deer-flow/backend/docs/API.md

# API Reference

This document provides a complete reference for the DeerFlow backend APIs.

## Overview

DeerFlow backend exposes two sets of APIs:

1. **LangGraph-compatible API** - Agent interactions, threads, and streaming (`/api/langgraph/*`)
2. **Gateway API** - Models, MCP, skills, uploads, and artifacts (`/api/*`)

All APIs are accessed through the Nginx reverse proxy at port 2026.

For agent conversations, clients can either pre-create a thread
(`POST /api/langgraph/threads`) or start immediately with the stateless stream
endpoint (`POST /api/langgraph/runs/stream`). The latter auto-creates a thread
and returns `thread_id` and `run_id` in the response `Content-Location` header.

## Authentication

Browser sessions authenticate with the `access_token` session cookie issued at
login. Programmatic clients can instead use a **personal access token (PAT)**
sent as a Bearer credential:

```http
POST /api/threads/search
Authorization: Bearer dfp_...
Content-Type: application/json

{}
```

PATs require a configured database backend (SQLite/PostgreSQL) — on the
memory-only backend, Bearer credentials are rejected and PAT management routes
return `503`.

### Account Preferences

`GET /api/v1/auth/preferences` returns the signed-in browser user's five
preferences. `PATCH` updates only explicitly supplied fields and returns `204`.
Both require `X-Expected-User-Id` matching the session user; PATCH also requires
the normal `X-CSRF-Token` header. The expected ID is a stale-tab guard, not an
authorization credential. PAT, internal, and auth-disabled callers receive
`403`; a different session user receives `409`.

```json
{
  "notification_enabled": false,
  "model_name": "my-model",
  "mode": "pro",
  "reasoning_effort": "high",
  "locale": "zh-CN"
}
```

All five fields accept `null` to restore the default. `mode` accepts `flash`,
`thinking`, `pro`, or `ultra`; `reasoning_effort` accepts `minimal`, `low`,
`medium`, or `high`; `locale` accepts `en-US` or `zh-CN`; model names are at
most 200 characters. Unknown fields and
invalid values return `422`. Missing preferences read as `null`. Separate-field
patches preserve each other's changes, and same-field writes are last-commit-wins.
Storage requires SQLite or PostgreSQL (`503` when unavailable). Browser
notification permission remains device-local and is not changed by this API.
`locale` is the web UI language; the web app writes it after sign-in and on
every language switch, and scheduled-task IM notices are written in it (without
it they use `channel_connections.notification_locale`).

### Personal Access Tokens

Base URL: `/api/v1/auth`

PAT management requires an **interactive session** (a PAT cannot manage PATs
or change passwords, so a leaked automation token cannot mint fresh
credentials). The raw token is returned **exactly once** at creation; only its
SHA-256 digest is stored server-side.

#### Create Token

```http
POST /api/v1/auth/pats
Content-Type: application/json
```

**Request Body:**
```json
{
  "name": "ci-runner",
  "scopes": ["threads:read", "runs:create", "runs:read"],
  "expires_in_days": 90
}
```

- `scopes` — subset of the route permissions: `threads:read`, `threads:write`,
  `threads:delete`, `runs:create`, `runs:read`, `runs:cancel`. A PAT can only
  *narrow* its owning user's permissions, never widen them.
- `expires_in_days` — optional (`1`–`365`); omitted means the token never expires.

**Response (`201`):**
```json
{
  "id": "0f0c6e6a-...",
  "name": "ci-runner",
  "scopes": ["runs:create", "runs:read", "threads:read"],
  "expires_at": "2026-11-25T10:30:00Z",
  "created_at": "2026-08-27T10:30:00Z",
  "token": "dfp_..."
}
```

Save `token` immediately — it cannot be retrieved again.

#### List Tokens

```http
GET /api/v1/auth/pats
```

Returns the caller's tokens with `last_used_at` / `revoked_at` audit fields;
never returns digests or raw tokens.

#### Revoke Token

```http
DELETE /api/v1/auth/pats/{pat_id}
```

Revocation is immediate.

### PAT Constraints

- A request carrying an `Authorization` header that fails validation gets a
  hard `401` — it never falls back to the session cookie.
- **Cancel capability requires `runs:cancel` on every request dimension that
  carries it**, not just the dedicated cancel route: `?action=interrupt|rollback`
  on `POST /api/threads/{thread_id}/runs/{run_id}/stream` (action-less joins
  stay at `runs:read`), and `multitask_strategy=interrupt|rollback` on run
  creation (the default `reject` stays at `runs:create`). Joining a run's
  stream is pure observation — an observer disconnecting never cancels the run.
- **Route-level default-deny:** PAT requests are admitted only to the
  thread/run lifecycle routes the v1 scopes govern — `POST /api/threads`
  (create), `POST /api/threads/search` (list), `GET/PATCH/DELETE
  /api/threads/{thread_id}`, the thread `goal`/`state`/`compact`/`history`/
  `branches` subroutes, and exactly the implemented `/runs` subroutes
  (`GET|POST /api/threads/{thread_id}/runs`, the POST-only `stream`, `wait`,
  `regenerate/prepare`, and `edit-regenerate/prepare` collection endpoints,
  `GET /api/threads/{thread_id}/runs/{run_id}` plus its `cancel` (POST),
  `join`/`messages`/`events`/`workspace-changes` (GET), and
  `GET|POST .../runs/{run_id}/stream`), plus `POST /api/runs/stream|wait` and
  `GET /api/runs/{run_id}/messages|feedback`. A route added under `/runs` is
  denied until explicitly added to the policy.
  Every other authenticated route — memory, agents, models, MCP/skills
  config, integrations, channels, uploads — answers `403` to PAT callers
  regardless of scopes. Scope enforcement alone only constrains
  permission-decorated routes, so the allowlist is the outer boundary;
  session-cookie callers are unaffected.
- PAT credentials never carry admin capability, even when the owning user is
  an admin. This includes extension-contributed admin routes: the extension
  principal projection suppresses every admin signal for PAT callers.
- Revoking or deleting the owning user invalidates their PATs on the next
  request.

## LangGraph-compatible API

Base URL: `/api/langgraph`

The public LangGraph-compatible API follows LangGraph SDK conventions. In the unified nginx deployment, Gateway owns `/api/langgraph/*` and translates those paths to its native `/api/*` run, thread, and streaming routers.

### Threads

#### Create Thread

```http
POST /api/langgraph/threads
Content-Type: application/json
```

**Request Body:**
```json
{
  "metadata": {}
}
```

Optional `assistant_id` binds the thread to `lead_agent` or a custom agent.
After trimming, lowercasing and mapping `_` to `-`, a custom name must contain
only letters, digits and hyphens; any other value is rejected with `422`.
Branches preserve valid stored IDs, including their original spelling. If a
legacy source has an invalid ID, its new branch stores `null` and uses the lead agent.

**Response:**
```json
{
  "thread_id": "abc123",
  "created_at": "2024-01-15T10:30:00Z",
  "metadata": {}
}
```

#### Get Thread State

```http
GET /api/langgraph/threads/{thread_id}/state
```

**Response:**
```json
{
  "values": {
    "messages": [...],
    "sandbox": {...},
    "artifacts": [...],
    "thread_data": {...},
    "title": "Conversation Title"
  },
  "next": [],
  "config": {...}
}
```

#### Get Thread History

```http
POST /api/langgraph/threads/{thread_id}/history
```

Only the newest entry carries `messages`, the `goal` while it is active, and
`goal_outcome` (the latest met goal, kept until the next goal change). Older
entries carry only `title` and `thread_data`. `goal_outcome` is server-owned:
state updates and run input that include it are rejected with 400.

External state updates and run input that supply a non-null `goal` must include
a string `objective` that is non-empty after whitespace normalization and at
most 4,000 normalized characters. Invalid values return 400 before state writes
or run admission. Optional goal timestamps and counters are not required or reset.
When present, `continuation_count`, `max_continuations`, `no_progress_count` and
`max_no_progress_continuations` must be non-negative integers; booleans, numeric
strings, floats and null values are rejected with 400. Valid counters and limits
are preserved, including an already-used continuation budget.
Goal payloads must be state values, not reducer control wrappers: the reserved
`__overwrite__` key and `type: "__overwrite__"` are rejected with 400 at both
external boundaries. Ordinary `type`/`value` fields remain compatible.
`goal: null` retains its existing semantics: it clears a state update's goal,
but does not replace a goal through run input.

Replaying messages from an earlier root checkpoint, including response regeneration, preserves
the thread's current goal and continuation budget. A cleared or achieved goal
is not restored from history. A valid goal explicitly supplied in run input
still replaces it. Historical replay clears the old `goal_outcome`, because
the previous answer's verdict does not certify the replacement answer.

### Runs

#### Create Run

Execute the agent with input.

```http
POST /api/langgraph/threads/{thread_id}/runs
Content-Type: application/json
Idempotency-Key: <unique key for this logical request>  # optional
```

The thread-scoped create, stream, and wait endpoints accept an optional
`Idempotency-Key` header. Retrying with the same authenticated user, `thread_id`,
and key reuses the existing run instead of executing the input again. The key is
shared across `/runs`, `/runs/stream`, and `/runs/wait` for a given user and
thread, so the same key string cannot back two different calls even across those
endpoints. Reuse is bound to the original `input`, `assistant_id` and
`conversation_references`; a retry that changes them returns 409. When
`command.resume` takes precedence over `input`, reuse instead requires the same
resume value, assistant and conversation references. Resume values must be strict
JSON with finite numbers, with or without an idempotency key; booleans, integers
and floats remain distinct when comparing keyed retries. Generate a new key for every intentional user
action; reuse a key only when retrying that same action after an uncertain HTTP
result. Keys may be at most 255 characters. Trusted internal callers (the IM
channel manager) get keys scoped by user only: the same key sent for another
thread returns 409 instead of starting a run, because a redelivered IM message
may be requested on a new thread after `/new`. Stateless `/api/langgraph/runs/*`
endpoints do not support this header because requests without an explicit thread
create a new temporary conversation.

When an `input.messages` entry has no usable `id`, the Gateway assigns one
before persisting the run input and invoking the graph. The same ID then appears
in the run record and checkpoint, so a client rejoining an active run can match
the two copies. Caller-supplied nonempty IDs are preserved. For keyed requests,
the generated IDs are stable for retries with the same thread and key; unkeyed
runs receive fresh IDs. During a rolling upgrade, route keyed requests with
ID-less input and their retries to upgraded Gateway workers until all workers
are upgraded. An older worker may reject a retry against a newly stamped run
record with 409.

**Rolling upgrades and keyed resume:** route all initial submissions and retries
containing both `command.resume` and `Idempotency-Key` to upgraded Gateway workers
only, across `/runs`, `/runs/stream` and `/runs/wait`, until every worker serving
these endpoints is upgraded. Older workers ignore `runs.idempotency_request_json`
and compare only `input`, so with `input: null` they can reuse an `approve` run
for a conflicting `deny` request. If version-aware routing is unavailable, pause
keyed resume traffic until the rollout completes, or replace all old workers
before accepting it again. The additive migration preserves old-reader run-history
compatibility, not mixed-version resume admission safety. Upgraded workers return
409 for identity-less legacy resume rows, including identical decisions; inspect
the original run and current thread state before submitting a new intentional
action with a fresh key. Removing the header is not a retry workaround: it makes
the request non-idempotent.

