Skip to content

release: free-tier deepseek serving option + Polar go-live evidence - #678

Merged
devswha merged 17 commits into
mainfrom
dev
Aug 3, 2026
Merged

release: free-tier deepseek serving option + Polar go-live evidence#678
devswha merged 17 commits into
mainfrom
dev

Conversation

@devswha

@devswha devswha commented Aug 3, 2026

Copy link
Copy Markdown
Owner

Ships PRs #671#677 to production: the free-tier deepseek reasoning controls (rewrite low-cut scoped to free+deepseek, scorer cut extended to deepseek, streaming extraBody passthrough) plus the go-live evidence docs. Enables the production env flip of the free tier to deepseek-v4-flash (1/8 cost, ~45s, decouples the free tier from the exhausted Gemini spend cap). Checkout remains disabled. Local gates: tests 1666/0, lint clean, hold valid; preview smoke 46.6s MPS 100/fidelity 100.

devswha added 17 commits August 3, 2026 05:40
PR #670 shipped the Polar binding to main; the alias serves the six-field
disabled launch shape, the Polar gate answers 401/403 fail-closed on 7.0.0,
and the preview build's designed failure against the stale LS enable flag is
recorded as live fail-closed proof. Open incident: the free-tier Gemini key
is over its monthly spend cap (429) — owner billing action before Gate-B
health evidence.
ops: DEP_PROD_DISABLED evidence — v7.0.0 live with checkout off
Meaning-gutting is gone (ko-news-01: MPS 24 -> 100) but five fixtures now
fail the fidelity floor; 15/22 vs the shipped engine's 20/22. Not a Pro swap
today; candidate for the free tier and for re-measurement on the next flash
update.
ops: deepseek-v4-flash-0731 remeasurement
Regenerated four failing deliveries: three violate the [BODY]/[SELF_AUDIT]
output contract (audit blocks and orphan tags reach the delivered text), and
ko-blog-01 fabricates a research-finding attribution absent from the source.
Contract drift is partially mitigable; fabrication is disqualifying for Pro.
ops: deepseek-0731 fidelity root cause
…he apparatus

The 15/22 run disabled thinking while the gemini baseline ran default
reasoning. With the override removed, 0731 scores 20 pass / 2 warn / 0 error
on the same 22 fixtures and judge — quality parity at ~8x lower cost. The
remaining tradeoff is latency (~60-90s vs ~8s per rewrite). n=1 caveat and
the frozen-default process both stand.
ops: deepseek-0731 correction — apparatus artifact, quality parity at 1/8 cost
budget_tokens is ignored; reasoning_effort works. off=4s/broken,
low=44s/20 pass with two marketing MPS drops, default=90s/clean. Decode is
~125 tok/s and the 22k prompt prefix cache-hits at 99.9%, so reasoning
volume is the only latency lever.
ops: deepseek-0731 reasoning-dial measurements
deepseek-v4-flash-0731 matches the live-quality gate at default thinking and
holds 20/22 at reasoning_effort low (~44s vs ~90s). Wire the measured dial:
scoringExtraBody now covers deepseek (its scorer thinking would dominate
latency/cost), and a new rewriteExtraBody cuts rewrite reasoning to low ONLY
for free-tier deepseek requests — BYOK and Pro rewrites keep their provider
defaults (the gemini amputation lesson). streaming-api gains extraBody
passthrough mirroring api.js semantics. .env.example documents the option;
its frozen hash is re-frozen.
…aBody

Review finding on #677: with modelRejectsTemperature cached true, the
spread-first extraBody could reintroduce temperature and replay the known
400. Both the streaming and buffered clients now strip it.
feat(web): free-tier deepseek reasoning controls
@vercel

vercel Bot commented Aug 3, 2026

Copy link
Copy Markdown
Contributor

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
patina Ready Ready Preview Aug 3, 2026 2:47am

Request Review

@devswha
devswha merged commit 19285e5 into main Aug 3, 2026
8 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant