aichat

Package agent implements a tool-calling agent loop over ai.LLMProvider: it streams a model's response, executes any tool calls the model asked for via caller-supplied Handlers, feeds the results back, and repeats until the model stops calling tools or a configured limit is hit.

Package aiconfig loads and applies the product-level configuration for which LLM and decision providers a product wires up: cloud or BYOK (bring your own key, connecting directly to the provider, never via the cloud), and which deciders run before/instead of the cloud's decision service.

Package anthropic is an ai.LLMProvider for the Anthropic Messages API, implemented directly over net/http -- no vendor SDK.

Package cloud is the client for the cloudproto protocol (see ai/cloudproto): it implements ai.LLMProvider (chat) and exposes a separate decision.Provider via Decider(), plus a Usage lookup.

Package cloudproto is the small, product-neutral wire protocol between clients and an AI cloud boundary.

Package ctxmgr selects which ai.ContextBlock values go into a ChatRequest, balancing two goals: keep the LLM within a token budget, and keep the provider-cached prefix (the leading static blocks a provider such as Anthropic prompt-caches) as stable as possible across turns so caching actually pays off.

Package decision defines the DecisionProvider contract and the Chain that runs providers in order until one decides.

Package llmdecider is a decision.Provider that makes ONE structured inference against an ai.LLMProvider to produce a decision.Decision.

Package rules is a deterministic, table-driven decision.Provider.

Package diag is the shared diagnostics record for one conversational turn: which path handled it, what the decision chain did, which LLM answered, timing and usage.

Package openaicompat is an ai.LLMProvider for the OpenAI Chat Completions streaming API (and any API-compatible endpoint), implemented directly over net/http -- no vendor SDK.

Package openairesponses is an ai.LLMProvider over OpenAI's Responses API (POST {base}/responses, stream:true), implemented directly over net/http -- no vendor SDK.

Package session holds the product-neutral conversational working state that helps resolve requests such as "move it to Friday": the focused entity, the current selection, sidebar pins, and the pending and previous actions.

Package chatshell composes tui/transcript, tui/sidebar, tui/focus and tui/stream into the reusable chat screen every aichat product runs: history + composer + sidebar + top/status bars, with a split pane when the terminal is wide enough (>= 104 columns, matching DataTug's splitEnabled).

Package focus is a small, pure focus-ring state machine shared by every chatshell screen.

Package grid is a product-neutral result-grid transcript block: a bubble-table-backed table with sort, per-cell column selection, a scrollbar, style presets, and slots for a product's own secondary views (charts, a current-row card, raw responses, ...) and split-pane layout.

Package sidebar is the right-hand working-context panel shared by every aichat chatshell: an ordered list of session.EntityRef, rendered by a product-supplied renderer, with a cursor, remove ("x"/Delete), open (Enter) and a resizable split width — generalised from DataTug chat's F6 workspace dock and Ctrl+←/→ split handling (ui.go splitEnabled/ resizeChatPane, workspace_ui.go).

Package stream pumps an ai.LLMProvider's iter.Seq2[ai.Event, error] into Bubble Tea messages without buffering the response: each event arrives as an EventMsg carrying the tea.Cmd that re-arms the pump for the next event (the "channel re-arm" pattern DataTug chat used for HTTP/LLM streaming).

Package transcript renders the scrolling chat history shared by every aichat product: plain user/assistant messages, streamed assistant text and rich Block entries (e.g.