 Matthew Meszaros
|
bb040828b1
|
feat: rework automation AI steps to mirror campaign step types with one agentic AI step plus an AI switch, per-case switch handles on the canvas, a Stop node, and a step-type add menu
|
2026-07-22 10:58:35 +02:00 |
|
|
Matthew Meszaros
|
4e576daddf
|
feat: stream the dashboard agent's text token by token: the OpenAI-compatible provider now runs chat completions with stream=true whenever a listener is attached, forwarding content deltas as a new EventTextDelta (SSE text_delta) while stitching tool-call fragments by index, capturing usage via stream_options with a sticky compatibility fallback, tolerating backends that ignore streaming, and always closing each block with the authoritative full-text event (now also emitted for commentary alongside tool calls)
|
2026-07-19 10:41:45 +02:00 |
|
|
Matthew Meszaros
|
e64b0b161d
|
feat: sweep stale OPENAI_API_KEY/AI_LOCAL_MODEL mentions out of comments, the provider-not-configured and warmup-admin error strings, env.example, and the deployment guide now that AI_* is the only config surface
|
2026-07-16 06:39:07 +02:00 |
|
|
Matthew Meszaros
|
84424fec8f
|
feat: add a free/local model flag to the generation provider - ProviderConfig.Local, Provider.IsLocal(), and a local default model (llama3.1) for OpenAI-compatible endpoints
|
2026-07-15 18:51:51 +02:00 |
|
|
Matthew Meszaros
|
a067bf5922
|
feat: AI action nodes in automations - ai_classify/ai_extract/ai_generate run one generation.Provider completion per node over the event data and merge the result back as a variable (ai_class, output_keys[], ai_text) so downstream conditions branch on it, charging 1 credit per node with a run_id+node_id idempotency key, refund on provider failure, out-of-credits fails only that node while the run continues, and a per-automation ai_credit_failures counter (migration 000064) auto-pauses a flow after 20 consecutive credit misses; classify/extract run deterministic (temperature 0), dry-run executes AI nodes for real and is charged but never touches the auto-pause lifecycle, provider + credit ledger wired into both backend and consumer integration services, plus the reply-classifier Layer 3 ported off its bespoke OpenAI client onto the platform provider via an injected ModelClassifyFunc (platform-paid, deterministic), a CompletionRequest.Temperature primitive threaded through the OpenAI and Anthropic providers, the flow builder palette/config panels/validation for the three AI steps, and a new AI-steps-in-automations guide
|
2026-07-14 05:30:55 +02:00 |
|
|
Matthew Meszaros
|
8d208d8c41
|
feat: org voice profile grounding every AI writing surface plus context-grounded unibox reply drafts - organizations gains product_description/icp_notes/voice_profile (000059) with a manage_settings workspace settings section, generation.BuildVoiceRules and BuildReplyRules fold the humanizer rules and org grounding into every prompt, WritingGenerator.GenerateWriting now takes a VoiceContext and a provider.Complete primitive powers single completions, POST /unibox/reply/draft assembles thread history plus the counterpart contact (custom fields and campaigns) plus the voice profile into a reply-tuned prompt that charges 2 credits with refund-on-failure and a client idempotency key and never sends, a Draft reply button fills the composer, and docs cover the voice profile and AI reply drafts
|
2026-07-13 19:04:57 +02:00 |
|
|
Matthew Meszaros
|
7f425c1624
|
feat: dashboard-wide AI assistant with streamed tool-use runs, per-action approvals, and per-iteration credits - agent_sessions/agent_messages/ai_tool_policies migration, aiagent service runs the M2 registry through the provider loop over a resumable jsonb transcript, streams text/tool-step/approval/done events over SSE, gates write tools behind approve/deny/always-allow (send always per-action) and charges 1 credit per iteration (budget 20, refund on provider failure, out-of-credits 402 insufficient_credits, cap 429 usage_cap_exceeded) with a resume-safe idempotency namespace, provider PreIteration budget hook, /ai/sessions endpoints (cursor list + two SSE runs) gated by membership with tools bound to the member's org-permission bits, APIPermAIAgent bit 22, ai_session audit entity + spine, and a right-side resizable panel (Cmd+I / sparkle button) with streamed text, collapsible tool steps, inline approval and draft-artifact deep-link cards, stop, new chat, and a credits/budget meter; tenancy enforced at the SQL layer and docs for the assistant, endpoints, permissions, and error codes
|
2026-07-13 18:46:43 +02:00 |
|
|
Matthew Meszaros
|
e07c33d3cd
|
feat: provider-agnostic LLM agent-loop layer (OpenAI-first, Anthropic self-host connector) with RunAgent tool-use loop, approval/resume contract, pluggable web search, and extracted voice-rules builder - generation.Provider.RunAgent executes Go tool handlers over a resumable transcript with a read-auto/write-approve gate, openAIProvider drives it via a lean HTTP client retargetable with OPENAI_BASE_URL for any OpenAI-compatible endpoint and also serves the writing assistant, anthropicProvider mirrors it with tool_use/tool_result multi-turn plus hosted web_search, NewProvider prefers OPENAI_API_KEY, SearchClient supports Serper/SearXNG with a clean not-configured error, and BuildVoiceRules extracts the humanizer prompt for reuse; refreshed stale model ids to sonnet-4-6 and haiku-4-5
|
2026-07-13 17:52:09 +02:00 |
|