Retrying a still-running run that this worker cannot stream returns 409 from
`/runs/stream` (`Run ... is not active on this worker and cannot be streamed`)
with no `Retry-After`. The same shape on `/runs/wait` returns 200
`{"status": "<durable status>", "error": ...}` without blocking for a final
state. Retrying a finished run through `/runs/wait` also returns that durable
status payload rather than the latest thread checkpoint: a later run on the
same thread may have advanced the head, and `/wait` does not claim that head
as this run's result. That status is the durable row after completion, not
the hydrated record from admission time. The original creating `/wait` still
returns this run's checkpoint even if a retry overlaps while it is waiting. Retrying a finished run whose SSE log is gone emits a `gap` frame
(`stream_replay_gap`, `recovery: reload_durable_state`) on the creating
`/runs/stream` endpoint and closes without an `end` frame; reload durable
thread/run state instead of treating the stream as empty. Observer joins of
that same run still end with `end`. Stateless `/api/langgraph/runs/stream`
does not accept this header and keeps the existing missing-stream close of
`end`; the `gap` signal is only on a thread-scoped creating retry.

**Request Body:**
```json
{
  "input": {
    "messages": [
      {
        "role": "user",
        "content": "Hello, can you help me?"
      }
    ]
  },
  "config": {
    "recursion_limit": 100,
    "configurable": {
      "model_name": "gpt-4",
      "thinking_enabled": false,
      "is_plan_mode": false
    }
  },
  "stream_mode": ["values", "messages-tuple", "custom"]
}
```

**Stream Mode Compatibility:**
- Use: `values`, `messages-tuple`, `custom`, `updates`, `debug`, `tasks`, `checkpoints`
- Unsupported modes, including `messages`, `events`, and `tools`, return `422` before a run is created. DeerFlow never substitutes `values` for an unsupported mode.

**Run Option Compatibility:**
- Supported concurrency strategies: `reject`, `rollback`, and `interrupt`
- Compatibility default: `if_not_exists="create"`; this matches DeerFlow's current behavior
- Artifact delivery is enforced automatically when a run creates or modifies regular files under `/mnt/user-data/outputs`. `present_files` must present at least one path produced by the current run (or a directory containing it), and the terminal receipt must be persisted; presenting only an unrelated file does not satisfy delivery. Runs without changed outputs retain ordinary conversational behavior. `artifact_delivery` is not a client-settable run option.
- Unsupported options return `422`: `webhook`, `stream_resumable=true`, `after_seconds`, `feedback_keys`, any non-null `on_completion` value (including the SDK values `"complete"` and `"continue"`), `if_not_exists="reject"`, and `multitask_strategy="enqueue"`
- `stream_resumable=false` is accepted: it is the LangGraph SDK's default and requests the non-resumable stream DeerFlow already serves
- Undeclared SDK options, including `checkpoint_during` and `durability`, also return `422` instead of being silently discarded

**Request Validation:**
- `config.configurable` and `config.context` must be objects or `null`; any other value returns `422` (`invalid_run_config_section`)
- An `assistant_id` that does not normalize to letters, digits and hyphens returns `422`, as does a non-string `agent_name` in `config.configurable`, `config.context` or `context`
- A message in `input.messages` that cannot be parsed returns `400` naming its index
- These checks run before a run is created

When outputs changed during the run, `run.delivery` events retain the Slice 1
facts (`presented`, `paths`, and `by_tool`) and add `produced_paths`,
`presented_paths`, `matched_paths`, plus an explicit verdict: `verification`,
`stage` (`presented`, `mismatched`, or `not_started`), and `satisfied`. Receipts
for runs without changed outputs keep their existing shape.

**Recursion Limit:**

`config.recursion_limit` caps the number of graph steps LangGraph will execute
in a single run. The unified Gateway path uses the top-level `recursion_limit`
from `config.yaml` (default `100`) when a request does not provide one. Clients
can still set `recursion_limit` explicitly in the request body, and a valid
request value takes precedence. Scheduled-task launches do not take a client body: they
use `scheduler.recursion_limit` from `config.yaml` (default `1000`, matching
the web UI). For safety, the Gateway clamps any supplied
or configured value to a server ceiling (`max_recursion_limit` in `config.yaml`,
default `1000`) so a single run cannot execute unbounded graph steps (runaway
LLM cost / DoS); invalid or non-positive request values fall back to the
configured default. Both top-level fields are read per run, so edits apply to
the next request without restarting the Gateway. IM channel runs get the same
default unless a channel `session.config` sets `recursion_limit` (GitHub channel
runs have their own floor; see
[CONFIGURATION.md](CONFIGURATION.md#recursion-limits)); embedded
`DeerFlowClient` runs retain their own defaults and override paths.

**Configurable Options:**
- `model_name` (string): Override the default model
- `thinking_enabled` (boolean): Enable extended thinking for supported models
- `is_plan_mode` (boolean): Enable TodoList middleware for task tracking

**Client Context:**

Top-level `context.locale` accepts `en-US` or `zh-CN` as the current run's
preferred response and automatic DeerMem-writing language. Explicit user
language requests take precedence. Unsupported values are ignored. The value
is runtime-only and is not a checkpoint `configurable` default; legacy
`config.context.locale` is also validated, with an explicit null/invalid
top-level value clearing it. Omitting a locale retains language inference.
Scheduled runs and MCP task notification runs currently omit this preference;
the owner's saved locale localizes scheduled IM notices only, not those runs'
replies or memory writes.

A custom client can describe what it can display under top-level
`context.client`, so the agent and its skills can choose an output shape that
fits (for example a CSV download or a chart).

```json
{
  "input": {
    "messages": [
      {
        "role": "user",
        "content": "Analyze this dataset"
      }
    ]
  },
  "context": {
    "client": {
      "name": "custom-analytics-frontend",
      "capabilities": {
        "artifacts": true,
        "csv_download": true,
        "charts": true
      },
      "preferences": {
        "csv": "present",
        "chart": "present"
      }
    }
  }
}
```

The Gateway keeps only `name` (at most 80 characters), `capabilities` (boolean
flags) and `preferences` (string, number or boolean values; strings at most 200
characters, integers at most 2^53 − 1 in magnitude), reading only the first 32
entries of each map. Keys must match `^[A-Za-z0-9][A-Za-z0-9_.:-]{0,63}$`;
whitespace is collapsed, and other fields, nested values and non-finite numbers
are dropped. Dropped fields are logged at debug level without their values. The
result reaches `runtime.context["client"]` only, never `configurable` or the
checkpoint. The legacy `config.context.client` input is sanitized the same way;
`config.configurable.client` is ignored. Both nested config locations omit
`client` when saving a run record or returning its configuration through the
runs API. Cleanup leaves the caller's request object unchanged. Reading an
older record also omits these fields without rewriting the stored record.

On every model call the lead agent receives the current value as one hidden,
user-role `<client_context>` block placed after the system prompt, with the
values framed by the same `--- BEGIN/END USER INPUT ---` markers as the user's
own message. It is never system text and never written to the thread history.
A run without a usable client value in either supported input location has
no block; only the current request's value applies.

- Capabilities and preferences are advisory hints, not enforcement. A client
  that must never receive some output still has to enforce that outside the
  model, in server-side policy or in the client itself.
- `name` is informational only. Never use it for authorization or identity;
  trust comes from the authenticated principal.
- Send the same `client` object on each run from the same client. Placement
  after the system messages keeps an unchanged client ahead of the growing
  conversation, allowing the message prefix to remain stable across runs.
  Changing or omitting it can invalidate reuse of the following conversation
  prefix. Actual cache hits, cost and latency depend on the model provider;
  offline message-order tests do not establish those outcomes.
- Subagents and embedded `DeerFlowClient` runs do not receive it; the lead
  agent decides the user-facing output shape. Runs the server starts
  (schedules, IM channels, GitHub agents, MCP notifications, and extensions
  unless the plugin passes a `client`) carry none, even on a thread a custom
  client opened. A thread does not retain a "last-seen client" for future
  tasks: a launcher must explicitly supply its own current hints if needed.
  Existing channel-specific response formatting is unchanged.

**Response:** Server-Sent Events (SSE) stream

```
event: values
data: {"messages": [...], "title": "..."}

event: messages
data: {"content": "Hello! I'd be happy to help.", "role": "assistant"}

event: end
data: {}
```

#### Referencing a previous conversation

With `read_conversation` enabled in `config.yaml` (see [configuration](CONFIGURATION.md#reading-referenced-conversations)),
Gateway API callers can attach up to three explicit references to create/stream/wait requests:

```json
{
  "input": {"messages": [{"role": "user", "content": "Use the requirements agreed in the referenced conversation."}]},
  "conversation_references": ["https://deerflow.example/workspace/chats/source-thread"]
}
```

A reference is a valid thread ID or an absolute `/workspace/chats/{thread_id}` URL
(also `/workspace/agents/{agent_name}/chats/{thread_id}` for custom agents)
with the same scheme and authority as the run request, without query or fragment.
URLs are parsed as local selectors and are never fetched. For split-origin clients
or internal proxies, pass the thread ID. The field is separate from message text:
links in pasted documents, tool results, or previous messages grant no access.
The server supplies source IDs to the model as background user-role data and
binds the reader to this run's references and authenticated identity.

Clients that cannot add top-level fields to a run request (the LangGraph JS SDK
builds a fixed body and drops unknown keys) may send the same list as
`context.conversation_references`:

```json
{
  "input": {"messages": [{"role": "user", "content": "Use the requirements agreed in the referenced conversation."}]},
  "context": {"conversation_references": ["https://deerflow.example/workspace/chats/source-thread"]}
}
```

The Gateway lifts the key out of `context` before the run context is assembled,
so it has the same bounds and error locations as the top-level field, is
recorded on the run in the same way, and never reaches the merged run context
or the checkpointed `configurable`. Sending the top-level field and the context
key together returns 422. `GET /api/features` reports
`conversation_references.enabled` (the tool is configured) and `max_references`,
so a client can hide its entry point on deployments without the tool.

