Adds `AssistantMessage.responseModel` on the openai-completions path:
surfaces the concrete `chunk.model` when it differs from the requested
id (e.g. OpenRouter `auto` -> `anthropic/...`).
* feat(ai): add Cloudflare Workers AI as a provider
Cloudflare Workers AI hosts open-weight LLMs (Kimi K2.6, GPT-OSS,
GLM-4.7, Llama 4, Gemma 4, Nemotron 3) on Cloudflare's GPU network with
an OpenAI-compatible endpoint. Reuses the openai-completions API
protocol; the per-account URL contains a {CLOUDFLARE_ACCOUNT_ID}
placeholder resolved at request time by a small helper.
Pi automatically sets x-session-affinity for prefix caching:
https://developers.cloudflare.com/workers-ai/features/prompt-caching/
Auth: CLOUDFLARE_API_KEY (matches pi's *_API_KEY convention) +
CLOUDFLARE_ACCOUNT_ID. The User-Agent identifies traffic as
'pi-coding-agent' in Cloudflare analytics.
Verified end-to-end against a real Cloudflare account: 17 e2e tests
pass across stream/empty/tokens/unicode/tool-call-without-result/
total-tokens against @cf/moonshotai/kimi-k2.6.
Cloudflare AI Gateway is a separate, larger change (it requires routing
through provider-specific subpaths with the matching API protocol per
upstream) and will land in a follow-up PR.
* refactor(ai): move Cloudflare User-Agent and session-affinity flag to per-model metadata
Instead of conditionally setting them in openai-completions.ts based on
provider detection, declare them as model-level fields in the catalog
(headers + compat). This is consistent with how the github-copilot and
kimi-coding entries already declare their static headers.
packages/ai/scripts/generate-models.ts: emit headers and compat fields
on each cloudflare-workers-ai entry (CLOUDFLARE_STATIC_HEADERS).
packages/ai/src/providers/openai-completions.ts: drop the
isCloudflareProvider conditional that injected User-Agent and the
isCloudflareWorkersAI override of sendSessionAffinityHeaders.
packages/ai/src/models.generated.ts: re-spliced 8 cloudflare-workers-ai
entries with headers + compat.
Behavior is unchanged - verified via fetch interceptor that User-Agent
and x-session-affinity / session_id / x-client-request-id are still sent
on outbound requests. 5/5 e2e tests pass.
DashScope / Aliyun Qwen (OpenAI-compatible) rejects `tools: []`
with HTTP 400 `"[] is too short - 'tools'"`. Five providers used
a truthy check (`if (context.tools)`) that treated an empty array
as "send tools", so `pi --no-tools` produced `tools: []` in the
request body. Matching the Google provider's pattern, we now
guard on `context.tools.length > 0`:
- openai-completions.ts
- openai-responses.ts
- openai-codex-responses.ts
- azure-openai-responses.ts
- anthropic.ts
The openai-completions fallback that emits `tools: []` when the
conversation has tool history (required by LiteLLM / Anthropic
proxies) is preserved via the existing `else if (hasToolHistory)`
branch.
closes#3649
Co-authored-by: 槐聚 <huaiju@zbyte-inc.com>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: Mario Zechner <badlogicgames@gmail.com>
Mutate persisted tool-call blocks in place on function_call completion,
remove partialJson, and emit the same reference on toolcall_end.
Add regression coverage for persisted block cleanup and event identity.
fixes#3078
Map unknown finish_reason values (e.g. "end" from Ollama/LM Studio) to
"stop" instead of throwing, since assistant content is already produced.
fixes#2142
Some OpenAI-compatible providers (e.g., Moonshot/Kimi) return usage
data in chunk.choices[0].usage instead of the standard chunk.usage.
Extract usage parsing into a helper and check choice.usage as fallback.
closes#2017
The OpenAI Chat Completions API standard format for assistant message
content is a plain string. Sending it as an array of
{type:"text", text:"..."} objects causes some models (notably DeepSeek
V3.2 via NVIDIA NIM) to mirror the content-block structure literally
in their output. This produces recursive nesting where each turn wraps
the previous content blocks deeper:
[{'type':'text','text':'[{\'type\':\'text\',\'text\':...}]'}]
The fix unifies the assistant content serialization to always use a
joined string — the same approach already used for the github-copilot
provider — for all openai-completions backends.
Affected models observed: deepseek-ai/deepseek-v3.2 (nvidia provider).
Models like GLM-5, GPT-4, Claude were unaffected as they tolerate
array content, but sending a standard string is safer for all.
Co-authored-by: geraldoaax <geraldoaax@users.noreply.github.com>
Z.ai uses the same enable_thinking: boolean parameter as Qwen to control reasoning, not thinking: { type: "enabled" | "disabled" }.
The wrong parameter name means Z.ai ignores the disable request and always runs with thinking enabled, wasting tokens and adding latency.
Merge the Z.ai and Qwen branches since they use the same format.
PR by @okuyam2y
Mario Zechner