TOpenCode (OpenCode Zen)

Source: source/classes/topencode.prg

TOpenCode talks to the OpenCode Zen chat API (https://opencode.ai/zen/v1/chat/completions) with Authorization: Bearer public. No API key. Free models (mimo-v2.6-flash-free, space-bunny-free, big-pickle, beam-free, glm-5.3-flash-free, deepseek-v4.1-flash-free; checked 2026-10-17). Uses hb_curl. Async methods need the Harbour MT VM (build_new.bat <sample> hb32 mt).

Transport: hbcurl over a static libcurl (Schannel) — no libcurl.dll to ship. Curl sources and build scripts live in 3rdparty/ (outputs: lib/libcurl_static*.{lib,a} for bcc32/bcc64/msvc32/ msvc64/gcc).

Architecture

flowchart LR A[TOpenCode] --> B[hb_curl] B --> C["opencode.ai/zen/v1/chat/completions"] A -->|SendAsync| D[hb_threadStart] D --> E[oc_*.json]

Key DATA

DATA	Default	Description
`cModel`	`"mimo-v2.6-flash-free"`	Zen free model
`cUrl`	Zen chat completions	Endpoint
`nTemperature`	`0.7`	0.0–2.0
`aFallbackModels`	other free models	Used on rate limit
`lVerifySSL`	`.T.`	Verify the server certificate. Leave it on; set `.F.` only to diagnose a certificate problem.
`lRunning`	`.F.`	Async in flight
`cBaseUrl`	`""`	`opencode serve` URL (local server mode); empty = Zen
`cProviderID`	`""`	Server mode: provider for `cModel` (empty = server default)
`nTimeout`	`90000`	Per-request timeout, ms (`SetTimeout()`)
`lHistory`	`.T.`	`Send()` uses and updates the `aMessages` conversation

Methods

Method	Description
`New( cModel )`	Own curl handle per instance (safe to run several in parallel).
`Send( cPrompt )`	Sync. Retries with `cFallbackModel()` on rate limit.
`SendWithSystem( cSys, cUser )`	Sync with a system prompt (agent role).
`SendAsync( cPrompt )`	Returns at once. Poll `IsRunning()`.
`SendWithSystemAsync( cSys, cUser )`	Async + system prompt.
`IsRunning()`	`.T.` until the worker finishes. If the worker ends without a result, it returns `.F.` and `GetValue()` reports an error instead of waiting forever.
`GetValue()`	Assistant text from JSON; UTF-8 → ACP.
`Wait() / End()`	Join the thread; cleanup curl.
`SetModel() / SetTemperature()`	Change model / temperature.
`SetServer( cBaseUrl [, cProviderID [, cModel ]] )`	Use a local `opencode serve` server instead of Zen (empty URL = back to Zen).
`SessionStart() / SessionEnd()`	Manage the server session by hand (automatic on `Send()` / `End()`).
`SetTimeout( nMs )`	Per-request timeout in milliseconds.
`ClearMessages()`	Forget the conversation; in server mode the next `Send()` starts a fresh session.

Async writes oc_<ms>_<seq>.json via a .part file then rename, so three SendAsync() calls in the same tick do not collide and IsRunning() never reads a half-written body.

Local server mode (`opencode serve`)

With SetServer() the class talks to a local opencode serve instance (default http://127.0.0.1:4096) instead of OpenCode Zen. The class keeps a persistent session (POST /session), so the server remembers the whole conversation, and any provider/model configured in opencode can be used (set cProviderID and cModel; empty = server default). If the session is lost (HTTP 404) it is recreated once automatically, and End() deletes it (DELETE /session/{id}). Basic auth is supported through cServerUser / cServerPass (OPENCODE_SERVER_PASSWORD). In this mode the SendAsync() methods run synchronously. This backend follows the approach of Charly 9000's TOpenCode class (MIT license).

o := TOpenCode():New()
o:SetServer( "http://127.0.0.1:4096" )
o:Send( "Hello" )
o:Send( "What did I just say?" )   // the server remembers it
o:End()

Conversation history

By default (lHistory := .T.) every successful Send()/SendWithSystem() appends the turn to aMessages and the next request carries the whole conversation, so the model has memory. Call ClearMessages() to start a new conversation, or set lHistory := .F. for stateless requests.

Example

#include "FiveWin.ch"

function Main()
   local o := TOpenCode():New( "mimo-v2.5-free" )
   o:Send( "Say hello in one sentence." )
   MsgInfo( o:GetValue() )
   o:End()
return nil

Samples

Sample	Description
`samples/ai/opencode1.prg`	Three agents, each with its own TOpenCode, SendAsync + timer. Build `hb32 mt`.
`samples/ai/opencode2.prg`	TWebView2 UI (WhatsApp layout): agents left, chat right, add/edit roles. Persists `opencode2.json`. Needs `opencode2.html` next to the exe. Build `hb32 mt`.

cd samples\ai
build_new.bat opencode1 hb32 mt
build_new.bat opencode2 hb32 mt
:: run from this folder (libcurl.dll)

Notes

Best effort: the free tier of OpenCode Zen only answers requests that look like they come from the OpenCode client, so TOpenCode sends the headers and tool list that tier expects. Zen may change this or the free model list at any time; if a model stops answering, call SetModel() with another free model.
No API key. OAuth / auth.json on TOpenCode is unused (Zen rejects sk- keys).
Async requires hbvmmt — pass mt to build_new.bat.
32-bit samples need the x86 libcurl.dll beside the exe (samples\ai).
Do not call the Zen API from inside TWebView2:bOnBind; queue the work and run it from a timer (see opencode2).