deepseek

Shared Rust client for the DeepSeek API with pluggable HTTP transport, agentic tool-use loop, and in-memory caching.

Modules

Feature Flags

Key Types

Models & Effort

pub enum DeepSeekModel { Reasoner, Chat }  // aliases: "r1"/"deep" → Reasoner, "v3"/"fast" → Chat
pub enum EffortLevel { Low, Medium, High, Max }
// Low: 0.1 temp / 2048 tokens → Max: 1.0 temp / 16384 tokens

Messages

pub struct ChatMessage { role, content, reasoning_content?, tool_calls?, tool_call_id?, name? }
pub struct ChatRequest  { model, messages, tools?, tool_choice?, temperature?, max_tokens?, stream? }
pub struct ChatResponse { id, choices, usage? }

// Constructors
pub fn system_msg(s: &str) -> ChatMessage
pub fn user_msg(s: &str) -> ChatMessage
pub fn assistant_msg(s: &str) -> ChatMessage
pub fn tool_result_msg(tool_call_id: &str, content: &str) -> ChatMessage

Request Builder (free function)

pub fn build_request(model, messages, tools?, effort) -> ChatRequest

Usable without a client instance — builds a typed ChatRequest with effort-mapped temperature/max_tokens.

Client

pub trait HttpClient: Send + Sync {
    async fn post_json(&self, url: &str, token: &str, body: ChatRequest) -> Result<ChatResponse>;
}

pub struct DeepSeekClient<H: HttpClient> { .. }
impl<H: HttpClient> DeepSeekClient<H> {
    pub fn new(api_key, http) -> Self
    pub fn with_base_url(self, url) -> Self
    pub async fn chat(&self, req: &ChatRequest) -> Result<ChatResponse>
}

// reqwest-client feature
pub fn client_from_env() -> Result<DeepSeekClient<ReqwestClient>>  // reads DEEPSEEK_API_KEY
pub async fn reason(client, system, user) -> Result<ReasonerOutput>
pub async fn reason_with_retry(client, system, user) -> Result<ReasonerOutput>  // 3 attempts, exp backoff

Agent (feature agent)

pub trait Tool: Send + Sync {
    fn name(&self) -> &str;
    fn definition(&self) -> ToolDefinition;
    async fn call_json(&self, args: Value) -> Result<String, String>;
}

pub struct AgentBuilder<H: HttpClient> { .. }
impl<H: HttpClient> AgentBuilder<H> {
    pub fn new() -> Self
    pub fn preamble(self, s: &str) -> Self
    pub fn tool(self, t: impl Tool + 'static) -> Self
    pub fn build(self) -> DeepSeekAgent<H>
}

impl<H: HttpClient> DeepSeekAgent<H> {
    pub async fn prompt(&self, input: String) -> Result<String>  // legacy: collects to text
    pub fn run(&self, input: String) -> impl Stream<Item = SdkMessage>  // Claude-Code shape
}

Streaming loop — Claude Code shape

The run() function (and the DeepSeekAgent::run() method) yields a typed
sequence of [SdkMessage] events that mirror the Claude Agent SDK's loop:

System{Init}                   ← session metadata, fresh uuid session id
Assistant{ ToolUse… }          ← per turn the model requests tools
User{ ToolResult… }            ← combined results fed back to the model
…repeat until no tool_use…
Assistant{ Text }              ← final answer
Result{ subtype, result, total_cost_usd, usage, num_turns, session_id, stop_reason }

RunOptions lets you configure: model, system_prompt, allowed_tools,
disallowed_tools, max_turns, max_budget_usd, effort,
permission_mode, pre_tool_hook, session_id, base_url.

Read-only tools (those whose read_only_hint() == true) are dispatched in
parallel within a turn; mutating tools run sequentially. PermissionMode
mirrors the SDK: Default, AcceptEdits, Plan, DontAsk,
BypassPermissions. Hitting max_turns / max_budget_usd yields a
Result with the matching error subtype.

CLI (feature cli)

The binary streams SdkMessage events as NDJSON (one JSON object per line) on stdout. Pipe it into jq or any line-oriented consumer.

# One-shot run
cargo run -p deepseek-loop --features cli --bin deepseek-loop -- \
    "list rust files under crates/deepseek-loop"

# Read-only review (Plan mode blocks Bash/Write/Edit)
cargo run -p deepseek-loop --features cli --bin deepseek-loop -- \
    --permission-mode plan --max-turns 4 --max-budget-usd 0.05 \
    "summarize the agent loop"

Scheduler / /loop (feature scheduler, on by default)

Models Claude Code's /loop scheduler — see the upstream spec at https://code.claude.com/docs/en/scheduled-tasks. The crate ships four built-in tools (CronCreate, CronList, CronDelete, Monitor) and three CLI invocation shapes.

