ob_ai_call is the only input to V2 pricing, and it was dropping the two
numbers needed to interpret its own cost column.
tokens_reasoning was captured from the provider response but never written to
the table, so the cost of thinking mode was invisible once it left the
process. Folding it into tokens_out was not an option either -- it is billed
at the output rate, and mixing the two hides exactly the difference we now
have to watch.
rate_tier records which price list applied. Without it a row's cost_cents
cannot be rechecked: DeepSeek's peak rate is double off-peak from 2026-08-17,
so recomputing an old row at "the current rate" silently produces a different
answer than what was actually billed.
Migration 0002 adds both columns. tokens_reasoning defaults to 0 rather than
NULL so cost aggregates do not collapse to NULL on historical rows;
rate_tier stays nullable because rows written before this change genuinely
have an unknown tier, and NULL says that more honestly than a guessed 'flat'.
The smoke script now asserts both: reasoning tokens must be 0 (a non-zero
value means a provider default re-enabled thinking mode behind our backs) and
rate_tier must be present. Verified against the real API -- tokens_reasoning=0,
rate_tier=off_peak, 3.9s round trip.
Two defects that only surface against the real DeepSeek API, both of which
break the cost model rather than the code:
deepseek-v4-flash enables thinking mode by default. The same single-turn
question took 13.5s with it on and 1.8s with it off, and the thinking tokens
are billed at the output rate. That misses NFR-PERF-02b (15s per turn) and
inflates every conversational call. SceneConfig now carries an explicit
`thinking` flag, all four V1 scenes set it to false, and the provider always
sends the parameter rather than relying on the upstream default. Reasoning
tokens are recorded separately in Usage.tokens_reasoning so the cost of ever
turning it back on is visible.
DeepSeek moves to peak/off-peak pricing on 2026-08-17, with peak (Beijing
09:00-12:00 and 14:00-18:00) at twice the off-peak rate. Reports are generated
mostly during peak hours, so a single blended rate would leave ob_ai_call --
the only input to V2 pricing -- off by nearly a factor of two.
PeakAwareRateTable resolves the rate at call time; cost is computed and stored
then, never recomputed later at whatever the current rate happens to be.
Also here:
- Replace the live API key in the redaction test with a synthetic one. The
test that proves we redact secrets was itself a secret leak waiting for its
first commit.
- Correct LLM_COST_RATES_JSON, which held placeholder rates.
- Make ids.SnowflakeGenerator take an injectable clock so rollover and
clock-rollback are testable without patching a slots class.
- Rename exceptions to the -Error suffix (N818) and clear the remaining ruff,
mypy and import-linter findings, including two dead lines: a duplicate
pytestmark assignment and a no-op assert on a None return.
Gates: ruff, mypy (96 files incl. tests), 4/4 import contracts,
detect-secrets, 166 unit + 3 integration tests, migration up/down round-trip.
docker-compose brings up PostgreSQL 16 with pgvector and Redis 7 on
non-standard host ports so it never collides with an existing local database.
scripts/smoke.sh is the proof: compose, migrate, boot the API, issue a token,
call the real DeepSeek API and print the resulting ob_ai_call and bl_ledger
rows. services/vision and services/matting are placeholders noting that the
local RTX 5090 makes BiRefNet and RTMPose developable without cloud GPU.
Covers routing, failover, the circuit breaker, structured-output retries, the
cache-prefix validator, cost computation including cached-token pricing, the
daily budget breaker, the quota protocol, the task state machine and log
redaction. HTTP is mocked -- no unit test spends money. The integration test
runs the real slice against PostgreSQL and Redis with a faked provider.
One request exercises every layer: JWT auth, quota reservation, a versioned
prompt through the LLM gateway, a cost row, quota commit or refund, and the
unified envelope. The trivial ping task proves the queue routing and the
tk_task transitions before any real long-running task is written.
Routes are mounted at the /api/v1 base path of 架构文档 §9.1; /v1/* is
rewritten to it at the ASGI layer so both spellings work from one route table.
Creates the vector extension plus id_user, id_user_identity, id_consent,
bl_quota_rule, bl_quota, bl_ledger, tk_task and ob_ai_call using the domain
prefixes of 架构文档 §6.1, and seeds the quota rules. Reversible: verified by
an upgrade/downgrade/upgrade round-trip, which CI now runs on every push.
Billing implements reserve -> commit -> refund against bl_quota and bl_ledger
with row-level locking and idempotency keys; the refund commits on its own
transaction so a failed request cannot silently keep a user's quota. Task
implements the unified task centre over the state machine. Identity issues and
resolves users. Remaining domains are placeholders whose READMEs state what
belongs there and what must not.
DeepSeek (OpenAI-compatible) as primary with a 智谱 GLM slot that is a real
routing path but raises until the 算法备案 model list is confirmed. Adds scene
routing, failover, a per-provider circuit breaker, structured output with
retry-on-validation-failure, prompts loaded by id+version from prompts/, and
cost attribution written to ob_ai_call for every attempt -- successes,
retries and failures alike.
Two controls exist because they are cheap now and expensive later: a
cache-prefix validator that rejects timestamps, uuids and nicknames in the
cached system prefix (a volatile prefix silently multiplies input cost), and
a hard platform-wide daily budget that rejects non-critical scenes instead of
merely alerting, since a retry storm outruns any human on-call.
Typed pydantic-settings config, structlog JSON logging with request/trace id
propagation and redaction of phone numbers, tokens and face keys, the unified
response envelope and error-code table from 架构文档 §9.2, async SQLAlchemy
session management, ORM models for the eight foundation tables, and the
OpenTelemetry/Prometheus wiring.
packages/domain stays pure logic -- snowflake ids, the quota three-phase
protocol and the tk_task state machine -- so it can be tested without a
database and is held to mypy strict.
uv-managed PEP 621 project with ruff, mypy (strict on packages/domain and
packages/llm), pytest with coverage, and import-linter contracts that
mechanically enforce the dependency rules of 架构文档 §4.2 -- including the
two forbidden edges (compliance and identity must not depend on business
domains). CI, pre-commit and a detect-secrets baseline gate every push;
.env is gitignored before anything is ever staged.
xsl