The request requires `runs:read` as well as the normal run-creation permission.
The tool rechecks source ownership on each read; foreign, deleted and unowned
legacy threads are unavailable. `read_conversation(thread_id, cursor?, limit?)`
reads newest-first pages (messages within each page are chronological), at most
50 visible user/assistant messages, 4,000 characters per message and 20,000 text
characters per page. Each page also stays within the tool-output budget that
applies to `read_conversation` (`tool_output.tool_overrides.read_conversation`,
else `externalize_min_chars`, and `fallback_max_chars`; 12,000 serialized
characters by default), so results reach the model inline instead of being
externalized to a file. A message that does not fit starts the next page intact.
Only a message longer than 4,000 characters, or one whose serialized form alone
exceeds the budget, is truncated. Such a message carries
`continuation: {"message_seq", "offset"}`; `read_conversation(thread_id,
message_seq=..., offset=...)` without a cursor returns the next part of that one
message (at most 20,000 text characters, sized to the same budget) with its
`offset`, `text_length` and, while text remains, a new continuation. Offsets
refer to the source's current text: an offset past its end returns
`invalid_request`, and a message that is no longer visible is unavailable. If the
`read_conversation` budget is too small to return any text (below roughly 800
serialized characters), the result is `output_budget_too_small` rather than a
continuation that makes no progress.
Results include message IDs, sequence numbers, continuation, truncation and
unavailability. Hidden messages, reasoning blocks, raw tool
results and subagent internals are excluded. Source data is not changed.

**Live reads and retained copies.** Each call reads the source's current visible
history. Editing or regenerating the source can change subsequent reads, including
later pages; a reference does not pin an immutable transcript. Text already returned
to the destination is a copy and is not automatically refreshed by source changes.

Read permission lasts only for this run, including its internal continuation steps.
Every new run, including resume, regenerate or edit replay, must submit references
again; checkpoints and old hints never restore permission. A resume can reuse
IDs already visible in the interrupted conversation, but needs the explicit
request field again. Missing/expired transcripts are not reconstructed from
checkpoints or memory.

Permission expiry does not erase excerpts already stored in the destination
conversation or conclusions derived from them. Deleting the source does not
retroactively erase those copies either; they follow the destination's own
retention and deletion behavior. Once the source is unavailable, further source
reads report unavailability rather than reconstructing it from destination copies.

**Incomplete requirements.** When `truncated` is true, the tool's notice tells
the agent to read the rest through each cut message's continuation before relying
on it, and to acknowledge the omission and request the missing material if that
read is unavailable. `has_more: false` means there are no older messages to page
through, not that every returned message is complete. This is model guidance, not
a new confirmation mechanism or a guarantee of model compliance.

This first version adds no frontend picker or link-to-reference conversion. The
tool is unavailable to bootstrap agents, subagents and embedded clients without
a host-provided reader. Active tool/skill policies continue to apply.

#### Get Run History

```http
GET /api/langgraph/threads/{thread_id}/runs
```

**Response:**
```json
{
  "runs": [
    {
      "run_id": "run123",
      "status": "success",
      "created_at": "2024-01-15T10:30:00Z"
    }
  ]
}
```

#### Stream Run

Stream responses in real-time.

```http
POST /api/langgraph/threads/{thread_id}/runs/stream
Content-Type: application/json
Idempotency-Key: <unique key for this logical request>  # optional
```

Same request body as Create Run. Returns SSE stream.

#### Stateless Stream Run

Start a conversation without creating a thread first. Gateway auto-creates a
thread when `config.configurable.thread_id` is omitted, and returns both
identifiers in the response `Content-Location` header.

```http
POST /api/langgraph/runs/stream
Content-Type: application/json
Accept: text/event-stream
```

Through Nginx, `/api/langgraph/runs/stream` is rewritten to the native Gateway
path `POST /api/runs/stream`.

**Request Body:** Same as [Create Run](#create-run). Omit `thread_id` to start a
new conversation; include it to continue an existing one:

```json
{
  "input": {
    "messages": [
      {
        "role": "user",
        "content": "Hello, can you help me?"
      }
    ]
  },
  "config": {
    "recursion_limit": 100,
    "configurable": {
      "model_name": "gpt-4",
      "thinking_enabled": false,
      "is_plan_mode": false
    }
  },
  "stream_mode": ["values", "messages-tuple", "custom"]
}
```

**Response:** Server-Sent Events (SSE) stream with a `Content-Location` header:

```http
Content-Location: /api/threads/{thread_id}/runs/{run_id}
```

Clients should parse `thread_id` and `run_id` from this header (the path ends
with `/runs/{run_id}`). Persist `thread_id` and send it back on the next turn
via `config.configurable.thread_id` to keep conversation history.

**Continuing a conversation:**

```json
{
  "input": {
    "messages": [
      {
        "role": "user",
        "content": "What did I just ask?"
      }
    ]
  },
  "config": {
    "configurable": {
      "thread_id": "abc123",
      "model_name": "gpt-4"
    }
  },
  "stream_mode": ["values", "messages-tuple", "custom"]
}
```

---

## Gateway API

Base URL: `/api`

### Custom Agent portability

`GET /api/agents/{name}/export` downloads a version-1 JSON package for a
caller-owned Custom Agent. The package uses `format: "deerflow.custom-agent"`
and contains the portable Agent configuration plus SOUL. It excludes memory
contents, conversations, credentials, and deployment-owned GitHub bindings.

`POST /api/agents/import` creates the packaged Agent for the current user.
Pass `?name=<new-name>` to choose a different local identifier. The document
schema rejects unknown fields and unsupported format/version values; invalid
names or models return 422, and an existing name returns 409 without changing
the existing Agent. Import is create-only and never restores runtime state.

```json
{
  "format": "deerflow.custom-agent",
  "version": 1,
  "agent": {
    "name": "research-lead",
    "description": "Coordinates parallel research",
    "model": "deepseek-v3",
    "tool_groups": ["web"],
    "skills": ["literature-review"],
    "allowed_subagents": ["researcher", "reporter"],
    "memory_enabled": true,
    "soul": "Delegate independent searches, then synthesize evidence."
  }
}
```

### Custom Agent errors

**Error behavior:**
- `/api/agents` and `/api/user-profile` routes answer unexpected failures with a fixed `500` detail such as `{"detail": "Failed to get agent"}`; an unparsable agent `config.yaml` or a filesystem error names host paths, so the full exception stays in server logs
- Scheduled-task create and update return `invalid_assistant` (422) with a fixed message when the selected agent's config cannot be loaded

### Models

#### List Models

Get all available LLM models from configuration.

```http
GET /api/models
```

**Response:**
```json
{
  "models": [
    {
      "name": "gpt-4",
      "display_name": "GPT-4",
      "supports_thinking": false,
      "supports_reasoning_effort": false,
      "reasoning": {"thinking": "unsupported", "effort": null, "history": null, "source": "legacy"}
    },
    {
      "name": "deepseek-v3",
      "display_name": "DeepSeek V3",
      "supports_thinking": true,
      "supports_reasoning_effort": true,
      "reasoning": {
        "thinking": "optional",
        "effort": {"values": ["minimal", "low", "medium", "high"], "default": null, "aliases": {}},
        "history": null,
        "source": "legacy"
      }
    },
    {
      "name": "glm-5.3-flash",
      "display_name": "GLM-5.3-Flash",
      "supports_thinking": true,
      "supports_reasoning_effort": true,
      "reasoning": {
        "thinking": "required",
        "effort": {"values": ["low", "high", "max"], "default": "high", "aliases": {"minimal": "low", "medium": "high"}},
        "history": "clear",
        "source": "contract"
      }
    }
  ]
}
```

`supports_thinking` and `supports_reasoning_effort` are deprecated projections of
`reasoning`. `reasoning.source` is `legacy` when the profile declares no
`reasoning:` block (the booleans were normalized) and `contract` when it does.

#### Get Model Details

```http
GET /api/models/{model_name}
```

**Response:**
```json
{
  "name": "gpt-4",
  "display_name": "GPT-4",
  "model": "gpt-4",
  "max_tokens": 4096,
  "supports_thinking": false,
  "supports_reasoning_effort": false,
  "reasoning": {"thinking": "unsupported", "effort": null, "history": null, "source": "legacy"}
}
```

### MCP Configuration

#### Get MCP Config

Get current MCP server configurations.

```http
GET /api/mcp/config
```

Requires an authenticated admin session. Sensitive env/header/OAuth secret
values are masked in the response. Environment placeholders outside secret
containers are returned in their raw form so editing cannot expose or persist
their expanded values. Invalid operator-authored JSON/config shapes return
`400` instead of being reported as a Gateway fault.

**Response:**
```json
{
  "mcp_servers": {
    "github": {
      "enabled": true,
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "***"
      },
      "description": "GitHub operations"
    }
  }
}
```

#### Update MCP Config

Update MCP server configurations.

```http
PUT /api/mcp/config
Content-Type: application/json
```

Requires an authenticated admin session. API-managed `stdio` MCP servers may
only use allowed executable names for `command` (default: `npx`, `uvx`). Set
`DEER_FLOW_MCP_STDIO_COMMAND_ALLOWLIST` to a comma-separated list when a
deployment needs additional trusted launchers.

**Request Body:**
```json
{
  "mcp_servers": {
    "github": {
      "enabled": true,
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "$GITHUB_TOKEN"
      },
      "description": "GitHub operations"
    }
  }
}
```

**Response:**
```json
{
  "mcp_servers": {
    "github": {
      "enabled": true,
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "***"
      },
      "description": "GitHub operations"
    }
  }
}
```

#### Update One MCP Server State

Enable or disable one configured MCP server without replacing the full
extensions configuration.

```http
PATCH /api/mcp/config
Content-Type: application/json
```

Requires an authenticated admin session. Enabling a `stdio` server validates
that server's `command` against the same allowlist used by the full `PUT`
endpoint. Disabling a server does not require its command to be allowlisted, and
invalid commands on other servers do not block the update. The endpoint
preserves secrets, environment-variable placeholders, skills, custom server
fields, and other top-level extensions config. SSE/HTTP targets may use either
DeerFlow's `type` field or the MCP-spec `transport` field.

**Request Body:**
```json
{
  "server_name": "semantic-scholar",
  "enabled": false
}
```

The response is the full masked MCP configuration, matching `GET` and `PUT`.
An unknown `server_name` returns `404`; attempting to enable a server with a
disallowed `stdio` command returns `400`.

#### Add MCP Servers

Add one or more servers without replacing existing entries. The Gateway
re-reads the file under the shared configuration lock, so concurrent sibling
changes are preserved. Existing names return `409`.

```http
POST /api/mcp/config/servers
Content-Type: application/json
```

The request body uses the same `mcp_servers` map as the full `PUT` endpoint.

#### Replace One MCP Server

Completely replace one existing server while preserving sibling entries.
Omitted ordinary fields are deleted or reset; explicit `***` placeholders
restore the corresponding stored secret.

A disabled `stdio` replacement may keep a syntactically valid command outside
the allowlist for offline editing. Command-shape and code-injecting environment
variable checks still run when saving; the allowlist and executable-argument
policy run when the server is enabled.

```http
PUT /api/mcp/config/server
Content-Type: application/json
```

```json
{
  "server_name": "github",
  "server": {
    "enabled": true,
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-github"],
    "env": {"GITHUB_TOKEN": "***"}
  }
}
```

#### Delete One MCP Server

Delete one server without replacing sibling entries. The server name is a
path parameter and the DELETE request has no body. Percent-encode names before
placing them in the URL; the path converter also keeps legacy empty and
slash-containing names addressable.

```http
DELETE /api/mcp/config/servers/{server_name}
```

All targeted mutations return the full masked MCP configuration. Before any
write, the Gateway resolves environment variables in a copy and validates the
same expanded document the runtime will load while persisting the original raw
placeholders.