Three invocation shapes

# Fixed cron interval, capped at 10 fires
deepseek-loop --loop-every 5m --loop-prompt 'check the deploy and report' \
    --loop-max-iterations 10

# Dynamic interval
deepseek-loop --loop --loop-prompt 'check whether CI passed and address review comments'

# Bare /loop with the built-in maintenance prompt (or your loop.md)
deepseek-loop --loop

# Resume a prior session's scheduled tasks
deepseek-loop --resume <session-id> --loop

Times are interpreted in the host's local timezone as of v0.3.4 (was UTC in 0.3.3 and earlier). A cron expression like 0 9 * * * fires at 9 AM local. Stored fire times are still serialized as UTC so persisted state is timezone-independent.

loop.md discovery

When you run /loop (bare or interval-only) and don't supply a prompt, the bin resolves the prompt from these locations in order:

Files larger than 25 KB are truncated. When neither file exists, the bin falls back to the built-in maintenance prompt.

Built-in maintenance prompt

A bare --loop (no prompt, no loop.md) runs a maintenance prompt that, on each iteration, in order:

1. Continues any unfinished work from the conversation.
2. Tends to the current branch's pull request: review comments, failed CI runs, merge conflicts.
3. If nothing else is pending, runs a cleanup pass — bug hunts or simplification.

It does not start new initiatives outside that scope, and irreversible actions only proceed when they continue something the transcript already authorized.

Manage tasks

Ask the agent in natural language during a run:

what scheduled tasks do I have?
cancel the deploy check job
remind me at 3pm to push the release branch
in 45 minutes, check whether the integration tests passed

Under the hood, the agent calls these tools:

Each task has an 8-character base32 ID. A session can hold up to 50 scheduled tasks at once.

Cron expression reference

5-field standard cron: minute hour day-of-month month day-of-week. All fields support wildcards (*), single values (5), steps (*/15), ranges (1-5), and comma-separated lists (1,15,30).

Day-of-week uses 0 or 7 for Sunday through 6 for Saturday. Extended syntax (L, W, ?, MON, JAN) is not supported. When both day-of-month and day-of-week are constrained, a date matches if either field matches (vixie-cron semantics).

Timezone note (since v0.3.4): cron expressions are evaluated in the host's local timezone. 0 9 * * * means 9 AM local, not 9 AM UTC. Fire times are stored as UTC so persisted task state is portable.

Jitter & expiry

The scheduler adds a deterministic offset to fire times so independent sessions don't hit downstream APIs at the same wall-clock moment:

• Recurring tasks fire up to 30 minutes after the scheduled time, or up to half the interval for sub-hourly tasks. An hourly job at :00 may fire anywhere up to :30.
• One-shot tasks scheduled at the top or bottom of the hour fire up to 90 seconds early.

The offset is derived from the task ID, so the same task always gets the same offset. If exact timing matters, pick a minute that isn't :00 or :30 (e.g. 3 9 * * * instead of 0 9 * * *).

Recurring tasks expire 7 days after creation: they fire one final time, then delete themselves. This bounds how long a forgotten loop can run.

Persistence

Tasks are written to ${cache_dir}/deepseek-loop/sessions/<session_id>/tasks.json after each mutation. Restore them with --resume <session_id>. Background Bash and Monitor tasks are not restored.

Disable

Set either env var to 1 to disable the scheduler entirely. Cron tools and the --loop flags become unavailable, and any already-scheduled tasks stop firing:

CLAUDE_CODE_DISABLE_CRON=1   # canonical; matches Claude Code
DEEPSEEK_LOOP_DISABLE_CRON=1 # crate-specific alias

Limitations

• Tasks only fire while deepseek-loop is running and idle. Closing the terminal or letting the bin exit stops them firing.
• No catch-up for missed fires — if a task's scheduled time passes while the bin is busy on a long request, it fires once when the bin becomes idle, not once per missed interval.
• Dynamic mode currently uses a fixed 60 s floor between iterations. The upstream spec's adaptive 60–3600 s delay lands in v0.4 (requires a host↔agent protocol for the agent to signal a chosen delay).
• For unattended cron-driven automation that must survive without the bin running, drive deepseek-loop from system cron / launchd / GitHub Actions instead.

Cache (feature cache)

pub struct Cache { .. }          // entry-level TTL, evicts expired on insert
pub struct InFlightDedup<T> { .. } // broadcast-based — first caller inserts, others subscribe

Retry Logic

reason_with_retry() retries only on 5xx / network errors with exponential backoff (1s → 2s → 4s). 4xx errors fail immediately.

Dependencies

Standalone library — no sibling crate dependencies. Used by sdd, research, and genesis.