Skip to content

feat(oathbreaker): add DeepSeek V4.1 Flash models - #13837

Open
jIrwinCline wants to merge 1 commit into
anomalyco:devfrom
jIrwinCline:oathbreaker-deepseek-models
Open

jIrwinCline wants to merge 1 commit into
anomalyco:devfrom
jIrwinCline:oathbreaker-deepseek-models

Conversation

@jIrwinCline

@jIrwinCline jIrwinCline commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

Adds Oathbreaker's two DeepSeek V4.1 Flash models next to its existing GLM entries. Both are override-only on deepseek/deepseek-v4.1-flash-1017, in the same format as the GLM files.

Oathbreaker meters DeepSeek requests in bytes, so limit.context is a conservative 32,768-token estimate of its 131,072-unit budget, and the 262,144-byte request cap is in limit.request_bytes. The file headers explain both.

Sources: docs, pricing, llms.txt

bun validate passes. bun run v1:check shows only the two new models.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CfDjYwPPXJYCHDNNQTBkbk

@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [medium] [possible mistake] providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash.toml:41 (same at oathbreaker-deepseek-v4.1-flash-fast.toml:43) - Check: limit.context must state the host's effective token context window, and a byte-denominated request cap belongs in limit.request_bytes. Why: The file's own header (lines 13–15) says the 131,072 budget is counted in UTF-8 input bytes plus framing plus requested output, and that "131,072 is an upper bound and real prompts reach it at fewer tokens." A byte budget of 131,072 admits roughly a quarter of that in tokens for typical English text, so publishing it as limit.context knowingly overstates the window clients use to size prompts, and they will hit admission errors well below the published limit. It also contradicts the same gateway's existing entries, which record the identical 131,072 budget in tokens (oathbreaker-glm-5.3.toml: "131,072 GLM tokens shared by input, safety headroom and requested output") — at least one of the two unit interpretations is wrong. Action: Verify the unit of Oathbreaker's admission budget in its docs. If it is tokens, fix the header comments. If it is bytes, record the byte cap as limit.request_bytes (as the Cerebras and Snowflake Cortex entries do) and set limit.context to a token-basis value from the docs or a documented conservative token estimate, not the loose byte upper bound.

  • [low] [possible mistake] providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash.toml:16 (same at oathbreaker-deepseek-v4.1-flash-fast.toml:16) - Check: The reasoning header must state the host's real default-effort behavior unambiguously, and a documented default belongs in capabilities.reasoning.default. Why: "when omitted the gateway adds none" reads two ways: the gateway injects no effort parameter (the PR body's claim: "adds no effort when the field is omitted"), or it injects reasoning_effort = "none" — which would make none an accepted/default level that the effort = ["low", "high", "max"] list and the absent default both currently deny. The sibling GLM entries on this gateway record a documented default (default = { type = "effort", value = "low" }), so a per-model default may be documented for these aliases too. A future editor reading the file alone could add none to the effort list or set the wrong default. Action: Verify in Oathbreaker's docs what happens when reasoning_effort is omitted for these aliases; record default = { type = "effort", value = … } if a default is documented, and reword the header (e.g. "the gateway adds no effort parameter when the field is omitted") so it cannot be parsed as effort none.

Add oathbreaker-deepseek-v4.1-flash and oathbreaker-deepseek-v4.1-flash-fast,
based on deepseek/deepseek-v4.1-flash-1017, with Oathbreaker's text-only
modalities, byte-based limits, low/high/max effort and USD prices.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CfDjYwPPXJYCHDNNQTBkbk
@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [medium] [possible mistake] providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash.toml:44 (same at providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash-fast.toml:46) - Check: Published limit.context vs the PR description, and whether Oathbreaker's 131,072-unit admission budget counts input as bytes or tokens. Why: The PR body's table states limit.context = 131_072 and says "That makes 131,072 an upper bound... noted in the file headers", but both files actually publish context = 32_768 and the headers derive it as a bytes→tokens estimate (131,072 / ~4) — a 4× contradiction inside the PR itself. The existing Oathbreaker GLM entries treat the same 131,072 figure as a token budget (context = 131_072 in oathbreaker-glm-5.3*.toml), so at most one interpretation matches the host's docs; if the budget is tokens, these files understate context 4×, and if it is bytes, the PR body misdescribes what is being merged. Action: Verify from https://oathbreaker.ai/docs whether the budget for these aliases counts UTF-8 input bytes (keep 32_768) or tokens (use 131_072 like the GLM siblings), then align the two files and the PR body so they state the same value and semantics.
  • [low] [violation] providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash.toml:33 (same at providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash-fast.toml:35) - Check: Override-only provider files after base_model (AGENTS.md → Override-only provider files / README → "do not restate the base's ... modalities ... unless this host actually differs"). Why: output = ["text"] restates the base model's identical output modality — models/deepseek/deepseek-v4.1-flash-1017.toml already has output = ["text"]. Only input = ["text"] is a real delta here; [modalities] deep-merges, so the inherited output is unchanged either way. Action: Drop the output line from both files and keep only the text-only input override.

This branch was previously deployed

1 inactive deployment
release — 3499dd9c Deployed Nov 16, 2026 by jIrwinCline via review #19252
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