feat(friendli): add GLM-5.3-Flash provider support

[Lee-Si-Yoon]

Summary

Adds Friendli as a provider host for zai-org/GLM-5.3-Flash (base_model = zhipuai/glm-5.3-flash, override-only).

available at https://api.friendli.ai/serverless/v1/models

Reasoning

reasoning_effort values low / high / max per chat_template.jinja:

{%- set effective_reasoning_effort = reasoning_effort if reasoning_effort is defined and reasoning_effort in ['low', 'high'] else 'max' -%}

Reasoning is always enabled (no toggle) — unrecognized/omitted values fall back to max.

Cost (USD/MTok)

• input: $0.15
• output: $0.5
• cache_read: $0.03

Checklist

• Non-lab host uses base_model
• Provider file is override-only (cost + reasoning_options + interleaved only)
• reasoning = true ⇒ reasoning_options set
• Costs in USD/MTok
• bun validate passes

[github-actions]

Action items

• [low] [possible mistake] providers/friendli/models/zai-org/GLM-5.3-Flash.toml:10 - Check: Data-changing PRs should cite direct provider pricing (or equivalent) mapped to the claimed values. Why: The file sets Friendli cost at input 0.15 / output 0.5 / cache_read 0.03 USD/MTok, and the PR body restates those numbers, but neither points to a Friendli pricing/docs/API source for this host. Those figures match common list rates on other hosts and cannot be confirmed as Friendli’s rates from the PR alone. Action: Add a Friendli-first-party pricing (or catalog) citation in the PR body and/or a leading TOML comment stating what it supports, or correct the costs if Friendli differs.

[github-actions]

The PR has a concrete modalities contradiction that needs fixing before merge.

## Action items
- **[medium] [possible mistake]** `providers/friendli/models/zai-org/GLM-5.3-Flash.toml:11` - **Check:** Provider modality overrides must reflect what this host actually serves, not only the upstream lab. **Why:** Patch 2 cited Friendli `GET /serverless/v1/models` (2026-08-27) reporting `input_modalities: ["text"]` for `zai-org/GLM-5.3-Flash`, then patch 3 set `input = ["text", "image"]` with only a future-tense claim that Friendli “will serve” image input and no Friendli-side confirmation. That can mislabel the host as vision-capable when the catalog API said text-only. **Action:** Verify Friendli’s live modalities for this model ID. If the API still reports text-only, set `input = ["text"]` (and `attachment = false` if attachments are not served). If image is actually supported, keep the override and add a leading top-of-file comment citing the Friendli source that confirms image input (not only the lab checkpoint).
- **[low] [possible mistake]** `providers/friendli/models/zai-org/GLM-5.3-Flash.toml:14` - **Check:** Material cost changes should cite a direct provider pricing source. **Why:** Costs (`0.15` / `0.5` / `0.03`) match common third-party list prices but the PR body does not point to Friendli pricing or docs for this model, so host-specific pricing cannot be reviewed from the submission alone. **Action:** Add a Friendli pricing/docs citation (PR body and/or leading TOML comment) that maps to these USD/MTok values, or confirm they are intentional list prices from a named Friendli source.

[github-actions]

Action items

• [medium] [possible mistake] providers/friendli/models/zai-org/GLM-5.3-Flash.toml:11 - Check: Provider [modalities] must match what this host actually accepts, not the lab defaults or planned support. Why: The PR first overrode to text-only citing Friendli GET /serverless/v1/models (input_modalities: ["text"]), then added image and video in later commits while dropping that API citation and giving no new Friendli evidence (only “upstream is multimodal” / “will serve”). Lab base includes pdf; peers disagree (text+image vs text+image+video). Wrong modalities mislead clients on attachments. Check: Confirm live Friendli model metadata or docs for zai-org/GLM-5.3-Flash, set input to the verified set, drop the override entirely only if it fully matches the lab entry, and set attachment if the resolved input is text-only.

[Lee-Si-Yoon]

Pushed 7f28c9573 addressing all three bot comments with live evidence:

1. Pricing citation: leading comment now cites live GET /serverless/v1/models for cost — confirmed input=0.00000015, output=0.0000005, input_cache_read=0.00000003 USD-per-token, converting to input=0.15, output=0.5, cache_read=0.03 USD/MTok exactly as authored.
2. Modality contradiction: verified live input_modalities: ["text","image","video"] for zai-org/GLM-5.3-Flash — matches the file's [modalities].input, so no change needed there, just the citation.
3. Added budget_tokens (min=-1, max=1_048_576) and a [limit] override (context=1_048_576, output=1_048_576) — live context_length/max_completion_tokens exceed the inherited lab file's rounded 1_000_000, and reasoning_budget is a documented request field per the chat-completions OpenAPI docs.

bun validate exit 0.

[github-actions]

No actionable findings.