#### Reset MCP Tools Cache

Publish a shared cache generation, then clear cached MCP tools and persistent
MCP sessions in the handling process. Every Gateway worker sharing the writable
extensions-config directory observes that generation and reloads tools on its
next agent run or tool lookup. This also refreshes remote `tools/list` changes
that did not modify `extensions_config.json`.

```http
POST /api/mcp/cache/reset
```

Requires an authenticated admin session.

**Response:**
```json
{
  "success": true,
  "scope": "shared_config",
  "message": "MCP tools cache reset published through the shared config directory. Tools will reload on next use."
}
```

`shared_config` means every worker mounting that same directory observes the
generation; it does not claim a deployment-wide broadcast when replicas use
independent filesystems. When no extensions-config path can be resolved, the
request still resets the current worker and returns `"scope": "process"`.

### Skills

#### List Skills

Get all available skills, including public and custom skills. When authorization
is enabled, only skills visible to the caller's role are returned.

```http
GET /api/skills
```

**Response:**
```json
{
  "skills": [
    {
      "name": "pdf-processing",
      "description": "Handle PDF documents efficiently",
      "license": "MIT",
      "category": "public",
      "enabled": true,
      "editable": false
    },
    {
      "name": "frontend-design",
      "description": "Design and build frontend interfaces",
      "license": "MIT",
      "category": "public",
      "enabled": false,
      "editable": false
    }
  ]
}
```

#### Get Skill Details

Return one skill's metadata. The response does not include `content`; read a
custom skill's `SKILL.md` with [Get Custom Skill Content](#get-custom-skill-content).
When authorization is enabled, a skill hidden from the caller's role returns
`404`, indistinguishable from a missing skill.

```http
GET /api/skills/{skill_name}
```

**Response:**
```json
{
  "name": "pdf-processing",
  "description": "Handle PDF documents efficiently",
  "license": "MIT",
  "category": "public",
  "enabled": true,
  "editable": false
}
```

#### Enable or Disable Skill

```http
PUT /api/skills/{skill_name}
Content-Type: application/json
```

Requires an authenticated admin session.

**Request Body:**
```json
{
  "enabled": false
}
```

**Response:** the updated skill. An unknown `skill_name` returns `404`.
```json
{
  "name": "pdf-processing",
  "description": "Handle PDF documents efficiently",
  "license": "MIT",
  "category": "public",
  "enabled": false,
  "editable": false
}
```

#### Install Skill

Install a `.skill` archive that already exists in a thread's user-data directory.
Requires an authenticated administrator.

```http
POST /api/skills/install
Content-Type: application/json
```

**Request Body:**
```json
{
  "thread_id": "<thread id>",
  "path": "mnt/user-data/outputs/my-skill.skill"
}
```

**Response:**
```json
{
  "success": true,
  "skill_name": "my-skill",
  "message": "Skill 'my-skill' installed successfully"
}
```

#### Upload and Install Skill

Upload a local `.skill` archive and install it for the current user. Requires an
authenticated administrator.

```http
POST /api/skills/install/upload
Content-Type: multipart/form-data
```

**Request Body:**
- `archive`: The `.skill` file to install. The filename must end with `.skill`.

**Response:** the same shape as [Install Skill](#install-skill). An upload beyond
the archive limit returns `413`. A missing `archive` field returns `422`; a
non-file `archive` part is rejected during multipart parsing and returns `400`.

#### List Custom Skills

List only the caller's user-owned custom skills. Legacy shared skills are
read-only and appear only in [List Skills](#list-skills). Available to normal
users; no administrator role is required. When authorization is enabled, only
skills visible to the caller's role are returned.

```http
GET /api/skills/custom
```

**Response:** the same shape as [List Skills](#list-skills).

#### Get Custom Skill Content

Return a custom skill's metadata plus its raw `SKILL.md` content. Requires an
authenticated administrator.

```http
GET /api/skills/custom/{skill_name}
```

**Response:**
```json
{
  "name": "my-skill",
  "description": "My custom skill",
  "license": null,
  "category": "custom",
  "enabled": true,
  "editable": true,
  "content": "---\nname: my-skill\ndescription: My custom skill\n---\n\n# My Skill\n\nInstructions for the agent..."
}
```

#### Edit Custom Skill

Replace a custom skill's `SKILL.md`. The content is validated and security
scanned; a blocked scan returns `400` and nothing is written. Each accepted edit
appends a history entry. Requires an authenticated administrator.

```http
PUT /api/skills/custom/{skill_name}
Content-Type: application/json
```

**Request Body:**
```json
{
  "content": "---\nname: my-skill\ndescription: My custom skill\n---\n\n# My Skill\n\nUpdated instructions..."
}
```

The submitted content replaces the complete `SKILL.md`, so it must include YAML
frontmatter with a nonempty `description` and a `name` that matches
`{skill_name}`. Content without frontmatter returns `400`.

**Response:** the updated skill with `content`, the same shape as
[Get Custom Skill Content](#get-custom-skill-content).

#### Delete Custom Skill

Delete a custom skill and record the deletion in its history. Requires an
authenticated administrator.

```http
DELETE /api/skills/custom/{skill_name}
```

**Response:**
```json
{
  "success": true
}
```

#### Get Custom Skill History

Return the custom skill's recorded changes, oldest first. Requires an
authenticated administrator.

```http
GET /api/skills/custom/{skill_name}/history
```

**Response:**
```json
{
  "history": [
    {
      "ts": "2026-01-01T00:00:00+00:00",
      "action": "human_edit",
      "author": "human",
      "thread_id": null,
      "file_path": "SKILL.md",
      "prev_content": "...",
      "new_content": "...",
      "scanner": {"decision": "allow", "reason": "..."}
    }
  ]
}
```

An invalid skill name returns `400`; a skill with neither a `SKILL.md` nor
recorded history returns `404`.

#### Rollback Custom Skill

Restore a custom skill's `SKILL.md` from a history entry. The restored content is
validated and security scanned before it is written, and the rollback is itself
recorded as a history entry. Requires an authenticated administrator.

```http
POST /api/skills/custom/{skill_name}/rollback
Content-Type: application/json
```

**Request Body:**
```json
{
  "history_index": 0
}
```

`history_index` indexes the list returned by
[Get Custom Skill History](#get-custom-skill-history) and defaults to `-1` (the
latest change). An out-of-range index returns `400`.

**Response:** the updated skill with `content`, the same shape as
[Get Custom Skill Content](#get-custom-skill-content).

#### Export a Custom Skill

Admin session authentication is required for both requests. PAT credentials cannot export. Only the current user's custom skill is eligible; public, legacy and integration fallback is never used. A disabled custom skill remains eligible.

1. `GET /api/skills/custom/{skill_name}/export-manifest` returns `skill_name`, `revision` (SHA-256 or null), `can_export`, `file_count`, `directory_count`, `total_bytes`, `files` (`path`, `type`, `size`, `executable`), `requirements` (`compatibility`, `allowed_tools`, `required_secrets` names and optional flags), and structured `warnings`/`blockers`. Paths are relative; `.` is the package root, counted in directory/entry totals. Structural blockers return a non-downloadable manifest. Declarations are not credential values or dependency verification.
2. `GET /api/skills/custom/{skill_name}/export?expected_revision=<64 lowercase hex characters>` recaptures content and rejects stale previews with 409 before sending ZIP headers. Successful responses carry `application/zip`, attachment `<skill_name>.skill`, accurate `Content-Length`, `Cache-Control: private, no-store`, and `X-Content-Type-Options: nosniff`.

Error `detail` contains a safe `code`, `message`, and optional relative `path`. Codes/statuses: `skill_not_found` 404, `skill_changed` 409, `skill_export_limit_exceeded` 413, `skill_export_unsupported` 422, `skill_export_busy` 429, `skill_export_timeout` 503, `skill_export_failed` 500; existing 401/403 auth behavior applies. Limits are 4096 entries including directories, 64 MiB/file, 100 MiB raw/ZIP, 1 MiB frontmatter, 1024 UTF-8 bytes per ZIP path and depth 32. Frontmatter preflight rejects YAML aliases and bounds structure to 32 nesting levels / 16384 parser events before constructing YAML objects. A 5-second lock wait and 60-second cooperative worker deadline bound work; blocking OS calls cannot be forcibly interrupted. Two export slots are shared across all users in each Gateway process; both previews and downloads use them, and 429 means that process-wide capacity is occupied. Slots remain held through worker drain and temporary-file cleanup. The streaming phase has a separate 120-second inactivity deadline, reset after each successful ASGI send. A continuously progressing transfer may exceed 120 seconds overall; a stalled send does not reset the deadline. Expiry aborts the incomplete download (no replacement JSON after ZIP headers); clients must retry. Client disconnect during preparation cancels and drains the worker, then exits the handler normally rather than leaking a synthetic task cancellation. No export cache, persistent job or sharing URL is created.

Raw skill files, sidecars and empty directories are preserved. No hooks/scripts run during export and no secrets are redacted from package files. Import still uses normal security scanning and conflict checks. Export requires no-follow descriptor-relative host filesystem operations; unsupported platforms receive 422 rather than following links unsafely.

#### Reload Skills

Invalidate the skill prompt caches for every user, then publish a reset marker
beside the shared extensions config so every Gateway process mounting that
directory rescans the configured public, custom, and legacy skill directories
before its next prompt build. Runs that have already started keep their
existing skill snapshot.

```http
POST /api/skills/reload
```

The request has no body and requires an authenticated administrator. For a
cookie-authenticated request, send the CSRF cookie value in the matching header:

```bash
curl -X POST http://localhost:2026/api/skills/reload \
  -b cookies.txt \
  -H "X-CSRF-Token: <csrf_token-cookie-value>"
```

**Response:**

```json
{
  "success": true,
  "scope": "shared_config",
  "message": "Skill cache reset published through the shared config directory; subsequent runs in every Gateway process sharing it will rescan the latest skills."
}
```

`success` confirms cache invalidation, not that every file on disk was valid:
malformed skills retain the existing parser behavior of being skipped and
logged. The endpoint returns `401` for unauthenticated callers, `403` for
non-admin users, and a generic `500` if the invalidation mechanism itself
fails, the handling process's background scan does not finish within the cache
refresh timeout, or the reset marker cannot be written. A loader-level failure,
such as an unavailable mounted root, does not publish an empty catalog: the
last successfully loaded process cache remains available. A timed-out scan
continues in its daemon worker and can still populate the process cache when
it finishes.

