[missing-model] volcengine: add GLM-5.3-Flash (glm-5-3-flash-260828)

[fenglingQAQ233]

Summary

Add the GLM-5.3-Flash model to the Volcengine Ark provider. Volcengine serves it under the date-suffixed API ID glm-5-3-flash-260828 (same convention as the existing glm-5-2-260617 entry).

Upstream metadata already exists at models/zhipuai/glm-5.3-flash.toml (context 1M, output 131,072, text/image/video/pdf input, reasoning + tool_call + structured_output), so the provider TOML only needs base_model plus provider-specific fields.

Why this matters

This model is live on Volcengine Ark (verified against POST /api/v3/chat/completions, 2026-09-16) and already in use by real clients. Without a catalog entry, tools that derive context window from models.dev (e.g. pi/rikkahub auto-compaction threshold) fall back to a 128K default instead of the real 1M window, causing premature conversation compaction.

Changes

• providers/volcengine/models/glm-5-3-flash-260828.toml: new file with base_model = "zhipuai/glm-5.3-flash", reasoning options and interleaved field mirroring the existing glm-5-2-260617.toml.

Open questions

1. Cost: Volcengine bills GLM per-token; the sibling glm-5-2-260617.toml has a [cost] block. Left [cost] out to avoid guessing — happy to add if maintainers share the list price from https://www.volcengine.com/docs/82379/1544106.
2. Modalities: upstream lists text/image/video/pdf input; override with a [modalities] block if the Volcengine serving differs.

[github-actions]

Action items

• [high] [violation] providers/volcengine/models/glm-5-3-flash-260828.toml:4 - Check: Reasoning options must follow the lab + same-model peer baseline for this host, not a different generation’s Volcengine entry. Why: First-party providers/zhipuai/models/glm-5.3-flash.toml and same-model peers (e.g. tinfoil, tokengo, bothub) use always-on reasoning with effort low|high|max and no toggle. This file copies glm-5-2-260617 (toggle + minimal|low|medium|high), which is a different model and invents controls the 5.3-Flash lab surface does not expose. Action: Replace with lab/peer-aligned options (typically effort ["low", "high", "max"] only). Keep toggle or other effort values only with Volcengine-specific evidence that this model ID accepts them.
• [high] [violation] providers/volcengine/models/glm-5-3-flash-260828.toml:4 - Check: Every toggle requires a leading top-of-file wire-path comment. Why: The file declares type = "toggle" but the header only notes model-ID verification; AGENTS.md requires the exact request field (e.g. sibling’s # Toggle: thinking.type = enabled|disabled). Action: If toggle remains after the baseline fix, add the Volcengine wire comment above the first key; otherwise remove toggle and drop any toggle comment.
• [low] [possible mistake] providers/volcengine/models/glm-5-3-flash-260828.toml:1 - Check: Provider models should include USD/MTok cost when public list pricing exists. Why: Every other Volcengine model (including sibling glm-5-2-260617.toml) authors [cost] from https://www.volcengine.com/docs/82379/1544106 with a CNY→USD rate note; omitting cost leaves catalog consumers without pricing for a billed model. Action: Add [cost] from that page (convert CNY→USD with rate/date in the leading comment), or state in the PR why this endpoint is intentionally unpriced.

[fenglingQAQ233]

All review items addressed in the latest commit:

1. Reasoning options — replaced with the same-model baseline: effort = ["low", "high", "max"] only (no toggle). This matches the Volcengine docs for glm-5-3-flash-260828, which state thinking is always enabled and cannot be disabled (https://docs.volcengine.com/docs/82379/1449737), consistent with the zhipuai lab entry.

2. Toggle wire comment — no longer needed since toggle was removed.

3. Cost — added [cost] from the public Volcengine CNY list price (input ¥0.8/MTok, output ¥2.8/MTok, cache hit ¥0.23/MTok), converted at 6.737012 CNY/USD (2026-08-26) with source comments at the top of the file.

[github-actions]

No actionable findings.

[fenglingQAQ233]

Hi @oliver-mee — gentle nudge on this one 🙏

This PR adds GLM-5.3-Flash to the Volcengine Ark provider. Quick recap of why it matters and why it's ready to merge:

Why it matters

• The model is live on Volcengine Ark (verified against POST /api/v3/chat/completions, 2026-09-16) and already in production use by real clients.
• Without a catalog entry, tools that derive context windows from models.dev (e.g. pi/rikkahub's context-limit lookup and auto-compaction threshold) fall back to a 128K default instead of the real 1M window — causing premature conversation compaction for real users.

Ready to merge

• ✅ CI/bot validation passed ("No actionable findings", 2026-09-16 08:28 UTC)
• ✅ base_model = "zhipuai/glm-5.3-flash" references existing upstream metadata (context 1M, output 131,072)
• ✅ Reasoning options aligned with the same-model baseline: effort = ["low", "high", "max"] only (Volcengine docs confirm thinking is always-on for this model ID)
• ✅ [cost] added from the public Volcengine CNY list price (converted at 6.737012 CNY/USD, sources in-file)

I noticed GLM-related model PRs (e.g. #7183, #7208) were merged today — this one is in the same category. Happy to make any further adjustments. Thanks for your time!