API.md

DFlash Server API Reference

HTTP server exposing OpenAI-compatible, Anthropic, and Responses API endpoints for inference with speculative decoding. Model-independent — works with any backend (qwen35, qwen3, gemma4, laguna).

---

Endpoints

Method	Path	Description	Status
GET	/health, /	Health check	✅
GET	/v1/models	List available models	✅
POST	/v1/chat/completions	OpenAI Chat Completions	✅
POST	/v1/messages	Anthropic Messages	✅
POST	/v1/responses	OpenAI Responses API	✅

---

Agent Turn Cache

The in-memory prefix cache extends through generated tool calls. After the model emits a valid tool call, the server keeps the state the generation left behind as a checkpoint of the prompt plus the generated turn, as far as the next request renders that turn with the same tokens (tool memory replays the generated text, so normally all of it). On the next OpenAI Chat Completions or Responses request, only the appended tool result and the new suffix need prefill. When that request is the next one the server runs, it continues the live state without copying the checkpoint back.

This is a server-wide optimization for single-sequence serving; request bodies do not change. It requires --prefix-cache-slots to be nonzero. Compressed or token-rewritten prompts fall back to ordinary prefix caching.

--agent-turn-cache adds a fallback for turns the next request renders differently: the server reuses its deepest compatible prefix checkpoint, replays the uncached tail once while idle, and saves the canonical completed turn. The replay moves prefill work out of the follow-up request when tool execution is long enough to overlap it; it does not eliminate that work. Paged attention and --max-concurrency do not yet support shared prefix blocks, so they cannot be combined with --agent-turn-cache.

The exact full-prompt cache may remain enabled for identical-request hits. Disable it only when benchmarking the incremental Agent Turn Cache benefit.

Successful requests expose the measured result under usage.timings:

• agent_turn_cache_hit: the restored prefix includes a generated agent turn.
• cached_prefix_tokens: backend-confirmed tokens restored from KV state.
• prefilled_tokens: prompt tokens computed for this request.

---

POST /v1/chat/completions (OpenAI-compatible)

Supported Request Parameters

Parameter	Type	Default	Description	Status
model	string	server default	Model identifier	✅
messages	array	required	Conversation messages	✅
stream	bool	false	SSE streaming	✅
max_tokens	int	4096	Max output tokens	✅
max_output_tokens	int	4096	Alias for max_tokens	✅
max_completion_tokens	int	4096	Alias for max_tokens	✅
temperature	float	0.0	Sampling temperature (0 = greedy)	✅
top_p	float	1.0	Nucleus sampling threshold	✅
top_k	int	0	Top-k filtering (0 = disabled)	✅
seed	int	0	Random seed for reproducibility	✅
frequency_penalty	float	0.0	Penalize frequent tokens (range: -2.0 to 2.0)	✅
presence_penalty	float	0.0	Penalize present tokens (range: -2.0 to 2.0)	✅
repetition_penalty	float	1.0	HF-style multiplicative penalty (>1 penalizes)	✅
rep_pen	float	1.0	Alias for repetition_penalty	✅
rep_window	int	256	Token lookback window for penalties	✅
tools	array	none	Tool/function definitions	✅
reasoning	object	—	Reasoning effort control ({"effort":"medium"})	✅
chat_template_kwargs	object	—	Direct template control ({"enable_thinking":true}). Jinja templates (e.g. qwen4exp) also accept preserve_thinking (bool): whether earlier assistant turns replay their recorded reasoning_content inside <think>...</think> (default: template's own default, typically true)	✅
stop	string/array	—	Stop sequences	✅
n	int	—	Number of completions	❌ TODO
logprobs	bool	—	Return log probabilities	❌ TODO
top_logprobs	int	—	Number of top logprobs per token	❌ TODO
response_format	object	—	Structured output: json_object, or json_schema with a schema the reply follows exactly (grammar-constrained decoding)	✅
tool_choice	string/object	—	Tool choice; required or a named function forces a tool call	✅
logit_bias	object	—	Per-token logit adjustments	❌ TODO
user	string	—	End-user identifier (tracking)	❌ TODO
stream_options	object	—	Streaming options (e.g. include_usage)	❌ TODO 🔴

Response Fields

Field	Status
id	✅ chatcmpl_<hex>
object	✅ "chat.completion" / "chat.completion.chunk"
model	✅
choices[].message.role	✅
choices[].message.content	✅
choices[].message.tool_calls	✅
choices[].finish_reason	✅ (stop, length, tool_calls)
choices[].delta (streaming)	✅
usage.prompt_tokens	✅
usage.completion_tokens	✅
usage.total_tokens	✅
choices[].logprobs	❌ TODO

---

POST /v1/messages (Anthropic-compatible)

Supported Request Parameters

Parameter	Type	Default	Description	Status
model	string	server default	Model identifier	✅
messages	array	required	Conversation messages	✅
system	string/array	—	System prompt (top-level)	✅
stream	bool	false	SSE streaming	✅
max_tokens	int	4096	Max output tokens	✅
temperature	float	0.0	Sampling temperature	✅
top_p	float	1.0	Nucleus sampling	✅
top_k	int	0	Top-k filtering	✅
seed	int	—	Random seed	✅
frequency_penalty	float	0.0	Penalize frequent tokens	✅
presence_penalty	float	0.0	Penalize present tokens	✅
thinking	object	—	Thinking mode ({"type":"enabled"})	✅
tools	array	—	Tool definitions	✅
tool_choice	string/object	—	Tool choice; any or a named tool forces a tool call	✅
output_format	object	—	Structured output ({"type":"json_schema","schema":{...}})	✅
stop_sequences	array	—	Stop sequences	✅
metadata	object	—	Request metadata (tracing)	❌ TODO