`shared_config` means the handling process refreshed itself and then atomically
replaced `.<extensions config name>.skills-cache-reset.json` next to the
resolved `extensions_config.json`. Every Gateway worker or Pod mounting that
directory compares the marker's signature (at most once per second) before it
serves a cached skill list, so one request reaches every replica sharing the
volume; it does not claim a deployment-wide broadcast when replicas use
independent filesystems. The skill install, edit, delete, rollback and
enable/disable endpoints publish the same marker after their own change is
durable, scoped to the calling user for custom skills; if that publication
fails after the change was applied, they return their generic `500` (never
`404`), and a client that disconnects mid-request does not skip it. When no extensions
config path can be resolved there is no shared directory to publish into; the
request still refreshes the current process and returns `"scope": "process"`
with the message `Skill caches invalidated; subsequent runs in this Gateway
process will rescan the latest skills.` External MinIO/NFS/CSI writes bypass
the validation, SkillScan, and history used by the install/edit APIs, so the
mounted directory must be writable only by trusted operators.

### File Uploads

#### Upload Files

Upload one or more files to a thread.

Automatic Office/PDF-to-Markdown conversion is disabled by default. To enable
host-side parsing in a trusted deployment, set
`uploads.auto_convert_documents: true` in `config.yaml`. See
[document conversion configuration](FILE_UPLOAD.md#支持的文档格式).

```http
POST /api/threads/{thread_id}/uploads
Content-Type: multipart/form-data
```

**Request Body:**
- `files`: One or more files to upload

**Response** (example with automatic conversion enabled and successful):
```json
{
  "success": true,
  "files": [
    {
      "filename": "document.pdf",
      "size": 1234567,
      "path": ".deer-flow/threads/abc123/user-data/uploads/document.pdf",
      "virtual_path": "/mnt/user-data/uploads/document.pdf",
      "artifact_url": "/api/threads/abc123/artifacts/mnt/user-data/uploads/document.pdf",
      "markdown_file": "document.md",
      "markdown_path": ".deer-flow/threads/abc123/user-data/uploads/document.md",
      "markdown_virtual_path": "/mnt/user-data/uploads/document.md",
      "markdown_artifact_url": "/api/threads/abc123/artifacts/mnt/user-data/uploads/document.md"
    }
  ],
  "message": "Successfully uploaded 1 file(s)"
}

```

**Error behavior:** an unexpected storage failure answers `500` with `{"detail": "Failed to upload <filename>."}`; the underlying error, which can name host paths, stays in server logs.

**Name collisions:** filenames are claimed unique against the thread's existing uploads and reserved atomically — a same-name upload never replaces the existing file; it lands as `document_1.pdf` (the response's `filename`/`original_filename` reflect the claimed name). Use the artifacts `PUT` endpoint for sanctioned in-place updates.

**Supported Document Formats** (converted to Markdown only when automatic conversion is enabled):
- PDF (`.pdf`)
- PowerPoint (`.ppt`, `.pptx`)
- Excel (`.xls`, `.xlsx`)
- Word (`.doc`, `.docx`)

#### List Uploaded Files

```http
GET /api/threads/{thread_id}/uploads/list
```

**Response:**
```json
{
  "files": [
    {
      "filename": "document.pdf",
      "size": 1234567,
      "path": ".deer-flow/threads/abc123/user-data/uploads/document.pdf",
      "virtual_path": "/mnt/user-data/uploads/document.pdf",
      "artifact_url": "/api/threads/abc123/artifacts/mnt/user-data/uploads/document.pdf",
      "extension": ".pdf",
      "modified": 1705997600.0
    }
  ],
  "count": 1
}
```

#### Delete File

```http
DELETE /api/threads/{thread_id}/uploads/{filename}
```

**Response:**
```json
{
  "success": true,
  "message": "Deleted document.pdf"
}
```

**Error behavior:** an unexpected storage failure answers `500` with `{"detail": "Failed to delete <filename>."}`; the underlying error stays in server logs.

### Thread Cleanup

Remove DeerFlow-managed local thread files under `.deer-flow/threads/{thread_id}` after the LangGraph thread itself has been deleted.

```http
DELETE /api/threads/{thread_id}
```

**Response:**
```json
{
  "success": true,
  "message": "Deleted local thread data for abc123"
}
```

**Error behavior:**
- `422` for invalid thread IDs
- `500` returns a generic `{"detail": "Failed to delete local thread data."}` response while full exception details stay in server logs

Deleting a thread also removes every user's read marker for it.

### Thread Origin, Activity and Unread State

Runs the server starts for a user (a schedule, an IM channel, a GitHub agent,
an extension, an MCP notification) carry a server-owned origin in their
metadata, and so does a thread the server creates for one:

```json
{ "deerflow_origin": { "kind": "im_channel", "provider": "feishu" } }
```

`kind` is one of `schedule`, `im_channel`, `github`, `extension`,
`mcp_notification` (`contracts/thread_origin_contract.json`); `provider`
(IM/GitHub) and `namespace` (extension plugin) are optional. Clients cannot
set it: `POST /api/threads` and `PATCH /api/threads/{thread_id}` strip it, run
admission drops it from `metadata` and `config.metadata`, and only server-side
launchers and the internal channel caller may stamp it. The run's kind is also
stored as `runs.origin_kind` (`null` for interactive and pre-upgrade runs). IM
threads keep `metadata.channel_source` as their marker. This key is unrelated to
the message-level `additional_kwargs.deerflow_scheduled_origin` of a scheduled
prompt.

**Activity feed.** Requires SQL persistence (`503` on the memory backend);
`GET /api/features` reports `{"thread_activity": {"available": true}}`.

```http
GET /api/thread-activity?cursor=1834:run-42&limit=200
```

```json
{
  "cursor": "1840:run-57",
  "threads": [{ "thread_id": "…", "origin_kind": "schedule", "status": "success" }],
  "truncated": false,
  "read_version": 17
}
```

- Without `cursor` the call only seeds: it returns the caller's current
  position (`"0:"` for a user with no runs) and no threads.
- With a cursor it pages the caller's run changes in order (`limit` 1-500,
  default 200). `threads` lists each thread with a change from a
  server-originated run of the caller in the page, once, with the status of
  its latest such change. The caller's own interactive runs advance the cursor
  but are not listed.
- `cursor` is the position of the last change used, so a `truncated` page
  continues on the next poll without gaps. Treat it as opaque.
- `read_version` is the caller's read clock; it changes when a thread is
  marked read on any device.
- A malformed cursor returns `422` with `detail.code` `invalid_cursor`.
- Every query is scoped to the caller; another user's cursor reveals nothing.

**Unread state.** A thread is unread for a user while one of that user's
server-originated runs in it changed after the user last read it. The user's
own interactive runs, other users' runs in a shared thread and pre-upgrade runs
never make a thread unread.

```http
POST /api/threads/{thread_id}/read
```

```json
{ "unread": false, "read_version": 18 }
```

The read position never moves backwards. A thread already read up to its
latest run writes nothing and returns the unchanged `read_version`. Items of
`POST /api/threads/search` carry `unread` (`true`/`false`; `null` without SQL
persistence); other thread responses return `null`.

### Projects

#### Get Projects Config

```http
GET /api/projects/config
```

The `projects` config-block knobs the UI needs for client-side validation. Requires the `projects:read` scope (PATs included).

**Response:**
```json
{
  "instructions_max_bytes": 8192,
  "trash_retention_days": 30
}
```

Values come from `projects.instructions_max_bytes` and `projects.trash_retention_days` in `config.yaml`; the documented defaults apply when the block is absent.

### Project Documents

Per-project document shelf (Projects Phase 2). All routes fail closed: a missing or foreign project/document is `404` (never `403`); uploads and individual trash require an active project — archived projects answer `404` for those while keeping reads; a memory-backend deployment answers `503` `"Projects not available"`. Trashed rows are invisible to every route.

#### List Documents

```http
GET /api/projects/{project_id}/documents?limit=100&offset=0
```

**Query Parameters:** `limit` (default 100, 1..1000), `offset` (default 0) — out-of-bounds values are `422`.

**Response:** `{"documents": [{"id", "name", "size_bytes", "sha256", "source_thread_id", "source_kind", "source_name", "created_at", "updated_at", "summary", "content_missing"}], "total", "limit", "offset"}` in `updated_at DESC, id ASC` order. `content_missing` is read-time truth (never persisted): `true` when the document's immutable original is missing or size-mismatched (external interference); the derived `converted.md` companion is not the integrity anchor. `summary` is a best-effort LLM-generated one-line description (nullable; see `projects.summaries_enabled` in `config.example.yaml`) — null when generation is disabled, pending, failed, or the source is ineligible; it is LLM-processed untrusted content and is tag-neutralized at every model-visible exit.

#### Upload Document

```http
POST /api/projects/{project_id}/documents
Content-Type: multipart/form-data
```

Exactly one file per request (`file` part), plus an optional `name` form field (defaults to the multipart filename). The name is rejected with `400` when empty after normalization, separator-bearing, or over 255 UTF-8 bytes; empty files are `400`; files over `uploads.max_file_size` are `413`.

**Response:** `201 Created` with `{"document": {...}, "deduplicated": false}`. Re-uploading identical content returns the existing row with `200 OK` and `"deduplicated": true` — the first writer's name wins. Re-upload after trash creates a fresh row.

#### Save Thread File to Shelf (from-thread)

```http
POST /api/projects/{project_id}/documents/from-thread
Content-Type: application/json
```

```json
{"thread_id": "abc123", "kind": "upload", "name": "report.pdf", "shelf_name": "q3-report.pdf"}
```

Copies one file from the thread's own uploads (`"kind": "upload"`) or outputs (`"kind": "output"`) directory into the shelf as a project-owned snapshot; the source file is never moved. `name` locates the source file inside the thread; `shelf_name` is optional and defaults to the source name, following upload-name validation (`400`). A source that does not resolve inside that thread's directory — separator-bearing names, escapes, missing files, or a missing/foreign thread — is `404`, indistinguishable from absence. Files over `uploads.max_file_size` are `413`; empty files are `400`.

**Response:** same as Upload Document — `201 Created` with `{"document": {...}, "deduplicated": false}`, or `200 OK` on a content dedup hit. The created row records `source_thread_id` / `source_kind` / `source_name` provenance.

#### Attach Document to Thread

```http
POST /api/projects/{project_id}/documents/{document_id}/attach-to-thread/{thread_id}
```

Materializes an independent copy of a live shelf document into the target thread's uploads directory through the same ingestion pipeline as an ordinary upload (filename claiming, size checks, optional conversion under `uploads.auto_convert_documents`, sandbox-readable permissions, and sandbox sync for non-mounted providers; a caller denied `sandbox:execute` keeps the host upload without allocating a sandbox). Reading an archived source project's shelf is allowed and does not mutate it; a missing/foreign document, or a target thread the caller cannot write, is `404`. A row whose original bytes are missing or size-mismatched answers `409` `"content_missing"`.

**Response:** `200 OK` with `{"filename", "size_bytes", "virtual_path", "artifact_url"}` — returned only after ingestion succeeds. An unexpected ingestion failure answers `500` with `{"detail": "Failed to upload <name>."}`; the underlying error stays in server logs.

