feat: add tool calling, file upload, reasoning stream

[shlaji]

• Implement JSON tool call parsing and extraction from streaming responses

• Add file attachment upload via Dify /files/upload API with base64/Uint8Array support

• Parse ... tags for reasoning content in streams (reasoning-start/end events)

---

Summary by cubic

Adds tool calling, file uploads, reasoning streaming, and logger support to the Dify chat provider, with improved prompt building, stream parsing, and event handling. This enables richer interactions (tools + files), clearer reasoning/text streams, better debugging, accurate token usage, and robust file/header handling.

• New Features

  • Tool calling: parses JSON {name, arguments} in blocking/stream, emits tool-input-* and tool-call, finishReason can be "tool-calls".
  • Reasoning/text stream: splits into reasoning-* and text-*; ignores agent_thought; handles message_replace/message_file/ping/error; emits response-metadata; corrects usage for workflow_finished vs message_end; correctly ignores /</think> inside JSON strings (e.g., tool args).
  • File uploads: detects file parts, uploads via /files/upload, attaches IDs; supports base64/Uint8Array; logs failures; ignores both "content-type" and "Content-Type"; sets image/document types.
  • Prompt building: formats system prompt (, , identity); optional tool prompt injection; combines tool results + user text; detects conversation context from headers or message metadata.
  • Logging: pluggable logger and logMessages; structured logs for requests, chunks, and errors; default logger; disable via logger=false.

• Migration

  • Send "user-id" header (required) and optional "chat-id" for multi-turn.
  • Provide tools in call options; set injectToolsPrompt=false to disable prompt injection.
  • Attach files using content parts of type "file" with mediaType/mimeType and base64 or Uint8Array.
  • Handle new stream parts (reasoning-, tool-input-, tool-call) and the "tool-calls" finishReason; configure logger/logMessages as needed.

^{Written for commit 87b6d1e. Summary will update on new commits.}

[cubic-dev-ai]

3 issues found across 6 files

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="src/stream-parser.ts">

<violation number="1" location="src/stream-parser.ts:148">
P2: Tool-call JSON extraction can fail after a stray closing brace because depth goes negative and startIdx is never set for the next JSON object.</violation>

<violation number="2" location="src/stream-parser.ts:291">
P2: Content-Type header removal is case-sensitive, leaving lowercase content-type intact and breaking multipart boundary handling</violation>
</file>

<file name="src/dify-provider.ts">

<violation number="1" location="src/dify-provider.ts:49">
P2: Custom fetch provided via DifyProviderSettings is never forwarded to the model config, so it is ignored during API calls.</violation>
</file>

---

Since this is your first cubic review, here's how it works:

• cubic automatically reviews your code and comments on bugs and improvements
• Teach cubic by replying to its comments. cubic learns from your replies and gets better over time
• Add one-off context when rerunning by tagging @cubic-dev-ai with guidance or docs links (including llms.txt)
• Ask questions if you need clarification on any suggestion

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[warmwind]

PR Review: Real API Testing + Code Review

Real API Test Results (Agent Chat App)

Feature	Status	Notes
Streaming (streamText)	✅ WORKS	Correct text output, proper text-start → text-delta → text-end → finish lifecycle
Conversation continuity (chat-id)	✅ WORKS	Bot correctly remembers previous messages across turns
agent_thought handling	✅ WORKS	Properly ignored to avoid duplicate content from agent_message
Empty agent_message events	✅ WORKS	Properly skipped (no empty deltas emitted)
Token usage extraction	✅ WORKS	message_end correctly extracts from metadata.usage
providerMetadata in stream events	✅ WORKS	conversationId/messageId/taskId attached to all stream parts
Build (npm run build)	✅ PASSES	Clean TypeScript compilation
Blocking mode	⚠️ NOT TESTED	Test app is Agent Chat type (doesn't support blocking)

Issues Found

1. 🔴 2 Unit Tests Fail

npm run test → 2 failed | 9 passed (11)

• should handle message_end and agent_message events (line 447): Expected outputTokens to be 0, got 25. The PR changed message_end to use metadata.usage.completion_tokens instead of returning 0.
• should handle message_end with usage tokens from data field (line 490): Expected outputTokens to be 50, got 25. The PR now uses metadata.usage for message_end and data.total_tokens for workflow_finished — test expectations need updating to match.

2. 🔴 Content-Type Header Casing Bug (stream-parser.ts:292)

const { "content-type": _, ...authHeaders } = headers();

The headers() function in dify-provider.ts returns "Content-Type" (Pascal case), but the destructuring uses "content-type" (lowercase). JavaScript object destructuring is case-sensitive, so Content-Type: application/json will leak into the FormData upload request, breaking multipart boundary handling.

Fix: Change to "Content-Type" to match the headers function.

3. 🟡 fetch Not Forwarded (dify-provider.ts:49)

Pre-existing bug (not introduced by this PR): options.fetch from DifyProviderSettings is never passed to the model config, so custom fetch implementations are silently ignored.

4. 🟡 No Tests for New Code

stream-parser.ts adds 360 lines of new logic (ThinkTagParser, parseToolCalls, formatToolsPrompt, uploadFileToDify, extractFileAttachments) with zero test coverage.

5. 🟢 Minor: Hardcoded TOOL_SHORT_DESCRIPTIONS (stream-parser.ts:212-224)

Contains tool names specific to Claude Code (bash, glob, grep, edit, write, etc.) rather than generic Dify tools. Consider removing or making this configurable.

Verdict

Core streaming features work correctly against a real Dify API. Please fix the 2 failing tests and the Content-Type casing bug before merge.

[cubic-dev-ai]

1 issue found across 2 files (changes from recent commits).

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="src/stream-parser.ts">

<violation number="1" location="src/stream-parser.ts:33">
P2: Quoted content inside `<think>` blocks is always emitted as text because the new inString logic runs before the state machine and hardcodes `type: "text"`, so reasoning content within quotes is misclassified and can leak.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}