_prefix|>": 100258,
"<|fim_middle|>": 100259,
"<|fim_suffix|>": 100260,
"<|endofprompt|>": 100276,
}
def _mergeable_ranks() -> dict[bytes, int]:
try:
data = BPE_PATH.read_bytes()
except OSError as exc: # pragma: no cover - packaging fault, not a run fault
raise RuntimeError(
f"The vendored {ENCODING_NAME} table is missing at {BPE_PATH}. "
f"Re-download it from {SOURCE_URL} (SHA-256 {BPE_SHA256})."
) from exc
digest = hashlib.sha256(data).hexdigest()
if digest != BPE_SHA256:
raise RuntimeError(
f"The vendored {ENCODING_NAME} table at {BPE_PATH} has SHA-256 "
f"{digest}, expected {BPE_SHA256}."
)
# Same parse as tiktoken.load.load_tiktoken_bpe, minus its fetch/cache step.
ranks: dict[bytes, int] = {}
for line in data.splitlines():
if not line:
continue
try:
token, rank = line.split()
ranks[base64.b64decode(token)] = int(rank)
except Exception as exc:
raise ValueError(f"Error parsing line {line!r} in {BPE_PATH}") from exc
return ranks
@functools.lru_cache(maxsize=1)
def get_encoding() -> tiktoken.Encoding:
"""The `cl100k_base` encoding, built from the vendored table."""
return tiktoken.Encoding(
name=ENCODING_NAME,
pat_str=_PAT_STR,
mergeable_ranks=_mergeable_ranks(),
special_tokens=_SPECIAL_TOKENS,
)