#### Get Document Content

```http
GET /api/projects/{project_id}/documents/{document_id}/content?download=false
```
Serves the converted-markdown companion when present, else inline text when the original samples as text, else an attachment; `download=true` always attaches. Active content (`text/html`, `text/xml`, `application/xml`, `text/xsl`, any `+xml` type such as XHTML/SVG) is always forced to an attachment regardless of `download`, mirroring the artifacts router, so it never executes script on the application origin. A row whose original bytes are missing or size-mismatched answers `409` `"content_missing"`.

#### Delete Document (move to trash)

```http
DELETE /api/projects/{project_id}/documents/{document_id}
```

**Response:** `204`. Recoverable trash: the row keeps its bytes and a `{project_id, project_name}` origin snapshot. Deleting a project moves its whole shelf to trash in the same transaction. Restore/purge endpoints land with the trash-completion slice.

### Project Thread Files

Read-only conversation-files view over a project's member threads (Projects Phase 2) — the discovery route for Save Thread File to Shelf. Archived projects keep read access; a missing or foreign project is `404`.

#### List Thread Files

```http
GET /api/projects/{project_id}/thread-files?offset=0&thread_limit=20&file_limit=50
```

**Query Parameters:** `offset` (member-thread cursor, default 0), `thread_limit` (default 20, 1..50), `file_limit` (per-thread file cap, default 50, 1..200) — out-of-bounds values are `422`.

**Response:** `{"groups": [{"thread_id", "display_name", "updated_at", "truncated", "files": [{"kind": "upload"|"output", "name", "size_bytes", "modified_at", "artifact_url"}]}], "next_offset", "truncated"}`. Member threads are paged in the same non-archived order as the project thread list; `next_offset` is `null` when no threads remain. Each thread contributes up to `file_limit` files across its uploads and outputs; a group's `truncated` is `true` when that thread's listing was cut, and the envelope `truncated` is the OR over the page's groups. Entries disappear when their thread is deleted — the view keeps no storage of its own.


### Trash

Recoverable deletion tier for project shelf documents (Projects Phase 2). Trashed rows keep their bytes and a `{project_id, project_name}` origin snapshot for `projects.trash_retention_days` (default 30) before the retention sweep may purge them; permanent purge is a separate action. All routes fail closed: a missing or foreign document/project is `404` (never `403`), restoring into an archived or foreign target is the same `404`, and a memory-backend deployment answers `503` `"Projects not available"`. Purge endpoints carry no confirmation parameter — the "this cannot be undone" step is a UI contract, not a server-enforced handshake.

#### List Trashed Documents

```http
GET /api/trash/documents?limit=100&offset=0
```

**Query Parameters:** `limit` (default 100, 1..1000), `offset` (default 0) — out-of-bounds values are `422`.

**Response:** `{"documents": [{"id", "name", "size_bytes", "sha256", "source_thread_id", "source_kind", "source_name", "created_at", "updated_at", "trashed_at", "trash_origin": {"project_id", "project_name"} | null}], "total", "limit", "offset"}`, most recently trashed first. The retention sweep runs lazily before the listing (a sweep failure is logged and never blocks it).

#### Restore Document

```http
POST /api/trash/documents/{document_id}/restore
Content-Type: application/json

{"project_id": "…"}
```

**Body:** `project_id` optional. Target = the body value, else `trash_origin.project_id` when that project still exists, is owned, and is active; otherwise `404` (the UI offers the project picker). A foreign or archived target is the same `404` as a missing one.

**Response:** `{"outcome": "restored" | "merged", "document": <ProjectDocumentResponse>}`. `merged` means the target already had an active row with identical bytes: the trash row is deleted and `document` is the surviving active row. Restore re-points the row without moving any file. Missing or size-mismatched content answers `409` `"content_missing"` and leaves the row trashed.

#### Purge Document

```http
POST /api/trash/documents/{document_id}/purge
```

**Response:** `204`. Permanently unlinks the original and `derived/converted.md`, then deletes the row, in one row-locked transaction. Already-absent content counts as removed; any other file-cleanup failure rolls back, keeps the trashed row, and answers `500` with a retryable message.

#### Empty Trash

```http
POST /api/trash/purge
```

**Response:** `{"purged": <int>}` — permanently deletes every trashed document of the caller, regardless of age: the confirmation covers the whole listing, so the retention cutoff never gates this route. Each row goes through the same guarded row-locked transaction as the single-document purge — bytes first, then the row. A file-cleanup failure other than already-absent content answers `500` with a retryable message, leaving that row and every row not yet visited trashed. Retention expiry is enforced only by the sweep (lazily before `GET /api/trash/documents` and once at gateway startup).

### Artifacts

#### Get Artifact

Download or view an artifact generated by the agent.

```http
GET /api/threads/{thread_id}/artifacts/{path}
```

**Path Examples:**
- `/api/threads/abc123/artifacts/mnt/user-data/outputs/result.txt`
- `/api/threads/abc123/artifacts/mnt/user-data/uploads/document.pdf`

**Query Parameters:**
- `download` (boolean): If `true`, force download with Content-Disposition header

**Response:** File content with appropriate Content-Type. HTML and XML documents (`.html`, `.xml`, `.xhtml`, `.svg`, and other `+xml` types) are always returned as attachments, regardless of `download`, so generated markup never renders in the application origin.

---

### Scheduled Tasks

Owner-scoped task management for `/workspace/scheduled-tasks`. Reads need
`threads:read`; mutations need `threads:write`, and create, update, resume and
trigger also need `runs:create`.

```http
GET    /api/scheduled-tasks
POST   /api/scheduled-tasks
GET    /api/scheduled-tasks/{task_id}
PATCH  /api/scheduled-tasks/{task_id}
DELETE /api/scheduled-tasks/{task_id}
POST   /api/scheduled-tasks/{task_id}/pause
POST   /api/scheduled-tasks/{task_id}/resume
POST   /api/scheduled-tasks/{task_id}/trigger
GET    /api/scheduled-tasks/{task_id}/runs?status=&limit=&offset=
GET    /api/threads/{thread_id}/scheduled-tasks
GET    /api/threads/{thread_id}/scheduled-task-events?limit=50
POST   /api/scheduled-tasks/preview-cron
```

**Task fields.** Create and PATCH accept `title`, `prompt`, `schedule_type`
(`once` | `cron` | `interval`), `schedule_spec`, `timezone`, `context_mode`,
`thread_id`, `assistant_id`, plus:

| Field | Type | Notes |
|---|---|---|
| `goal_objective` | string \| null | What one run must achieve (at most 4000 characters after whitespace normalization); requires `fresh_thread_per_run`. |
| `max_runs` | integer \| null | Safety cap: automatic runs over the task's lifetime (trial runs excluded); at least 1. |
| `end_at` | date-time \| null | Safety cap: no automatic run after this time. Without a UTC offset it is wall-clock time in the task's timezone. |
| `stop_condition` | string \| null | The user's "stop when …" rule, stored in its own field. Whitespace collapses to single spaces; at most 500 characters; blank means none. It is never part of `prompt`: each launch appends it as an instruction to call `stop_scheduled_task` (or, with `scheduler.tool_enabled` off, to report a met rule). |
| `notification_group_connection_id` | string \| null | One additional connected Feishu group belonging to the task owner. Select a group binding returned by `GET /api/channels/connections`; it must record `chat_type=group` and a matching chat/workspace ID. Omitted in PATCH preserves the target; null clears it. Invalid choices return `invalid_notification_group` (422), lookup outages return `scheduler_unavailable` (503). |

For one-time `schedule_spec.run_at` and `end_at` without an offset, a local time
skipped by a timezone clock change is rejected: `422 invalid_schedule` for
`run_at`, or `422 invalid_request` for `end_at`, including an end-time renewal
on Resume. This applies to REST and conversation scheduling. Ambiguous local
times retain the earlier occurrence (`fold=0`); explicit offsets preserve the
specified instant. Unchanged live once definitions keep their previously
accepted execution time; a changed definition or a PATCH that rearms a terminal
task is validated again. Ordinary Resume retains the admitted stored schedule
without revalidating its wall time; a renewed `end_at` is still checked.

In a PATCH, sending `null` clears `goal_objective`, `max_runs`, `end_at` or
`stop_condition` or `notification_group_connection_id`; for every other field `null` means unchanged (`assistant_id:
null` restores `lead_agent`). PATCH may change `schedule_type` together with
`schedule_spec`. Changing the goal, `prompt` or `stop_condition` starts a new
count for the automatic pause after three unmet runs. A PATCH that changes the
schedule of a finished task re-arms it (then its safety cap applies); a PATCH
that only changes the cap leaves a finished task finished.

Group-target edits alone never change the schedule or re-arm a finished task.
Personal recipients are retained, and pending outbox rows keep their original
target. `/api/features.scheduled_tasks.group_notifications` reports support;
the conversation scheduling tool's inputs are unchanged in this version.

Task responses (list, get, create, PATCH, pause, resume) return the stored task
plus `automatic_runs_used` (integer) and `active_run_status` (`queued` |
`launching` | `running` | null). A recurring task's `status` stays `enabled`
while its run executes, so check `active_run_status` for "is a run active".
`GET /api/threads/{thread_id}/scheduled-tasks` rows add `thread_relation`
(`origin`: created in the chat, `reuse`: runs in the chat, `run`: the chat is
one of its runs) and `thread_run` (`{"run_number", "trigger", "scheduled_for",
"status"}` for `run`, else null).

`GET /api/threads/{thread_id}/scheduled-task-events` (thread owner check;
`limit` 1–200, default 50) returns the caller's lifecycle events for a chat
that created tasks, oldest first:

```json
{ "events": [ { "id": "evt-…", "task_id": "task-…", "event": "task_stopped", "reason_code": "agent_stop",
  "task_title": "Check the release checklist", "stop_condition": "all items are ticked",
  "run_thread_id": "…", "run_agent_name": "lead_agent", "run_number": 4, "run_status": "success", "max_runs": null, "end_at": null,
  "schedule_type": "cron", "after_run_id": "…", "created_at": "2026-10-06T09:00:03+00:00" } ] }
```

`event` is `task_stopped` (the agent paused its own schedule, reason
`agent_stop`), `task_paused` (automatic pause, `consecutive_unmet`) or
`task_finished` (`max_runs`, `end_at`, or a one-time task's `once_done` /
`once_failed`); the vocabulary is pinned in
`contracts/scheduled_goal_notes_contract.json`. Each row is written in the
same transaction as the state change it reports and is unique per task,
transition and event, so recovery never adds a second one. `run_status` is the
last run's outcome (a stop or finish after a failed run says so);
`run_thread_id` is null when the occurrence never launched, and
`run_agent_name` is the agent that ran it (its run chat's route); `after_run_id` is
the newest run of the chat when the event was written (the chat shows the line
after that turn, else at the end). Rows keep a title snapshot and stay after
the task is deleted; deleting the chat removes them. Tasks created on the
Scheduled tasks page have no originating chat and no rows. IDs are for links
only.