Response Structure

Follows Anthropic Messages API structure with content blocks:

• type: "text" — text content
• type: "thinking" — reasoning/thinking content
• type: "tool_use" — tool call

---

POST /v1/responses (OpenAI Responses API)

Supported Request Parameters

Parameter	Type	Default	Description	Status
model	string	server default	Model identifier	✅
input	string/array	required	Input messages or text	✅
instructions	string	—	System instructions	✅
stream	bool	false	SSE streaming	✅
max_output_tokens	int	4096	Max output tokens	✅
temperature	float	0.0	Sampling temperature	✅
top_p	float	1.0	Nucleus sampling	✅
seed	int	—	Random seed	✅
frequency_penalty	float	0.0	Penalize frequent tokens	✅
presence_penalty	float	0.0	Penalize present tokens	✅
reasoning	object	—	Reasoning effort	✅
tools	array	—	Tool definitions	✅
tool_choice	string/object	—	Tool choice / force tool usage	✅
parallel_tool_calls	bool	—	Allow parallel tool calls	❌ TODO 🔴
store	bool	—	Persist response	❌ TODO 🔴
include	array	—	Include extra response fields	❌ TODO 🔴
text	object	—	Structured output ({"format":{"type":"json_schema","schema":{...}}})	✅
service_tier	string	—	Routing hint	❌ TODO 🔴
previous_response_id	string	—	Multi-turn chaining	❌ TODO

Tool-call output and replay

Both JSON responses and the stream's response.completed.output contain an assistant message first (its text may be empty), followed by one function_call item per call. The stream closes the message at output_index 0 before opening calls at indices 1, 2, and so on. Each call emits response.output_item.added, response.function_call_arguments.delta, response.function_call_arguments.done, then response.output_item.done, with a consistent item ID and output index.

For a tool follow-up, append the completed output items to input, then append one function_call_output per call, using its call_id. The assistant message and immediately following calls normalize to one assistant turn. Tool memory replays the original generated turn once when available; otherwise the server keeps the prose and structured calls together for the model's chat template to render in its native format.

---

Sampling Chain

The sampler applies penalties and sampling in this order:

logits (from GPU)
  → repetition_penalty (multiplicative, HF-style)
  → frequency_penalty + presence_penalty (additive, OpenAI-style)
  → top_k filtering
  → temperature softmax
  → top_p (nucleus) filtering
  → random draw (or argmax if temp=0)

All penalties respect rep_window (default 256 tokens lookback).

When temperature = 0:

• Penalties are still applied to logits
• Argmax is used (deterministic, no random sampling)
• Speculative decode is enabled only when no logit processing is needed

---

Streaming

• OpenAI format: text/event-stream with data: {...}\n\n chunks, terminated by data: [DONE]\n\n
• Anthropic format: SSE with event: message_start, content_block_start, content_block_delta, message_stop
• Responses format: SSE with response.created, response.output_item.*, response.completed
• During long prefill or cleanup phases with no token data, the server emits a valid SSE comment (: keep-alive) every 15 seconds. Clients ignore the comment as payload while HTTP body-idle timers remain active. If the peer closes, the server propagates cancellation into the backend at its next safe prefill or decode boundary.

---

Features

Feature	Status	Notes
Multi-turn conversation	✅	Full message history
Tool/function calling	✅	XML and JSON tool parsing
Stop sequences	✅	OpenAI stop and Anthropic stop_sequences
Thinking/reasoning	✅	OpenAI reasoning.effort, Anthropic thinking.type
Prefix cache (memory)	✅	Automatic KV cache reuse
Prefix cache (disk)	✅	Persistent across restarts
PFlash (speculative prefill)	✅	Compresses long prompts
Client disconnect detection	✅	Aborts generation on disconnect
CORS	✅	Enabled by default
Tool memory	✅	Caches tool call results

---

TODO

🔴 High Priority (used by Codex and/or Claude Code)

These parameters are sent by real Codex CLI or Claude Code clients. Missing support causes errors or silent feature degradation.

Feature	Used By	Notes
parallel_tool_calls	Codex (Responses API)	Codex always sends true. Can accept and ignore (we serialize calls).
store	Codex (Responses API)	Controls response persistence. Accept field; can be no-op locally.
include	Codex (Responses API)	Controls what's included in response events. Accept field.
service_tier	Codex (Responses API)	Routing hint (e.g., "default"). Accept and ignore.

🟡 Medium Priority

Feature	Used By	Notes
metadata	Claude Code (Anthropic)	Request metadata for tracing. Accept and ignore.
stream_options	Some OpenAI clients	{"include_usage": true} — usage in final streaming chunk.
Input validation	—	Clamp penalty ranges [-2,2], reject invalid params with 400.
previous_response_id	Codex (Responses API)	Multi-turn response chaining (we handle via input already).

🟢 Low Priority

Feature	Notes
logprobs / top_logprobs	Token probabilities in response. Debugging/analysis only.
n (multiple completions)	Generate N choices per request. No known agent uses this.
logit_bias	Per-token logit adjustments.
user	End-user identifier (tracking only).
prompt_cache_key	Codex sends this for server-side caching hints. We have our own prefix cache.