**Create** returns `409 scheduler_not_running` (after the request validates)
while this Gateway process's scheduler is not running; the tasks page's
Duplicate is a create.

**Pause** of a `completed`, `failed` or `cancelled` task returns `409
task_finished`.

**Resume** accepts an optional renewal body:

```json
{ "max_runs": 70, "end_at": "2026-12-31T18:00:00" }
```

Each field is optional; `null` clears that cap (a chat-created task that runs
more often than hourly must keep one: `422 frequent_requires_limit`). The next
run is the stored one when still ahead, otherwise computed from now, so no
catch-up run happens; a one-time task whose time passed returns `422
once_time_passed`. Resuming an `enabled` task changes nothing. A task whose
safety cap is used up returns `409 limits_exhausted` unless the same request
renews the limit named in `params.limit`: for `max_runs`, a `max_runs` above
`params.used` or `null`; for `end_at`, a later `end_at` or `null`. A later
`end_at` alone does not renew a used-up `max_runs`.
A future `end_at` that falls before the next run returns `422
end_at_before_first_run` (also on a PATCH that changes the schedule).

**Trigger** (one trial run) returns:

```json
{ "id": "task-…", "triggered": true, "outcome": "launched", "existing": false, "thread_id": "…" }
```

`outcome` is `launched` or `queued`; `existing: true` means a run was already
waiting and no trial was added.

**Runs** rows add `run_number` (the automatic-run number counted like
`max_runs`; null for trials and runs that never launched), `total_tokens` of the
launched run (null if none), and `summary` (first line of the agent's final
reply as plain text; when that line ends with a colon, the list items that
follow it are appended, joined with `；` for CJK text and `; ` otherwise; at
most 160 characters).

**Errors (breaking change).** Every error these routes raise for a well-typed
request is coded:

```json
{ "detail": { "code": "limits_exhausted", "message": "All 5 automatic runs are used. Raise max_runs above 5 or clear it (max_runs: null) in the same request to reactivate.", "params": { "limit": "max_runs", "used": 5, "max_runs": 5, "end_at": null } } }
```

`message` stays English for API clients; `params` is omitted when empty.
Clients that read `detail` as a string must read `detail.message` instead.
Three errors keep their old shape: a `403` from route permissions is the plain
string `"Permission denied: <permission>"`, FastAPI's own `422` for malformed
JSON or wrong types (for example `"max_runs": "abc"`) keeps its list `detail`,
and the shared `503` `"Thread metadata store not available"` (a Gateway without
a thread store) stays a plain string. Codes the web UI translates: `invalid_request`, `invalid_schedule`, `invalid_schedule_type`, `invalid_timezone`, `interval_too_short`, `interval_too_long`, `once_in_past`, `once_too_soon`, `once_time_passed`, `invalid_context_mode`, `reuse_thread_requires_thread`, `thread_not_found`, `invalid_assistant`, `unknown_assistant`, `invalid_goal`, `goal_requires_fresh_thread`, `invalid_stop_condition`, `invalid_notification_group`, `invalid_max_runs`, `end_at_in_past`, `end_at_before_first_run`, `frequent_requires_limit`, `max_runs_not_above_used`, `limits_exhausted`, `task_not_found`, `task_running`, `run_queued`, `task_changed`, `task_finished`, `task_quota_exceeded`, `scheduler_not_running`, `scheduler_unavailable` (503: this Gateway has no scheduled-task persistence), `trigger_failed`. Codes only the chat capability
returns: `timezone_required`, `authentication_required`, `permission_denied`, `scheduler_tools_disabled`, `authority_expired`, `conversation_not_found`, `interactive_run_required`, `unsupported_action`, `unsupported_fields`, `task_id_required`, `note_not_verbatim`, `note_limit_reached`, `trial_requires_direct_request`, `no_stop_authority`, `occurrence_not_active`. The machine-readable list is
`contracts/scheduled_task_errors_contract.json`.

**Feature flag.** `GET /api/features` includes:

```json
{ "scheduled_tasks": { "available": true, "running": true, "tool_enabled": false, "min_interval_seconds": 60 } }
```

`available`: the task APIs have persistence. `running`: this process's
scheduler poller runs (tasks fire and can be created). `tool_enabled`: chats
can manage tasks and scheduled runs can stop their own schedule.
`min_interval_seconds`: shortest interval and earliest one-time delay.

**Browser timezone for chat-created tasks.** A run request may carry
`context.client_timezone` (an IANA name, at most 64 characters, for example
`"Asia/Shanghai"`). It is read only as the default timezone of tasks the
`schedule_task` tool creates in that turn; an unknown or malformed value is
ignored, and it never reaches the run config, the checkpoint or the prompt.

**Scheduled run messages.** A scheduled run's prompt message has the id
`scheduled-<task_run_id>` and carries `additional_kwargs.deerflow_scheduled_origin`
(`task_id`, `task_run_id`, `trigger`, `run_number`, `scheduled_for`, `timezone`,
`schedule_type`, `task_title`, `instructions`, `stop_condition`,
`standing_notes`), the user-written parts a client shows instead of the
launched text. Show an `interval` run's time in the viewer's zone (its stored
`timezone` may be the `"UTC"` placeholder of a task created without one) and
other runs in `timezone`. A new run conversation is titled
`"{task title} · MM-DD HH:MM"` in the task's zone, or `"{task title} · #{run}"`
when that zone is only the placeholder. The key is server-owned: it is
stripped from client-supplied messages and state updates.

**IM notices ("scheduled task updates").** With `channel_connections.enabled`,
each item of `GET /api/channels/providers` carries
`proactive_notifications` (bool): whether scheduled-task updates are pushed to
that app. Only providers with proactive push (WeCom, Feishu/Lark, WeChat and QQ private chat today) get notices; the
others get none, and Settings says so. An occurrence sends at most one message,
in the owner's `locale` preference (see Account Preferences), else
`channel_connections.notification_locale`. Notices carry no IDs and no links;
see `backend/docs/CONFIGURATION.md` for the events and the merge rule.

```json
{ "provider": "wecom", "display_name": "WeCom", "connection_status": "connected", "proactive_notifications": true }
```

**Runtime bot credentials.** `POST /api/channels/{provider}/runtime-config`
(body `{"values": {...}}`, keyed by the provider's `credential_fields`) and
`DELETE /api/channels/{provider}/runtime-config` are admin-only; they save or
disconnect the deployment's bot for that provider and answer with its provider
item. Password values come back masked (`********`); sending the mask back
keeps the stored value. A save whose channel fails to start on the serving
instance, or a disconnect whose channel fails to stop there, answers `400` and
stores nothing (under `channels.leader_lease` a non-holder stores the save and
the holder applies it on its next renewal tick). With a SQLite or
PostgreSQL database the values are stored encrypted with
`DEER_FLOW_CREDENTIALS_KEY` and shared by every Gateway instance; without a
usable key a save answers `503`. See `backend/docs/IM_CHANNEL_CONNECTIONS.md`.

---

## Error Responses

Most APIs return errors in this format (scheduled-task routes and skill export
return a coded object instead; see [Scheduled Tasks](#scheduled-tasks)):

```json
{
  "detail": "Error message describing what went wrong"
}
```

**HTTP Status Codes:**
- `400` - Bad Request: Invalid input
- `404` - Not Found: Resource not found
- `422` - Validation Error: Request validation failed
- `500` - Internal Server Error: Server-side error

---

## Authentication

DeerFlow supports four HTTP identity sources. They share the same thread/run isolation rules but differ in whether a row is created in `users` and how external identities are mapped. See [AUTH_DESIGN.md](AUTH_DESIGN.md) for the full design.

| Model | Entry | `users` table | Isolation key |
|---|---|---|---|
| Browser session | `access_token` cookie after login/register | Yes | `users.id` |
| OIDC / SSO | OAuth callback → cookie | Yes | `users.id` (see [SSO.md](SSO.md)) |
| IM channel binding | Connect code + `channel_connections` | Bound to registered user | `channel_connections.owner_user_id` |
| **Internal Auth** | `X-DeerFlow-Internal-Token` + `X-DeerFlow-Owner-User-Id` | **No** | Owner string on `threads_meta.user_id` |

**IM channel binding** and **Internal Auth** are both *platform-trust* integrations: DeerFlow trusts the channel/platform to authenticate end users. IM bindings persist the mapping in `channel_connections` / `channel_conversations` and require a DeerFlow `users` row. Internal Auth lets a platform call the Gateway API directly with a deployment-shared token and a per-request owner header—no `users` row, but thread/run/checkpoint isolation works the same way.

### Browser session (default)

DeerFlow enforces authentication for all non-public HTTP routes. Public routes are limited to health/docs metadata and these public auth endpoints:

- `POST /api/v1/auth/initialize` creates the first admin account when no admin exists.
- `POST /api/v1/auth/login/local` logs in with email/password and sets an HttpOnly `access_token` cookie.
- `POST /api/v1/auth/register` creates a regular `user` account and sets the session cookie.
- `POST /api/v1/auth/logout` clears the session cookie.
- `GET /api/v1/auth/setup-status` reports whether the first admin still needs to be created.

The authenticated auth endpoints are:

- `GET /api/v1/auth/me` returns the current user.
- `POST /api/v1/auth/change-password` changes password, optionally changes email during setup, increments `token_version`, and reissues the cookie.

Protected state-changing requests also require the CSRF double-submit token: send the `csrf_token` cookie value as the `X-CSRF-Token` header. Login/register/initialize/logout are bootstrap auth endpoints: they are exempt from the double-submit token but still reject hostile browser `Origin` headers.

User isolation is enforced from the authenticated user context:

- Thread metadata is scoped by `threads_meta.user_id`; search/read/write/delete APIs only expose the current user's threads.
- Thread files live under `{base_dir}/users/{user_id}/threads/{thread_id}/user-data/` and are exposed inside the sandbox as `/mnt/user-data/`.
- Memory and custom agents are stored under `{base_dir}/users/{user_id}/...`.

Note: MCP outbound connections can still use OAuth for configured HTTP/SSE MCP servers; that is separate from DeerFlow API authentication.

### Internal Auth (platform HTTP integration)

For server-to-server integrations (e.g. a Feishu or WeCom/Enterprise WeChat bot backend), configure:

```bash
export DEER_FLOW_INTERNAL_AUTH_TOKEN="<long-random-secret>"
```

| Header | Required | Description |
|---|---|---|
| `X-DeerFlow-Internal-Token` | Yes | Must match `DEER_FLOW_INTERNAL_AUTH_TOKEN`; missing/invalid → `401` |
| `X-DeerFlow-Owner-User-Id` | Yes for per-user isolation | Platform user id (e.g. `feishu_ou_alice`, `wecom_user_bob`); omit → `default` bucket |

Does **not** use browser cookies or CSRF tokens. Does **not** insert into `users`; sets `threads_meta.user_id` / `runs.user_id` from the owner header. DeerFlow validates only the platform token—not whether the owner id represents a real end user; user validity is entirely the platform's responsibility. See [AUTH_DESIGN.md — Internal Auth](AUTH_DESIGN.md#internal-auth-direct-http) for trust boundaries, persistence, and security notes.

Use the standard Gateway thread/run endpoints (`POST /api/threads`, `POST /api/threads/{thread_id}/runs/stream`, etc.) with the headers above on every request.

---

## Rate Limiting

No rate limiting is implemented by default. For production deployments, configure rate limiting in Nginx:

```nginx
limit_req_zone $binary_remote_addr zone=api:10m rate=10r/s;

location /api/ {
    limit_req zone=api burst=20 nodelay;
    proxy_pass http://backend;
}
```

---

## Streaming Support

Gateway's LangGraph-compatible API streams run events with Server-Sent Events (SSE).

**Thread-scoped streaming** (thread must exist):

```http
POST /api/langgraph/threads/{thread_id}/runs/stream
Accept: text/event-stream
```

**Stateless streaming** (no pre-created thread; Gateway auto-creates one):

```http
POST /api/langgraph/runs/stream
Accept: text/event-stream
```

Both endpoints return `Content-Location: /api/threads/{thread_id}/runs/{run_id}`.
The DeerFlow web UI and LangGraph SDK clients rely on this header to discover the
assigned `thread_id` and `run_id` on the first message of a new chat.

Every SSE response (both create endpoints above, `GET /api/threads/{thread_id}/runs/{run_id}/join`
and `GET`/`POST /api/threads/{thread_id}/runs/{run_id}/stream`) carries the same headers:

```http
Content-Type: text/event-stream
Cache-Control: no-cache, no-transform
Connection: keep-alive
X-Accel-Buffering: no
```

`no-transform` keeps compressing proxies (for example the Next.js rewrite proxy
when the frontend runs with `pnpm start` and no nginx) from gzipping and therefore
buffering the stream; `X-Accel-Buffering: no` does the same for nginx. A reverse
proxy you put in front of the Gateway should not compress `text/event-stream`.

### SSE replay retention and gaps

Clients may reconnect to a run stream with `Last-Event-ID`. Replay history is
bounded by `stream_bridge.queue_maxsize` (default `256`) and, for Redis, by the
rolling `stream_ttl_seconds`. A retained cursor resumes after that event with no
additional control frame.

The Memory bridge also remembers its exact last evicted event ID. Reconnecting
from that known predecessor replays every retained successor without a `gap`,
because none of the events after the cursor has been lost. This uses one ID per
run and does not increase the retained event count. Redis keeps its conservative
trimmed-cursor policy.

For other syntactically valid cursors older than the retained watermark, the
server sends exactly one `gap` event before any retained data and closes that
subscription without an `end` event:

```text
event: gap
data: {"code":"stream_replay_gap","run_id":"run-123","requested_event_id":"1718000000000-1","earliest_available_event_id":"1718000000100-42","latest_available_event_id":"1718000000200-84","recovery":"reload_durable_state"}

```

The frame deliberately has no SSE `id:`. Both `earliest_available_event_id` and
`latest_available_event_id` are `string | null` (they are `null` when no events
are retained in the buffer). Consumers must reload durable thread state and
persisted run events/messages, then may reconnect from `latest_available_event_id`
to follow newer live events, or rejoin without a cursor when the buffer is empty
(`latest_available_event_id` is `null`). A gap does not cancel the active run.
The same signal applies when a no-cursor subscriber has already established an
empty-stream wait but the first Redis wake-up falls behind before delivery; in
that case `requested_event_id` is `null`. Malformed cursor handling is
backend-specific and is not the same as a valid cursor that was evicted.

---

## SDK Usage

### Python (LangGraph SDK)

```python
from langgraph_sdk import get_client

client = get_client(url="http://localhost:2026/api/langgraph")
run_meta: dict[str, str] = {}


def on_run_created(meta) -> None:
    # langgraph-sdk 0.3.x parses Content-Location only when this callback is set.
    if meta.thread_id:
        run_meta["thread_id"] = meta.thread_id
    run_meta["run_id"] = meta.run_id


# Option A: stateless stream — no thread pre-creation
# Gateway auto-creates a thread and returns thread_id/run_id in Content-Location.
async for event in client.runs.stream(
    None,
    "lead_agent",
    input={"messages": [{"role": "user", "content": "Hello"}]},
    config={"configurable": {"model_name": "gpt-4"}},
    stream_mode=["values", "messages-tuple", "custom"],
    on_run_created=on_run_created,
):
    print(event)

thread_id = run_meta["thread_id"]  # persist before the next turn

# Option A (continued): same thread on the next turn
async for event in client.runs.stream(
    None,
    "lead_agent",
    input={"messages": [{"role": "user", "content": "What did I just ask?"}]},
    config={"configurable": {"thread_id": thread_id, "model_name": "gpt-4"}},
    stream_mode=["values", "messages-tuple", "custom"],
    on_run_created=on_run_created,
):
    print(event)

# Option B: thread-scoped stream — create thread first, then stream
thread = await client.threads.create()
async for event in client.runs.stream(
    thread["thread_id"],
    "lead_agent",
    input={"messages": [{"role": "user", "content": "Hello"}]},
    config={"configurable": {"model_name": "gpt-4"}},
    stream_mode=["values", "messages-tuple", "custom"],
    on_run_created=on_run_created,
):
    print(event)
```

### JavaScript/TypeScript

```typescript
// Using fetch for Gateway API
const response = await fetch('/api/models');
const data = await response.json();
console.log(data.models);

function parseRunLocation(contentLocation: string | null) {
  if (!contentLocation) return null;
  const match = /\/threads\/([^/]+)\/runs\/([^/]+)/.exec(contentLocation);
  if (!match) return null;
  return { threadId: match[1], runId: match[2] };
}

// Option A: stateless stream — no thread pre-creation
let threadId: string | undefined;
const firstResponse = await fetch("/api/langgraph/runs/stream", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    Accept: "text/event-stream",
  },
  body: JSON.stringify({
    input: { messages: [{ role: "user", content: "Hello" }] },
    stream_mode: ["values", "messages-tuple", "custom"],
  }),
});

const created = parseRunLocation(firstResponse.headers.get("Content-Location"));
threadId = created?.threadId;
console.log("thread_id:", created?.threadId, "run_id:", created?.runId);

// Option B: continue the same thread on the next turn
const followUpResponse = await fetch("/api/langgraph/runs/stream", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    Accept: "text/event-stream",
  },
  body: JSON.stringify({
    input: { messages: [{ role: "user", content: "What did I just ask?" }] },
    config: { configurable: { thread_id: threadId } },
    stream_mode: ["values", "messages-tuple", "custom"],
  }),
});

// Option C: thread-scoped stream when you already have a thread_id
const streamResponse = await fetch(`/api/langgraph/threads/${threadId}/runs/stream`, {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    Accept: "text/event-stream",
  },
  body: JSON.stringify({
    input: { messages: [{ role: "user", content: "Hello" }] },
    stream_mode: ["values", "messages-tuple", "custom"],
  }),
});

const reader = streamResponse.body?.getReader();
// Decode and parse SSE frames from reader in your client code.
```

### cURL Examples

```bash
# List models
curl http://localhost:2026/api/models

# Get MCP config
curl http://localhost:2026/api/mcp/config

# Upload file
curl -X POST http://localhost:2026/api/threads/abc123/uploads \
  -F "files=@document.pdf"

# Enable skill
curl -X PUT http://localhost:2026/api/skills/pdf-processing \
  -H "Content-Type: application/json" \
  -d '{"enabled": true}'

# Stateless stream — no thread pre-creation
curl -s -D - -N -X POST http://localhost:2026/api/langgraph/runs/stream \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{
    "input": {"messages": [{"role": "user", "content": "Hello"}]},
    "config": {
      "recursion_limit": 100,
      "configurable": {"model_name": "gpt-4"}
    },
    "stream_mode": ["values", "messages-tuple", "custom"]
  }'
# Read Content-Location: /api/threads/{thread_id}/runs/{run_id} from the headers.

# Continue the same thread on the next turn
curl -s -N -X POST http://localhost:2026/api/langgraph/runs/stream \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{
    "input": {"messages": [{"role": "user", "content": "What did I just ask?"}]},
    "config": {
      "configurable": {"thread_id": "abc123", "model_name": "gpt-4"}
    },
    "stream_mode": ["values", "messages-tuple", "custom"]
  }'

# Thread-scoped flow — create thread first, then stream
curl -X POST http://localhost:2026/api/langgraph/threads \
  -H "Content-Type: application/json" \
  -d '{}'

curl -X POST http://localhost:2026/api/langgraph/threads/abc123/runs/stream \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{
    "input": {"messages": [{"role": "user", "content": "Hello"}]},
    "config": {
      "recursion_limit": 100,
      "configurable": {"model_name": "gpt-4"}
    },
    "stream_mode": ["values", "messages-tuple", "custom"]
  }'
```

> The unified Gateway path defaults `config.recursion_limit` to 100 for
> plan-mode and subagent-heavy runs. Clients may still set
> `config.recursion_limit` explicitly — see the [Create Run](#create-run)
> section for details. Scheduled-task launches use
> `scheduler.recursion_limit` from `config.yaml` instead of a client body.

## Chat archive and restore

`POST /api/threads/search` accepts `archived: true` for archived chats or
`archived: false` for recent chats (including legacy rows without an archive flag).
Omit the field or use null to include both. Filtering applies before `limit` and
`offset` and is scoped to the authenticated user. Combine it with the existing
`metadata` and `status` filters when needed.

Archive with `PATCH /api/threads/{thread_id}` and body
`{"metadata":{"deerflow_archived":true}}`; use false to restore. The flag must be
a JSON boolean. Writes containing only boolean pin/archive flags preserve
`updated_at` and all other metadata. The owner-checked endpoint returns the normal
thread metadata response; original thread and artifact URLs remain available.
Archiving does not cancel runs, pause schedules, or change retention.

### Proactive IM notification preferences

`GET/PATCH /api/channels/notification-preferences` is authenticated and owner-scoped.
Both require `X-Expected-User-Id` matching the authenticated viewer; a stale
account returns 409. Session mutations retain normal CSRF protection. The local
auth-disabled default identity is supported without creating an auth account.

GET and PATCH return `{ "feishu": true, "wecom": true, "wechat": true, "qq": true }`
with the owner's effective booleans. PATCH accepts only supplied provider fields,
strict booleans or `null` (reset to enabled); other fields are rejected. Missing
preferences preserve the existing connected-channel default. Browser notification
preferences and chat connection state remain independent.

Opt-out and cancellation of that owner's pending/sending outbox rows for the
provider commit together; the internal terminal reason is `disabled_by_owner`.
Retries consume no additional attempts, and opting back in cannot reopen the
old rows. The worker rechecks preferences and claim ownership before starting a
provider send. A provider call already started may still complete; it cannot be
retracted. All Gateway workers must be upgraded to honor the shared preference.
Migration `0041_im_notice_preferences` adds only a small owner/provider preference
table; downgrade removes these choices, so subsequent upgrades use defaults.